program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}})]
{
    func main<ios16>(tensor<fp16, [1]> conditioning_scale, tensor<fp16, [1, 3, 1024, 1024]> controlnet_cond, tensor<fp16, [1, 77, 2048]> encoder_hidden_states, tensor<fp16, [1, 4, 128, 128]> sample, tensor<fp16, [1, 1280]> text_embeds, tensor<fp16, [1, 6]> time_ids, tensor<fp16, [1]> timestep) {
            tensor<int32, []> var_45 = const()[name = tensor<string, []>("op_45"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> var_61_axes_0 = const()[name = tensor<string, []>("op_61_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1]> var_61_cast_fp16 = expand_dims(axes = var_61_axes_0, x = timestep)[name = tensor<string, []>("op_61_cast_fp16")];
            tensor<fp16, [1, 160]> var_63_to_fp16 = const()[name = tensor<string, []>("op_63_to_fp16"), val = tensor<fp16, [1, 160]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [1, 160]> emb_3_cast_fp16 = mul(x = var_61_cast_fp16, y = var_63_to_fp16)[name = tensor<string, []>("emb_3_cast_fp16")];
            tensor<fp16, [1, 160]> var_68_cast_fp16 = sin(x = emb_3_cast_fp16)[name = tensor<string, []>("op_68_cast_fp16")];
            tensor<fp16, [1, 160]> var_69_cast_fp16 = cos(x = emb_3_cast_fp16)[name = tensor<string, []>("op_69_cast_fp16")];
            tensor<bool, []> emb_7_interleave_0 = const()[name = tensor<string, []>("emb_7_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 320]> emb_7_cast_fp16 = concat(axis = var_45, interleave = emb_7_interleave_0, values = (var_68_cast_fp16, var_69_cast_fp16))[name = tensor<string, []>("emb_7_cast_fp16")];
            tensor<int32, [2]> var_73_begin_0 = const()[name = tensor<string, []>("op_73_begin_0"), val = tensor<int32, [2]>([0, 160])];
            tensor<int32, [2]> var_73_end_0 = const()[name = tensor<string, []>("op_73_end_0"), val = tensor<int32, [2]>([1, 320])];
            tensor<bool, [2]> var_73_end_mask_0 = const()[name = tensor<string, []>("op_73_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<fp16, [1, 160]> var_73_cast_fp16 = slice_by_index(begin = var_73_begin_0, end = var_73_end_0, end_mask = var_73_end_mask_0, x = emb_7_cast_fp16)[name = tensor<string, []>("op_73_cast_fp16")];
            tensor<int32, [2]> var_75_begin_0 = const()[name = tensor<string, []>("op_75_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> var_75_end_0 = const()[name = tensor<string, []>("op_75_end_0"), val = tensor<int32, [2]>([1, 160])];
            tensor<bool, [2]> var_75_end_mask_0 = const()[name = tensor<string, []>("op_75_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [1, 160]> var_75_cast_fp16 = slice_by_index(begin = var_75_begin_0, end = var_75_end_0, end_mask = var_75_end_mask_0, x = emb_7_cast_fp16)[name = tensor<string, []>("op_75_cast_fp16")];
            tensor<bool, []> t_emb_interleave_0 = const()[name = tensor<string, []>("t_emb_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 320]> t_emb_cast_fp16 = concat(axis = var_45, interleave = t_emb_interleave_0, values = (var_73_cast_fp16, var_75_cast_fp16))[name = tensor<string, []>("t_emb_cast_fp16")];
            tensor<fp16, [1280, 320]> time_embedding_linear_1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307712))), name = tensor<string, []>("time_embedding_linear_1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 320])];
            tensor<fp16, [1280]> time_embedding_linear_1_bias_to_fp16 = const()[name = tensor<string, []>("time_embedding_linear_1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307904)))];
            tensor<fp16, [1, 1280]> linear_0_cast_fp16 = linear(bias = time_embedding_linear_1_bias_to_fp16, weight = time_embedding_linear_1_weight_to_fp16_palettized, x = t_emb_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<fp16, [1, 1280]> input_5_cast_fp16 = silu(x = linear_0_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<fp16, [1280, 1280]> time_embedding_linear_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310528))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1539392))), name = tensor<string, []>("time_embedding_linear_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> time_embedding_linear_2_bias_to_fp16 = const()[name = tensor<string, []>("time_embedding_linear_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1539584)))];
            tensor<fp16, [1, 1280]> linear_1_cast_fp16 = linear(bias = time_embedding_linear_2_bias_to_fp16, weight = time_embedding_linear_2_weight_to_fp16_palettized, x = input_5_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<int32, [1]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<int32, [1]>([6])];
            tensor<fp16, [6]> timesteps_cast_fp16 = reshape(shape = concat_0, x = time_ids)[name = tensor<string, []>("timesteps_cast_fp16")];
            tensor<int32, []> var_95 = const()[name = tensor<string, []>("op_95"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> var_111_axes_0 = const()[name = tensor<string, []>("op_111_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [6, 1]> var_111_cast_fp16 = expand_dims(axes = var_111_axes_0, x = timesteps_cast_fp16)[name = tensor<string, []>("op_111_cast_fp16")];
            tensor<fp16, [1, 128]> var_113_to_fp16 = const()[name = tensor<string, []>("op_113_to_fp16"), val = tensor<fp16, [1, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1542208)))];
            tensor<fp16, [6, 128]> emb_11_cast_fp16 = mul(x = var_111_cast_fp16, y = var_113_to_fp16)[name = tensor<string, []>("emb_11_cast_fp16")];
            tensor<fp16, [6, 128]> var_118_cast_fp16 = sin(x = emb_11_cast_fp16)[name = tensor<string, []>("op_118_cast_fp16")];
            tensor<fp16, [6, 128]> var_119_cast_fp16 = cos(x = emb_11_cast_fp16)[name = tensor<string, []>("op_119_cast_fp16")];
            tensor<bool, []> emb_15_interleave_0 = const()[name = tensor<string, []>("emb_15_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [6, 256]> emb_15_cast_fp16 = concat(axis = var_95, interleave = emb_15_interleave_0, values = (var_118_cast_fp16, var_119_cast_fp16))[name = tensor<string, []>("emb_15_cast_fp16")];
            tensor<int32, [2]> var_123_begin_0 = const()[name = tensor<string, []>("op_123_begin_0"), val = tensor<int32, [2]>([0, 128])];
            tensor<int32, [2]> var_123_end_0 = const()[name = tensor<string, []>("op_123_end_0"), val = tensor<int32, [2]>([6, 256])];
            tensor<bool, [2]> var_123_end_mask_0 = const()[name = tensor<string, []>("op_123_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<fp16, [6, 128]> var_123_cast_fp16 = slice_by_index(begin = var_123_begin_0, end = var_123_end_0, end_mask = var_123_end_mask_0, x = emb_15_cast_fp16)[name = tensor<string, []>("op_123_cast_fp16")];
            tensor<int32, [2]> var_125_begin_0 = const()[name = tensor<string, []>("op_125_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> var_125_end_0 = const()[name = tensor<string, []>("op_125_end_0"), val = tensor<int32, [2]>([6, 128])];
            tensor<bool, [2]> var_125_end_mask_0 = const()[name = tensor<string, []>("op_125_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [6, 128]> var_125_cast_fp16 = slice_by_index(begin = var_125_begin_0, end = var_125_end_0, end_mask = var_125_end_mask_0, x = emb_15_cast_fp16)[name = tensor<string, []>("op_125_cast_fp16")];
            tensor<bool, []> time_embeds_1_interleave_0 = const()[name = tensor<string, []>("time_embeds_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [6, 256]> time_embeds_1_cast_fp16 = concat(axis = var_95, interleave = time_embeds_1_interleave_0, values = (var_123_cast_fp16, var_125_cast_fp16))[name = tensor<string, []>("time_embeds_1_cast_fp16")];
            tensor<int32, [2]> var_133 = const()[name = tensor<string, []>("op_133"), val = tensor<int32, [2]>([1, -1])];
            tensor<fp16, [1, 1536]> time_embeds_cast_fp16 = reshape(shape = var_133, x = time_embeds_1_cast_fp16)[name = tensor<string, []>("time_embeds_cast_fp16")];
            tensor<int32, []> var_136 = const()[name = tensor<string, []>("op_136"), val = tensor<int32, []>(-1)];
            tensor<bool, []> add_embeds_interleave_0 = const()[name = tensor<string, []>("add_embeds_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 2816]> add_embeds_cast_fp16 = concat(axis = var_136, interleave = add_embeds_interleave_0, values = (text_embeds, time_embeds_cast_fp16))[name = tensor<string, []>("add_embeds_cast_fp16")];
            tensor<fp16, [1280, 2816]> add_embedding_linear_1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2703360]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1542528))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4245952))), name = tensor<string, []>("add_embedding_linear_1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2816])];
            tensor<fp16, [1280]> add_embedding_linear_1_bias_to_fp16 = const()[name = tensor<string, []>("add_embedding_linear_1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4246144)))];
            tensor<fp16, [1, 1280]> linear_2_cast_fp16 = linear(bias = add_embedding_linear_1_bias_to_fp16, weight = add_embedding_linear_1_weight_to_fp16_palettized, x = add_embeds_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<fp16, [1, 1280]> input_11_cast_fp16 = silu(x = linear_2_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<fp16, [1280, 1280]> add_embedding_linear_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4248768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5477632))), name = tensor<string, []>("add_embedding_linear_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> add_embedding_linear_2_bias_to_fp16 = const()[name = tensor<string, []>("add_embedding_linear_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5477824)))];
            tensor<fp16, [1, 1280]> linear_3_cast_fp16 = linear(bias = add_embedding_linear_2_bias_to_fp16, weight = add_embedding_linear_2_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<fp16, [1, 1280]> input_47_cast_fp16 = add(x = linear_1_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<int32, []> var_157 = const()[name = tensor<string, []>("op_157"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_160 = const()[name = tensor<string, []>("op_160"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_162 = const()[name = tensor<string, []>("op_162"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_3_pad_type_0 = const()[name = tensor<string, []>("sample_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_3_pad_0 = const()[name = tensor<string, []>("sample_3_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [320, 4, 3, 3]> conv_in_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [8640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5480448))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5489152))), name = tensor<string, []>("conv_in_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 4, 3, 3])];
            tensor<fp16, [320]> conv_in_bias_to_fp16 = const()[name = tensor<string, []>("conv_in_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5489344)))];
            tensor<fp16, [1, 320, 128, 128]> sample_3_cast_fp16 = conv(bias = conv_in_bias_to_fp16, dilations = var_162, groups = var_157, pad = sample_3_pad_0, pad_type = sample_3_pad_type_0, strides = var_160, weight = conv_in_weight_to_fp16_palettized, x = sample)[name = tensor<string, []>("sample_3_cast_fp16")];
            tensor<int32, []> var_166 = const()[name = tensor<string, []>("op_166"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_186 = const()[name = tensor<string, []>("op_186"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_188 = const()[name = tensor<string, []>("op_188"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [16, 3, 3, 3]> controlnet_cond_embedding_conv_in_weight_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_conv_in_weight_to_fp16"), val = tensor<fp16, [16, 3, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5490048)))];
            tensor<fp16, [16]> controlnet_cond_embedding_conv_in_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_conv_in_bias_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5491008)))];
            tensor<fp16, [1, 16, 1024, 1024]> input_13_cast_fp16 = conv(bias = controlnet_cond_embedding_conv_in_bias_to_fp16, dilations = var_188, groups = var_166, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = var_186, weight = controlnet_cond_embedding_conv_in_weight_to_fp16, x = controlnet_cond)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<int32, [2]> var_194 = const()[name = tensor<string, []>("op_194"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_196 = const()[name = tensor<string, []>("op_196"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [16, 16, 3, 3]> controlnet_cond_embedding_blocks_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5491136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5492928))), name = tensor<string, []>("controlnet_cond_embedding_blocks_0_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([16, 16, 3, 3])];
            tensor<fp16, [16]> controlnet_cond_embedding_blocks_0_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_blocks_0_bias_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5493120)))];
            tensor<fp16, [1, 16, 1024, 1024]> input_17_cast_fp16 = conv(bias = controlnet_cond_embedding_blocks_0_bias_to_fp16, dilations = var_196, groups = var_166, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_194, weight = controlnet_cond_embedding_blocks_0_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> input_19_cast_fp16 = silu(x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<int32, [2]> var_202 = const()[name = tensor<string, []>("op_202"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_204 = const()[name = tensor<string, []>("op_204"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_21_pad_type_0 = const()[name = tensor<string, []>("input_21_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = tensor<string, []>("input_21_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [32, 16, 3, 3]> controlnet_cond_embedding_blocks_1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5493248))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5496768))), name = tensor<string, []>("controlnet_cond_embedding_blocks_1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([32, 16, 3, 3])];
            tensor<fp16, [32]> controlnet_cond_embedding_blocks_1_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_blocks_1_bias_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5496960)))];
            tensor<fp16, [1, 32, 512, 512]> input_21_cast_fp16 = conv(bias = controlnet_cond_embedding_blocks_1_bias_to_fp16, dilations = var_204, groups = var_166, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = var_202, weight = controlnet_cond_embedding_blocks_1_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<fp16, [1, 32, 512, 512]> input_23_cast_fp16 = silu(x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<int32, [2]> var_210 = const()[name = tensor<string, []>("op_210"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_212 = const()[name = tensor<string, []>("op_212"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_25_pad_type_0 = const()[name = tensor<string, []>("input_25_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = tensor<string, []>("input_25_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [32, 32, 3, 3]> controlnet_cond_embedding_blocks_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5497088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5504064))), name = tensor<string, []>("controlnet_cond_embedding_blocks_2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([32, 32, 3, 3])];
            tensor<fp16, [32]> controlnet_cond_embedding_blocks_2_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_blocks_2_bias_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5504256)))];
            tensor<fp16, [1, 32, 512, 512]> input_25_cast_fp16 = conv(bias = controlnet_cond_embedding_blocks_2_bias_to_fp16, dilations = var_212, groups = var_166, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = var_210, weight = controlnet_cond_embedding_blocks_2_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<fp16, [1, 32, 512, 512]> input_27_cast_fp16 = silu(x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<int32, [2]> var_218 = const()[name = tensor<string, []>("op_218"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_220 = const()[name = tensor<string, []>("op_220"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [96, 32, 3, 3]> controlnet_cond_embedding_blocks_3_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [20736]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5504384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5525184))), name = tensor<string, []>("controlnet_cond_embedding_blocks_3_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([96, 32, 3, 3])];
            tensor<fp16, [96]> controlnet_cond_embedding_blocks_3_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_blocks_3_bias_to_fp16"), val = tensor<fp16, [96]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5525376)))];
            tensor<fp16, [1, 96, 256, 256]> input_29_cast_fp16 = conv(bias = controlnet_cond_embedding_blocks_3_bias_to_fp16, dilations = var_220, groups = var_166, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = var_218, weight = controlnet_cond_embedding_blocks_3_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<fp16, [1, 96, 256, 256]> input_31_cast_fp16 = silu(x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<int32, [2]> var_226 = const()[name = tensor<string, []>("op_226"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_228 = const()[name = tensor<string, []>("op_228"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_33_pad_type_0 = const()[name = tensor<string, []>("input_33_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_33_pad_0 = const()[name = tensor<string, []>("input_33_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [96, 96, 3, 3]> controlnet_cond_embedding_blocks_4_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [62208]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5525632))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5587904))), name = tensor<string, []>("controlnet_cond_embedding_blocks_4_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([96, 96, 3, 3])];
            tensor<fp16, [96]> controlnet_cond_embedding_blocks_4_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_blocks_4_bias_to_fp16"), val = tensor<fp16, [96]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5588096)))];
            tensor<fp16, [1, 96, 256, 256]> input_33_cast_fp16 = conv(bias = controlnet_cond_embedding_blocks_4_bias_to_fp16, dilations = var_228, groups = var_166, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = var_226, weight = controlnet_cond_embedding_blocks_4_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<fp16, [1, 96, 256, 256]> input_35_cast_fp16 = silu(x = input_33_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<int32, [2]> var_234 = const()[name = tensor<string, []>("op_234"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_236 = const()[name = tensor<string, []>("op_236"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [256, 96, 3, 3]> controlnet_cond_embedding_blocks_5_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [165888]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5588352))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5754304))), name = tensor<string, []>("controlnet_cond_embedding_blocks_5_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([256, 96, 3, 3])];
            tensor<fp16, [256]> controlnet_cond_embedding_blocks_5_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_blocks_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5754496)))];
            tensor<fp16, [1, 256, 128, 128]> input_37_cast_fp16 = conv(bias = controlnet_cond_embedding_blocks_5_bias_to_fp16, dilations = var_236, groups = var_166, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_234, weight = controlnet_cond_embedding_blocks_5_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<fp16, [1, 256, 128, 128]> input_39_cast_fp16 = silu(x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<int32, [2]> var_242 = const()[name = tensor<string, []>("op_242"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_244 = const()[name = tensor<string, []>("op_244"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> controlnet_cond_pad_type_0 = const()[name = tensor<string, []>("controlnet_cond_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> controlnet_cond_pad_0 = const()[name = tensor<string, []>("controlnet_cond_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [320, 256, 3, 3]> controlnet_cond_embedding_conv_out_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [552960]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5755072))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6308096))), name = tensor<string, []>("controlnet_cond_embedding_conv_out_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 256, 3, 3])];
            tensor<fp16, [320]> controlnet_cond_embedding_conv_out_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_cond_embedding_conv_out_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6308288)))];
            tensor<fp16, [1, 320, 128, 128]> controlnet_cond_cast_fp16 = conv(bias = controlnet_cond_embedding_conv_out_bias_to_fp16, dilations = var_244, groups = var_166, pad = controlnet_cond_pad_0, pad_type = controlnet_cond_pad_type_0, strides = var_242, weight = controlnet_cond_embedding_conv_out_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = tensor<string, []>("controlnet_cond_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> input_41_cast_fp16 = add(x = sample_3_cast_fp16, y = controlnet_cond_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<int32, []> var_254 = const()[name = tensor<string, []>("op_254"), val = tensor<int32, []>(1)];
            tensor<int32, [5]> reshape_0_shape_0 = const()[name = tensor<string, []>("reshape_0_shape_0"), val = tensor<int32, [5]>([1, 32, 10, 128, 128])];
            tensor<fp16, [1, 32, 10, 128, 128]> reshape_0_cast_fp16 = reshape(shape = reshape_0_shape_0, x = input_41_cast_fp16)[name = tensor<string, []>("reshape_0_cast_fp16")];
            tensor<int32, [3]> reduce_mean_0_axes_0 = const()[name = tensor<string, []>("reduce_mean_0_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_0_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_0_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = reshape_0_cast_fp16)[name = tensor<string, []>("reduce_mean_0_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> sub_0_cast_fp16 = sub(x = reshape_0_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor<string, []>("sub_0_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor<string, []>("square_0_cast_fp16")];
            tensor<int32, [3]> reduce_mean_2_axes_0 = const()[name = tensor<string, []>("reduce_mean_2_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_2_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_2_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = square_0_cast_fp16)[name = tensor<string, []>("reduce_mean_2_cast_fp16")];
            tensor<fp16, []> add_0_y_0_to_fp16 = const()[name = tensor<string, []>("add_0_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_0_cast_fp16 = add(x = reduce_mean_2_cast_fp16, y = add_0_y_0_to_fp16)[name = tensor<string, []>("add_0_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_0_cast_fp16 = sqrt(x = add_0_cast_fp16)[name = tensor<string, []>("sqrt_0_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> real_div_0_cast_fp16 = real_div(x = sub_0_cast_fp16, y = sqrt_0_cast_fp16)[name = tensor<string, []>("real_div_0_cast_fp16")];
            tensor<int32, [4]> reshape_1_shape_0 = const()[name = tensor<string, []>("reshape_1_shape_0"), val = tensor<int32, [4]>([1, 320, 128, 128])];
            tensor<fp16, [1, 320, 128, 128]> reshape_1_cast_fp16 = reshape(shape = reshape_1_shape_0, x = real_div_0_cast_fp16)[name = tensor<string, []>("reshape_1_cast_fp16")];
            tensor<fp16, [320]> add_1_mean_0_to_fp16 = const()[name = tensor<string, []>("add_1_mean_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6308992)))];
            tensor<fp16, [320]> add_1_variance_0_to_fp16 = const()[name = tensor<string, []>("add_1_variance_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6309696)))];
            tensor<fp16, [320]> add_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_1_gamma_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6310400)))];
            tensor<fp16, [320]> add_1_beta_0_to_fp16 = const()[name = tensor<string, []>("add_1_beta_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6311104)))];
            tensor<fp16, []> add_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 320, 128, 128]> add_1_cast_fp16 = batch_norm(beta = add_1_beta_0_to_fp16, epsilon = add_1_epsilon_0_to_fp16, gamma = add_1_gamma_0_to_fp16, mean = add_1_mean_0_to_fp16, variance = add_1_variance_0_to_fp16, x = reshape_1_cast_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> input_45_cast_fp16 = silu(x = add_1_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<int32, [2]> var_276 = const()[name = tensor<string, []>("op_276"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_278 = const()[name = tensor<string, []>("op_278"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_1_pad_type_0 = const()[name = tensor<string, []>("hidden_states_1_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_1_pad_0 = const()[name = tensor<string, []>("hidden_states_1_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [320, 320, 3, 3]> down_blocks_0_resnets_0_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [691200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6311808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7003072))), name = tensor<string, []>("down_blocks_0_resnets_0_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 3, 3])];
            tensor<fp16, [320]> down_blocks_0_resnets_0_conv1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_0_resnets_0_conv1_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7003264)))];
            tensor<fp16, [1, 320, 128, 128]> hidden_states_1_cast_fp16 = conv(bias = down_blocks_0_resnets_0_conv1_bias_to_fp16, dilations = var_278, groups = var_254, pad = hidden_states_1_pad_0, pad_type = hidden_states_1_pad_type_0, strides = var_276, weight = down_blocks_0_resnets_0_conv1_weight_to_fp16_palettized, x = input_45_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
            tensor<fp16, [1, 1280]> input_49_cast_fp16 = silu(x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<fp16, [320, 1280]> down_blocks_0_resnets_0_time_emb_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7003968))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7311232))), name = tensor<string, []>("down_blocks_0_resnets_0_time_emb_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([320, 1280])];
            tensor<fp16, [320]> down_blocks_0_resnets_0_time_emb_proj_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_0_resnets_0_time_emb_proj_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7311424)))];
            tensor<fp16, [1, 320]> linear_4_cast_fp16 = linear(bias = down_blocks_0_resnets_0_time_emb_proj_bias_to_fp16, weight = down_blocks_0_resnets_0_time_emb_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = tensor<string, []>("op_287_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 320, 1]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = linear_4_cast_fp16)[name = tensor<string, []>("op_287_cast_fp16")];
            tensor<int32, [1]> temb_1_axes_0 = const()[name = tensor<string, []>("temb_1_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 320, 1, 1]> temb_1_cast_fp16 = expand_dims(axes = temb_1_axes_0, x = var_287_cast_fp16)[name = tensor<string, []>("temb_1_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> input_51_cast_fp16 = add(x = hidden_states_1_cast_fp16, y = temb_1_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<int32, [5]> reshape_4_shape_0 = const()[name = tensor<string, []>("reshape_4_shape_0"), val = tensor<int32, [5]>([1, 32, 10, 128, 128])];
            tensor<fp16, [1, 32, 10, 128, 128]> reshape_4_cast_fp16 = reshape(shape = reshape_4_shape_0, x = input_51_cast_fp16)[name = tensor<string, []>("reshape_4_cast_fp16")];
            tensor<int32, [3]> reduce_mean_3_axes_0 = const()[name = tensor<string, []>("reduce_mean_3_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_3_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_3_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = reshape_4_cast_fp16)[name = tensor<string, []>("reduce_mean_3_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> sub_2_cast_fp16 = sub(x = reshape_4_cast_fp16, y = reduce_mean_3_cast_fp16)[name = tensor<string, []>("sub_2_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor<string, []>("square_1_cast_fp16")];
            tensor<int32, [3]> reduce_mean_5_axes_0 = const()[name = tensor<string, []>("reduce_mean_5_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_5_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_5_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_1_cast_fp16)[name = tensor<string, []>("reduce_mean_5_cast_fp16")];
            tensor<fp16, []> add_2_y_0_to_fp16 = const()[name = tensor<string, []>("add_2_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_2_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = add_2_y_0_to_fp16)[name = tensor<string, []>("add_2_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_1_cast_fp16 = sqrt(x = add_2_cast_fp16)[name = tensor<string, []>("sqrt_1_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> real_div_1_cast_fp16 = real_div(x = sub_2_cast_fp16, y = sqrt_1_cast_fp16)[name = tensor<string, []>("real_div_1_cast_fp16")];
            tensor<int32, [4]> reshape_5_shape_0 = const()[name = tensor<string, []>("reshape_5_shape_0"), val = tensor<int32, [4]>([1, 320, 128, 128])];
            tensor<fp16, [1, 320, 128, 128]> reshape_5_cast_fp16 = reshape(shape = reshape_5_shape_0, x = real_div_1_cast_fp16)[name = tensor<string, []>("reshape_5_cast_fp16")];
            tensor<fp16, [320]> add_3_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_3_gamma_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7312128)))];
            tensor<fp16, [320]> add_3_beta_0_to_fp16 = const()[name = tensor<string, []>("add_3_beta_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7312832)))];
            tensor<fp16, []> add_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 320, 128, 128]> add_3_cast_fp16 = batch_norm(beta = add_3_beta_0_to_fp16, epsilon = add_3_epsilon_0_to_fp16, gamma = add_3_gamma_0_to_fp16, mean = add_1_mean_0_to_fp16, variance = add_1_variance_0_to_fp16, x = reshape_5_cast_fp16)[name = tensor<string, []>("add_3_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> input_55_cast_fp16 = silu(x = add_3_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<int32, [2]> var_297 = const()[name = tensor<string, []>("op_297"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_299 = const()[name = tensor<string, []>("op_299"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_3_pad_type_0 = const()[name = tensor<string, []>("hidden_states_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = tensor<string, []>("hidden_states_3_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [320, 320, 3, 3]> down_blocks_0_resnets_0_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [691200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7313536))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8004800))), name = tensor<string, []>("down_blocks_0_resnets_0_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 3, 3])];
            tensor<fp16, [320]> down_blocks_0_resnets_0_conv2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_0_resnets_0_conv2_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8004992)))];
            tensor<fp16, [1, 320, 128, 128]> hidden_states_3_cast_fp16 = conv(bias = down_blocks_0_resnets_0_conv2_bias_to_fp16, dilations = var_299, groups = var_254, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_297, weight = down_blocks_0_resnets_0_conv2_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> var_302_cast_fp16 = add(x = input_41_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("op_302_cast_fp16")];
            tensor<int32, [5]> reshape_8_shape_0 = const()[name = tensor<string, []>("reshape_8_shape_0"), val = tensor<int32, [5]>([1, 32, 10, 128, 128])];
            tensor<fp16, [1, 32, 10, 128, 128]> reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = var_302_cast_fp16)[name = tensor<string, []>("reshape_8_cast_fp16")];
            tensor<int32, [3]> reduce_mean_6_axes_0 = const()[name = tensor<string, []>("reduce_mean_6_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_6_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_6_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_6_cast_fp16 = reduce_mean(axes = reduce_mean_6_axes_0, keep_dims = reduce_mean_6_keep_dims_0, x = reshape_8_cast_fp16)[name = tensor<string, []>("reduce_mean_6_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> sub_4_cast_fp16 = sub(x = reshape_8_cast_fp16, y = reduce_mean_6_cast_fp16)[name = tensor<string, []>("sub_4_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor<string, []>("square_2_cast_fp16")];
            tensor<int32, [3]> reduce_mean_8_axes_0 = const()[name = tensor<string, []>("reduce_mean_8_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_8_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_8_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_8_cast_fp16 = reduce_mean(axes = reduce_mean_8_axes_0, keep_dims = reduce_mean_8_keep_dims_0, x = square_2_cast_fp16)[name = tensor<string, []>("reduce_mean_8_cast_fp16")];
            tensor<fp16, []> add_4_y_0_to_fp16 = const()[name = tensor<string, []>("add_4_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_4_cast_fp16 = add(x = reduce_mean_8_cast_fp16, y = add_4_y_0_to_fp16)[name = tensor<string, []>("add_4_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_2_cast_fp16 = sqrt(x = add_4_cast_fp16)[name = tensor<string, []>("sqrt_2_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> real_div_2_cast_fp16 = real_div(x = sub_4_cast_fp16, y = sqrt_2_cast_fp16)[name = tensor<string, []>("real_div_2_cast_fp16")];
            tensor<int32, [4]> reshape_9_shape_0 = const()[name = tensor<string, []>("reshape_9_shape_0"), val = tensor<int32, [4]>([1, 320, 128, 128])];
            tensor<fp16, [1, 320, 128, 128]> reshape_9_cast_fp16 = reshape(shape = reshape_9_shape_0, x = real_div_2_cast_fp16)[name = tensor<string, []>("reshape_9_cast_fp16")];
            tensor<fp16, [320]> add_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_5_gamma_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8005696)))];
            tensor<fp16, [320]> add_5_beta_0_to_fp16 = const()[name = tensor<string, []>("add_5_beta_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8006400)))];
            tensor<fp16, []> add_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 320, 128, 128]> add_5_cast_fp16 = batch_norm(beta = add_5_beta_0_to_fp16, epsilon = add_5_epsilon_0_to_fp16, gamma = add_5_gamma_0_to_fp16, mean = add_1_mean_0_to_fp16, variance = add_1_variance_0_to_fp16, x = reshape_9_cast_fp16)[name = tensor<string, []>("add_5_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> input_63_cast_fp16 = silu(x = add_5_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<int32, [2]> var_316 = const()[name = tensor<string, []>("op_316"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_318 = const()[name = tensor<string, []>("op_318"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [320, 320, 3, 3]> down_blocks_0_resnets_1_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [691200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8007104))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8698368))), name = tensor<string, []>("down_blocks_0_resnets_1_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 3, 3])];
            tensor<fp16, [320]> down_blocks_0_resnets_1_conv1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_0_resnets_1_conv1_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8698560)))];
            tensor<fp16, [1, 320, 128, 128]> hidden_states_5_cast_fp16 = conv(bias = down_blocks_0_resnets_1_conv1_bias_to_fp16, dilations = var_318, groups = var_254, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_316, weight = down_blocks_0_resnets_1_conv1_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
            tensor<fp16, [320, 1280]> down_blocks_0_resnets_1_time_emb_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8699264))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9006528))), name = tensor<string, []>("down_blocks_0_resnets_1_time_emb_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([320, 1280])];
            tensor<fp16, [320]> down_blocks_0_resnets_1_time_emb_proj_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_0_resnets_1_time_emb_proj_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9006720)))];
            tensor<fp16, [1, 320]> linear_5_cast_fp16 = linear(bias = down_blocks_0_resnets_1_time_emb_proj_bias_to_fp16, weight = down_blocks_0_resnets_1_time_emb_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<int32, [1]> var_327_axes_0 = const()[name = tensor<string, []>("op_327_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 320, 1]> var_327_cast_fp16 = expand_dims(axes = var_327_axes_0, x = linear_5_cast_fp16)[name = tensor<string, []>("op_327_cast_fp16")];
            tensor<int32, [1]> temb_3_axes_0 = const()[name = tensor<string, []>("temb_3_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 320, 1, 1]> temb_3_cast_fp16 = expand_dims(axes = temb_3_axes_0, x = var_327_cast_fp16)[name = tensor<string, []>("temb_3_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> input_67_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = temb_3_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<int32, [5]> reshape_12_shape_0 = const()[name = tensor<string, []>("reshape_12_shape_0"), val = tensor<int32, [5]>([1, 32, 10, 128, 128])];
            tensor<fp16, [1, 32, 10, 128, 128]> reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = input_67_cast_fp16)[name = tensor<string, []>("reshape_12_cast_fp16")];
            tensor<int32, [3]> reduce_mean_9_axes_0 = const()[name = tensor<string, []>("reduce_mean_9_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_9_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_9_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = reshape_12_cast_fp16)[name = tensor<string, []>("reduce_mean_9_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> sub_6_cast_fp16 = sub(x = reshape_12_cast_fp16, y = reduce_mean_9_cast_fp16)[name = tensor<string, []>("sub_6_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> square_3_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor<string, []>("square_3_cast_fp16")];
            tensor<int32, [3]> reduce_mean_11_axes_0 = const()[name = tensor<string, []>("reduce_mean_11_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_11_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_11_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_3_cast_fp16)[name = tensor<string, []>("reduce_mean_11_cast_fp16")];
            tensor<fp16, []> add_6_y_0_to_fp16 = const()[name = tensor<string, []>("add_6_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_6_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = add_6_y_0_to_fp16)[name = tensor<string, []>("add_6_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_3_cast_fp16 = sqrt(x = add_6_cast_fp16)[name = tensor<string, []>("sqrt_3_cast_fp16")];
            tensor<fp16, [1, 32, 10, 128, 128]> real_div_3_cast_fp16 = real_div(x = sub_6_cast_fp16, y = sqrt_3_cast_fp16)[name = tensor<string, []>("real_div_3_cast_fp16")];
            tensor<int32, [4]> reshape_13_shape_0 = const()[name = tensor<string, []>("reshape_13_shape_0"), val = tensor<int32, [4]>([1, 320, 128, 128])];
            tensor<fp16, [1, 320, 128, 128]> reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = real_div_3_cast_fp16)[name = tensor<string, []>("reshape_13_cast_fp16")];
            tensor<fp16, [320]> add_7_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_7_gamma_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9007424)))];
            tensor<fp16, [320]> add_7_beta_0_to_fp16 = const()[name = tensor<string, []>("add_7_beta_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9008128)))];
            tensor<fp16, []> add_7_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_7_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 320, 128, 128]> add_7_cast_fp16 = batch_norm(beta = add_7_beta_0_to_fp16, epsilon = add_7_epsilon_0_to_fp16, gamma = add_7_gamma_0_to_fp16, mean = add_1_mean_0_to_fp16, variance = add_1_variance_0_to_fp16, x = reshape_13_cast_fp16)[name = tensor<string, []>("add_7_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> input_71_cast_fp16 = silu(x = add_7_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<int32, [2]> var_337 = const()[name = tensor<string, []>("op_337"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_339 = const()[name = tensor<string, []>("op_339"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [320, 320, 3, 3]> down_blocks_0_resnets_1_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [691200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9008832))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9700096))), name = tensor<string, []>("down_blocks_0_resnets_1_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 3, 3])];
            tensor<fp16, [320]> down_blocks_0_resnets_1_conv2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_0_resnets_1_conv2_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9700288)))];
            tensor<fp16, [1, 320, 128, 128]> hidden_states_7_cast_fp16 = conv(bias = down_blocks_0_resnets_1_conv2_bias_to_fp16, dilations = var_339, groups = var_254, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_337, weight = down_blocks_0_resnets_1_conv2_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> var_342_cast_fp16 = add(x = var_302_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("op_342_cast_fp16")];
            tensor<int32, [2]> var_348 = const()[name = tensor<string, []>("op_348"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_75_pad_type_0 = const()[name = tensor<string, []>("input_75_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_75_pad_0 = const()[name = tensor<string, []>("input_75_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [320, 320, 3, 3]> down_blocks_0_downsamplers_0_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [691200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9700992))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10392256))), name = tensor<string, []>("down_blocks_0_downsamplers_0_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 3, 3])];
            tensor<fp16, [320]> down_blocks_0_downsamplers_0_conv_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_0_downsamplers_0_conv_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10392448)))];
            tensor<fp16, [1, 320, 64, 64]> input_75_cast_fp16 = conv(bias = down_blocks_0_downsamplers_0_conv_bias_to_fp16, dilations = var_350, groups = var_254, pad = input_75_pad_0, pad_type = input_75_pad_type_0, strides = var_348, weight = down_blocks_0_downsamplers_0_conv_weight_to_fp16_palettized, x = var_342_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<int32, []> var_370 = const()[name = tensor<string, []>("op_370"), val = tensor<int32, []>(1)];
            tensor<int32, [5]> reshape_16_shape_0 = const()[name = tensor<string, []>("reshape_16_shape_0"), val = tensor<int32, [5]>([1, 32, 10, 64, 64])];
            tensor<fp16, [1, 32, 10, 64, 64]> reshape_16_cast_fp16 = reshape(shape = reshape_16_shape_0, x = input_75_cast_fp16)[name = tensor<string, []>("reshape_16_cast_fp16")];
            tensor<int32, [3]> reduce_mean_12_axes_0 = const()[name = tensor<string, []>("reduce_mean_12_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_12_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_12_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_12_cast_fp16 = reduce_mean(axes = reduce_mean_12_axes_0, keep_dims = reduce_mean_12_keep_dims_0, x = reshape_16_cast_fp16)[name = tensor<string, []>("reduce_mean_12_cast_fp16")];
            tensor<fp16, [1, 32, 10, 64, 64]> sub_8_cast_fp16 = sub(x = reshape_16_cast_fp16, y = reduce_mean_12_cast_fp16)[name = tensor<string, []>("sub_8_cast_fp16")];
            tensor<fp16, [1, 32, 10, 64, 64]> square_4_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor<string, []>("square_4_cast_fp16")];
            tensor<int32, [3]> reduce_mean_14_axes_0 = const()[name = tensor<string, []>("reduce_mean_14_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_14_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_14_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_14_cast_fp16 = reduce_mean(axes = reduce_mean_14_axes_0, keep_dims = reduce_mean_14_keep_dims_0, x = square_4_cast_fp16)[name = tensor<string, []>("reduce_mean_14_cast_fp16")];
            tensor<fp16, []> add_8_y_0_to_fp16 = const()[name = tensor<string, []>("add_8_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_8_cast_fp16 = add(x = reduce_mean_14_cast_fp16, y = add_8_y_0_to_fp16)[name = tensor<string, []>("add_8_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_4_cast_fp16 = sqrt(x = add_8_cast_fp16)[name = tensor<string, []>("sqrt_4_cast_fp16")];
            tensor<fp16, [1, 32, 10, 64, 64]> real_div_4_cast_fp16 = real_div(x = sub_8_cast_fp16, y = sqrt_4_cast_fp16)[name = tensor<string, []>("real_div_4_cast_fp16")];
            tensor<int32, [4]> reshape_17_shape_0 = const()[name = tensor<string, []>("reshape_17_shape_0"), val = tensor<int32, [4]>([1, 320, 64, 64])];
            tensor<fp16, [1, 320, 64, 64]> reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = real_div_4_cast_fp16)[name = tensor<string, []>("reshape_17_cast_fp16")];
            tensor<fp16, [320]> add_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_9_gamma_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10393152)))];
            tensor<fp16, [320]> add_9_beta_0_to_fp16 = const()[name = tensor<string, []>("add_9_beta_0_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10393856)))];
            tensor<fp16, []> add_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 320, 64, 64]> add_9_cast_fp16 = batch_norm(beta = add_9_beta_0_to_fp16, epsilon = add_9_epsilon_0_to_fp16, gamma = add_9_gamma_0_to_fp16, mean = add_1_mean_0_to_fp16, variance = add_1_variance_0_to_fp16, x = reshape_17_cast_fp16)[name = tensor<string, []>("add_9_cast_fp16")];
            tensor<fp16, [1, 320, 64, 64]> input_79_cast_fp16 = silu(x = add_9_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<int32, [2]> var_397 = const()[name = tensor<string, []>("op_397"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_399 = const()[name = tensor<string, []>("op_399"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_11_pad_type_0 = const()[name = tensor<string, []>("hidden_states_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = tensor<string, []>("hidden_states_11_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [640, 320, 3, 3]> down_blocks_1_resnets_0_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1382400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10394560))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11777024))), name = tensor<string, []>("down_blocks_1_resnets_0_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 320, 3, 3])];
            tensor<fp16, [640]> down_blocks_1_resnets_0_conv1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_resnets_0_conv1_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11777216)))];
            tensor<fp16, [1, 640, 64, 64]> hidden_states_11_cast_fp16 = conv(bias = down_blocks_1_resnets_0_conv1_bias_to_fp16, dilations = var_399, groups = var_370, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = var_397, weight = down_blocks_1_resnets_0_conv1_weight_to_fp16_palettized, x = input_79_cast_fp16)[name = tensor<string, []>("hidden_states_11_cast_fp16")];
            tensor<fp16, [640, 1280]> down_blocks_1_resnets_0_time_emb_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [614400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11778560))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12393024))), name = tensor<string, []>("down_blocks_1_resnets_0_time_emb_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 1280])];
            tensor<fp16, [640]> down_blocks_1_resnets_0_time_emb_proj_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_resnets_0_time_emb_proj_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12393216)))];
            tensor<fp16, [1, 640]> linear_6_cast_fp16 = linear(bias = down_blocks_1_resnets_0_time_emb_proj_bias_to_fp16, weight = down_blocks_1_resnets_0_time_emb_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<int32, [1]> var_408_axes_0 = const()[name = tensor<string, []>("op_408_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 640, 1]> var_408_cast_fp16 = expand_dims(axes = var_408_axes_0, x = linear_6_cast_fp16)[name = tensor<string, []>("op_408_cast_fp16")];
            tensor<int32, [1]> temb_5_axes_0 = const()[name = tensor<string, []>("temb_5_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 640, 1, 1]> temb_5_cast_fp16 = expand_dims(axes = temb_5_axes_0, x = var_408_cast_fp16)[name = tensor<string, []>("temb_5_cast_fp16")];
            tensor<fp16, [1, 640, 64, 64]> input_83_cast_fp16 = add(x = hidden_states_11_cast_fp16, y = temb_5_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<int32, [5]> reshape_20_shape_0 = const()[name = tensor<string, []>("reshape_20_shape_0"), val = tensor<int32, [5]>([1, 32, 20, 64, 64])];
            tensor<fp16, [1, 32, 20, 64, 64]> reshape_20_cast_fp16 = reshape(shape = reshape_20_shape_0, x = input_83_cast_fp16)[name = tensor<string, []>("reshape_20_cast_fp16")];
            tensor<int32, [3]> reduce_mean_15_axes_0 = const()[name = tensor<string, []>("reduce_mean_15_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_15_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_15_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = reshape_20_cast_fp16)[name = tensor<string, []>("reduce_mean_15_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> sub_10_cast_fp16 = sub(x = reshape_20_cast_fp16, y = reduce_mean_15_cast_fp16)[name = tensor<string, []>("sub_10_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> square_5_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor<string, []>("square_5_cast_fp16")];
            tensor<int32, [3]> reduce_mean_17_axes_0 = const()[name = tensor<string, []>("reduce_mean_17_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_17_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_17_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_5_cast_fp16)[name = tensor<string, []>("reduce_mean_17_cast_fp16")];
            tensor<fp16, []> add_10_y_0_to_fp16 = const()[name = tensor<string, []>("add_10_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_10_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = add_10_y_0_to_fp16)[name = tensor<string, []>("add_10_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_5_cast_fp16 = sqrt(x = add_10_cast_fp16)[name = tensor<string, []>("sqrt_5_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> real_div_5_cast_fp16 = real_div(x = sub_10_cast_fp16, y = sqrt_5_cast_fp16)[name = tensor<string, []>("real_div_5_cast_fp16")];
            tensor<int32, [4]> reshape_21_shape_0 = const()[name = tensor<string, []>("reshape_21_shape_0"), val = tensor<int32, [4]>([1, 640, 64, 64])];
            tensor<fp16, [1, 640, 64, 64]> reshape_21_cast_fp16 = reshape(shape = reshape_21_shape_0, x = real_div_5_cast_fp16)[name = tensor<string, []>("reshape_21_cast_fp16")];
            tensor<fp16, [640]> add_11_mean_0_to_fp16 = const()[name = tensor<string, []>("add_11_mean_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12394560)))];
            tensor<fp16, [640]> add_11_variance_0_to_fp16 = const()[name = tensor<string, []>("add_11_variance_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12395904)))];
            tensor<fp16, [640]> add_11_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_11_gamma_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12397248)))];
            tensor<fp16, [640]> add_11_beta_0_to_fp16 = const()[name = tensor<string, []>("add_11_beta_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12398592)))];
            tensor<fp16, []> add_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 640, 64, 64]> add_11_cast_fp16 = batch_norm(beta = add_11_beta_0_to_fp16, epsilon = add_11_epsilon_0_to_fp16, gamma = add_11_gamma_0_to_fp16, mean = add_11_mean_0_to_fp16, variance = add_11_variance_0_to_fp16, x = reshape_21_cast_fp16)[name = tensor<string, []>("add_11_cast_fp16")];
            tensor<fp16, [1, 640, 64, 64]> input_87_cast_fp16 = silu(x = add_11_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<int32, [2]> var_418 = const()[name = tensor<string, []>("op_418"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_420 = const()[name = tensor<string, []>("op_420"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_13_pad_type_0 = const()[name = tensor<string, []>("hidden_states_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = tensor<string, []>("hidden_states_13_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [640, 640, 3, 3]> down_blocks_1_resnets_0_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2764800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12399936))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15164800))), name = tensor<string, []>("down_blocks_1_resnets_0_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 640, 3, 3])];
            tensor<fp16, [640]> down_blocks_1_resnets_0_conv2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_resnets_0_conv2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15164992)))];
            tensor<fp16, [1, 640, 64, 64]> hidden_states_13_cast_fp16 = conv(bias = down_blocks_1_resnets_0_conv2_bias_to_fp16, dilations = var_420, groups = var_370, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = var_418, weight = down_blocks_1_resnets_0_conv2_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")];
            tensor<int32, [2]> var_425 = const()[name = tensor<string, []>("op_425"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_427 = const()[name = tensor<string, []>("op_427"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_tensor_1_pad_type_0 = const()[name = tensor<string, []>("input_tensor_1_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_tensor_1_pad_0 = const()[name = tensor<string, []>("input_tensor_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [640, 320, 1, 1]> down_blocks_1_resnets_0_conv_shortcut_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [153600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15166336))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15320000))), name = tensor<string, []>("down_blocks_1_resnets_0_conv_shortcut_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 320, 1, 1])];
            tensor<fp16, [640]> down_blocks_1_resnets_0_conv_shortcut_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_resnets_0_conv_shortcut_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15320192)))];
            tensor<fp16, [1, 640, 64, 64]> input_tensor_1_cast_fp16 = conv(bias = down_blocks_1_resnets_0_conv_shortcut_bias_to_fp16, dilations = var_427, groups = var_370, pad = input_tensor_1_pad_0, pad_type = input_tensor_1_pad_type_0, strides = var_425, weight = down_blocks_1_resnets_0_conv_shortcut_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = tensor<string, []>("input_tensor_1_cast_fp16")];
            tensor<fp16, [1, 640, 64, 64]> var_430_cast_fp16 = add(x = input_tensor_1_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor<string, []>("op_430_cast_fp16")];
            tensor<int32, [5]> reshape_24_shape_0 = const()[name = tensor<string, []>("reshape_24_shape_0"), val = tensor<int32, [5]>([1, 32, 20, 64, 64])];
            tensor<fp16, [1, 32, 20, 64, 64]> reshape_24_cast_fp16 = reshape(shape = reshape_24_shape_0, x = var_430_cast_fp16)[name = tensor<string, []>("reshape_24_cast_fp16")];
            tensor<int32, [3]> reduce_mean_18_axes_0 = const()[name = tensor<string, []>("reduce_mean_18_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_18_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_18_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = reshape_24_cast_fp16)[name = tensor<string, []>("reduce_mean_18_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> sub_12_cast_fp16 = sub(x = reshape_24_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor<string, []>("sub_12_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> square_6_cast_fp16 = square(x = sub_12_cast_fp16)[name = tensor<string, []>("square_6_cast_fp16")];
            tensor<int32, [3]> reduce_mean_20_axes_0 = const()[name = tensor<string, []>("reduce_mean_20_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_20_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_20_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = square_6_cast_fp16)[name = tensor<string, []>("reduce_mean_20_cast_fp16")];
            tensor<fp16, []> add_12_y_0_to_fp16 = const()[name = tensor<string, []>("add_12_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_12_cast_fp16 = add(x = reduce_mean_20_cast_fp16, y = add_12_y_0_to_fp16)[name = tensor<string, []>("add_12_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_6_cast_fp16 = sqrt(x = add_12_cast_fp16)[name = tensor<string, []>("sqrt_6_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> real_div_6_cast_fp16 = real_div(x = sub_12_cast_fp16, y = sqrt_6_cast_fp16)[name = tensor<string, []>("real_div_6_cast_fp16")];
            tensor<int32, [4]> reshape_25_shape_0 = const()[name = tensor<string, []>("reshape_25_shape_0"), val = tensor<int32, [4]>([1, 640, 64, 64])];
            tensor<fp16, [1, 640, 64, 64]> reshape_25_cast_fp16 = reshape(shape = reshape_25_shape_0, x = real_div_6_cast_fp16)[name = tensor<string, []>("reshape_25_cast_fp16")];
            tensor<fp16, [640]> add_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_13_gamma_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15321536)))];
            tensor<fp16, [640]> add_13_beta_0_to_fp16 = const()[name = tensor<string, []>("add_13_beta_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15322880)))];
            tensor<fp16, []> add_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 640, 64, 64]> add_13_cast_fp16 = batch_norm(beta = add_13_beta_0_to_fp16, epsilon = add_13_epsilon_0_to_fp16, gamma = add_13_gamma_0_to_fp16, mean = add_11_mean_0_to_fp16, variance = add_11_variance_0_to_fp16, x = reshape_25_cast_fp16)[name = tensor<string, []>("add_13_cast_fp16")];
            tensor<int32, [4]> var_452 = const()[name = tensor<string, []>("op_452"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_456 = const()[name = tensor<string, []>("op_456"), val = tensor<int32, [3]>([1, 4096, 640])];
            tensor<fp16, [1, 64, 64, 640]> transpose_689 = transpose(perm = var_452, x = add_13_cast_fp16)[name = tensor<string, []>("transpose_689")];
            tensor<fp16, [1, 4096, 640]> input_91_cast_fp16 = reshape(shape = var_456, x = transpose_689)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_proj_in_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15324224))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15631488))), name = tensor<string, []>("down_blocks_1_attentions_0_proj_in_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_proj_in_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_proj_in_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15631680)))];
            tensor<fp16, [1, 4096, 640]> linear_7_cast_fp16 = linear(bias = down_blocks_1_attentions_0_proj_in_bias_to_fp16, weight = down_blocks_1_attentions_0_proj_in_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<int32, [1]> hidden_states_21_axes_0 = const()[name = tensor<string, []>("hidden_states_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_norm1_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15633024)))];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_norm1_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15634368)))];
            tensor<fp16, []> var_366_to_fp16 = const()[name = tensor<string, []>("op_366_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 4096, 640]> hidden_states_21_cast_fp16 = layer_norm(axes = hidden_states_21_axes_0, beta = down_blocks_1_attentions_0_transformer_blocks_0_norm1_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_0_transformer_blocks_0_norm1_weight_to_fp16, x = linear_7_cast_fp16)[name = tensor<string, []>("hidden_states_21_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15635712))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15942976))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_8_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_21_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15943168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16250432))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_9_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_21_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16250624))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16557888))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_10_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_21_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<int32, [4]> var_489 = const()[name = tensor<string, []>("op_489"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_490_cast_fp16 = reshape(shape = var_489, x = linear_8_cast_fp16)[name = tensor<string, []>("op_490_cast_fp16")];
            tensor<int32, [4]> var_492 = const()[name = tensor<string, []>("op_492"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_493_cast_fp16 = reshape(shape = var_492, x = linear_9_cast_fp16)[name = tensor<string, []>("op_493_cast_fp16")];
            tensor<int32, [4]> var_495 = const()[name = tensor<string, []>("op_495"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_496_cast_fp16 = reshape(shape = var_495, x = linear_10_cast_fp16)[name = tensor<string, []>("op_496_cast_fp16")];
            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_7_y_0_to_fp16 = const()[name = tensor<string, []>("mul_7_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4096, 10, 64]> mul_7_cast_fp16 = mul(x = var_490_cast_fp16, y = mul_7_y_0_to_fp16)[name = tensor<string, []>("mul_7_cast_fp16")];
            tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_272_perm_0 = const()[name = tensor<string, []>("transpose_272_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_273_perm_0 = const()[name = tensor<string, []>("transpose_273_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 10, 4096, 64]> transpose_687 = transpose(perm = transpose_273_perm_0, x = var_493_cast_fp16)[name = tensor<string, []>("transpose_687")];
            tensor<fp16, [1, 10, 4096, 64]> transpose_688 = transpose(perm = transpose_272_perm_0, x = mul_7_cast_fp16)[name = tensor<string, []>("transpose_688")];
            tensor<fp16, [1, 10, 4096, 4096]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = transpose_688, y = transpose_687)[name = tensor<string, []>("matmul_0_cast_fp16")];
            tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 10, 4096, 4096]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = matmul_0_cast_fp16)[name = tensor<string, []>("softmax_0_cast_fp16")];
            tensor<bool, []> hidden_states_23_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_23_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_23_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 10, 4096, 64]> transpose_686 = transpose(perm = value_3_perm_0, x = var_496_cast_fp16)[name = tensor<string, []>("transpose_686")];
            tensor<fp16, [1, 10, 4096, 64]> hidden_states_23_cast_fp16 = matmul(transpose_x = hidden_states_23_transpose_x_0, transpose_y = hidden_states_23_transpose_y_0, x = softmax_0_cast_fp16, y = transpose_686)[name = tensor<string, []>("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_499_perm_0 = const()[name = tensor<string, []>("op_499_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_503 = const()[name = tensor<string, []>("op_503"), val = tensor<int32, [3]>([1, -1, 640])];
            tensor<fp16, [1, 4096, 10, 64]> transpose_685 = transpose(perm = var_499_perm_0, x = hidden_states_23_cast_fp16)[name = tensor<string, []>("transpose_685")];
            tensor<fp16, [1, 4096, 640]> hidden_states_25_cast_fp16 = reshape(shape = var_503, x = transpose_685)[name = tensor<string, []>("hidden_states_25_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16558080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16865344))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16865536)))];
            tensor<fp16, [1, 4096, 640]> linear_11_cast_fp16 = linear(bias = down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_25_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_97_cast_fp16 = add(x = linear_11_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
            tensor<int32, [1]> input_99_axes_0 = const()[name = tensor<string, []>("input_99_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_norm2_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16866880)))];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_norm2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16868224)))];
            tensor<fp16, [1, 4096, 640]> input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = down_blocks_1_attentions_0_transformer_blocks_0_norm2_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_0_transformer_blocks_0_norm2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16869568))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17176832))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_12_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<fp16, [640, 2048]> down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [983040]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17177024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18160128))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2048])];
            tensor<fp16, [1, 77, 640]> linear_13_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<fp16, [640, 2048]> down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [983040]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18160320))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19143424))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2048])];
            tensor<fp16, [1, 77, 640]> linear_14_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<int32, [4]> var_535 = const()[name = tensor<string, []>("op_535"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_536_cast_fp16 = reshape(shape = var_535, x = linear_12_cast_fp16)[name = tensor<string, []>("op_536_cast_fp16")];
            tensor<int32, [4]> var_538 = const()[name = tensor<string, []>("op_538"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 77, 10, 64]> var_539_cast_fp16 = reshape(shape = var_538, x = linear_13_cast_fp16)[name = tensor<string, []>("op_539_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = tensor<string, []>("op_541"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 77, 10, 64]> var_542_cast_fp16 = reshape(shape = var_541, x = linear_14_cast_fp16)[name = tensor<string, []>("op_542_cast_fp16")];
            tensor<int32, [4]> value_7_perm_0 = const()[name = tensor<string, []>("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_8_y_0_to_fp16 = const()[name = tensor<string, []>("mul_8_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4096, 10, 64]> mul_8_cast_fp16 = mul(x = var_536_cast_fp16, y = mul_8_y_0_to_fp16)[name = tensor<string, []>("mul_8_cast_fp16")];
            tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_274_perm_0 = const()[name = tensor<string, []>("transpose_274_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_275_perm_0 = const()[name = tensor<string, []>("transpose_275_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 10, 77, 64]> transpose_683 = transpose(perm = transpose_275_perm_0, x = var_539_cast_fp16)[name = tensor<string, []>("transpose_683")];
            tensor<fp16, [1, 10, 4096, 64]> transpose_684 = transpose(perm = transpose_274_perm_0, x = mul_8_cast_fp16)[name = tensor<string, []>("transpose_684")];
            tensor<fp16, [1, 10, 4096, 77]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_684, y = transpose_683)[name = tensor<string, []>("matmul_1_cast_fp16")];
            tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 10, 4096, 77]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = matmul_1_cast_fp16)[name = tensor<string, []>("softmax_1_cast_fp16")];
            tensor<bool, []> hidden_states_29_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_29_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 10, 77, 64]> transpose_682 = transpose(perm = value_7_perm_0, x = var_542_cast_fp16)[name = tensor<string, []>("transpose_682")];
            tensor<fp16, [1, 10, 4096, 64]> hidden_states_29_cast_fp16 = matmul(transpose_x = hidden_states_29_transpose_x_0, transpose_y = hidden_states_29_transpose_y_0, x = softmax_1_cast_fp16, y = transpose_682)[name = tensor<string, []>("hidden_states_29_cast_fp16")];
            tensor<int32, [4]> var_545_perm_0 = const()[name = tensor<string, []>("op_545_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_549 = const()[name = tensor<string, []>("op_549"), val = tensor<int32, [3]>([1, -1, 640])];
            tensor<fp16, [1, 4096, 10, 64]> transpose_681 = transpose(perm = var_545_perm_0, x = hidden_states_29_cast_fp16)[name = tensor<string, []>("transpose_681")];
            tensor<fp16, [1, 4096, 640]> hidden_states_31_cast_fp16 = reshape(shape = var_549, x = transpose_681)[name = tensor<string, []>("hidden_states_31_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19143616))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19450880))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19451072)))];
            tensor<fp16, [1, 4096, 640]> linear_15_cast_fp16 = linear(bias = down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_31_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_105_cast_fp16 = add(x = linear_15_cast_fp16, y = input_97_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<int32, [1]> input_107_axes_0 = const()[name = tensor<string, []>("input_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_norm3_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19452416)))];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_norm3_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19453760)))];
            tensor<fp16, [1, 4096, 640]> input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = down_blocks_1_attentions_0_transformer_blocks_0_norm3_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_0_transformer_blocks_0_norm3_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<fp16, [5120, 640]> down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2457600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19455104))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21912768))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([5120, 640])];
            tensor<fp16, [5120]> down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3840]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21912960))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21916864))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([5120])];
            tensor<fp16, [1, 4096, 5120]> linear_16_cast_fp16 = linear(bias = down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<int32, [2]> var_571_split_sizes_0 = const()[name = tensor<string, []>("op_571_split_sizes_0"), val = tensor<int32, [2]>([2560, 2560])];
            tensor<int32, []> var_571_axis_0 = const()[name = tensor<string, []>("op_571_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4096, 2560]> var_571_cast_fp16_0, tensor<fp16, [1, 4096, 2560]> var_571_cast_fp16_1 = split(axis = var_571_axis_0, split_sizes = var_571_split_sizes_0, x = linear_16_cast_fp16)[name = tensor<string, []>("op_571_cast_fp16")];
            tensor<string, []> var_573_mode_0 = const()[name = tensor<string, []>("op_573_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 2560]> var_573_cast_fp16 = gelu(mode = var_573_mode_0, x = var_571_cast_fp16_1)[name = tensor<string, []>("op_573_cast_fp16")];
            tensor<fp16, [1, 4096, 2560]> input_109_cast_fp16 = mul(x = var_571_cast_fp16_0, y = var_573_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<fp16, [640, 2560]> down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21917056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23145920))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2560])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23146112)))];
            tensor<fp16, [1, 4096, 640]> linear_17_cast_fp16 = linear(bias = down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized, x = input_109_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> hidden_states_39_cast_fp16 = add(x = linear_17_cast_fp16, y = input_105_cast_fp16)[name = tensor<string, []>("hidden_states_39_cast_fp16")];
            tensor<int32, [1]> hidden_states_41_axes_0 = const()[name = tensor<string, []>("hidden_states_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_norm1_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23147456)))];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_norm1_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23148800)))];
            tensor<fp16, [1, 4096, 640]> hidden_states_41_cast_fp16 = layer_norm(axes = hidden_states_41_axes_0, beta = down_blocks_1_attentions_0_transformer_blocks_1_norm1_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_0_transformer_blocks_1_norm1_weight_to_fp16, x = hidden_states_39_cast_fp16)[name = tensor<string, []>("hidden_states_41_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23150144))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23457408))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_18_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_41_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23457600))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23764864))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_19_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_41_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23765056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24072320))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_20_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_41_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<int32, [4]> var_608 = const()[name = tensor<string, []>("op_608"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_609_cast_fp16 = reshape(shape = var_608, x = linear_18_cast_fp16)[name = tensor<string, []>("op_609_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = tensor<string, []>("op_611"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_19_cast_fp16)[name = tensor<string, []>("op_612_cast_fp16")];
            tensor<int32, [4]> var_614 = const()[name = tensor<string, []>("op_614"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_615_cast_fp16 = reshape(shape = var_614, x = linear_20_cast_fp16)[name = tensor<string, []>("op_615_cast_fp16")];
            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<string, []>("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_9_y_0_to_fp16 = const()[name = tensor<string, []>("mul_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4096, 10, 64]> mul_9_cast_fp16 = mul(x = var_609_cast_fp16, y = mul_9_y_0_to_fp16)[name = tensor<string, []>("mul_9_cast_fp16")];
            tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_276_perm_0 = const()[name = tensor<string, []>("transpose_276_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_277_perm_0 = const()[name = tensor<string, []>("transpose_277_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 10, 4096, 64]> transpose_679 = transpose(perm = transpose_277_perm_0, x = var_612_cast_fp16)[name = tensor<string, []>("transpose_679")];
            tensor<fp16, [1, 10, 4096, 64]> transpose_680 = transpose(perm = transpose_276_perm_0, x = mul_9_cast_fp16)[name = tensor<string, []>("transpose_680")];
            tensor<fp16, [1, 10, 4096, 4096]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = transpose_680, y = transpose_679)[name = tensor<string, []>("matmul_2_cast_fp16")];
            tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 10, 4096, 4096]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = matmul_2_cast_fp16)[name = tensor<string, []>("softmax_2_cast_fp16")];
            tensor<bool, []> hidden_states_43_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_43_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_43_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_43_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 10, 4096, 64]> transpose_678 = transpose(perm = value_11_perm_0, x = var_615_cast_fp16)[name = tensor<string, []>("transpose_678")];
            tensor<fp16, [1, 10, 4096, 64]> hidden_states_43_cast_fp16 = matmul(transpose_x = hidden_states_43_transpose_x_0, transpose_y = hidden_states_43_transpose_y_0, x = softmax_2_cast_fp16, y = transpose_678)[name = tensor<string, []>("hidden_states_43_cast_fp16")];
            tensor<int32, [4]> var_618_perm_0 = const()[name = tensor<string, []>("op_618_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_622 = const()[name = tensor<string, []>("op_622"), val = tensor<int32, [3]>([1, -1, 640])];
            tensor<fp16, [1, 4096, 10, 64]> transpose_677 = transpose(perm = var_618_perm_0, x = hidden_states_43_cast_fp16)[name = tensor<string, []>("transpose_677")];
            tensor<fp16, [1, 4096, 640]> hidden_states_45_cast_fp16 = reshape(shape = var_622, x = transpose_677)[name = tensor<string, []>("hidden_states_45_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24072512))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24379776))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24379968)))];
            tensor<fp16, [1, 4096, 640]> linear_21_cast_fp16 = linear(bias = down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_45_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_117_cast_fp16 = add(x = linear_21_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<int32, [1]> input_119_axes_0 = const()[name = tensor<string, []>("input_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_norm2_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24381312)))];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_norm2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24382656)))];
            tensor<fp16, [1, 4096, 640]> input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = down_blocks_1_attentions_0_transformer_blocks_1_norm2_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_0_transformer_blocks_1_norm2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24384000))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24691264))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_22_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized, x = input_119_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<fp16, [640, 2048]> down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [983040]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24691456))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25674560))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2048])];
            tensor<fp16, [1, 77, 640]> linear_23_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<fp16, [640, 2048]> down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [983040]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25674752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26657856))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2048])];
            tensor<fp16, [1, 77, 640]> linear_24_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<int32, [4]> var_654 = const()[name = tensor<string, []>("op_654"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_655_cast_fp16 = reshape(shape = var_654, x = linear_22_cast_fp16)[name = tensor<string, []>("op_655_cast_fp16")];
            tensor<int32, [4]> var_657 = const()[name = tensor<string, []>("op_657"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 77, 10, 64]> var_658_cast_fp16 = reshape(shape = var_657, x = linear_23_cast_fp16)[name = tensor<string, []>("op_658_cast_fp16")];
            tensor<int32, [4]> var_660 = const()[name = tensor<string, []>("op_660"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 77, 10, 64]> var_661_cast_fp16 = reshape(shape = var_660, x = linear_24_cast_fp16)[name = tensor<string, []>("op_661_cast_fp16")];
            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<string, []>("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_10_y_0_to_fp16 = const()[name = tensor<string, []>("mul_10_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4096, 10, 64]> mul_10_cast_fp16 = mul(x = var_655_cast_fp16, y = mul_10_y_0_to_fp16)[name = tensor<string, []>("mul_10_cast_fp16")];
            tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_278_perm_0 = const()[name = tensor<string, []>("transpose_278_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_279_perm_0 = const()[name = tensor<string, []>("transpose_279_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 10, 77, 64]> transpose_675 = transpose(perm = transpose_279_perm_0, x = var_658_cast_fp16)[name = tensor<string, []>("transpose_675")];
            tensor<fp16, [1, 10, 4096, 64]> transpose_676 = transpose(perm = transpose_278_perm_0, x = mul_10_cast_fp16)[name = tensor<string, []>("transpose_676")];
            tensor<fp16, [1, 10, 4096, 77]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = transpose_676, y = transpose_675)[name = tensor<string, []>("matmul_3_cast_fp16")];
            tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 10, 4096, 77]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = matmul_3_cast_fp16)[name = tensor<string, []>("softmax_3_cast_fp16")];
            tensor<bool, []> hidden_states_49_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_49_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_49_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_49_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 10, 77, 64]> transpose_674 = transpose(perm = value_15_perm_0, x = var_661_cast_fp16)[name = tensor<string, []>("transpose_674")];
            tensor<fp16, [1, 10, 4096, 64]> hidden_states_49_cast_fp16 = matmul(transpose_x = hidden_states_49_transpose_x_0, transpose_y = hidden_states_49_transpose_y_0, x = softmax_3_cast_fp16, y = transpose_674)[name = tensor<string, []>("hidden_states_49_cast_fp16")];
            tensor<int32, [4]> var_664_perm_0 = const()[name = tensor<string, []>("op_664_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_668 = const()[name = tensor<string, []>("op_668"), val = tensor<int32, [3]>([1, -1, 640])];
            tensor<fp16, [1, 4096, 10, 64]> transpose_673 = transpose(perm = var_664_perm_0, x = hidden_states_49_cast_fp16)[name = tensor<string, []>("transpose_673")];
            tensor<fp16, [1, 4096, 640]> hidden_states_51_cast_fp16 = reshape(shape = var_668, x = transpose_673)[name = tensor<string, []>("hidden_states_51_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26658048))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26965312))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26965504)))];
            tensor<fp16, [1, 4096, 640]> linear_25_cast_fp16 = linear(bias = down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_51_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_125_cast_fp16 = add(x = linear_25_cast_fp16, y = input_117_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
            tensor<int32, [1]> input_127_axes_0 = const()[name = tensor<string, []>("input_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_norm3_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26966848)))];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_norm3_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26968192)))];
            tensor<fp16, [1, 4096, 640]> input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = down_blocks_1_attentions_0_transformer_blocks_1_norm3_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_0_transformer_blocks_1_norm3_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<fp16, [5120, 640]> down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2457600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26969536))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29427200))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([5120, 640])];
            tensor<fp16, [5120]> down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3840]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29427392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29431296))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([5120])];
            tensor<fp16, [1, 4096, 5120]> linear_26_cast_fp16 = linear(bias = down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<int32, [2]> var_690_split_sizes_0 = const()[name = tensor<string, []>("op_690_split_sizes_0"), val = tensor<int32, [2]>([2560, 2560])];
            tensor<int32, []> var_690_axis_0 = const()[name = tensor<string, []>("op_690_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4096, 2560]> var_690_cast_fp16_0, tensor<fp16, [1, 4096, 2560]> var_690_cast_fp16_1 = split(axis = var_690_axis_0, split_sizes = var_690_split_sizes_0, x = linear_26_cast_fp16)[name = tensor<string, []>("op_690_cast_fp16")];
            tensor<string, []> var_692_mode_0 = const()[name = tensor<string, []>("op_692_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 2560]> var_692_cast_fp16 = gelu(mode = var_692_mode_0, x = var_690_cast_fp16_1)[name = tensor<string, []>("op_692_cast_fp16")];
            tensor<fp16, [1, 4096, 2560]> input_129_cast_fp16 = mul(x = var_690_cast_fp16_0, y = var_692_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<fp16, [640, 2560]> down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29431488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30660352))), name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2560])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30660544)))];
            tensor<fp16, [1, 4096, 640]> linear_27_cast_fp16 = linear(bias = down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16, weight = down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized, x = input_129_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_133_cast_fp16 = add(x = linear_27_cast_fp16, y = input_125_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_0_proj_out_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30661888))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30969152))), name = tensor<string, []>("down_blocks_1_attentions_0_proj_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_0_proj_out_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_0_proj_out_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30969344)))];
            tensor<fp16, [1, 4096, 640]> linear_28_cast_fp16 = linear(bias = down_blocks_1_attentions_0_proj_out_bias_to_fp16, weight = down_blocks_1_attentions_0_proj_out_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<int32, [4]> var_702 = const()[name = tensor<string, []>("op_702"), val = tensor<int32, [4]>([1, 64, 64, 640])];
            tensor<fp16, [1, 64, 64, 640]> var_703_cast_fp16 = reshape(shape = var_702, x = linear_28_cast_fp16)[name = tensor<string, []>("op_703_cast_fp16")];
            tensor<int32, [4]> var_704 = const()[name = tensor<string, []>("op_704"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 640, 64, 64]> transpose_672 = transpose(perm = var_704, x = var_703_cast_fp16)[name = tensor<string, []>("transpose_672")];
            tensor<fp16, [1, 640, 64, 64]> input_135_cast_fp16 = add(x = transpose_672, y = var_430_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<int32, [5]> reshape_28_shape_0 = const()[name = tensor<string, []>("reshape_28_shape_0"), val = tensor<int32, [5]>([1, 32, 20, 64, 64])];
            tensor<fp16, [1, 32, 20, 64, 64]> reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = input_135_cast_fp16)[name = tensor<string, []>("reshape_28_cast_fp16")];
            tensor<int32, [3]> reduce_mean_21_axes_0 = const()[name = tensor<string, []>("reduce_mean_21_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_21_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_21_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = reshape_28_cast_fp16)[name = tensor<string, []>("reduce_mean_21_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> sub_14_cast_fp16 = sub(x = reshape_28_cast_fp16, y = reduce_mean_21_cast_fp16)[name = tensor<string, []>("sub_14_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> square_7_cast_fp16 = square(x = sub_14_cast_fp16)[name = tensor<string, []>("square_7_cast_fp16")];
            tensor<int32, [3]> reduce_mean_23_axes_0 = const()[name = tensor<string, []>("reduce_mean_23_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_23_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_23_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_7_cast_fp16)[name = tensor<string, []>("reduce_mean_23_cast_fp16")];
            tensor<fp16, []> add_14_y_0_to_fp16 = const()[name = tensor<string, []>("add_14_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_14_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = add_14_y_0_to_fp16)[name = tensor<string, []>("add_14_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_7_cast_fp16 = sqrt(x = add_14_cast_fp16)[name = tensor<string, []>("sqrt_7_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> real_div_7_cast_fp16 = real_div(x = sub_14_cast_fp16, y = sqrt_7_cast_fp16)[name = tensor<string, []>("real_div_7_cast_fp16")];
            tensor<int32, [4]> reshape_29_shape_0 = const()[name = tensor<string, []>("reshape_29_shape_0"), val = tensor<int32, [4]>([1, 640, 64, 64])];
            tensor<fp16, [1, 640, 64, 64]> reshape_29_cast_fp16 = reshape(shape = reshape_29_shape_0, x = real_div_7_cast_fp16)[name = tensor<string, []>("reshape_29_cast_fp16")];
            tensor<fp16, [640]> add_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_15_gamma_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30970688)))];
            tensor<fp16, [640]> add_15_beta_0_to_fp16 = const()[name = tensor<string, []>("add_15_beta_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30972032)))];
            tensor<fp16, []> add_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 640, 64, 64]> add_15_cast_fp16 = batch_norm(beta = add_15_beta_0_to_fp16, epsilon = add_15_epsilon_0_to_fp16, gamma = add_15_gamma_0_to_fp16, mean = add_11_mean_0_to_fp16, variance = add_11_variance_0_to_fp16, x = reshape_29_cast_fp16)[name = tensor<string, []>("add_15_cast_fp16")];
            tensor<fp16, [1, 640, 64, 64]> input_139_cast_fp16 = silu(x = add_15_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
            tensor<int32, [2]> var_719 = const()[name = tensor<string, []>("op_719"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_721 = const()[name = tensor<string, []>("op_721"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_63_pad_type_0 = const()[name = tensor<string, []>("hidden_states_63_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = tensor<string, []>("hidden_states_63_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [640, 640, 3, 3]> down_blocks_1_resnets_1_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2764800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30973376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33738240))), name = tensor<string, []>("down_blocks_1_resnets_1_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 640, 3, 3])];
            tensor<fp16, [640]> down_blocks_1_resnets_1_conv1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_resnets_1_conv1_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33738432)))];
            tensor<fp16, [1, 640, 64, 64]> hidden_states_63_cast_fp16 = conv(bias = down_blocks_1_resnets_1_conv1_bias_to_fp16, dilations = var_721, groups = var_370, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = var_719, weight = down_blocks_1_resnets_1_conv1_weight_to_fp16_palettized, x = input_139_cast_fp16)[name = tensor<string, []>("hidden_states_63_cast_fp16")];
            tensor<fp16, [640, 1280]> down_blocks_1_resnets_1_time_emb_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [614400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33739776))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34354240))), name = tensor<string, []>("down_blocks_1_resnets_1_time_emb_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 1280])];
            tensor<fp16, [640]> down_blocks_1_resnets_1_time_emb_proj_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_resnets_1_time_emb_proj_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34354432)))];
            tensor<fp16, [1, 640]> linear_29_cast_fp16 = linear(bias = down_blocks_1_resnets_1_time_emb_proj_bias_to_fp16, weight = down_blocks_1_resnets_1_time_emb_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<int32, [1]> var_730_axes_0 = const()[name = tensor<string, []>("op_730_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 640, 1]> var_730_cast_fp16 = expand_dims(axes = var_730_axes_0, x = linear_29_cast_fp16)[name = tensor<string, []>("op_730_cast_fp16")];
            tensor<int32, [1]> temb_7_axes_0 = const()[name = tensor<string, []>("temb_7_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 640, 1, 1]> temb_7_cast_fp16 = expand_dims(axes = temb_7_axes_0, x = var_730_cast_fp16)[name = tensor<string, []>("temb_7_cast_fp16")];
            tensor<fp16, [1, 640, 64, 64]> input_143_cast_fp16 = add(x = hidden_states_63_cast_fp16, y = temb_7_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<int32, [5]> reshape_32_shape_0 = const()[name = tensor<string, []>("reshape_32_shape_0"), val = tensor<int32, [5]>([1, 32, 20, 64, 64])];
            tensor<fp16, [1, 32, 20, 64, 64]> reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = input_143_cast_fp16)[name = tensor<string, []>("reshape_32_cast_fp16")];
            tensor<int32, [3]> reduce_mean_24_axes_0 = const()[name = tensor<string, []>("reduce_mean_24_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_24_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_24_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_24_cast_fp16 = reduce_mean(axes = reduce_mean_24_axes_0, keep_dims = reduce_mean_24_keep_dims_0, x = reshape_32_cast_fp16)[name = tensor<string, []>("reduce_mean_24_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> sub_16_cast_fp16 = sub(x = reshape_32_cast_fp16, y = reduce_mean_24_cast_fp16)[name = tensor<string, []>("sub_16_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> square_8_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor<string, []>("square_8_cast_fp16")];
            tensor<int32, [3]> reduce_mean_26_axes_0 = const()[name = tensor<string, []>("reduce_mean_26_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_26_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_26_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_26_cast_fp16 = reduce_mean(axes = reduce_mean_26_axes_0, keep_dims = reduce_mean_26_keep_dims_0, x = square_8_cast_fp16)[name = tensor<string, []>("reduce_mean_26_cast_fp16")];
            tensor<fp16, []> add_16_y_0_to_fp16 = const()[name = tensor<string, []>("add_16_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_16_cast_fp16 = add(x = reduce_mean_26_cast_fp16, y = add_16_y_0_to_fp16)[name = tensor<string, []>("add_16_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_8_cast_fp16 = sqrt(x = add_16_cast_fp16)[name = tensor<string, []>("sqrt_8_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> real_div_8_cast_fp16 = real_div(x = sub_16_cast_fp16, y = sqrt_8_cast_fp16)[name = tensor<string, []>("real_div_8_cast_fp16")];
            tensor<int32, [4]> reshape_33_shape_0 = const()[name = tensor<string, []>("reshape_33_shape_0"), val = tensor<int32, [4]>([1, 640, 64, 64])];
            tensor<fp16, [1, 640, 64, 64]> reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = real_div_8_cast_fp16)[name = tensor<string, []>("reshape_33_cast_fp16")];
            tensor<fp16, [640]> add_17_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_17_gamma_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34355776)))];
            tensor<fp16, [640]> add_17_beta_0_to_fp16 = const()[name = tensor<string, []>("add_17_beta_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34357120)))];
            tensor<fp16, []> add_17_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_17_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 640, 64, 64]> add_17_cast_fp16 = batch_norm(beta = add_17_beta_0_to_fp16, epsilon = add_17_epsilon_0_to_fp16, gamma = add_17_gamma_0_to_fp16, mean = add_11_mean_0_to_fp16, variance = add_11_variance_0_to_fp16, x = reshape_33_cast_fp16)[name = tensor<string, []>("add_17_cast_fp16")];
            tensor<fp16, [1, 640, 64, 64]> input_147_cast_fp16 = silu(x = add_17_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<int32, [2]> var_740 = const()[name = tensor<string, []>("op_740"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_742 = const()[name = tensor<string, []>("op_742"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_65_pad_type_0 = const()[name = tensor<string, []>("hidden_states_65_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_65_pad_0 = const()[name = tensor<string, []>("hidden_states_65_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [640, 640, 3, 3]> down_blocks_1_resnets_1_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2764800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34358464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37123328))), name = tensor<string, []>("down_blocks_1_resnets_1_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 640, 3, 3])];
            tensor<fp16, [640]> down_blocks_1_resnets_1_conv2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_resnets_1_conv2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37123520)))];
            tensor<fp16, [1, 640, 64, 64]> hidden_states_65_cast_fp16 = conv(bias = down_blocks_1_resnets_1_conv2_bias_to_fp16, dilations = var_742, groups = var_370, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = var_740, weight = down_blocks_1_resnets_1_conv2_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = tensor<string, []>("hidden_states_65_cast_fp16")];
            tensor<fp16, [1, 640, 64, 64]> var_745_cast_fp16 = add(x = input_135_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor<string, []>("op_745_cast_fp16")];
            tensor<int32, [5]> reshape_36_shape_0 = const()[name = tensor<string, []>("reshape_36_shape_0"), val = tensor<int32, [5]>([1, 32, 20, 64, 64])];
            tensor<fp16, [1, 32, 20, 64, 64]> reshape_36_cast_fp16 = reshape(shape = reshape_36_shape_0, x = var_745_cast_fp16)[name = tensor<string, []>("reshape_36_cast_fp16")];
            tensor<int32, [3]> reduce_mean_27_axes_0 = const()[name = tensor<string, []>("reduce_mean_27_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_27_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_27_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = reshape_36_cast_fp16)[name = tensor<string, []>("reduce_mean_27_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> sub_18_cast_fp16 = sub(x = reshape_36_cast_fp16, y = reduce_mean_27_cast_fp16)[name = tensor<string, []>("sub_18_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> square_9_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor<string, []>("square_9_cast_fp16")];
            tensor<int32, [3]> reduce_mean_29_axes_0 = const()[name = tensor<string, []>("reduce_mean_29_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_29_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_29_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_9_cast_fp16)[name = tensor<string, []>("reduce_mean_29_cast_fp16")];
            tensor<fp16, []> add_18_y_0_to_fp16 = const()[name = tensor<string, []>("add_18_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_18_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = add_18_y_0_to_fp16)[name = tensor<string, []>("add_18_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_9_cast_fp16 = sqrt(x = add_18_cast_fp16)[name = tensor<string, []>("sqrt_9_cast_fp16")];
            tensor<fp16, [1, 32, 20, 64, 64]> real_div_9_cast_fp16 = real_div(x = sub_18_cast_fp16, y = sqrt_9_cast_fp16)[name = tensor<string, []>("real_div_9_cast_fp16")];
            tensor<int32, [4]> reshape_37_shape_0 = const()[name = tensor<string, []>("reshape_37_shape_0"), val = tensor<int32, [4]>([1, 640, 64, 64])];
            tensor<fp16, [1, 640, 64, 64]> reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = real_div_9_cast_fp16)[name = tensor<string, []>("reshape_37_cast_fp16")];
            tensor<fp16, [640]> add_19_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_19_gamma_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37124864)))];
            tensor<fp16, [640]> add_19_beta_0_to_fp16 = const()[name = tensor<string, []>("add_19_beta_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37126208)))];
            tensor<fp16, []> add_19_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_19_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 640, 64, 64]> add_19_cast_fp16 = batch_norm(beta = add_19_beta_0_to_fp16, epsilon = add_19_epsilon_0_to_fp16, gamma = add_19_gamma_0_to_fp16, mean = add_11_mean_0_to_fp16, variance = add_11_variance_0_to_fp16, x = reshape_37_cast_fp16)[name = tensor<string, []>("add_19_cast_fp16")];
            tensor<int32, [4]> var_767 = const()[name = tensor<string, []>("op_767"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_771 = const()[name = tensor<string, []>("op_771"), val = tensor<int32, [3]>([1, 4096, 640])];
            tensor<fp16, [1, 64, 64, 640]> transpose_671 = transpose(perm = var_767, x = add_19_cast_fp16)[name = tensor<string, []>("transpose_671")];
            tensor<fp16, [1, 4096, 640]> input_151_cast_fp16 = reshape(shape = var_771, x = transpose_671)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_proj_in_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37127552))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37434816))), name = tensor<string, []>("down_blocks_1_attentions_1_proj_in_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_proj_in_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_proj_in_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37435008)))];
            tensor<fp16, [1, 4096, 640]> linear_30_cast_fp16 = linear(bias = down_blocks_1_attentions_1_proj_in_bias_to_fp16, weight = down_blocks_1_attentions_1_proj_in_weight_to_fp16_palettized, x = input_151_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<int32, [1]> hidden_states_73_axes_0 = const()[name = tensor<string, []>("hidden_states_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_norm1_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37436352)))];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_norm1_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37437696)))];
            tensor<fp16, [1, 4096, 640]> hidden_states_73_cast_fp16 = layer_norm(axes = hidden_states_73_axes_0, beta = down_blocks_1_attentions_1_transformer_blocks_0_norm1_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_1_transformer_blocks_0_norm1_weight_to_fp16, x = linear_30_cast_fp16)[name = tensor<string, []>("hidden_states_73_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37439040))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37746304))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_31_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_73_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37746496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38053760))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_32_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_73_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38053952))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38361216))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_33_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_73_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<int32, [4]> var_804 = const()[name = tensor<string, []>("op_804"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_805_cast_fp16 = reshape(shape = var_804, x = linear_31_cast_fp16)[name = tensor<string, []>("op_805_cast_fp16")];
            tensor<int32, [4]> var_807 = const()[name = tensor<string, []>("op_807"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_808_cast_fp16 = reshape(shape = var_807, x = linear_32_cast_fp16)[name = tensor<string, []>("op_808_cast_fp16")];
            tensor<int32, [4]> var_810 = const()[name = tensor<string, []>("op_810"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_811_cast_fp16 = reshape(shape = var_810, x = linear_33_cast_fp16)[name = tensor<string, []>("op_811_cast_fp16")];
            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<string, []>("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_14_y_0_to_fp16 = const()[name = tensor<string, []>("mul_14_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4096, 10, 64]> mul_14_cast_fp16 = mul(x = var_805_cast_fp16, y = mul_14_y_0_to_fp16)[name = tensor<string, []>("mul_14_cast_fp16")];
            tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_280_perm_0 = const()[name = tensor<string, []>("transpose_280_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_281_perm_0 = const()[name = tensor<string, []>("transpose_281_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 10, 4096, 64]> transpose_669 = transpose(perm = transpose_281_perm_0, x = var_808_cast_fp16)[name = tensor<string, []>("transpose_669")];
            tensor<fp16, [1, 10, 4096, 64]> transpose_670 = transpose(perm = transpose_280_perm_0, x = mul_14_cast_fp16)[name = tensor<string, []>("transpose_670")];
            tensor<fp16, [1, 10, 4096, 4096]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = transpose_670, y = transpose_669)[name = tensor<string, []>("matmul_4_cast_fp16")];
            tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 10, 4096, 4096]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = matmul_4_cast_fp16)[name = tensor<string, []>("softmax_4_cast_fp16")];
            tensor<bool, []> hidden_states_75_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_75_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_75_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_75_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 10, 4096, 64]> transpose_668 = transpose(perm = value_19_perm_0, x = var_811_cast_fp16)[name = tensor<string, []>("transpose_668")];
            tensor<fp16, [1, 10, 4096, 64]> hidden_states_75_cast_fp16 = matmul(transpose_x = hidden_states_75_transpose_x_0, transpose_y = hidden_states_75_transpose_y_0, x = softmax_4_cast_fp16, y = transpose_668)[name = tensor<string, []>("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_814_perm_0 = const()[name = tensor<string, []>("op_814_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, [3]>([1, -1, 640])];
            tensor<fp16, [1, 4096, 10, 64]> transpose_667 = transpose(perm = var_814_perm_0, x = hidden_states_75_cast_fp16)[name = tensor<string, []>("transpose_667")];
            tensor<fp16, [1, 4096, 640]> hidden_states_77_cast_fp16 = reshape(shape = var_818, x = transpose_667)[name = tensor<string, []>("hidden_states_77_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38361408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38668672))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38668864)))];
            tensor<fp16, [1, 4096, 640]> linear_34_cast_fp16 = linear(bias = down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0_bias_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_77_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_157_cast_fp16 = add(x = linear_34_cast_fp16, y = linear_30_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<int32, [1]> input_159_axes_0 = const()[name = tensor<string, []>("input_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_norm2_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38670208)))];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_norm2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38671552)))];
            tensor<fp16, [1, 4096, 640]> input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = down_blocks_1_attentions_1_transformer_blocks_0_norm2_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_1_transformer_blocks_0_norm2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38672896))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38980160))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_35_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized, x = input_159_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<fp16, [640, 2048]> down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [983040]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38980352))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39963456))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2048])];
            tensor<fp16, [1, 77, 640]> linear_36_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<fp16, [640, 2048]> down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [983040]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39963648))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40946752))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2048])];
            tensor<fp16, [1, 77, 640]> linear_37_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<int32, [4]> var_850 = const()[name = tensor<string, []>("op_850"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_851_cast_fp16 = reshape(shape = var_850, x = linear_35_cast_fp16)[name = tensor<string, []>("op_851_cast_fp16")];
            tensor<int32, [4]> var_853 = const()[name = tensor<string, []>("op_853"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 77, 10, 64]> var_854_cast_fp16 = reshape(shape = var_853, x = linear_36_cast_fp16)[name = tensor<string, []>("op_854_cast_fp16")];
            tensor<int32, [4]> var_856 = const()[name = tensor<string, []>("op_856"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 77, 10, 64]> var_857_cast_fp16 = reshape(shape = var_856, x = linear_37_cast_fp16)[name = tensor<string, []>("op_857_cast_fp16")];
            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<string, []>("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_15_y_0_to_fp16 = const()[name = tensor<string, []>("mul_15_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4096, 10, 64]> mul_15_cast_fp16 = mul(x = var_851_cast_fp16, y = mul_15_y_0_to_fp16)[name = tensor<string, []>("mul_15_cast_fp16")];
            tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_282_perm_0 = const()[name = tensor<string, []>("transpose_282_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_283_perm_0 = const()[name = tensor<string, []>("transpose_283_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 10, 77, 64]> transpose_665 = transpose(perm = transpose_283_perm_0, x = var_854_cast_fp16)[name = tensor<string, []>("transpose_665")];
            tensor<fp16, [1, 10, 4096, 64]> transpose_666 = transpose(perm = transpose_282_perm_0, x = mul_15_cast_fp16)[name = tensor<string, []>("transpose_666")];
            tensor<fp16, [1, 10, 4096, 77]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = transpose_666, y = transpose_665)[name = tensor<string, []>("matmul_5_cast_fp16")];
            tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 10, 4096, 77]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = matmul_5_cast_fp16)[name = tensor<string, []>("softmax_5_cast_fp16")];
            tensor<bool, []> hidden_states_81_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_81_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_81_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_81_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 10, 77, 64]> transpose_664 = transpose(perm = value_23_perm_0, x = var_857_cast_fp16)[name = tensor<string, []>("transpose_664")];
            tensor<fp16, [1, 10, 4096, 64]> hidden_states_81_cast_fp16 = matmul(transpose_x = hidden_states_81_transpose_x_0, transpose_y = hidden_states_81_transpose_y_0, x = softmax_5_cast_fp16, y = transpose_664)[name = tensor<string, []>("hidden_states_81_cast_fp16")];
            tensor<int32, [4]> var_860_perm_0 = const()[name = tensor<string, []>("op_860_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_864 = const()[name = tensor<string, []>("op_864"), val = tensor<int32, [3]>([1, -1, 640])];
            tensor<fp16, [1, 4096, 10, 64]> transpose_663 = transpose(perm = var_860_perm_0, x = hidden_states_81_cast_fp16)[name = tensor<string, []>("transpose_663")];
            tensor<fp16, [1, 4096, 640]> hidden_states_83_cast_fp16 = reshape(shape = var_864, x = transpose_663)[name = tensor<string, []>("hidden_states_83_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40946944))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41254208))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41254400)))];
            tensor<fp16, [1, 4096, 640]> linear_38_cast_fp16 = linear(bias = down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0_bias_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_83_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_165_cast_fp16 = add(x = linear_38_cast_fp16, y = input_157_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<int32, [1]> input_167_axes_0 = const()[name = tensor<string, []>("input_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_norm3_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41255744)))];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_norm3_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41257088)))];
            tensor<fp16, [1, 4096, 640]> input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = down_blocks_1_attentions_1_transformer_blocks_0_norm3_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_1_transformer_blocks_0_norm3_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<fp16, [5120, 640]> down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2457600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41258432))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43716096))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([5120, 640])];
            tensor<fp16, [5120]> down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3840]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43716288))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43720192))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([5120])];
            tensor<fp16, [1, 4096, 5120]> linear_39_cast_fp16 = linear(bias = down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<int32, [2]> var_886_split_sizes_0 = const()[name = tensor<string, []>("op_886_split_sizes_0"), val = tensor<int32, [2]>([2560, 2560])];
            tensor<int32, []> var_886_axis_0 = const()[name = tensor<string, []>("op_886_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4096, 2560]> var_886_cast_fp16_0, tensor<fp16, [1, 4096, 2560]> var_886_cast_fp16_1 = split(axis = var_886_axis_0, split_sizes = var_886_split_sizes_0, x = linear_39_cast_fp16)[name = tensor<string, []>("op_886_cast_fp16")];
            tensor<string, []> var_888_mode_0 = const()[name = tensor<string, []>("op_888_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 2560]> var_888_cast_fp16 = gelu(mode = var_888_mode_0, x = var_886_cast_fp16_1)[name = tensor<string, []>("op_888_cast_fp16")];
            tensor<fp16, [1, 4096, 2560]> input_169_cast_fp16 = mul(x = var_886_cast_fp16_0, y = var_888_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<fp16, [640, 2560]> down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43720384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44949248))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2560])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44949440)))];
            tensor<fp16, [1, 4096, 640]> linear_40_cast_fp16 = linear(bias = down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2_bias_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> hidden_states_91_cast_fp16 = add(x = linear_40_cast_fp16, y = input_165_cast_fp16)[name = tensor<string, []>("hidden_states_91_cast_fp16")];
            tensor<int32, [1]> hidden_states_93_axes_0 = const()[name = tensor<string, []>("hidden_states_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_norm1_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44950784)))];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_norm1_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44952128)))];
            tensor<fp16, [1, 4096, 640]> hidden_states_93_cast_fp16 = layer_norm(axes = hidden_states_93_axes_0, beta = down_blocks_1_attentions_1_transformer_blocks_1_norm1_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_1_transformer_blocks_1_norm1_weight_to_fp16, x = hidden_states_91_cast_fp16)[name = tensor<string, []>("hidden_states_93_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44953472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45260736))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_41_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_93_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45260928))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45568192))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_42_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_93_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45568384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45875648))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_43_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_93_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<int32, [4]> var_923 = const()[name = tensor<string, []>("op_923"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_924_cast_fp16 = reshape(shape = var_923, x = linear_41_cast_fp16)[name = tensor<string, []>("op_924_cast_fp16")];
            tensor<int32, [4]> var_926 = const()[name = tensor<string, []>("op_926"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_927_cast_fp16 = reshape(shape = var_926, x = linear_42_cast_fp16)[name = tensor<string, []>("op_927_cast_fp16")];
            tensor<int32, [4]> var_929 = const()[name = tensor<string, []>("op_929"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_930_cast_fp16 = reshape(shape = var_929, x = linear_43_cast_fp16)[name = tensor<string, []>("op_930_cast_fp16")];
            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<string, []>("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_16_y_0_to_fp16 = const()[name = tensor<string, []>("mul_16_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4096, 10, 64]> mul_16_cast_fp16 = mul(x = var_924_cast_fp16, y = mul_16_y_0_to_fp16)[name = tensor<string, []>("mul_16_cast_fp16")];
            tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_284_perm_0 = const()[name = tensor<string, []>("transpose_284_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_285_perm_0 = const()[name = tensor<string, []>("transpose_285_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 10, 4096, 64]> transpose_661 = transpose(perm = transpose_285_perm_0, x = var_927_cast_fp16)[name = tensor<string, []>("transpose_661")];
            tensor<fp16, [1, 10, 4096, 64]> transpose_662 = transpose(perm = transpose_284_perm_0, x = mul_16_cast_fp16)[name = tensor<string, []>("transpose_662")];
            tensor<fp16, [1, 10, 4096, 4096]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = transpose_662, y = transpose_661)[name = tensor<string, []>("matmul_6_cast_fp16")];
            tensor<int32, []> softmax_6_axis_0 = const()[name = tensor<string, []>("softmax_6_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 10, 4096, 4096]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = matmul_6_cast_fp16)[name = tensor<string, []>("softmax_6_cast_fp16")];
            tensor<bool, []> hidden_states_95_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_95_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_95_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_95_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 10, 4096, 64]> transpose_660 = transpose(perm = value_27_perm_0, x = var_930_cast_fp16)[name = tensor<string, []>("transpose_660")];
            tensor<fp16, [1, 10, 4096, 64]> hidden_states_95_cast_fp16 = matmul(transpose_x = hidden_states_95_transpose_x_0, transpose_y = hidden_states_95_transpose_y_0, x = softmax_6_cast_fp16, y = transpose_660)[name = tensor<string, []>("hidden_states_95_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = tensor<string, []>("op_933_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_937 = const()[name = tensor<string, []>("op_937"), val = tensor<int32, [3]>([1, -1, 640])];
            tensor<fp16, [1, 4096, 10, 64]> transpose_659 = transpose(perm = var_933_perm_0, x = hidden_states_95_cast_fp16)[name = tensor<string, []>("transpose_659")];
            tensor<fp16, [1, 4096, 640]> hidden_states_97_cast_fp16 = reshape(shape = var_937, x = transpose_659)[name = tensor<string, []>("hidden_states_97_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45875840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46183104))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46183296)))];
            tensor<fp16, [1, 4096, 640]> linear_44_cast_fp16 = linear(bias = down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0_bias_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_97_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_177_cast_fp16 = add(x = linear_44_cast_fp16, y = hidden_states_91_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
            tensor<int32, [1]> input_179_axes_0 = const()[name = tensor<string, []>("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_norm2_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46184640)))];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_norm2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46185984)))];
            tensor<fp16, [1, 4096, 640]> input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = down_blocks_1_attentions_1_transformer_blocks_1_norm2_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_1_transformer_blocks_1_norm2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46187328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46494592))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [1, 4096, 640]> linear_45_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized, x = input_179_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<fp16, [640, 2048]> down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [983040]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46494784))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47477888))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2048])];
            tensor<fp16, [1, 77, 640]> linear_46_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<fp16, [640, 2048]> down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [983040]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47478080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48461184))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2048])];
            tensor<fp16, [1, 77, 640]> linear_47_cast_fp16 = linear(bias = add_11_mean_0_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<int32, [4]> var_969 = const()[name = tensor<string, []>("op_969"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 4096, 10, 64]> var_970_cast_fp16 = reshape(shape = var_969, x = linear_45_cast_fp16)[name = tensor<string, []>("op_970_cast_fp16")];
            tensor<int32, [4]> var_972 = const()[name = tensor<string, []>("op_972"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 77, 10, 64]> var_973_cast_fp16 = reshape(shape = var_972, x = linear_46_cast_fp16)[name = tensor<string, []>("op_973_cast_fp16")];
            tensor<int32, [4]> var_975 = const()[name = tensor<string, []>("op_975"), val = tensor<int32, [4]>([1, -1, 10, 64])];
            tensor<fp16, [1, 77, 10, 64]> var_976_cast_fp16 = reshape(shape = var_975, x = linear_47_cast_fp16)[name = tensor<string, []>("op_976_cast_fp16")];
            tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<string, []>("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_17_y_0_to_fp16 = const()[name = tensor<string, []>("mul_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4096, 10, 64]> mul_17_cast_fp16 = mul(x = var_970_cast_fp16, y = mul_17_y_0_to_fp16)[name = tensor<string, []>("mul_17_cast_fp16")];
            tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_286_perm_0 = const()[name = tensor<string, []>("transpose_286_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_287_perm_0 = const()[name = tensor<string, []>("transpose_287_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 10, 77, 64]> transpose_657 = transpose(perm = transpose_287_perm_0, x = var_973_cast_fp16)[name = tensor<string, []>("transpose_657")];
            tensor<fp16, [1, 10, 4096, 64]> transpose_658 = transpose(perm = transpose_286_perm_0, x = mul_17_cast_fp16)[name = tensor<string, []>("transpose_658")];
            tensor<fp16, [1, 10, 4096, 77]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = transpose_658, y = transpose_657)[name = tensor<string, []>("matmul_7_cast_fp16")];
            tensor<int32, []> softmax_7_axis_0 = const()[name = tensor<string, []>("softmax_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 10, 4096, 77]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = matmul_7_cast_fp16)[name = tensor<string, []>("softmax_7_cast_fp16")];
            tensor<bool, []> hidden_states_101_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_101_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_101_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_101_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 10, 77, 64]> transpose_656 = transpose(perm = value_31_perm_0, x = var_976_cast_fp16)[name = tensor<string, []>("transpose_656")];
            tensor<fp16, [1, 10, 4096, 64]> hidden_states_101_cast_fp16 = matmul(transpose_x = hidden_states_101_transpose_x_0, transpose_y = hidden_states_101_transpose_y_0, x = softmax_7_cast_fp16, y = transpose_656)[name = tensor<string, []>("hidden_states_101_cast_fp16")];
            tensor<int32, [4]> var_979_perm_0 = const()[name = tensor<string, []>("op_979_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_983 = const()[name = tensor<string, []>("op_983"), val = tensor<int32, [3]>([1, -1, 640])];
            tensor<fp16, [1, 4096, 10, 64]> transpose_655 = transpose(perm = var_979_perm_0, x = hidden_states_101_cast_fp16)[name = tensor<string, []>("transpose_655")];
            tensor<fp16, [1, 4096, 640]> hidden_states_103_cast_fp16 = reshape(shape = var_983, x = transpose_655)[name = tensor<string, []>("hidden_states_103_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48461376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48768640))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48768832)))];
            tensor<fp16, [1, 4096, 640]> linear_48_cast_fp16 = linear(bias = down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0_bias_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_103_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_185_cast_fp16 = add(x = linear_48_cast_fp16, y = input_177_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
            tensor<int32, [1]> input_187_axes_0 = const()[name = tensor<string, []>("input_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_norm3_weight_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48770176)))];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_norm3_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48771520)))];
            tensor<fp16, [1, 4096, 640]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = down_blocks_1_attentions_1_transformer_blocks_1_norm3_bias_to_fp16, epsilon = var_366_to_fp16, gamma = down_blocks_1_attentions_1_transformer_blocks_1_norm3_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<fp16, [5120, 640]> down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2457600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48772864))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51230528))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([5120, 640])];
            tensor<fp16, [5120]> down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3840]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51230720))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51234624))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([5120])];
            tensor<fp16, [1, 4096, 5120]> linear_49_cast_fp16 = linear(bias = down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<int32, [2]> var_1005_split_sizes_0 = const()[name = tensor<string, []>("op_1005_split_sizes_0"), val = tensor<int32, [2]>([2560, 2560])];
            tensor<int32, []> var_1005_axis_0 = const()[name = tensor<string, []>("op_1005_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4096, 2560]> var_1005_cast_fp16_0, tensor<fp16, [1, 4096, 2560]> var_1005_cast_fp16_1 = split(axis = var_1005_axis_0, split_sizes = var_1005_split_sizes_0, x = linear_49_cast_fp16)[name = tensor<string, []>("op_1005_cast_fp16")];
            tensor<string, []> var_1007_mode_0 = const()[name = tensor<string, []>("op_1007_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 2560]> var_1007_cast_fp16 = gelu(mode = var_1007_mode_0, x = var_1005_cast_fp16_1)[name = tensor<string, []>("op_1007_cast_fp16")];
            tensor<fp16, [1, 4096, 2560]> input_189_cast_fp16 = mul(x = var_1005_cast_fp16_0, y = var_1007_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<fp16, [640, 2560]> down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51234816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52463680))), name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 2560])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52463872)))];
            tensor<fp16, [1, 4096, 640]> linear_50_cast_fp16 = linear(bias = down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2_bias_to_fp16, weight = down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized, x = input_189_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<fp16, [1, 4096, 640]> input_193_cast_fp16 = add(x = linear_50_cast_fp16, y = input_185_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<fp16, [640, 640]> down_blocks_1_attentions_1_proj_out_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52465216))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52772480))), name = tensor<string, []>("down_blocks_1_attentions_1_proj_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([640, 640])];
            tensor<fp16, [640]> down_blocks_1_attentions_1_proj_out_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_attentions_1_proj_out_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52772672)))];
            tensor<fp16, [1, 4096, 640]> linear_51_cast_fp16 = linear(bias = down_blocks_1_attentions_1_proj_out_bias_to_fp16, weight = down_blocks_1_attentions_1_proj_out_weight_to_fp16_palettized, x = input_193_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1017 = const()[name = tensor<string, []>("op_1017"), val = tensor<int32, [4]>([1, 64, 64, 640])];
            tensor<fp16, [1, 64, 64, 640]> var_1018_cast_fp16 = reshape(shape = var_1017, x = linear_51_cast_fp16)[name = tensor<string, []>("op_1018_cast_fp16")];
            tensor<int32, [4]> var_1019 = const()[name = tensor<string, []>("op_1019"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 640, 64, 64]> transpose_654 = transpose(perm = var_1019, x = var_1018_cast_fp16)[name = tensor<string, []>("transpose_654")];
            tensor<fp16, [1, 640, 64, 64]> hidden_states_115_cast_fp16 = add(x = transpose_654, y = var_745_cast_fp16)[name = tensor<string, []>("hidden_states_115_cast_fp16")];
            tensor<int32, [2]> var_1026 = const()[name = tensor<string, []>("op_1026"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_1028 = const()[name = tensor<string, []>("op_1028"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_195_pad_type_0 = const()[name = tensor<string, []>("input_195_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_195_pad_0 = const()[name = tensor<string, []>("input_195_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [640, 640, 3, 3]> down_blocks_1_downsamplers_0_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2764800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52774016))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55538880))), name = tensor<string, []>("down_blocks_1_downsamplers_0_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 640, 3, 3])];
            tensor<fp16, [640]> down_blocks_1_downsamplers_0_conv_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_1_downsamplers_0_conv_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55539072)))];
            tensor<fp16, [1, 640, 32, 32]> input_195_cast_fp16 = conv(bias = down_blocks_1_downsamplers_0_conv_bias_to_fp16, dilations = var_1028, groups = var_370, pad = input_195_pad_0, pad_type = input_195_pad_type_0, strides = var_1026, weight = down_blocks_1_downsamplers_0_conv_weight_to_fp16_palettized, x = hidden_states_115_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<int32, []> var_1048 = const()[name = tensor<string, []>("op_1048"), val = tensor<int32, []>(1)];
            tensor<int32, [5]> reshape_40_shape_0 = const()[name = tensor<string, []>("reshape_40_shape_0"), val = tensor<int32, [5]>([1, 32, 20, 32, 32])];
            tensor<fp16, [1, 32, 20, 32, 32]> reshape_40_cast_fp16 = reshape(shape = reshape_40_shape_0, x = input_195_cast_fp16)[name = tensor<string, []>("reshape_40_cast_fp16")];
            tensor<int32, [3]> reduce_mean_30_axes_0 = const()[name = tensor<string, []>("reduce_mean_30_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_30_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_30_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_30_cast_fp16 = reduce_mean(axes = reduce_mean_30_axes_0, keep_dims = reduce_mean_30_keep_dims_0, x = reshape_40_cast_fp16)[name = tensor<string, []>("reduce_mean_30_cast_fp16")];
            tensor<fp16, [1, 32, 20, 32, 32]> sub_20_cast_fp16 = sub(x = reshape_40_cast_fp16, y = reduce_mean_30_cast_fp16)[name = tensor<string, []>("sub_20_cast_fp16")];
            tensor<fp16, [1, 32, 20, 32, 32]> square_10_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor<string, []>("square_10_cast_fp16")];
            tensor<int32, [3]> reduce_mean_32_axes_0 = const()[name = tensor<string, []>("reduce_mean_32_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_32_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_32_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_32_cast_fp16 = reduce_mean(axes = reduce_mean_32_axes_0, keep_dims = reduce_mean_32_keep_dims_0, x = square_10_cast_fp16)[name = tensor<string, []>("reduce_mean_32_cast_fp16")];
            tensor<fp16, []> add_20_y_0_to_fp16 = const()[name = tensor<string, []>("add_20_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_20_cast_fp16 = add(x = reduce_mean_32_cast_fp16, y = add_20_y_0_to_fp16)[name = tensor<string, []>("add_20_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_10_cast_fp16 = sqrt(x = add_20_cast_fp16)[name = tensor<string, []>("sqrt_10_cast_fp16")];
            tensor<fp16, [1, 32, 20, 32, 32]> real_div_10_cast_fp16 = real_div(x = sub_20_cast_fp16, y = sqrt_10_cast_fp16)[name = tensor<string, []>("real_div_10_cast_fp16")];
            tensor<int32, [4]> reshape_41_shape_0 = const()[name = tensor<string, []>("reshape_41_shape_0"), val = tensor<int32, [4]>([1, 640, 32, 32])];
            tensor<fp16, [1, 640, 32, 32]> reshape_41_cast_fp16 = reshape(shape = reshape_41_shape_0, x = real_div_10_cast_fp16)[name = tensor<string, []>("reshape_41_cast_fp16")];
            tensor<fp16, [640]> add_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_21_gamma_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55540416)))];
            tensor<fp16, [640]> add_21_beta_0_to_fp16 = const()[name = tensor<string, []>("add_21_beta_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55541760)))];
            tensor<fp16, []> add_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 640, 32, 32]> add_21_cast_fp16 = batch_norm(beta = add_21_beta_0_to_fp16, epsilon = add_21_epsilon_0_to_fp16, gamma = add_21_gamma_0_to_fp16, mean = add_11_mean_0_to_fp16, variance = add_11_variance_0_to_fp16, x = reshape_41_cast_fp16)[name = tensor<string, []>("add_21_cast_fp16")];
            tensor<fp16, [1, 640, 32, 32]> input_199_cast_fp16 = silu(x = add_21_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<int32, [2]> var_1073 = const()[name = tensor<string, []>("op_1073"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1075 = const()[name = tensor<string, []>("op_1075"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_117_pad_type_0 = const()[name = tensor<string, []>("hidden_states_117_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = tensor<string, []>("hidden_states_117_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [1280, 640, 3, 3]> down_blocks_2_resnets_0_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [5529600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55543104))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61072768))), name = tensor<string, []>("down_blocks_2_resnets_0_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 640, 3, 3])];
            tensor<fp16, [1280]> down_blocks_2_resnets_0_conv1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_resnets_0_conv1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61072960)))];
            tensor<fp16, [1, 1280, 32, 32]> hidden_states_117_cast_fp16 = conv(bias = down_blocks_2_resnets_0_conv1_bias_to_fp16, dilations = var_1075, groups = var_1048, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = var_1073, weight = down_blocks_2_resnets_0_conv1_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = tensor<string, []>("hidden_states_117_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_resnets_0_time_emb_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61075584))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62304448))), name = tensor<string, []>("down_blocks_2_resnets_0_time_emb_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_resnets_0_time_emb_proj_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_resnets_0_time_emb_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62304640)))];
            tensor<fp16, [1, 1280]> linear_52_cast_fp16 = linear(bias = down_blocks_2_resnets_0_time_emb_proj_bias_to_fp16, weight = down_blocks_2_resnets_0_time_emb_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<int32, [1]> var_1084_axes_0 = const()[name = tensor<string, []>("op_1084_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1280, 1]> var_1084_cast_fp16 = expand_dims(axes = var_1084_axes_0, x = linear_52_cast_fp16)[name = tensor<string, []>("op_1084_cast_fp16")];
            tensor<int32, [1]> temb_9_axes_0 = const()[name = tensor<string, []>("temb_9_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 1280, 1, 1]> temb_9_cast_fp16 = expand_dims(axes = temb_9_axes_0, x = var_1084_cast_fp16)[name = tensor<string, []>("temb_9_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_203_cast_fp16 = add(x = hidden_states_117_cast_fp16, y = temb_9_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<int32, [5]> reshape_44_shape_0 = const()[name = tensor<string, []>("reshape_44_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_44_cast_fp16 = reshape(shape = reshape_44_shape_0, x = input_203_cast_fp16)[name = tensor<string, []>("reshape_44_cast_fp16")];
            tensor<int32, [3]> reduce_mean_33_axes_0 = const()[name = tensor<string, []>("reduce_mean_33_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_33_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_33_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = reshape_44_cast_fp16)[name = tensor<string, []>("reduce_mean_33_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_22_cast_fp16 = sub(x = reshape_44_cast_fp16, y = reduce_mean_33_cast_fp16)[name = tensor<string, []>("sub_22_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_11_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor<string, []>("square_11_cast_fp16")];
            tensor<int32, [3]> reduce_mean_35_axes_0 = const()[name = tensor<string, []>("reduce_mean_35_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_35_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_35_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_11_cast_fp16)[name = tensor<string, []>("reduce_mean_35_cast_fp16")];
            tensor<fp16, []> add_22_y_0_to_fp16 = const()[name = tensor<string, []>("add_22_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_22_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = add_22_y_0_to_fp16)[name = tensor<string, []>("add_22_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_11_cast_fp16 = sqrt(x = add_22_cast_fp16)[name = tensor<string, []>("sqrt_11_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_11_cast_fp16 = real_div(x = sub_22_cast_fp16, y = sqrt_11_cast_fp16)[name = tensor<string, []>("real_div_11_cast_fp16")];
            tensor<int32, [4]> reshape_45_shape_0 = const()[name = tensor<string, []>("reshape_45_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_45_cast_fp16 = reshape(shape = reshape_45_shape_0, x = real_div_11_cast_fp16)[name = tensor<string, []>("reshape_45_cast_fp16")];
            tensor<fp16, [1280]> add_23_mean_0_to_fp16 = const()[name = tensor<string, []>("add_23_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62307264)))];
            tensor<fp16, [1280]> add_23_variance_0_to_fp16 = const()[name = tensor<string, []>("add_23_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62309888)))];
            tensor<fp16, [1280]> add_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_23_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62312512)))];
            tensor<fp16, [1280]> add_23_beta_0_to_fp16 = const()[name = tensor<string, []>("add_23_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62315136)))];
            tensor<fp16, []> add_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_23_cast_fp16 = batch_norm(beta = add_23_beta_0_to_fp16, epsilon = add_23_epsilon_0_to_fp16, gamma = add_23_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_45_cast_fp16)[name = tensor<string, []>("add_23_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_207_cast_fp16 = silu(x = add_23_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<int32, [2]> var_1094 = const()[name = tensor<string, []>("op_1094"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1096 = const()[name = tensor<string, []>("op_1096"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_119_pad_type_0 = const()[name = tensor<string, []>("hidden_states_119_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_119_pad_0 = const()[name = tensor<string, []>("hidden_states_119_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [1280, 1280, 3, 3]> down_blocks_2_resnets_0_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [11059200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62317760))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73377024))), name = tensor<string, []>("down_blocks_2_resnets_0_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 3, 3])];
            tensor<fp16, [1280]> down_blocks_2_resnets_0_conv2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_resnets_0_conv2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73377216)))];
            tensor<fp16, [1, 1280, 32, 32]> hidden_states_119_cast_fp16 = conv(bias = down_blocks_2_resnets_0_conv2_bias_to_fp16, dilations = var_1096, groups = var_1048, pad = hidden_states_119_pad_0, pad_type = hidden_states_119_pad_type_0, strides = var_1094, weight = down_blocks_2_resnets_0_conv2_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = tensor<string, []>("hidden_states_119_cast_fp16")];
            tensor<int32, [2]> var_1101 = const()[name = tensor<string, []>("op_1101"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1103 = const()[name = tensor<string, []>("op_1103"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_tensor_pad_type_0 = const()[name = tensor<string, []>("input_tensor_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_tensor_pad_0 = const()[name = tensor<string, []>("input_tensor_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 640, 1, 1]> down_blocks_2_resnets_0_conv_shortcut_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [614400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73379840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73994304))), name = tensor<string, []>("down_blocks_2_resnets_0_conv_shortcut_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 640, 1, 1])];
            tensor<fp16, [1280]> down_blocks_2_resnets_0_conv_shortcut_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_resnets_0_conv_shortcut_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73994496)))];
            tensor<fp16, [1, 1280, 32, 32]> input_tensor_cast_fp16 = conv(bias = down_blocks_2_resnets_0_conv_shortcut_bias_to_fp16, dilations = var_1103, groups = var_1048, pad = input_tensor_pad_0, pad_type = input_tensor_pad_type_0, strides = var_1101, weight = down_blocks_2_resnets_0_conv_shortcut_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = tensor<string, []>("input_tensor_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> var_1106_cast_fp16 = add(x = input_tensor_cast_fp16, y = hidden_states_119_cast_fp16)[name = tensor<string, []>("op_1106_cast_fp16")];
            tensor<int32, [5]> reshape_48_shape_0 = const()[name = tensor<string, []>("reshape_48_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = var_1106_cast_fp16)[name = tensor<string, []>("reshape_48_cast_fp16")];
            tensor<int32, [3]> reduce_mean_36_axes_0 = const()[name = tensor<string, []>("reduce_mean_36_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_36_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_36_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = reshape_48_cast_fp16)[name = tensor<string, []>("reduce_mean_36_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_24_cast_fp16 = sub(x = reshape_48_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor<string, []>("sub_24_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_12_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor<string, []>("square_12_cast_fp16")];
            tensor<int32, [3]> reduce_mean_38_axes_0 = const()[name = tensor<string, []>("reduce_mean_38_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_38_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_38_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = square_12_cast_fp16)[name = tensor<string, []>("reduce_mean_38_cast_fp16")];
            tensor<fp16, []> add_24_y_0_to_fp16 = const()[name = tensor<string, []>("add_24_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_24_cast_fp16 = add(x = reduce_mean_38_cast_fp16, y = add_24_y_0_to_fp16)[name = tensor<string, []>("add_24_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_12_cast_fp16 = sqrt(x = add_24_cast_fp16)[name = tensor<string, []>("sqrt_12_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_12_cast_fp16 = real_div(x = sub_24_cast_fp16, y = sqrt_12_cast_fp16)[name = tensor<string, []>("real_div_12_cast_fp16")];
            tensor<int32, [4]> reshape_49_shape_0 = const()[name = tensor<string, []>("reshape_49_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_49_cast_fp16 = reshape(shape = reshape_49_shape_0, x = real_div_12_cast_fp16)[name = tensor<string, []>("reshape_49_cast_fp16")];
            tensor<fp16, [1280]> add_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73997120)))];
            tensor<fp16, [1280]> add_25_beta_0_to_fp16 = const()[name = tensor<string, []>("add_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73999744)))];
            tensor<fp16, []> add_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_25_cast_fp16 = batch_norm(beta = add_25_beta_0_to_fp16, epsilon = add_25_epsilon_0_to_fp16, gamma = add_25_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_49_cast_fp16)[name = tensor<string, []>("add_25_cast_fp16")];
            tensor<int32, [4]> var_1144 = const()[name = tensor<string, []>("op_1144"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1148 = const()[name = tensor<string, []>("op_1148"), val = tensor<int32, [3]>([1, 1024, 1280])];
            tensor<fp16, [1, 32, 32, 1280]> transpose_653 = transpose(perm = var_1144, x = add_25_cast_fp16)[name = tensor<string, []>("transpose_653")];
            tensor<fp16, [1, 1024, 1280]> input_211_cast_fp16 = reshape(shape = var_1148, x = transpose_653)[name = tensor<string, []>("input_211_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_proj_in_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74002368))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75231232))), name = tensor<string, []>("down_blocks_2_attentions_0_proj_in_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_proj_in_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_proj_in_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75231424)))];
            tensor<fp16, [1, 1024, 1280]> linear_53_cast_fp16 = linear(bias = down_blocks_2_attentions_0_proj_in_bias_to_fp16, weight = down_blocks_2_attentions_0_proj_in_weight_to_fp16_palettized, x = input_211_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<int32, [1]> hidden_states_127_axes_0 = const()[name = tensor<string, []>("hidden_states_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75234048)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75236672)))];
            tensor<fp16, []> var_1044_to_fp16 = const()[name = tensor<string, []>("op_1044_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1280]> hidden_states_127_cast_fp16 = layer_norm(axes = hidden_states_127_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_0_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_0_norm1_weight_to_fp16, x = linear_53_cast_fp16)[name = tensor<string, []>("hidden_states_127_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75239296))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76468160))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_54_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_127_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76468352))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77697216))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_55_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_127_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77697408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78926272))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_56_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_127_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1181 = const()[name = tensor<string, []>("op_1181"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1182_cast_fp16 = reshape(shape = var_1181, x = linear_54_cast_fp16)[name = tensor<string, []>("op_1182_cast_fp16")];
            tensor<int32, [4]> var_1184 = const()[name = tensor<string, []>("op_1184"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1185_cast_fp16 = reshape(shape = var_1184, x = linear_55_cast_fp16)[name = tensor<string, []>("op_1185_cast_fp16")];
            tensor<int32, [4]> var_1187 = const()[name = tensor<string, []>("op_1187"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1188_cast_fp16 = reshape(shape = var_1187, x = linear_56_cast_fp16)[name = tensor<string, []>("op_1188_cast_fp16")];
            tensor<int32, [4]> value_35_perm_0 = const()[name = tensor<string, []>("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_21_y_0_to_fp16 = const()[name = tensor<string, []>("mul_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_21_cast_fp16 = mul(x = var_1182_cast_fp16, y = mul_21_y_0_to_fp16)[name = tensor<string, []>("mul_21_cast_fp16")];
            tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_288_perm_0 = const()[name = tensor<string, []>("transpose_288_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_289_perm_0 = const()[name = tensor<string, []>("transpose_289_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_651 = transpose(perm = transpose_289_perm_0, x = var_1185_cast_fp16)[name = tensor<string, []>("transpose_651")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_652 = transpose(perm = transpose_288_perm_0, x = mul_21_cast_fp16)[name = tensor<string, []>("transpose_652")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = transpose_652, y = transpose_651)[name = tensor<string, []>("matmul_8_cast_fp16")];
            tensor<int32, []> softmax_8_axis_0 = const()[name = tensor<string, []>("softmax_8_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = matmul_8_cast_fp16)[name = tensor<string, []>("softmax_8_cast_fp16")];
            tensor<bool, []> hidden_states_129_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_129_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_129_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_129_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_650 = transpose(perm = value_35_perm_0, x = var_1188_cast_fp16)[name = tensor<string, []>("transpose_650")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_129_cast_fp16 = matmul(transpose_x = hidden_states_129_transpose_x_0, transpose_y = hidden_states_129_transpose_y_0, x = softmax_8_cast_fp16, y = transpose_650)[name = tensor<string, []>("hidden_states_129_cast_fp16")];
            tensor<int32, [4]> var_1191_perm_0 = const()[name = tensor<string, []>("op_1191_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1195 = const()[name = tensor<string, []>("op_1195"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_649 = transpose(perm = var_1191_perm_0, x = hidden_states_129_cast_fp16)[name = tensor<string, []>("transpose_649")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_131_cast_fp16 = reshape(shape = var_1195, x = transpose_649)[name = tensor<string, []>("hidden_states_131_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78926464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80155328))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80155520)))];
            tensor<fp16, [1, 1024, 1280]> linear_57_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_131_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_217_cast_fp16 = add(x = linear_57_cast_fp16, y = linear_53_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
            tensor<int32, [1]> input_219_axes_0 = const()[name = tensor<string, []>("input_219_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80158144)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80160768)))];
            tensor<fp16, [1, 1024, 1280]> input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_0_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_0_norm2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80163392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81392256))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_58_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81392448))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83358592))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_59_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_59_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83358784))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85324928))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_60_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_60_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = tensor<string, []>("op_1227"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_58_cast_fp16)[name = tensor<string, []>("op_1228_cast_fp16")];
            tensor<int32, [4]> var_1230 = const()[name = tensor<string, []>("op_1230"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1231_cast_fp16 = reshape(shape = var_1230, x = linear_59_cast_fp16)[name = tensor<string, []>("op_1231_cast_fp16")];
            tensor<int32, [4]> var_1233 = const()[name = tensor<string, []>("op_1233"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1234_cast_fp16 = reshape(shape = var_1233, x = linear_60_cast_fp16)[name = tensor<string, []>("op_1234_cast_fp16")];
            tensor<int32, [4]> value_39_perm_0 = const()[name = tensor<string, []>("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_22_y_0_to_fp16 = const()[name = tensor<string, []>("mul_22_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_22_cast_fp16 = mul(x = var_1228_cast_fp16, y = mul_22_y_0_to_fp16)[name = tensor<string, []>("mul_22_cast_fp16")];
            tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_290_perm_0 = const()[name = tensor<string, []>("transpose_290_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_291_perm_0 = const()[name = tensor<string, []>("transpose_291_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_647 = transpose(perm = transpose_291_perm_0, x = var_1231_cast_fp16)[name = tensor<string, []>("transpose_647")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_648 = transpose(perm = transpose_290_perm_0, x = mul_22_cast_fp16)[name = tensor<string, []>("transpose_648")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = transpose_648, y = transpose_647)[name = tensor<string, []>("matmul_9_cast_fp16")];
            tensor<int32, []> softmax_9_axis_0 = const()[name = tensor<string, []>("softmax_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = matmul_9_cast_fp16)[name = tensor<string, []>("softmax_9_cast_fp16")];
            tensor<bool, []> hidden_states_135_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_135_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_135_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_135_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_646 = transpose(perm = value_39_perm_0, x = var_1234_cast_fp16)[name = tensor<string, []>("transpose_646")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_135_cast_fp16 = matmul(transpose_x = hidden_states_135_transpose_x_0, transpose_y = hidden_states_135_transpose_y_0, x = softmax_9_cast_fp16, y = transpose_646)[name = tensor<string, []>("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1237_perm_0 = const()[name = tensor<string, []>("op_1237_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1241 = const()[name = tensor<string, []>("op_1241"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_645 = transpose(perm = var_1237_perm_0, x = hidden_states_135_cast_fp16)[name = tensor<string, []>("transpose_645")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_137_cast_fp16 = reshape(shape = var_1241, x = transpose_645)[name = tensor<string, []>("hidden_states_137_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85325120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86553984))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86554176)))];
            tensor<fp16, [1, 1024, 1280]> linear_61_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_137_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_225_cast_fp16 = add(x = linear_61_cast_fp16, y = input_217_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
            tensor<int32, [1]> input_227_axes_0 = const()[name = tensor<string, []>("input_227_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86556800)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86559424)))];
            tensor<fp16, [1, 1024, 1280]> input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_0_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_0_norm3_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86562048))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96392512))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96392704))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96400448))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_62_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized, x = input_227_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
            tensor<int32, [2]> var_1263_split_sizes_0 = const()[name = tensor<string, []>("op_1263_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_1263_axis_0 = const()[name = tensor<string, []>("op_1263_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_1263_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_1263_cast_fp16_1 = split(axis = var_1263_axis_0, split_sizes = var_1263_split_sizes_0, x = linear_62_cast_fp16)[name = tensor<string, []>("op_1263_cast_fp16")];
            tensor<string, []> var_1265_mode_0 = const()[name = tensor<string, []>("op_1265_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_1265_cast_fp16 = gelu(mode = var_1265_mode_0, x = var_1263_cast_fp16_1)[name = tensor<string, []>("op_1265_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_229_cast_fp16 = mul(x = var_1263_cast_fp16_0, y = var_1265_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96400640))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101315904))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101316096)))];
            tensor<fp16, [1, 1024, 1280]> linear_63_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized, x = input_229_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_145_cast_fp16 = add(x = linear_63_cast_fp16, y = input_225_cast_fp16)[name = tensor<string, []>("hidden_states_145_cast_fp16")];
            tensor<int32, [1]> hidden_states_147_axes_0 = const()[name = tensor<string, []>("hidden_states_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101318720)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101321344)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_147_cast_fp16 = layer_norm(axes = hidden_states_147_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_1_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_1_norm1_weight_to_fp16, x = hidden_states_145_cast_fp16)[name = tensor<string, []>("hidden_states_147_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101323968))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102552832))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_64_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_147_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102553024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103781888))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_65_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_147_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103782080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105010944))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_66_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_147_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
            tensor<int32, [4]> var_1300 = const()[name = tensor<string, []>("op_1300"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1301_cast_fp16 = reshape(shape = var_1300, x = linear_64_cast_fp16)[name = tensor<string, []>("op_1301_cast_fp16")];
            tensor<int32, [4]> var_1303 = const()[name = tensor<string, []>("op_1303"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1304_cast_fp16 = reshape(shape = var_1303, x = linear_65_cast_fp16)[name = tensor<string, []>("op_1304_cast_fp16")];
            tensor<int32, [4]> var_1306 = const()[name = tensor<string, []>("op_1306"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1307_cast_fp16 = reshape(shape = var_1306, x = linear_66_cast_fp16)[name = tensor<string, []>("op_1307_cast_fp16")];
            tensor<int32, [4]> value_43_perm_0 = const()[name = tensor<string, []>("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_23_y_0_to_fp16 = const()[name = tensor<string, []>("mul_23_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_23_cast_fp16 = mul(x = var_1301_cast_fp16, y = mul_23_y_0_to_fp16)[name = tensor<string, []>("mul_23_cast_fp16")];
            tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_292_perm_0 = const()[name = tensor<string, []>("transpose_292_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_293_perm_0 = const()[name = tensor<string, []>("transpose_293_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_643 = transpose(perm = transpose_293_perm_0, x = var_1304_cast_fp16)[name = tensor<string, []>("transpose_643")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_644 = transpose(perm = transpose_292_perm_0, x = mul_23_cast_fp16)[name = tensor<string, []>("transpose_644")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = transpose_644, y = transpose_643)[name = tensor<string, []>("matmul_10_cast_fp16")];
            tensor<int32, []> softmax_10_axis_0 = const()[name = tensor<string, []>("softmax_10_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = matmul_10_cast_fp16)[name = tensor<string, []>("softmax_10_cast_fp16")];
            tensor<bool, []> hidden_states_149_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_149_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_149_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_149_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_642 = transpose(perm = value_43_perm_0, x = var_1307_cast_fp16)[name = tensor<string, []>("transpose_642")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_149_cast_fp16 = matmul(transpose_x = hidden_states_149_transpose_x_0, transpose_y = hidden_states_149_transpose_y_0, x = softmax_10_cast_fp16, y = transpose_642)[name = tensor<string, []>("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1310_perm_0 = const()[name = tensor<string, []>("op_1310_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1314 = const()[name = tensor<string, []>("op_1314"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_641 = transpose(perm = var_1310_perm_0, x = hidden_states_149_cast_fp16)[name = tensor<string, []>("transpose_641")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_151_cast_fp16 = reshape(shape = var_1314, x = transpose_641)[name = tensor<string, []>("hidden_states_151_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105011136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106240000))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106240192)))];
            tensor<fp16, [1, 1024, 1280]> linear_67_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_151_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_237_cast_fp16 = add(x = linear_67_cast_fp16, y = hidden_states_145_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")];
            tensor<int32, [1]> input_239_axes_0 = const()[name = tensor<string, []>("input_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106242816)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106245440)))];
            tensor<fp16, [1, 1024, 1280]> input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_1_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_1_norm2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106248064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107476928))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_68_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107477120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109443264))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_69_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_69_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109443456))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111409600))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_70_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1346 = const()[name = tensor<string, []>("op_1346"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1347_cast_fp16 = reshape(shape = var_1346, x = linear_68_cast_fp16)[name = tensor<string, []>("op_1347_cast_fp16")];
            tensor<int32, [4]> var_1349 = const()[name = tensor<string, []>("op_1349"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1350_cast_fp16 = reshape(shape = var_1349, x = linear_69_cast_fp16)[name = tensor<string, []>("op_1350_cast_fp16")];
            tensor<int32, [4]> var_1352 = const()[name = tensor<string, []>("op_1352"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1353_cast_fp16 = reshape(shape = var_1352, x = linear_70_cast_fp16)[name = tensor<string, []>("op_1353_cast_fp16")];
            tensor<int32, [4]> value_47_perm_0 = const()[name = tensor<string, []>("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_24_y_0_to_fp16 = const()[name = tensor<string, []>("mul_24_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_24_cast_fp16 = mul(x = var_1347_cast_fp16, y = mul_24_y_0_to_fp16)[name = tensor<string, []>("mul_24_cast_fp16")];
            tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_294_perm_0 = const()[name = tensor<string, []>("transpose_294_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_295_perm_0 = const()[name = tensor<string, []>("transpose_295_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_639 = transpose(perm = transpose_295_perm_0, x = var_1350_cast_fp16)[name = tensor<string, []>("transpose_639")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_640 = transpose(perm = transpose_294_perm_0, x = mul_24_cast_fp16)[name = tensor<string, []>("transpose_640")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = transpose_640, y = transpose_639)[name = tensor<string, []>("matmul_11_cast_fp16")];
            tensor<int32, []> softmax_11_axis_0 = const()[name = tensor<string, []>("softmax_11_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = matmul_11_cast_fp16)[name = tensor<string, []>("softmax_11_cast_fp16")];
            tensor<bool, []> hidden_states_155_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_155_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_155_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_155_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_638 = transpose(perm = value_47_perm_0, x = var_1353_cast_fp16)[name = tensor<string, []>("transpose_638")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_155_cast_fp16 = matmul(transpose_x = hidden_states_155_transpose_x_0, transpose_y = hidden_states_155_transpose_y_0, x = softmax_11_cast_fp16, y = transpose_638)[name = tensor<string, []>("hidden_states_155_cast_fp16")];
            tensor<int32, [4]> var_1356_perm_0 = const()[name = tensor<string, []>("op_1356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1360 = const()[name = tensor<string, []>("op_1360"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_637 = transpose(perm = var_1356_perm_0, x = hidden_states_155_cast_fp16)[name = tensor<string, []>("transpose_637")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_157_cast_fp16 = reshape(shape = var_1360, x = transpose_637)[name = tensor<string, []>("hidden_states_157_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111409792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112638656))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112638848)))];
            tensor<fp16, [1, 1024, 1280]> linear_71_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_157_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_245_cast_fp16 = add(x = linear_71_cast_fp16, y = input_237_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
            tensor<int32, [1]> input_247_axes_0 = const()[name = tensor<string, []>("input_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112641472)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112644096)))];
            tensor<fp16, [1, 1024, 1280]> input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_1_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_1_norm3_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112646720))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122477184))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122477376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122485120))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_72_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
            tensor<int32, [2]> var_1382_split_sizes_0 = const()[name = tensor<string, []>("op_1382_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_1382_axis_0 = const()[name = tensor<string, []>("op_1382_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_1382_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_1382_cast_fp16_1 = split(axis = var_1382_axis_0, split_sizes = var_1382_split_sizes_0, x = linear_72_cast_fp16)[name = tensor<string, []>("op_1382_cast_fp16")];
            tensor<string, []> var_1384_mode_0 = const()[name = tensor<string, []>("op_1384_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_1384_cast_fp16 = gelu(mode = var_1384_mode_0, x = var_1382_cast_fp16_1)[name = tensor<string, []>("op_1384_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_249_cast_fp16 = mul(x = var_1382_cast_fp16_0, y = var_1384_cast_fp16)[name = tensor<string, []>("input_249_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122485312))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127400576))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127400768)))];
            tensor<fp16, [1, 1024, 1280]> linear_73_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized, x = input_249_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_165_cast_fp16 = add(x = linear_73_cast_fp16, y = input_245_cast_fp16)[name = tensor<string, []>("hidden_states_165_cast_fp16")];
            tensor<int32, [1]> hidden_states_167_axes_0 = const()[name = tensor<string, []>("hidden_states_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127403392)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127406016)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_167_cast_fp16 = layer_norm(axes = hidden_states_167_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_2_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_2_norm1_weight_to_fp16, x = hidden_states_165_cast_fp16)[name = tensor<string, []>("hidden_states_167_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127408640))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128637504))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_74_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_167_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128637696))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129866560))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_75_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_167_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129866752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131095616))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_76_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_167_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
            tensor<int32, [4]> var_1419 = const()[name = tensor<string, []>("op_1419"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1420_cast_fp16 = reshape(shape = var_1419, x = linear_74_cast_fp16)[name = tensor<string, []>("op_1420_cast_fp16")];
            tensor<int32, [4]> var_1422 = const()[name = tensor<string, []>("op_1422"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1423_cast_fp16 = reshape(shape = var_1422, x = linear_75_cast_fp16)[name = tensor<string, []>("op_1423_cast_fp16")];
            tensor<int32, [4]> var_1425 = const()[name = tensor<string, []>("op_1425"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1426_cast_fp16 = reshape(shape = var_1425, x = linear_76_cast_fp16)[name = tensor<string, []>("op_1426_cast_fp16")];
            tensor<int32, [4]> value_51_perm_0 = const()[name = tensor<string, []>("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_25_y_0_to_fp16 = const()[name = tensor<string, []>("mul_25_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_25_cast_fp16 = mul(x = var_1420_cast_fp16, y = mul_25_y_0_to_fp16)[name = tensor<string, []>("mul_25_cast_fp16")];
            tensor<bool, []> matmul_12_transpose_y_0 = const()[name = tensor<string, []>("matmul_12_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_12_transpose_x_0 = const()[name = tensor<string, []>("matmul_12_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_296_perm_0 = const()[name = tensor<string, []>("transpose_296_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_297_perm_0 = const()[name = tensor<string, []>("transpose_297_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_635 = transpose(perm = transpose_297_perm_0, x = var_1423_cast_fp16)[name = tensor<string, []>("transpose_635")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_636 = transpose(perm = transpose_296_perm_0, x = mul_25_cast_fp16)[name = tensor<string, []>("transpose_636")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = transpose_636, y = transpose_635)[name = tensor<string, []>("matmul_12_cast_fp16")];
            tensor<int32, []> softmax_12_axis_0 = const()[name = tensor<string, []>("softmax_12_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = matmul_12_cast_fp16)[name = tensor<string, []>("softmax_12_cast_fp16")];
            tensor<bool, []> hidden_states_169_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_169_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_169_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_169_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_634 = transpose(perm = value_51_perm_0, x = var_1426_cast_fp16)[name = tensor<string, []>("transpose_634")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_169_cast_fp16 = matmul(transpose_x = hidden_states_169_transpose_x_0, transpose_y = hidden_states_169_transpose_y_0, x = softmax_12_cast_fp16, y = transpose_634)[name = tensor<string, []>("hidden_states_169_cast_fp16")];
            tensor<int32, [4]> var_1429_perm_0 = const()[name = tensor<string, []>("op_1429_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1433 = const()[name = tensor<string, []>("op_1433"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_633 = transpose(perm = var_1429_perm_0, x = hidden_states_169_cast_fp16)[name = tensor<string, []>("transpose_633")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_171_cast_fp16 = reshape(shape = var_1433, x = transpose_633)[name = tensor<string, []>("hidden_states_171_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131095808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132324672))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132324864)))];
            tensor<fp16, [1, 1024, 1280]> linear_77_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_171_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_257_cast_fp16 = add(x = linear_77_cast_fp16, y = hidden_states_165_cast_fp16)[name = tensor<string, []>("input_257_cast_fp16")];
            tensor<int32, [1]> input_259_axes_0 = const()[name = tensor<string, []>("input_259_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132327488)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132330112)))];
            tensor<fp16, [1, 1024, 1280]> input_259_cast_fp16 = layer_norm(axes = input_259_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_2_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_2_norm2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132332736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133561600))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_78_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = tensor<string, []>("linear_78_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133561792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135527936))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_79_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_79_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135528128))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137494272))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_80_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_80_cast_fp16")];
            tensor<int32, [4]> var_1465 = const()[name = tensor<string, []>("op_1465"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1466_cast_fp16 = reshape(shape = var_1465, x = linear_78_cast_fp16)[name = tensor<string, []>("op_1466_cast_fp16")];
            tensor<int32, [4]> var_1468 = const()[name = tensor<string, []>("op_1468"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1469_cast_fp16 = reshape(shape = var_1468, x = linear_79_cast_fp16)[name = tensor<string, []>("op_1469_cast_fp16")];
            tensor<int32, [4]> var_1471 = const()[name = tensor<string, []>("op_1471"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1472_cast_fp16 = reshape(shape = var_1471, x = linear_80_cast_fp16)[name = tensor<string, []>("op_1472_cast_fp16")];
            tensor<int32, [4]> value_55_perm_0 = const()[name = tensor<string, []>("value_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_26_y_0_to_fp16 = const()[name = tensor<string, []>("mul_26_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_26_cast_fp16 = mul(x = var_1466_cast_fp16, y = mul_26_y_0_to_fp16)[name = tensor<string, []>("mul_26_cast_fp16")];
            tensor<bool, []> matmul_13_transpose_y_0 = const()[name = tensor<string, []>("matmul_13_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_13_transpose_x_0 = const()[name = tensor<string, []>("matmul_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_298_perm_0 = const()[name = tensor<string, []>("transpose_298_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_299_perm_0 = const()[name = tensor<string, []>("transpose_299_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_631 = transpose(perm = transpose_299_perm_0, x = var_1469_cast_fp16)[name = tensor<string, []>("transpose_631")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_632 = transpose(perm = transpose_298_perm_0, x = mul_26_cast_fp16)[name = tensor<string, []>("transpose_632")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = transpose_632, y = transpose_631)[name = tensor<string, []>("matmul_13_cast_fp16")];
            tensor<int32, []> softmax_13_axis_0 = const()[name = tensor<string, []>("softmax_13_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = matmul_13_cast_fp16)[name = tensor<string, []>("softmax_13_cast_fp16")];
            tensor<bool, []> hidden_states_175_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_175_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_175_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_175_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_630 = transpose(perm = value_55_perm_0, x = var_1472_cast_fp16)[name = tensor<string, []>("transpose_630")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_175_cast_fp16 = matmul(transpose_x = hidden_states_175_transpose_x_0, transpose_y = hidden_states_175_transpose_y_0, x = softmax_13_cast_fp16, y = transpose_630)[name = tensor<string, []>("hidden_states_175_cast_fp16")];
            tensor<int32, [4]> var_1475_perm_0 = const()[name = tensor<string, []>("op_1475_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1479 = const()[name = tensor<string, []>("op_1479"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_629 = transpose(perm = var_1475_perm_0, x = hidden_states_175_cast_fp16)[name = tensor<string, []>("transpose_629")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_177_cast_fp16 = reshape(shape = var_1479, x = transpose_629)[name = tensor<string, []>("hidden_states_177_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137494464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138723328))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138723520)))];
            tensor<fp16, [1, 1024, 1280]> linear_81_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_177_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_265_cast_fp16 = add(x = linear_81_cast_fp16, y = input_257_cast_fp16)[name = tensor<string, []>("input_265_cast_fp16")];
            tensor<int32, [1]> input_267_axes_0 = const()[name = tensor<string, []>("input_267_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138726144)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138728768)))];
            tensor<fp16, [1, 1024, 1280]> input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_2_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_2_norm3_weight_to_fp16, x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138731392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148561856))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148562048))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148569792))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_82_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
            tensor<int32, [2]> var_1501_split_sizes_0 = const()[name = tensor<string, []>("op_1501_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_1501_axis_0 = const()[name = tensor<string, []>("op_1501_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_1501_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_1501_cast_fp16_1 = split(axis = var_1501_axis_0, split_sizes = var_1501_split_sizes_0, x = linear_82_cast_fp16)[name = tensor<string, []>("op_1501_cast_fp16")];
            tensor<string, []> var_1503_mode_0 = const()[name = tensor<string, []>("op_1503_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_1503_cast_fp16 = gelu(mode = var_1503_mode_0, x = var_1501_cast_fp16_1)[name = tensor<string, []>("op_1503_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_269_cast_fp16 = mul(x = var_1501_cast_fp16_0, y = var_1503_cast_fp16)[name = tensor<string, []>("input_269_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148569984))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153485248))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_2_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_2_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153485440)))];
            tensor<fp16, [1, 1024, 1280]> linear_83_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_2_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized, x = input_269_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_185_cast_fp16 = add(x = linear_83_cast_fp16, y = input_265_cast_fp16)[name = tensor<string, []>("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> hidden_states_187_axes_0 = const()[name = tensor<string, []>("hidden_states_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153488064)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153490688)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_187_cast_fp16 = layer_norm(axes = hidden_states_187_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_3_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_3_norm1_weight_to_fp16, x = hidden_states_185_cast_fp16)[name = tensor<string, []>("hidden_states_187_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153493312))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154722176))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_84_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_187_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154722368))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155951232))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_85_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_187_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155951424))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157180288))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_86_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_187_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1538 = const()[name = tensor<string, []>("op_1538"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1539_cast_fp16 = reshape(shape = var_1538, x = linear_84_cast_fp16)[name = tensor<string, []>("op_1539_cast_fp16")];
            tensor<int32, [4]> var_1541 = const()[name = tensor<string, []>("op_1541"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1542_cast_fp16 = reshape(shape = var_1541, x = linear_85_cast_fp16)[name = tensor<string, []>("op_1542_cast_fp16")];
            tensor<int32, [4]> var_1544 = const()[name = tensor<string, []>("op_1544"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1545_cast_fp16 = reshape(shape = var_1544, x = linear_86_cast_fp16)[name = tensor<string, []>("op_1545_cast_fp16")];
            tensor<int32, [4]> value_59_perm_0 = const()[name = tensor<string, []>("value_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_27_y_0_to_fp16 = const()[name = tensor<string, []>("mul_27_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_27_cast_fp16 = mul(x = var_1539_cast_fp16, y = mul_27_y_0_to_fp16)[name = tensor<string, []>("mul_27_cast_fp16")];
            tensor<bool, []> matmul_14_transpose_y_0 = const()[name = tensor<string, []>("matmul_14_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_14_transpose_x_0 = const()[name = tensor<string, []>("matmul_14_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_300_perm_0 = const()[name = tensor<string, []>("transpose_300_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_301_perm_0 = const()[name = tensor<string, []>("transpose_301_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_627 = transpose(perm = transpose_301_perm_0, x = var_1542_cast_fp16)[name = tensor<string, []>("transpose_627")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_628 = transpose(perm = transpose_300_perm_0, x = mul_27_cast_fp16)[name = tensor<string, []>("transpose_628")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = transpose_628, y = transpose_627)[name = tensor<string, []>("matmul_14_cast_fp16")];
            tensor<int32, []> softmax_14_axis_0 = const()[name = tensor<string, []>("softmax_14_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = matmul_14_cast_fp16)[name = tensor<string, []>("softmax_14_cast_fp16")];
            tensor<bool, []> hidden_states_189_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_189_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_189_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_189_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_626 = transpose(perm = value_59_perm_0, x = var_1545_cast_fp16)[name = tensor<string, []>("transpose_626")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_189_cast_fp16 = matmul(transpose_x = hidden_states_189_transpose_x_0, transpose_y = hidden_states_189_transpose_y_0, x = softmax_14_cast_fp16, y = transpose_626)[name = tensor<string, []>("hidden_states_189_cast_fp16")];
            tensor<int32, [4]> var_1548_perm_0 = const()[name = tensor<string, []>("op_1548_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1552 = const()[name = tensor<string, []>("op_1552"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_625 = transpose(perm = var_1548_perm_0, x = hidden_states_189_cast_fp16)[name = tensor<string, []>("transpose_625")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_191_cast_fp16 = reshape(shape = var_1552, x = transpose_625)[name = tensor<string, []>("hidden_states_191_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157180480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158409344))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158409536)))];
            tensor<fp16, [1, 1024, 1280]> linear_87_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_191_cast_fp16)[name = tensor<string, []>("linear_87_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_277_cast_fp16 = add(x = linear_87_cast_fp16, y = hidden_states_185_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
            tensor<int32, [1]> input_279_axes_0 = const()[name = tensor<string, []>("input_279_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158412160)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158414784)))];
            tensor<fp16, [1, 1024, 1280]> input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_3_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_3_norm2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("input_279_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158417408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159646272))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_88_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized, x = input_279_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159646464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161612608))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_89_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_89_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161612800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163578944))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_90_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_90_cast_fp16")];
            tensor<int32, [4]> var_1584 = const()[name = tensor<string, []>("op_1584"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1585_cast_fp16 = reshape(shape = var_1584, x = linear_88_cast_fp16)[name = tensor<string, []>("op_1585_cast_fp16")];
            tensor<int32, [4]> var_1587 = const()[name = tensor<string, []>("op_1587"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1588_cast_fp16 = reshape(shape = var_1587, x = linear_89_cast_fp16)[name = tensor<string, []>("op_1588_cast_fp16")];
            tensor<int32, [4]> var_1590 = const()[name = tensor<string, []>("op_1590"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1591_cast_fp16 = reshape(shape = var_1590, x = linear_90_cast_fp16)[name = tensor<string, []>("op_1591_cast_fp16")];
            tensor<int32, [4]> value_63_perm_0 = const()[name = tensor<string, []>("value_63_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_28_y_0_to_fp16 = const()[name = tensor<string, []>("mul_28_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_28_cast_fp16 = mul(x = var_1585_cast_fp16, y = mul_28_y_0_to_fp16)[name = tensor<string, []>("mul_28_cast_fp16")];
            tensor<bool, []> matmul_15_transpose_y_0 = const()[name = tensor<string, []>("matmul_15_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_15_transpose_x_0 = const()[name = tensor<string, []>("matmul_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_302_perm_0 = const()[name = tensor<string, []>("transpose_302_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_303_perm_0 = const()[name = tensor<string, []>("transpose_303_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_623 = transpose(perm = transpose_303_perm_0, x = var_1588_cast_fp16)[name = tensor<string, []>("transpose_623")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_624 = transpose(perm = transpose_302_perm_0, x = mul_28_cast_fp16)[name = tensor<string, []>("transpose_624")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = transpose_624, y = transpose_623)[name = tensor<string, []>("matmul_15_cast_fp16")];
            tensor<int32, []> softmax_15_axis_0 = const()[name = tensor<string, []>("softmax_15_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = matmul_15_cast_fp16)[name = tensor<string, []>("softmax_15_cast_fp16")];
            tensor<bool, []> hidden_states_195_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_195_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_195_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_195_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_622 = transpose(perm = value_63_perm_0, x = var_1591_cast_fp16)[name = tensor<string, []>("transpose_622")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_195_cast_fp16 = matmul(transpose_x = hidden_states_195_transpose_x_0, transpose_y = hidden_states_195_transpose_y_0, x = softmax_15_cast_fp16, y = transpose_622)[name = tensor<string, []>("hidden_states_195_cast_fp16")];
            tensor<int32, [4]> var_1594_perm_0 = const()[name = tensor<string, []>("op_1594_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1598 = const()[name = tensor<string, []>("op_1598"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_621 = transpose(perm = var_1594_perm_0, x = hidden_states_195_cast_fp16)[name = tensor<string, []>("transpose_621")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_197_cast_fp16 = reshape(shape = var_1598, x = transpose_621)[name = tensor<string, []>("hidden_states_197_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163579136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164808000))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164808192)))];
            tensor<fp16, [1, 1024, 1280]> linear_91_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_197_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_285_cast_fp16 = add(x = linear_91_cast_fp16, y = input_277_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
            tensor<int32, [1]> input_287_axes_0 = const()[name = tensor<string, []>("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164810816)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164813440)))];
            tensor<fp16, [1, 1024, 1280]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_3_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_3_norm3_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164816064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174646528))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174646720))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174654464))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_92_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")];
            tensor<int32, [2]> var_1620_split_sizes_0 = const()[name = tensor<string, []>("op_1620_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_1620_axis_0 = const()[name = tensor<string, []>("op_1620_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_1620_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_1620_cast_fp16_1 = split(axis = var_1620_axis_0, split_sizes = var_1620_split_sizes_0, x = linear_92_cast_fp16)[name = tensor<string, []>("op_1620_cast_fp16")];
            tensor<string, []> var_1622_mode_0 = const()[name = tensor<string, []>("op_1622_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_1622_cast_fp16 = gelu(mode = var_1622_mode_0, x = var_1620_cast_fp16_1)[name = tensor<string, []>("op_1622_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_289_cast_fp16 = mul(x = var_1620_cast_fp16_0, y = var_1622_cast_fp16)[name = tensor<string, []>("input_289_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174654656))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179569920))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_3_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_3_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179570112)))];
            tensor<fp16, [1, 1024, 1280]> linear_93_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_3_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized, x = input_289_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_205_cast_fp16 = add(x = linear_93_cast_fp16, y = input_285_cast_fp16)[name = tensor<string, []>("hidden_states_205_cast_fp16")];
            tensor<int32, [1]> hidden_states_207_axes_0 = const()[name = tensor<string, []>("hidden_states_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179572736)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179575360)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_207_cast_fp16 = layer_norm(axes = hidden_states_207_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_4_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_4_norm1_weight_to_fp16, x = hidden_states_205_cast_fp16)[name = tensor<string, []>("hidden_states_207_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179577984))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180806848))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_94_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_207_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180807040))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182035904))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_95_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_207_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182036096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183264960))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_96_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_207_cast_fp16)[name = tensor<string, []>("linear_96_cast_fp16")];
            tensor<int32, [4]> var_1657 = const()[name = tensor<string, []>("op_1657"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1658_cast_fp16 = reshape(shape = var_1657, x = linear_94_cast_fp16)[name = tensor<string, []>("op_1658_cast_fp16")];
            tensor<int32, [4]> var_1660 = const()[name = tensor<string, []>("op_1660"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1661_cast_fp16 = reshape(shape = var_1660, x = linear_95_cast_fp16)[name = tensor<string, []>("op_1661_cast_fp16")];
            tensor<int32, [4]> var_1663 = const()[name = tensor<string, []>("op_1663"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1664_cast_fp16 = reshape(shape = var_1663, x = linear_96_cast_fp16)[name = tensor<string, []>("op_1664_cast_fp16")];
            tensor<int32, [4]> value_67_perm_0 = const()[name = tensor<string, []>("value_67_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_29_y_0_to_fp16 = const()[name = tensor<string, []>("mul_29_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_29_cast_fp16 = mul(x = var_1658_cast_fp16, y = mul_29_y_0_to_fp16)[name = tensor<string, []>("mul_29_cast_fp16")];
            tensor<bool, []> matmul_16_transpose_y_0 = const()[name = tensor<string, []>("matmul_16_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_16_transpose_x_0 = const()[name = tensor<string, []>("matmul_16_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_304_perm_0 = const()[name = tensor<string, []>("transpose_304_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_305_perm_0 = const()[name = tensor<string, []>("transpose_305_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_619 = transpose(perm = transpose_305_perm_0, x = var_1661_cast_fp16)[name = tensor<string, []>("transpose_619")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_620 = transpose(perm = transpose_304_perm_0, x = mul_29_cast_fp16)[name = tensor<string, []>("transpose_620")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = transpose_620, y = transpose_619)[name = tensor<string, []>("matmul_16_cast_fp16")];
            tensor<int32, []> softmax_16_axis_0 = const()[name = tensor<string, []>("softmax_16_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = matmul_16_cast_fp16)[name = tensor<string, []>("softmax_16_cast_fp16")];
            tensor<bool, []> hidden_states_209_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_209_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_209_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_209_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_618 = transpose(perm = value_67_perm_0, x = var_1664_cast_fp16)[name = tensor<string, []>("transpose_618")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_209_cast_fp16 = matmul(transpose_x = hidden_states_209_transpose_x_0, transpose_y = hidden_states_209_transpose_y_0, x = softmax_16_cast_fp16, y = transpose_618)[name = tensor<string, []>("hidden_states_209_cast_fp16")];
            tensor<int32, [4]> var_1667_perm_0 = const()[name = tensor<string, []>("op_1667_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1671 = const()[name = tensor<string, []>("op_1671"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_617 = transpose(perm = var_1667_perm_0, x = hidden_states_209_cast_fp16)[name = tensor<string, []>("transpose_617")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_211_cast_fp16 = reshape(shape = var_1671, x = transpose_617)[name = tensor<string, []>("hidden_states_211_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183265152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184494016))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184494208)))];
            tensor<fp16, [1, 1024, 1280]> linear_97_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_211_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_297_cast_fp16 = add(x = linear_97_cast_fp16, y = hidden_states_205_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
            tensor<int32, [1]> input_299_axes_0 = const()[name = tensor<string, []>("input_299_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184496832)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184499456)))];
            tensor<fp16, [1, 1024, 1280]> input_299_cast_fp16 = layer_norm(axes = input_299_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_4_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_4_norm2_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("input_299_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184502080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185730944))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_98_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185731136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187697280))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_99_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_99_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187697472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189663616))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_100_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_100_cast_fp16")];
            tensor<int32, [4]> var_1703 = const()[name = tensor<string, []>("op_1703"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1704_cast_fp16 = reshape(shape = var_1703, x = linear_98_cast_fp16)[name = tensor<string, []>("op_1704_cast_fp16")];
            tensor<int32, [4]> var_1706 = const()[name = tensor<string, []>("op_1706"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1707_cast_fp16 = reshape(shape = var_1706, x = linear_99_cast_fp16)[name = tensor<string, []>("op_1707_cast_fp16")];
            tensor<int32, [4]> var_1709 = const()[name = tensor<string, []>("op_1709"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1710_cast_fp16 = reshape(shape = var_1709, x = linear_100_cast_fp16)[name = tensor<string, []>("op_1710_cast_fp16")];
            tensor<int32, [4]> value_71_perm_0 = const()[name = tensor<string, []>("value_71_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_30_y_0_to_fp16 = const()[name = tensor<string, []>("mul_30_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_30_cast_fp16 = mul(x = var_1704_cast_fp16, y = mul_30_y_0_to_fp16)[name = tensor<string, []>("mul_30_cast_fp16")];
            tensor<bool, []> matmul_17_transpose_y_0 = const()[name = tensor<string, []>("matmul_17_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_17_transpose_x_0 = const()[name = tensor<string, []>("matmul_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_306_perm_0 = const()[name = tensor<string, []>("transpose_306_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_307_perm_0 = const()[name = tensor<string, []>("transpose_307_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_615 = transpose(perm = transpose_307_perm_0, x = var_1707_cast_fp16)[name = tensor<string, []>("transpose_615")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_616 = transpose(perm = transpose_306_perm_0, x = mul_30_cast_fp16)[name = tensor<string, []>("transpose_616")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = transpose_616, y = transpose_615)[name = tensor<string, []>("matmul_17_cast_fp16")];
            tensor<int32, []> softmax_17_axis_0 = const()[name = tensor<string, []>("softmax_17_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = matmul_17_cast_fp16)[name = tensor<string, []>("softmax_17_cast_fp16")];
            tensor<bool, []> hidden_states_215_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_215_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_215_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_215_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_614 = transpose(perm = value_71_perm_0, x = var_1710_cast_fp16)[name = tensor<string, []>("transpose_614")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_215_cast_fp16 = matmul(transpose_x = hidden_states_215_transpose_x_0, transpose_y = hidden_states_215_transpose_y_0, x = softmax_17_cast_fp16, y = transpose_614)[name = tensor<string, []>("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_1713_perm_0 = const()[name = tensor<string, []>("op_1713_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1717 = const()[name = tensor<string, []>("op_1717"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_613 = transpose(perm = var_1713_perm_0, x = hidden_states_215_cast_fp16)[name = tensor<string, []>("transpose_613")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_217_cast_fp16 = reshape(shape = var_1717, x = transpose_613)[name = tensor<string, []>("hidden_states_217_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189663808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190892672))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190892864)))];
            tensor<fp16, [1, 1024, 1280]> linear_101_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_217_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_305_cast_fp16 = add(x = linear_101_cast_fp16, y = input_297_cast_fp16)[name = tensor<string, []>("input_305_cast_fp16")];
            tensor<int32, [1]> input_307_axes_0 = const()[name = tensor<string, []>("input_307_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190895488)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190898112)))];
            tensor<fp16, [1, 1024, 1280]> input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_4_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_4_norm3_weight_to_fp16, x = input_305_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190900736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200731200))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200731392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200739136))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_102_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized, x = input_307_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
            tensor<int32, [2]> var_1739_split_sizes_0 = const()[name = tensor<string, []>("op_1739_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_1739_axis_0 = const()[name = tensor<string, []>("op_1739_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_1739_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_1739_cast_fp16_1 = split(axis = var_1739_axis_0, split_sizes = var_1739_split_sizes_0, x = linear_102_cast_fp16)[name = tensor<string, []>("op_1739_cast_fp16")];
            tensor<string, []> var_1741_mode_0 = const()[name = tensor<string, []>("op_1741_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_1741_cast_fp16 = gelu(mode = var_1741_mode_0, x = var_1739_cast_fp16_1)[name = tensor<string, []>("op_1741_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_309_cast_fp16 = mul(x = var_1739_cast_fp16_0, y = var_1741_cast_fp16)[name = tensor<string, []>("input_309_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200739328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205654592))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_4_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_4_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205654784)))];
            tensor<fp16, [1, 1024, 1280]> linear_103_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_4_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_225_cast_fp16 = add(x = linear_103_cast_fp16, y = input_305_cast_fp16)[name = tensor<string, []>("hidden_states_225_cast_fp16")];
            tensor<int32, [1]> hidden_states_227_axes_0 = const()[name = tensor<string, []>("hidden_states_227_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205657408)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205660032)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_227_cast_fp16 = layer_norm(axes = hidden_states_227_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_5_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_5_norm1_weight_to_fp16, x = hidden_states_225_cast_fp16)[name = tensor<string, []>("hidden_states_227_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205662656))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206891520))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_104_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_227_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206891712))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208120576))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_105_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_227_cast_fp16)[name = tensor<string, []>("linear_105_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208120768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209349632))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_106_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_227_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
            tensor<int32, [4]> var_1776 = const()[name = tensor<string, []>("op_1776"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1777_cast_fp16 = reshape(shape = var_1776, x = linear_104_cast_fp16)[name = tensor<string, []>("op_1777_cast_fp16")];
            tensor<int32, [4]> var_1779 = const()[name = tensor<string, []>("op_1779"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1780_cast_fp16 = reshape(shape = var_1779, x = linear_105_cast_fp16)[name = tensor<string, []>("op_1780_cast_fp16")];
            tensor<int32, [4]> var_1782 = const()[name = tensor<string, []>("op_1782"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1783_cast_fp16 = reshape(shape = var_1782, x = linear_106_cast_fp16)[name = tensor<string, []>("op_1783_cast_fp16")];
            tensor<int32, [4]> value_75_perm_0 = const()[name = tensor<string, []>("value_75_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_31_y_0_to_fp16 = const()[name = tensor<string, []>("mul_31_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_31_cast_fp16 = mul(x = var_1777_cast_fp16, y = mul_31_y_0_to_fp16)[name = tensor<string, []>("mul_31_cast_fp16")];
            tensor<bool, []> matmul_18_transpose_y_0 = const()[name = tensor<string, []>("matmul_18_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_18_transpose_x_0 = const()[name = tensor<string, []>("matmul_18_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_308_perm_0 = const()[name = tensor<string, []>("transpose_308_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_309_perm_0 = const()[name = tensor<string, []>("transpose_309_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_611 = transpose(perm = transpose_309_perm_0, x = var_1780_cast_fp16)[name = tensor<string, []>("transpose_611")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_612 = transpose(perm = transpose_308_perm_0, x = mul_31_cast_fp16)[name = tensor<string, []>("transpose_612")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = transpose_612, y = transpose_611)[name = tensor<string, []>("matmul_18_cast_fp16")];
            tensor<int32, []> softmax_18_axis_0 = const()[name = tensor<string, []>("softmax_18_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = matmul_18_cast_fp16)[name = tensor<string, []>("softmax_18_cast_fp16")];
            tensor<bool, []> hidden_states_229_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_229_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_229_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_229_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_610 = transpose(perm = value_75_perm_0, x = var_1783_cast_fp16)[name = tensor<string, []>("transpose_610")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_229_cast_fp16 = matmul(transpose_x = hidden_states_229_transpose_x_0, transpose_y = hidden_states_229_transpose_y_0, x = softmax_18_cast_fp16, y = transpose_610)[name = tensor<string, []>("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_1786_perm_0 = const()[name = tensor<string, []>("op_1786_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1790 = const()[name = tensor<string, []>("op_1790"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_609 = transpose(perm = var_1786_perm_0, x = hidden_states_229_cast_fp16)[name = tensor<string, []>("transpose_609")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_231_cast_fp16 = reshape(shape = var_1790, x = transpose_609)[name = tensor<string, []>("hidden_states_231_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209349824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210578688))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210578880)))];
            tensor<fp16, [1, 1024, 1280]> linear_107_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_231_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_317_cast_fp16 = add(x = linear_107_cast_fp16, y = hidden_states_225_cast_fp16)[name = tensor<string, []>("input_317_cast_fp16")];
            tensor<int32, [1]> input_319_axes_0 = const()[name = tensor<string, []>("input_319_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210581504)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210584128)))];
            tensor<fp16, [1, 1024, 1280]> input_319_cast_fp16 = layer_norm(axes = input_319_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_5_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_5_norm2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210586752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211815616))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_108_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized, x = input_319_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211815808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213781952))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_109_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_109_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213782144))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215748288))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_110_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_110_cast_fp16")];
            tensor<int32, [4]> var_1822 = const()[name = tensor<string, []>("op_1822"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1823_cast_fp16 = reshape(shape = var_1822, x = linear_108_cast_fp16)[name = tensor<string, []>("op_1823_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = tensor<string, []>("op_1825"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_109_cast_fp16)[name = tensor<string, []>("op_1826_cast_fp16")];
            tensor<int32, [4]> var_1828 = const()[name = tensor<string, []>("op_1828"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1829_cast_fp16 = reshape(shape = var_1828, x = linear_110_cast_fp16)[name = tensor<string, []>("op_1829_cast_fp16")];
            tensor<int32, [4]> value_79_perm_0 = const()[name = tensor<string, []>("value_79_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_32_y_0_to_fp16 = const()[name = tensor<string, []>("mul_32_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_32_cast_fp16 = mul(x = var_1823_cast_fp16, y = mul_32_y_0_to_fp16)[name = tensor<string, []>("mul_32_cast_fp16")];
            tensor<bool, []> matmul_19_transpose_y_0 = const()[name = tensor<string, []>("matmul_19_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_19_transpose_x_0 = const()[name = tensor<string, []>("matmul_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_310_perm_0 = const()[name = tensor<string, []>("transpose_310_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_311_perm_0 = const()[name = tensor<string, []>("transpose_311_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_607 = transpose(perm = transpose_311_perm_0, x = var_1826_cast_fp16)[name = tensor<string, []>("transpose_607")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_608 = transpose(perm = transpose_310_perm_0, x = mul_32_cast_fp16)[name = tensor<string, []>("transpose_608")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = transpose_608, y = transpose_607)[name = tensor<string, []>("matmul_19_cast_fp16")];
            tensor<int32, []> softmax_19_axis_0 = const()[name = tensor<string, []>("softmax_19_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = matmul_19_cast_fp16)[name = tensor<string, []>("softmax_19_cast_fp16")];
            tensor<bool, []> hidden_states_235_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_235_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_235_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_235_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_606 = transpose(perm = value_79_perm_0, x = var_1829_cast_fp16)[name = tensor<string, []>("transpose_606")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_235_cast_fp16 = matmul(transpose_x = hidden_states_235_transpose_x_0, transpose_y = hidden_states_235_transpose_y_0, x = softmax_19_cast_fp16, y = transpose_606)[name = tensor<string, []>("hidden_states_235_cast_fp16")];
            tensor<int32, [4]> var_1832_perm_0 = const()[name = tensor<string, []>("op_1832_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1836 = const()[name = tensor<string, []>("op_1836"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_605 = transpose(perm = var_1832_perm_0, x = hidden_states_235_cast_fp16)[name = tensor<string, []>("transpose_605")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_237_cast_fp16 = reshape(shape = var_1836, x = transpose_605)[name = tensor<string, []>("hidden_states_237_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215748480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216977344))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216977536)))];
            tensor<fp16, [1, 1024, 1280]> linear_111_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_237_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_325_cast_fp16 = add(x = linear_111_cast_fp16, y = input_317_cast_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
            tensor<int32, [1]> input_327_axes_0 = const()[name = tensor<string, []>("input_327_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216980160)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216982784)))];
            tensor<fp16, [1, 1024, 1280]> input_327_cast_fp16 = layer_norm(axes = input_327_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_5_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_5_norm3_weight_to_fp16, x = input_325_cast_fp16)[name = tensor<string, []>("input_327_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216985408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226815872))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226816064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226823808))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_112_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized, x = input_327_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")];
            tensor<int32, [2]> var_1858_split_sizes_0 = const()[name = tensor<string, []>("op_1858_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_1858_axis_0 = const()[name = tensor<string, []>("op_1858_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_1858_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_1858_cast_fp16_1 = split(axis = var_1858_axis_0, split_sizes = var_1858_split_sizes_0, x = linear_112_cast_fp16)[name = tensor<string, []>("op_1858_cast_fp16")];
            tensor<string, []> var_1860_mode_0 = const()[name = tensor<string, []>("op_1860_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_1860_cast_fp16 = gelu(mode = var_1860_mode_0, x = var_1858_cast_fp16_1)[name = tensor<string, []>("op_1860_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_329_cast_fp16 = mul(x = var_1858_cast_fp16_0, y = var_1860_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226824000))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231739264))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_5_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_5_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231739456)))];
            tensor<fp16, [1, 1024, 1280]> linear_113_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_5_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_245_cast_fp16 = add(x = linear_113_cast_fp16, y = input_325_cast_fp16)[name = tensor<string, []>("hidden_states_245_cast_fp16")];
            tensor<int32, [1]> hidden_states_247_axes_0 = const()[name = tensor<string, []>("hidden_states_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231742080)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231744704)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_247_cast_fp16 = layer_norm(axes = hidden_states_247_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_6_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_6_norm1_weight_to_fp16, x = hidden_states_245_cast_fp16)[name = tensor<string, []>("hidden_states_247_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231747328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232976192))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_114_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_247_cast_fp16)[name = tensor<string, []>("linear_114_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232976384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234205248))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_115_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_247_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234205440))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235434304))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_116_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_247_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")];
            tensor<int32, [4]> var_1895 = const()[name = tensor<string, []>("op_1895"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1896_cast_fp16 = reshape(shape = var_1895, x = linear_114_cast_fp16)[name = tensor<string, []>("op_1896_cast_fp16")];
            tensor<int32, [4]> var_1898 = const()[name = tensor<string, []>("op_1898"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1899_cast_fp16 = reshape(shape = var_1898, x = linear_115_cast_fp16)[name = tensor<string, []>("op_1899_cast_fp16")];
            tensor<int32, [4]> var_1901 = const()[name = tensor<string, []>("op_1901"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1902_cast_fp16 = reshape(shape = var_1901, x = linear_116_cast_fp16)[name = tensor<string, []>("op_1902_cast_fp16")];
            tensor<int32, [4]> value_83_perm_0 = const()[name = tensor<string, []>("value_83_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_33_y_0_to_fp16 = const()[name = tensor<string, []>("mul_33_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_33_cast_fp16 = mul(x = var_1896_cast_fp16, y = mul_33_y_0_to_fp16)[name = tensor<string, []>("mul_33_cast_fp16")];
            tensor<bool, []> matmul_20_transpose_y_0 = const()[name = tensor<string, []>("matmul_20_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_20_transpose_x_0 = const()[name = tensor<string, []>("matmul_20_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_312_perm_0 = const()[name = tensor<string, []>("transpose_312_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_313_perm_0 = const()[name = tensor<string, []>("transpose_313_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_603 = transpose(perm = transpose_313_perm_0, x = var_1899_cast_fp16)[name = tensor<string, []>("transpose_603")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_604 = transpose(perm = transpose_312_perm_0, x = mul_33_cast_fp16)[name = tensor<string, []>("transpose_604")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = transpose_604, y = transpose_603)[name = tensor<string, []>("matmul_20_cast_fp16")];
            tensor<int32, []> softmax_20_axis_0 = const()[name = tensor<string, []>("softmax_20_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = matmul_20_cast_fp16)[name = tensor<string, []>("softmax_20_cast_fp16")];
            tensor<bool, []> hidden_states_249_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_249_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_249_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_249_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_602 = transpose(perm = value_83_perm_0, x = var_1902_cast_fp16)[name = tensor<string, []>("transpose_602")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_249_cast_fp16 = matmul(transpose_x = hidden_states_249_transpose_x_0, transpose_y = hidden_states_249_transpose_y_0, x = softmax_20_cast_fp16, y = transpose_602)[name = tensor<string, []>("hidden_states_249_cast_fp16")];
            tensor<int32, [4]> var_1905_perm_0 = const()[name = tensor<string, []>("op_1905_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1909 = const()[name = tensor<string, []>("op_1909"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_601 = transpose(perm = var_1905_perm_0, x = hidden_states_249_cast_fp16)[name = tensor<string, []>("transpose_601")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_251_cast_fp16 = reshape(shape = var_1909, x = transpose_601)[name = tensor<string, []>("hidden_states_251_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235434496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236663360))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236663552)))];
            tensor<fp16, [1, 1024, 1280]> linear_117_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_251_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_337_cast_fp16 = add(x = linear_117_cast_fp16, y = hidden_states_245_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
            tensor<int32, [1]> input_339_axes_0 = const()[name = tensor<string, []>("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236666176)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236668800)))];
            tensor<fp16, [1, 1024, 1280]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_6_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_6_norm2_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236671424))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237900288))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_118_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237900480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239866624))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_119_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_119_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239866816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241832960))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_120_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_120_cast_fp16")];
            tensor<int32, [4]> var_1941 = const()[name = tensor<string, []>("op_1941"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_1942_cast_fp16 = reshape(shape = var_1941, x = linear_118_cast_fp16)[name = tensor<string, []>("op_1942_cast_fp16")];
            tensor<int32, [4]> var_1944 = const()[name = tensor<string, []>("op_1944"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1945_cast_fp16 = reshape(shape = var_1944, x = linear_119_cast_fp16)[name = tensor<string, []>("op_1945_cast_fp16")];
            tensor<int32, [4]> var_1947 = const()[name = tensor<string, []>("op_1947"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_1948_cast_fp16 = reshape(shape = var_1947, x = linear_120_cast_fp16)[name = tensor<string, []>("op_1948_cast_fp16")];
            tensor<int32, [4]> value_87_perm_0 = const()[name = tensor<string, []>("value_87_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_34_y_0_to_fp16 = const()[name = tensor<string, []>("mul_34_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_34_cast_fp16 = mul(x = var_1942_cast_fp16, y = mul_34_y_0_to_fp16)[name = tensor<string, []>("mul_34_cast_fp16")];
            tensor<bool, []> matmul_21_transpose_y_0 = const()[name = tensor<string, []>("matmul_21_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_21_transpose_x_0 = const()[name = tensor<string, []>("matmul_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_314_perm_0 = const()[name = tensor<string, []>("transpose_314_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_315_perm_0 = const()[name = tensor<string, []>("transpose_315_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_599 = transpose(perm = transpose_315_perm_0, x = var_1945_cast_fp16)[name = tensor<string, []>("transpose_599")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_600 = transpose(perm = transpose_314_perm_0, x = mul_34_cast_fp16)[name = tensor<string, []>("transpose_600")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = transpose_600, y = transpose_599)[name = tensor<string, []>("matmul_21_cast_fp16")];
            tensor<int32, []> softmax_21_axis_0 = const()[name = tensor<string, []>("softmax_21_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = matmul_21_cast_fp16)[name = tensor<string, []>("softmax_21_cast_fp16")];
            tensor<bool, []> hidden_states_255_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_255_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_255_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_255_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_598 = transpose(perm = value_87_perm_0, x = var_1948_cast_fp16)[name = tensor<string, []>("transpose_598")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_255_cast_fp16 = matmul(transpose_x = hidden_states_255_transpose_x_0, transpose_y = hidden_states_255_transpose_y_0, x = softmax_21_cast_fp16, y = transpose_598)[name = tensor<string, []>("hidden_states_255_cast_fp16")];
            tensor<int32, [4]> var_1951_perm_0 = const()[name = tensor<string, []>("op_1951_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1955 = const()[name = tensor<string, []>("op_1955"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_597 = transpose(perm = var_1951_perm_0, x = hidden_states_255_cast_fp16)[name = tensor<string, []>("transpose_597")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_257_cast_fp16 = reshape(shape = var_1955, x = transpose_597)[name = tensor<string, []>("hidden_states_257_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241833152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243062016))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243062208)))];
            tensor<fp16, [1, 1024, 1280]> linear_121_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_257_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_345_cast_fp16 = add(x = linear_121_cast_fp16, y = input_337_cast_fp16)[name = tensor<string, []>("input_345_cast_fp16")];
            tensor<int32, [1]> input_347_axes_0 = const()[name = tensor<string, []>("input_347_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243064832)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243067456)))];
            tensor<fp16, [1, 1024, 1280]> input_347_cast_fp16 = layer_norm(axes = input_347_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_6_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_6_norm3_weight_to_fp16, x = input_345_cast_fp16)[name = tensor<string, []>("input_347_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243070080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252900544))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252900736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252908480))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_122_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized, x = input_347_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")];
            tensor<int32, [2]> var_1977_split_sizes_0 = const()[name = tensor<string, []>("op_1977_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_1977_axis_0 = const()[name = tensor<string, []>("op_1977_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_1977_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_1977_cast_fp16_1 = split(axis = var_1977_axis_0, split_sizes = var_1977_split_sizes_0, x = linear_122_cast_fp16)[name = tensor<string, []>("op_1977_cast_fp16")];
            tensor<string, []> var_1979_mode_0 = const()[name = tensor<string, []>("op_1979_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_1979_cast_fp16 = gelu(mode = var_1979_mode_0, x = var_1977_cast_fp16_1)[name = tensor<string, []>("op_1979_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_349_cast_fp16 = mul(x = var_1977_cast_fp16_0, y = var_1979_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252908672))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257823936))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_6_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_6_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257824128)))];
            tensor<fp16, [1, 1024, 1280]> linear_123_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_6_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized, x = input_349_cast_fp16)[name = tensor<string, []>("linear_123_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_265_cast_fp16 = add(x = linear_123_cast_fp16, y = input_345_cast_fp16)[name = tensor<string, []>("hidden_states_265_cast_fp16")];
            tensor<int32, [1]> hidden_states_267_axes_0 = const()[name = tensor<string, []>("hidden_states_267_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257826752)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257829376)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_267_cast_fp16 = layer_norm(axes = hidden_states_267_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_7_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_7_norm1_weight_to_fp16, x = hidden_states_265_cast_fp16)[name = tensor<string, []>("hidden_states_267_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257832000))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259060864))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_124_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_267_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259061056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260289920))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_125_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_267_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260290112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261518976))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_126_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_267_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2014 = const()[name = tensor<string, []>("op_2014"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2015_cast_fp16 = reshape(shape = var_2014, x = linear_124_cast_fp16)[name = tensor<string, []>("op_2015_cast_fp16")];
            tensor<int32, [4]> var_2017 = const()[name = tensor<string, []>("op_2017"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2018_cast_fp16 = reshape(shape = var_2017, x = linear_125_cast_fp16)[name = tensor<string, []>("op_2018_cast_fp16")];
            tensor<int32, [4]> var_2020 = const()[name = tensor<string, []>("op_2020"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2021_cast_fp16 = reshape(shape = var_2020, x = linear_126_cast_fp16)[name = tensor<string, []>("op_2021_cast_fp16")];
            tensor<int32, [4]> value_91_perm_0 = const()[name = tensor<string, []>("value_91_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_35_y_0_to_fp16 = const()[name = tensor<string, []>("mul_35_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_35_cast_fp16 = mul(x = var_2015_cast_fp16, y = mul_35_y_0_to_fp16)[name = tensor<string, []>("mul_35_cast_fp16")];
            tensor<bool, []> matmul_22_transpose_y_0 = const()[name = tensor<string, []>("matmul_22_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_22_transpose_x_0 = const()[name = tensor<string, []>("matmul_22_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_316_perm_0 = const()[name = tensor<string, []>("transpose_316_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_317_perm_0 = const()[name = tensor<string, []>("transpose_317_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_595 = transpose(perm = transpose_317_perm_0, x = var_2018_cast_fp16)[name = tensor<string, []>("transpose_595")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_596 = transpose(perm = transpose_316_perm_0, x = mul_35_cast_fp16)[name = tensor<string, []>("transpose_596")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = transpose_596, y = transpose_595)[name = tensor<string, []>("matmul_22_cast_fp16")];
            tensor<int32, []> softmax_22_axis_0 = const()[name = tensor<string, []>("softmax_22_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = matmul_22_cast_fp16)[name = tensor<string, []>("softmax_22_cast_fp16")];
            tensor<bool, []> hidden_states_269_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_269_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_269_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_269_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_594 = transpose(perm = value_91_perm_0, x = var_2021_cast_fp16)[name = tensor<string, []>("transpose_594")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_269_cast_fp16 = matmul(transpose_x = hidden_states_269_transpose_x_0, transpose_y = hidden_states_269_transpose_y_0, x = softmax_22_cast_fp16, y = transpose_594)[name = tensor<string, []>("hidden_states_269_cast_fp16")];
            tensor<int32, [4]> var_2024_perm_0 = const()[name = tensor<string, []>("op_2024_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2028 = const()[name = tensor<string, []>("op_2028"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_593 = transpose(perm = var_2024_perm_0, x = hidden_states_269_cast_fp16)[name = tensor<string, []>("transpose_593")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_271_cast_fp16 = reshape(shape = var_2028, x = transpose_593)[name = tensor<string, []>("hidden_states_271_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261519168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262748032))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262748224)))];
            tensor<fp16, [1, 1024, 1280]> linear_127_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_271_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_357_cast_fp16 = add(x = linear_127_cast_fp16, y = hidden_states_265_cast_fp16)[name = tensor<string, []>("input_357_cast_fp16")];
            tensor<int32, [1]> input_359_axes_0 = const()[name = tensor<string, []>("input_359_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262750848)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262753472)))];
            tensor<fp16, [1, 1024, 1280]> input_359_cast_fp16 = layer_norm(axes = input_359_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_7_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_7_norm2_weight_to_fp16, x = input_357_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262756096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263984960))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_128_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized, x = input_359_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263985152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265951296))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_129_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_129_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265951488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267917632))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_130_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_130_cast_fp16")];
            tensor<int32, [4]> var_2060 = const()[name = tensor<string, []>("op_2060"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2061_cast_fp16 = reshape(shape = var_2060, x = linear_128_cast_fp16)[name = tensor<string, []>("op_2061_cast_fp16")];
            tensor<int32, [4]> var_2063 = const()[name = tensor<string, []>("op_2063"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2064_cast_fp16 = reshape(shape = var_2063, x = linear_129_cast_fp16)[name = tensor<string, []>("op_2064_cast_fp16")];
            tensor<int32, [4]> var_2066 = const()[name = tensor<string, []>("op_2066"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2067_cast_fp16 = reshape(shape = var_2066, x = linear_130_cast_fp16)[name = tensor<string, []>("op_2067_cast_fp16")];
            tensor<int32, [4]> value_95_perm_0 = const()[name = tensor<string, []>("value_95_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_36_y_0_to_fp16 = const()[name = tensor<string, []>("mul_36_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_36_cast_fp16 = mul(x = var_2061_cast_fp16, y = mul_36_y_0_to_fp16)[name = tensor<string, []>("mul_36_cast_fp16")];
            tensor<bool, []> matmul_23_transpose_y_0 = const()[name = tensor<string, []>("matmul_23_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_23_transpose_x_0 = const()[name = tensor<string, []>("matmul_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_318_perm_0 = const()[name = tensor<string, []>("transpose_318_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_319_perm_0 = const()[name = tensor<string, []>("transpose_319_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_591 = transpose(perm = transpose_319_perm_0, x = var_2064_cast_fp16)[name = tensor<string, []>("transpose_591")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_592 = transpose(perm = transpose_318_perm_0, x = mul_36_cast_fp16)[name = tensor<string, []>("transpose_592")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = transpose_592, y = transpose_591)[name = tensor<string, []>("matmul_23_cast_fp16")];
            tensor<int32, []> softmax_23_axis_0 = const()[name = tensor<string, []>("softmax_23_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = matmul_23_cast_fp16)[name = tensor<string, []>("softmax_23_cast_fp16")];
            tensor<bool, []> hidden_states_275_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_275_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_275_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_275_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_590 = transpose(perm = value_95_perm_0, x = var_2067_cast_fp16)[name = tensor<string, []>("transpose_590")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_275_cast_fp16 = matmul(transpose_x = hidden_states_275_transpose_x_0, transpose_y = hidden_states_275_transpose_y_0, x = softmax_23_cast_fp16, y = transpose_590)[name = tensor<string, []>("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_2070_perm_0 = const()[name = tensor<string, []>("op_2070_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2074 = const()[name = tensor<string, []>("op_2074"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_589 = transpose(perm = var_2070_perm_0, x = hidden_states_275_cast_fp16)[name = tensor<string, []>("transpose_589")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_277_cast_fp16 = reshape(shape = var_2074, x = transpose_589)[name = tensor<string, []>("hidden_states_277_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267917824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269146688))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269146880)))];
            tensor<fp16, [1, 1024, 1280]> linear_131_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_277_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_365_cast_fp16 = add(x = linear_131_cast_fp16, y = input_357_cast_fp16)[name = tensor<string, []>("input_365_cast_fp16")];
            tensor<int32, [1]> input_367_axes_0 = const()[name = tensor<string, []>("input_367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269149504)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269152128)))];
            tensor<fp16, [1, 1024, 1280]> input_367_cast_fp16 = layer_norm(axes = input_367_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_7_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_7_norm3_weight_to_fp16, x = input_365_cast_fp16)[name = tensor<string, []>("input_367_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269154752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278985216))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278985408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278993152))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_132_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized, x = input_367_cast_fp16)[name = tensor<string, []>("linear_132_cast_fp16")];
            tensor<int32, [2]> var_2096_split_sizes_0 = const()[name = tensor<string, []>("op_2096_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_2096_axis_0 = const()[name = tensor<string, []>("op_2096_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_2096_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_2096_cast_fp16_1 = split(axis = var_2096_axis_0, split_sizes = var_2096_split_sizes_0, x = linear_132_cast_fp16)[name = tensor<string, []>("op_2096_cast_fp16")];
            tensor<string, []> var_2098_mode_0 = const()[name = tensor<string, []>("op_2098_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_2098_cast_fp16 = gelu(mode = var_2098_mode_0, x = var_2096_cast_fp16_1)[name = tensor<string, []>("op_2098_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_369_cast_fp16 = mul(x = var_2096_cast_fp16_0, y = var_2098_cast_fp16)[name = tensor<string, []>("input_369_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278993344))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283908608))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_7_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_7_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283908800)))];
            tensor<fp16, [1, 1024, 1280]> linear_133_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_7_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized, x = input_369_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_285_cast_fp16 = add(x = linear_133_cast_fp16, y = input_365_cast_fp16)[name = tensor<string, []>("hidden_states_285_cast_fp16")];
            tensor<int32, [1]> hidden_states_287_axes_0 = const()[name = tensor<string, []>("hidden_states_287_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283911424)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283914048)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_287_cast_fp16 = layer_norm(axes = hidden_states_287_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_8_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_8_norm1_weight_to_fp16, x = hidden_states_285_cast_fp16)[name = tensor<string, []>("hidden_states_287_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283916672))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285145536))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_134_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_287_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285145728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286374592))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_135_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_287_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286374784))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287603648))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_136_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_287_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = tensor<string, []>("op_2133"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_134_cast_fp16)[name = tensor<string, []>("op_2134_cast_fp16")];
            tensor<int32, [4]> var_2136 = const()[name = tensor<string, []>("op_2136"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2137_cast_fp16 = reshape(shape = var_2136, x = linear_135_cast_fp16)[name = tensor<string, []>("op_2137_cast_fp16")];
            tensor<int32, [4]> var_2139 = const()[name = tensor<string, []>("op_2139"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2140_cast_fp16 = reshape(shape = var_2139, x = linear_136_cast_fp16)[name = tensor<string, []>("op_2140_cast_fp16")];
            tensor<int32, [4]> value_99_perm_0 = const()[name = tensor<string, []>("value_99_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_37_y_0_to_fp16 = const()[name = tensor<string, []>("mul_37_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_37_cast_fp16 = mul(x = var_2134_cast_fp16, y = mul_37_y_0_to_fp16)[name = tensor<string, []>("mul_37_cast_fp16")];
            tensor<bool, []> matmul_24_transpose_y_0 = const()[name = tensor<string, []>("matmul_24_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_24_transpose_x_0 = const()[name = tensor<string, []>("matmul_24_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_320_perm_0 = const()[name = tensor<string, []>("transpose_320_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_321_perm_0 = const()[name = tensor<string, []>("transpose_321_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_587 = transpose(perm = transpose_321_perm_0, x = var_2137_cast_fp16)[name = tensor<string, []>("transpose_587")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_588 = transpose(perm = transpose_320_perm_0, x = mul_37_cast_fp16)[name = tensor<string, []>("transpose_588")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_24_cast_fp16 = matmul(transpose_x = matmul_24_transpose_x_0, transpose_y = matmul_24_transpose_y_0, x = transpose_588, y = transpose_587)[name = tensor<string, []>("matmul_24_cast_fp16")];
            tensor<int32, []> softmax_24_axis_0 = const()[name = tensor<string, []>("softmax_24_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_24_cast_fp16 = softmax(axis = softmax_24_axis_0, x = matmul_24_cast_fp16)[name = tensor<string, []>("softmax_24_cast_fp16")];
            tensor<bool, []> hidden_states_289_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_289_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_289_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_289_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_586 = transpose(perm = value_99_perm_0, x = var_2140_cast_fp16)[name = tensor<string, []>("transpose_586")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_289_cast_fp16 = matmul(transpose_x = hidden_states_289_transpose_x_0, transpose_y = hidden_states_289_transpose_y_0, x = softmax_24_cast_fp16, y = transpose_586)[name = tensor<string, []>("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_2143_perm_0 = const()[name = tensor<string, []>("op_2143_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2147 = const()[name = tensor<string, []>("op_2147"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_585 = transpose(perm = var_2143_perm_0, x = hidden_states_289_cast_fp16)[name = tensor<string, []>("transpose_585")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_291_cast_fp16 = reshape(shape = var_2147, x = transpose_585)[name = tensor<string, []>("hidden_states_291_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287603840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288832704))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288832896)))];
            tensor<fp16, [1, 1024, 1280]> linear_137_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_291_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_377_cast_fp16 = add(x = linear_137_cast_fp16, y = hidden_states_285_cast_fp16)[name = tensor<string, []>("input_377_cast_fp16")];
            tensor<int32, [1]> input_379_axes_0 = const()[name = tensor<string, []>("input_379_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288835520)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288838144)))];
            tensor<fp16, [1, 1024, 1280]> input_379_cast_fp16 = layer_norm(axes = input_379_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_8_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_8_norm2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor<string, []>("input_379_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288840768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290069632))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_138_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized, x = input_379_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290069824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292035968))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_139_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_139_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292036160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294002304))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_140_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_140_cast_fp16")];
            tensor<int32, [4]> var_2179 = const()[name = tensor<string, []>("op_2179"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2180_cast_fp16 = reshape(shape = var_2179, x = linear_138_cast_fp16)[name = tensor<string, []>("op_2180_cast_fp16")];
            tensor<int32, [4]> var_2182 = const()[name = tensor<string, []>("op_2182"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2183_cast_fp16 = reshape(shape = var_2182, x = linear_139_cast_fp16)[name = tensor<string, []>("op_2183_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = tensor<string, []>("op_2185"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2186_cast_fp16 = reshape(shape = var_2185, x = linear_140_cast_fp16)[name = tensor<string, []>("op_2186_cast_fp16")];
            tensor<int32, [4]> value_103_perm_0 = const()[name = tensor<string, []>("value_103_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_38_y_0_to_fp16 = const()[name = tensor<string, []>("mul_38_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_38_cast_fp16 = mul(x = var_2180_cast_fp16, y = mul_38_y_0_to_fp16)[name = tensor<string, []>("mul_38_cast_fp16")];
            tensor<bool, []> matmul_25_transpose_y_0 = const()[name = tensor<string, []>("matmul_25_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_25_transpose_x_0 = const()[name = tensor<string, []>("matmul_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_322_perm_0 = const()[name = tensor<string, []>("transpose_322_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_323_perm_0 = const()[name = tensor<string, []>("transpose_323_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_583 = transpose(perm = transpose_323_perm_0, x = var_2183_cast_fp16)[name = tensor<string, []>("transpose_583")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_584 = transpose(perm = transpose_322_perm_0, x = mul_38_cast_fp16)[name = tensor<string, []>("transpose_584")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_25_cast_fp16 = matmul(transpose_x = matmul_25_transpose_x_0, transpose_y = matmul_25_transpose_y_0, x = transpose_584, y = transpose_583)[name = tensor<string, []>("matmul_25_cast_fp16")];
            tensor<int32, []> softmax_25_axis_0 = const()[name = tensor<string, []>("softmax_25_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_25_cast_fp16 = softmax(axis = softmax_25_axis_0, x = matmul_25_cast_fp16)[name = tensor<string, []>("softmax_25_cast_fp16")];
            tensor<bool, []> hidden_states_295_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_295_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_295_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_295_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_582 = transpose(perm = value_103_perm_0, x = var_2186_cast_fp16)[name = tensor<string, []>("transpose_582")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_295_cast_fp16 = matmul(transpose_x = hidden_states_295_transpose_x_0, transpose_y = hidden_states_295_transpose_y_0, x = softmax_25_cast_fp16, y = transpose_582)[name = tensor<string, []>("hidden_states_295_cast_fp16")];
            tensor<int32, [4]> var_2189_perm_0 = const()[name = tensor<string, []>("op_2189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2193 = const()[name = tensor<string, []>("op_2193"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_581 = transpose(perm = var_2189_perm_0, x = hidden_states_295_cast_fp16)[name = tensor<string, []>("transpose_581")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_297_cast_fp16 = reshape(shape = var_2193, x = transpose_581)[name = tensor<string, []>("hidden_states_297_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294002496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295231360))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295231552)))];
            tensor<fp16, [1, 1024, 1280]> linear_141_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_297_cast_fp16)[name = tensor<string, []>("linear_141_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_385_cast_fp16 = add(x = linear_141_cast_fp16, y = input_377_cast_fp16)[name = tensor<string, []>("input_385_cast_fp16")];
            tensor<int32, [1]> input_387_axes_0 = const()[name = tensor<string, []>("input_387_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295234176)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295236800)))];
            tensor<fp16, [1, 1024, 1280]> input_387_cast_fp16 = layer_norm(axes = input_387_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_8_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_8_norm3_weight_to_fp16, x = input_385_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295239424))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305069888))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305070080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305077824))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_142_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized, x = input_387_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")];
            tensor<int32, [2]> var_2215_split_sizes_0 = const()[name = tensor<string, []>("op_2215_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_2215_axis_0 = const()[name = tensor<string, []>("op_2215_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_2215_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_2215_cast_fp16_1 = split(axis = var_2215_axis_0, split_sizes = var_2215_split_sizes_0, x = linear_142_cast_fp16)[name = tensor<string, []>("op_2215_cast_fp16")];
            tensor<string, []> var_2217_mode_0 = const()[name = tensor<string, []>("op_2217_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_2217_cast_fp16 = gelu(mode = var_2217_mode_0, x = var_2215_cast_fp16_1)[name = tensor<string, []>("op_2217_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_389_cast_fp16 = mul(x = var_2215_cast_fp16_0, y = var_2217_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305078016))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309993280))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_8_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_8_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309993472)))];
            tensor<fp16, [1, 1024, 1280]> linear_143_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_8_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized, x = input_389_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_305_cast_fp16 = add(x = linear_143_cast_fp16, y = input_385_cast_fp16)[name = tensor<string, []>("hidden_states_305_cast_fp16")];
            tensor<int32, [1]> hidden_states_307_axes_0 = const()[name = tensor<string, []>("hidden_states_307_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309996096)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309998720)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_307_cast_fp16 = layer_norm(axes = hidden_states_307_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_9_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_9_norm1_weight_to_fp16, x = hidden_states_305_cast_fp16)[name = tensor<string, []>("hidden_states_307_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310001344))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311230208))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_144_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_307_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311230400))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312459264))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_145_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_307_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312459456))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313688320))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_146_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_307_cast_fp16)[name = tensor<string, []>("linear_146_cast_fp16")];
            tensor<int32, [4]> var_2252 = const()[name = tensor<string, []>("op_2252"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2253_cast_fp16 = reshape(shape = var_2252, x = linear_144_cast_fp16)[name = tensor<string, []>("op_2253_cast_fp16")];
            tensor<int32, [4]> var_2255 = const()[name = tensor<string, []>("op_2255"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2256_cast_fp16 = reshape(shape = var_2255, x = linear_145_cast_fp16)[name = tensor<string, []>("op_2256_cast_fp16")];
            tensor<int32, [4]> var_2258 = const()[name = tensor<string, []>("op_2258"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2259_cast_fp16 = reshape(shape = var_2258, x = linear_146_cast_fp16)[name = tensor<string, []>("op_2259_cast_fp16")];
            tensor<int32, [4]> value_107_perm_0 = const()[name = tensor<string, []>("value_107_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_39_y_0_to_fp16 = const()[name = tensor<string, []>("mul_39_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_39_cast_fp16 = mul(x = var_2253_cast_fp16, y = mul_39_y_0_to_fp16)[name = tensor<string, []>("mul_39_cast_fp16")];
            tensor<bool, []> matmul_26_transpose_y_0 = const()[name = tensor<string, []>("matmul_26_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_26_transpose_x_0 = const()[name = tensor<string, []>("matmul_26_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_324_perm_0 = const()[name = tensor<string, []>("transpose_324_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_325_perm_0 = const()[name = tensor<string, []>("transpose_325_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_579 = transpose(perm = transpose_325_perm_0, x = var_2256_cast_fp16)[name = tensor<string, []>("transpose_579")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_580 = transpose(perm = transpose_324_perm_0, x = mul_39_cast_fp16)[name = tensor<string, []>("transpose_580")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_26_cast_fp16 = matmul(transpose_x = matmul_26_transpose_x_0, transpose_y = matmul_26_transpose_y_0, x = transpose_580, y = transpose_579)[name = tensor<string, []>("matmul_26_cast_fp16")];
            tensor<int32, []> softmax_26_axis_0 = const()[name = tensor<string, []>("softmax_26_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_26_cast_fp16 = softmax(axis = softmax_26_axis_0, x = matmul_26_cast_fp16)[name = tensor<string, []>("softmax_26_cast_fp16")];
            tensor<bool, []> hidden_states_309_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_309_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_309_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_309_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_578 = transpose(perm = value_107_perm_0, x = var_2259_cast_fp16)[name = tensor<string, []>("transpose_578")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_309_cast_fp16 = matmul(transpose_x = hidden_states_309_transpose_x_0, transpose_y = hidden_states_309_transpose_y_0, x = softmax_26_cast_fp16, y = transpose_578)[name = tensor<string, []>("hidden_states_309_cast_fp16")];
            tensor<int32, [4]> var_2262_perm_0 = const()[name = tensor<string, []>("op_2262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2266 = const()[name = tensor<string, []>("op_2266"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_577 = transpose(perm = var_2262_perm_0, x = hidden_states_309_cast_fp16)[name = tensor<string, []>("transpose_577")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_311_cast_fp16 = reshape(shape = var_2266, x = transpose_577)[name = tensor<string, []>("hidden_states_311_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313688512))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314917376))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314917568)))];
            tensor<fp16, [1, 1024, 1280]> linear_147_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_311_cast_fp16)[name = tensor<string, []>("linear_147_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_397_cast_fp16 = add(x = linear_147_cast_fp16, y = hidden_states_305_cast_fp16)[name = tensor<string, []>("input_397_cast_fp16")];
            tensor<int32, [1]> input_399_axes_0 = const()[name = tensor<string, []>("input_399_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314920192)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314922816)))];
            tensor<fp16, [1, 1024, 1280]> input_399_cast_fp16 = layer_norm(axes = input_399_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_9_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_9_norm2_weight_to_fp16, x = input_397_cast_fp16)[name = tensor<string, []>("input_399_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314925440))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316154304))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_148_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized, x = input_399_cast_fp16)[name = tensor<string, []>("linear_148_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316154496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318120640))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_149_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_149_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318120832))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320086976))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_150_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_150_cast_fp16")];
            tensor<int32, [4]> var_2298 = const()[name = tensor<string, []>("op_2298"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2299_cast_fp16 = reshape(shape = var_2298, x = linear_148_cast_fp16)[name = tensor<string, []>("op_2299_cast_fp16")];
            tensor<int32, [4]> var_2301 = const()[name = tensor<string, []>("op_2301"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2302_cast_fp16 = reshape(shape = var_2301, x = linear_149_cast_fp16)[name = tensor<string, []>("op_2302_cast_fp16")];
            tensor<int32, [4]> var_2304 = const()[name = tensor<string, []>("op_2304"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2305_cast_fp16 = reshape(shape = var_2304, x = linear_150_cast_fp16)[name = tensor<string, []>("op_2305_cast_fp16")];
            tensor<int32, [4]> value_111_perm_0 = const()[name = tensor<string, []>("value_111_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_40_y_0_to_fp16 = const()[name = tensor<string, []>("mul_40_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_40_cast_fp16 = mul(x = var_2299_cast_fp16, y = mul_40_y_0_to_fp16)[name = tensor<string, []>("mul_40_cast_fp16")];
            tensor<bool, []> matmul_27_transpose_y_0 = const()[name = tensor<string, []>("matmul_27_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_27_transpose_x_0 = const()[name = tensor<string, []>("matmul_27_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_326_perm_0 = const()[name = tensor<string, []>("transpose_326_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_327_perm_0 = const()[name = tensor<string, []>("transpose_327_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_575 = transpose(perm = transpose_327_perm_0, x = var_2302_cast_fp16)[name = tensor<string, []>("transpose_575")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_576 = transpose(perm = transpose_326_perm_0, x = mul_40_cast_fp16)[name = tensor<string, []>("transpose_576")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_27_cast_fp16 = matmul(transpose_x = matmul_27_transpose_x_0, transpose_y = matmul_27_transpose_y_0, x = transpose_576, y = transpose_575)[name = tensor<string, []>("matmul_27_cast_fp16")];
            tensor<int32, []> softmax_27_axis_0 = const()[name = tensor<string, []>("softmax_27_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_27_cast_fp16 = softmax(axis = softmax_27_axis_0, x = matmul_27_cast_fp16)[name = tensor<string, []>("softmax_27_cast_fp16")];
            tensor<bool, []> hidden_states_315_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_315_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_315_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_315_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_574 = transpose(perm = value_111_perm_0, x = var_2305_cast_fp16)[name = tensor<string, []>("transpose_574")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_315_cast_fp16 = matmul(transpose_x = hidden_states_315_transpose_x_0, transpose_y = hidden_states_315_transpose_y_0, x = softmax_27_cast_fp16, y = transpose_574)[name = tensor<string, []>("hidden_states_315_cast_fp16")];
            tensor<int32, [4]> var_2308_perm_0 = const()[name = tensor<string, []>("op_2308_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2312 = const()[name = tensor<string, []>("op_2312"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_573 = transpose(perm = var_2308_perm_0, x = hidden_states_315_cast_fp16)[name = tensor<string, []>("transpose_573")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_317_cast_fp16 = reshape(shape = var_2312, x = transpose_573)[name = tensor<string, []>("hidden_states_317_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320087168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321316032))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321316224)))];
            tensor<fp16, [1, 1024, 1280]> linear_151_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_317_cast_fp16)[name = tensor<string, []>("linear_151_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_405_cast_fp16 = add(x = linear_151_cast_fp16, y = input_397_cast_fp16)[name = tensor<string, []>("input_405_cast_fp16")];
            tensor<int32, [1]> input_407_axes_0 = const()[name = tensor<string, []>("input_407_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321318848)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321321472)))];
            tensor<fp16, [1, 1024, 1280]> input_407_cast_fp16 = layer_norm(axes = input_407_axes_0, beta = down_blocks_2_attentions_0_transformer_blocks_9_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_0_transformer_blocks_9_norm3_weight_to_fp16, x = input_405_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_0_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321324096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331154560))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_0_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331154752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331162496))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_152_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_0_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = tensor<string, []>("linear_152_cast_fp16")];
            tensor<int32, [2]> var_2334_split_sizes_0 = const()[name = tensor<string, []>("op_2334_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_2334_axis_0 = const()[name = tensor<string, []>("op_2334_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_2334_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_2334_cast_fp16_1 = split(axis = var_2334_axis_0, split_sizes = var_2334_split_sizes_0, x = linear_152_cast_fp16)[name = tensor<string, []>("op_2334_cast_fp16")];
            tensor<string, []> var_2336_mode_0 = const()[name = tensor<string, []>("op_2336_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_2336_cast_fp16 = gelu(mode = var_2336_mode_0, x = var_2334_cast_fp16_1)[name = tensor<string, []>("op_2336_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_409_cast_fp16 = mul(x = var_2334_cast_fp16_0, y = var_2336_cast_fp16)[name = tensor<string, []>("input_409_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_0_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331162688))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336077952))), name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_transformer_blocks_9_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_transformer_blocks_9_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336078144)))];
            tensor<fp16, [1, 1024, 1280]> linear_153_cast_fp16 = linear(bias = down_blocks_2_attentions_0_transformer_blocks_9_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_0_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized, x = input_409_cast_fp16)[name = tensor<string, []>("linear_153_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_413_cast_fp16 = add(x = linear_153_cast_fp16, y = input_405_cast_fp16)[name = tensor<string, []>("input_413_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_0_proj_out_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336080768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337309632))), name = tensor<string, []>("down_blocks_2_attentions_0_proj_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_0_proj_out_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_0_proj_out_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337309824)))];
            tensor<fp16, [1, 1024, 1280]> linear_154_cast_fp16 = linear(bias = down_blocks_2_attentions_0_proj_out_bias_to_fp16, weight = down_blocks_2_attentions_0_proj_out_weight_to_fp16_palettized, x = input_413_cast_fp16)[name = tensor<string, []>("linear_154_cast_fp16")];
            tensor<int32, [4]> var_2346 = const()[name = tensor<string, []>("op_2346"), val = tensor<int32, [4]>([1, 32, 32, 1280])];
            tensor<fp16, [1, 32, 32, 1280]> var_2347_cast_fp16 = reshape(shape = var_2346, x = linear_154_cast_fp16)[name = tensor<string, []>("op_2347_cast_fp16")];
            tensor<int32, [4]> var_2348 = const()[name = tensor<string, []>("op_2348"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 1280, 32, 32]> transpose_572 = transpose(perm = var_2348, x = var_2347_cast_fp16)[name = tensor<string, []>("transpose_572")];
            tensor<fp16, [1, 1280, 32, 32]> input_415_cast_fp16 = add(x = transpose_572, y = var_1106_cast_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
            tensor<int32, [5]> reshape_52_shape_0 = const()[name = tensor<string, []>("reshape_52_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = input_415_cast_fp16)[name = tensor<string, []>("reshape_52_cast_fp16")];
            tensor<int32, [3]> reduce_mean_39_axes_0 = const()[name = tensor<string, []>("reduce_mean_39_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_39_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_39_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = reshape_52_cast_fp16)[name = tensor<string, []>("reduce_mean_39_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_26_cast_fp16 = sub(x = reshape_52_cast_fp16, y = reduce_mean_39_cast_fp16)[name = tensor<string, []>("sub_26_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_13_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor<string, []>("square_13_cast_fp16")];
            tensor<int32, [3]> reduce_mean_41_axes_0 = const()[name = tensor<string, []>("reduce_mean_41_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_41_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_41_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_13_cast_fp16)[name = tensor<string, []>("reduce_mean_41_cast_fp16")];
            tensor<fp16, []> add_26_y_0_to_fp16 = const()[name = tensor<string, []>("add_26_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_26_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = add_26_y_0_to_fp16)[name = tensor<string, []>("add_26_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_13_cast_fp16 = sqrt(x = add_26_cast_fp16)[name = tensor<string, []>("sqrt_13_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_13_cast_fp16 = real_div(x = sub_26_cast_fp16, y = sqrt_13_cast_fp16)[name = tensor<string, []>("real_div_13_cast_fp16")];
            tensor<int32, [4]> reshape_53_shape_0 = const()[name = tensor<string, []>("reshape_53_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = real_div_13_cast_fp16)[name = tensor<string, []>("reshape_53_cast_fp16")];
            tensor<fp16, [1280]> add_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_27_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337312448)))];
            tensor<fp16, [1280]> add_27_beta_0_to_fp16 = const()[name = tensor<string, []>("add_27_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337315072)))];
            tensor<fp16, []> add_27_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_27_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_27_cast_fp16 = batch_norm(beta = add_27_beta_0_to_fp16, epsilon = add_27_epsilon_0_to_fp16, gamma = add_27_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_53_cast_fp16)[name = tensor<string, []>("add_27_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_419_cast_fp16 = silu(x = add_27_cast_fp16)[name = tensor<string, []>("input_419_cast_fp16")];
            tensor<int32, [2]> var_2363 = const()[name = tensor<string, []>("op_2363"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2365 = const()[name = tensor<string, []>("op_2365"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_329_pad_type_0 = const()[name = tensor<string, []>("hidden_states_329_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_329_pad_0 = const()[name = tensor<string, []>("hidden_states_329_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [1280, 1280, 3, 3]> down_blocks_2_resnets_1_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [11059200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337317696))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348376960))), name = tensor<string, []>("down_blocks_2_resnets_1_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 3, 3])];
            tensor<fp16, [1280]> down_blocks_2_resnets_1_conv1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_resnets_1_conv1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348377152)))];
            tensor<fp16, [1, 1280, 32, 32]> hidden_states_329_cast_fp16 = conv(bias = down_blocks_2_resnets_1_conv1_bias_to_fp16, dilations = var_2365, groups = var_1048, pad = hidden_states_329_pad_0, pad_type = hidden_states_329_pad_type_0, strides = var_2363, weight = down_blocks_2_resnets_1_conv1_weight_to_fp16_palettized, x = input_419_cast_fp16)[name = tensor<string, []>("hidden_states_329_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_resnets_1_time_emb_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348379776))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349608640))), name = tensor<string, []>("down_blocks_2_resnets_1_time_emb_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_resnets_1_time_emb_proj_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_resnets_1_time_emb_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349608832)))];
            tensor<fp16, [1, 1280]> linear_155_cast_fp16 = linear(bias = down_blocks_2_resnets_1_time_emb_proj_bias_to_fp16, weight = down_blocks_2_resnets_1_time_emb_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_155_cast_fp16")];
            tensor<int32, [1]> var_2374_axes_0 = const()[name = tensor<string, []>("op_2374_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1280, 1]> var_2374_cast_fp16 = expand_dims(axes = var_2374_axes_0, x = linear_155_cast_fp16)[name = tensor<string, []>("op_2374_cast_fp16")];
            tensor<int32, [1]> temb_11_axes_0 = const()[name = tensor<string, []>("temb_11_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 1280, 1, 1]> temb_11_cast_fp16 = expand_dims(axes = temb_11_axes_0, x = var_2374_cast_fp16)[name = tensor<string, []>("temb_11_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_423_cast_fp16 = add(x = hidden_states_329_cast_fp16, y = temb_11_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
            tensor<int32, [5]> reshape_56_shape_0 = const()[name = tensor<string, []>("reshape_56_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_56_cast_fp16 = reshape(shape = reshape_56_shape_0, x = input_423_cast_fp16)[name = tensor<string, []>("reshape_56_cast_fp16")];
            tensor<int32, [3]> reduce_mean_42_axes_0 = const()[name = tensor<string, []>("reduce_mean_42_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_42_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_42_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_42_cast_fp16 = reduce_mean(axes = reduce_mean_42_axes_0, keep_dims = reduce_mean_42_keep_dims_0, x = reshape_56_cast_fp16)[name = tensor<string, []>("reduce_mean_42_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_28_cast_fp16 = sub(x = reshape_56_cast_fp16, y = reduce_mean_42_cast_fp16)[name = tensor<string, []>("sub_28_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_14_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor<string, []>("square_14_cast_fp16")];
            tensor<int32, [3]> reduce_mean_44_axes_0 = const()[name = tensor<string, []>("reduce_mean_44_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_44_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_44_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_44_cast_fp16 = reduce_mean(axes = reduce_mean_44_axes_0, keep_dims = reduce_mean_44_keep_dims_0, x = square_14_cast_fp16)[name = tensor<string, []>("reduce_mean_44_cast_fp16")];
            tensor<fp16, []> add_28_y_0_to_fp16 = const()[name = tensor<string, []>("add_28_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_28_cast_fp16 = add(x = reduce_mean_44_cast_fp16, y = add_28_y_0_to_fp16)[name = tensor<string, []>("add_28_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_14_cast_fp16 = sqrt(x = add_28_cast_fp16)[name = tensor<string, []>("sqrt_14_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_14_cast_fp16 = real_div(x = sub_28_cast_fp16, y = sqrt_14_cast_fp16)[name = tensor<string, []>("real_div_14_cast_fp16")];
            tensor<int32, [4]> reshape_57_shape_0 = const()[name = tensor<string, []>("reshape_57_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = real_div_14_cast_fp16)[name = tensor<string, []>("reshape_57_cast_fp16")];
            tensor<fp16, [1280]> add_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_29_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349611456)))];
            tensor<fp16, [1280]> add_29_beta_0_to_fp16 = const()[name = tensor<string, []>("add_29_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349614080)))];
            tensor<fp16, []> add_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_29_cast_fp16 = batch_norm(beta = add_29_beta_0_to_fp16, epsilon = add_29_epsilon_0_to_fp16, gamma = add_29_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_57_cast_fp16)[name = tensor<string, []>("add_29_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_427_cast_fp16 = silu(x = add_29_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
            tensor<int32, [2]> var_2384 = const()[name = tensor<string, []>("op_2384"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2386 = const()[name = tensor<string, []>("op_2386"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_331_pad_type_0 = const()[name = tensor<string, []>("hidden_states_331_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_331_pad_0 = const()[name = tensor<string, []>("hidden_states_331_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [1280, 1280, 3, 3]> down_blocks_2_resnets_1_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [11059200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349616704))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360675968))), name = tensor<string, []>("down_blocks_2_resnets_1_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 3, 3])];
            tensor<fp16, [1280]> down_blocks_2_resnets_1_conv2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_resnets_1_conv2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360676160)))];
            tensor<fp16, [1, 1280, 32, 32]> hidden_states_331_cast_fp16 = conv(bias = down_blocks_2_resnets_1_conv2_bias_to_fp16, dilations = var_2386, groups = var_1048, pad = hidden_states_331_pad_0, pad_type = hidden_states_331_pad_type_0, strides = var_2384, weight = down_blocks_2_resnets_1_conv2_weight_to_fp16_palettized, x = input_427_cast_fp16)[name = tensor<string, []>("hidden_states_331_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> var_2389_cast_fp16 = add(x = input_415_cast_fp16, y = hidden_states_331_cast_fp16)[name = tensor<string, []>("op_2389_cast_fp16")];
            tensor<int32, [5]> reshape_60_shape_0 = const()[name = tensor<string, []>("reshape_60_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_60_cast_fp16 = reshape(shape = reshape_60_shape_0, x = var_2389_cast_fp16)[name = tensor<string, []>("reshape_60_cast_fp16")];
            tensor<int32, [3]> reduce_mean_45_axes_0 = const()[name = tensor<string, []>("reduce_mean_45_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_45_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_45_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = reshape_60_cast_fp16)[name = tensor<string, []>("reduce_mean_45_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_30_cast_fp16 = sub(x = reshape_60_cast_fp16, y = reduce_mean_45_cast_fp16)[name = tensor<string, []>("sub_30_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_15_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor<string, []>("square_15_cast_fp16")];
            tensor<int32, [3]> reduce_mean_47_axes_0 = const()[name = tensor<string, []>("reduce_mean_47_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_47_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_47_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_15_cast_fp16)[name = tensor<string, []>("reduce_mean_47_cast_fp16")];
            tensor<fp16, []> add_30_y_0_to_fp16 = const()[name = tensor<string, []>("add_30_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_30_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = add_30_y_0_to_fp16)[name = tensor<string, []>("add_30_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_15_cast_fp16 = sqrt(x = add_30_cast_fp16)[name = tensor<string, []>("sqrt_15_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_15_cast_fp16 = real_div(x = sub_30_cast_fp16, y = sqrt_15_cast_fp16)[name = tensor<string, []>("real_div_15_cast_fp16")];
            tensor<int32, [4]> reshape_61_shape_0 = const()[name = tensor<string, []>("reshape_61_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_61_cast_fp16 = reshape(shape = reshape_61_shape_0, x = real_div_15_cast_fp16)[name = tensor<string, []>("reshape_61_cast_fp16")];
            tensor<fp16, [1280]> add_31_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_31_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360678784)))];
            tensor<fp16, [1280]> add_31_beta_0_to_fp16 = const()[name = tensor<string, []>("add_31_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360681408)))];
            tensor<fp16, []> add_31_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_31_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_31_cast_fp16 = batch_norm(beta = add_31_beta_0_to_fp16, epsilon = add_31_epsilon_0_to_fp16, gamma = add_31_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_61_cast_fp16)[name = tensor<string, []>("add_31_cast_fp16")];
            tensor<int32, [4]> var_2427 = const()[name = tensor<string, []>("op_2427"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2431 = const()[name = tensor<string, []>("op_2431"), val = tensor<int32, [3]>([1, 1024, 1280])];
            tensor<fp16, [1, 32, 32, 1280]> transpose_571 = transpose(perm = var_2427, x = add_31_cast_fp16)[name = tensor<string, []>("transpose_571")];
            tensor<fp16, [1, 1024, 1280]> input_431_cast_fp16 = reshape(shape = var_2431, x = transpose_571)[name = tensor<string, []>("input_431_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_proj_in_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360684032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361912896))), name = tensor<string, []>("down_blocks_2_attentions_1_proj_in_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_proj_in_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_proj_in_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361913088)))];
            tensor<fp16, [1, 1024, 1280]> linear_156_cast_fp16 = linear(bias = down_blocks_2_attentions_1_proj_in_bias_to_fp16, weight = down_blocks_2_attentions_1_proj_in_weight_to_fp16_palettized, x = input_431_cast_fp16)[name = tensor<string, []>("linear_156_cast_fp16")];
            tensor<int32, [1]> hidden_states_339_axes_0 = const()[name = tensor<string, []>("hidden_states_339_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361915712)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361918336)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_339_cast_fp16 = layer_norm(axes = hidden_states_339_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_0_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_0_norm1_weight_to_fp16, x = linear_156_cast_fp16)[name = tensor<string, []>("hidden_states_339_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361920960))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(363149824))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_157_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_339_cast_fp16)[name = tensor<string, []>("linear_157_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(363150016))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(364378880))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_158_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_339_cast_fp16)[name = tensor<string, []>("linear_158_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(364379072))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(365607936))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_159_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_339_cast_fp16)[name = tensor<string, []>("linear_159_cast_fp16")];
            tensor<int32, [4]> var_2464 = const()[name = tensor<string, []>("op_2464"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2465_cast_fp16 = reshape(shape = var_2464, x = linear_157_cast_fp16)[name = tensor<string, []>("op_2465_cast_fp16")];
            tensor<int32, [4]> var_2467 = const()[name = tensor<string, []>("op_2467"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2468_cast_fp16 = reshape(shape = var_2467, x = linear_158_cast_fp16)[name = tensor<string, []>("op_2468_cast_fp16")];
            tensor<int32, [4]> var_2470 = const()[name = tensor<string, []>("op_2470"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2471_cast_fp16 = reshape(shape = var_2470, x = linear_159_cast_fp16)[name = tensor<string, []>("op_2471_cast_fp16")];
            tensor<int32, [4]> value_115_perm_0 = const()[name = tensor<string, []>("value_115_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_44_y_0_to_fp16 = const()[name = tensor<string, []>("mul_44_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_44_cast_fp16 = mul(x = var_2465_cast_fp16, y = mul_44_y_0_to_fp16)[name = tensor<string, []>("mul_44_cast_fp16")];
            tensor<bool, []> matmul_28_transpose_y_0 = const()[name = tensor<string, []>("matmul_28_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_28_transpose_x_0 = const()[name = tensor<string, []>("matmul_28_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_328_perm_0 = const()[name = tensor<string, []>("transpose_328_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_329_perm_0 = const()[name = tensor<string, []>("transpose_329_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_569 = transpose(perm = transpose_329_perm_0, x = var_2468_cast_fp16)[name = tensor<string, []>("transpose_569")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_570 = transpose(perm = transpose_328_perm_0, x = mul_44_cast_fp16)[name = tensor<string, []>("transpose_570")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_28_cast_fp16 = matmul(transpose_x = matmul_28_transpose_x_0, transpose_y = matmul_28_transpose_y_0, x = transpose_570, y = transpose_569)[name = tensor<string, []>("matmul_28_cast_fp16")];
            tensor<int32, []> softmax_28_axis_0 = const()[name = tensor<string, []>("softmax_28_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_28_cast_fp16 = softmax(axis = softmax_28_axis_0, x = matmul_28_cast_fp16)[name = tensor<string, []>("softmax_28_cast_fp16")];
            tensor<bool, []> hidden_states_341_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_341_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_341_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_341_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_568 = transpose(perm = value_115_perm_0, x = var_2471_cast_fp16)[name = tensor<string, []>("transpose_568")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_341_cast_fp16 = matmul(transpose_x = hidden_states_341_transpose_x_0, transpose_y = hidden_states_341_transpose_y_0, x = softmax_28_cast_fp16, y = transpose_568)[name = tensor<string, []>("hidden_states_341_cast_fp16")];
            tensor<int32, [4]> var_2474_perm_0 = const()[name = tensor<string, []>("op_2474_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2478 = const()[name = tensor<string, []>("op_2478"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_567 = transpose(perm = var_2474_perm_0, x = hidden_states_341_cast_fp16)[name = tensor<string, []>("transpose_567")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_343_cast_fp16 = reshape(shape = var_2478, x = transpose_567)[name = tensor<string, []>("hidden_states_343_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(365608128))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366836992))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366837184)))];
            tensor<fp16, [1, 1024, 1280]> linear_160_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_343_cast_fp16)[name = tensor<string, []>("linear_160_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_437_cast_fp16 = add(x = linear_160_cast_fp16, y = linear_156_cast_fp16)[name = tensor<string, []>("input_437_cast_fp16")];
            tensor<int32, [1]> input_439_axes_0 = const()[name = tensor<string, []>("input_439_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366839808)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366842432)))];
            tensor<fp16, [1, 1024, 1280]> input_439_cast_fp16 = layer_norm(axes = input_439_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_0_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_0_norm2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366845056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368073920))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_161_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized, x = input_439_cast_fp16)[name = tensor<string, []>("linear_161_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368074112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370040256))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_162_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_162_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370040448))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372006592))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_163_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_163_cast_fp16")];
            tensor<int32, [4]> var_2510 = const()[name = tensor<string, []>("op_2510"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2511_cast_fp16 = reshape(shape = var_2510, x = linear_161_cast_fp16)[name = tensor<string, []>("op_2511_cast_fp16")];
            tensor<int32, [4]> var_2513 = const()[name = tensor<string, []>("op_2513"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2514_cast_fp16 = reshape(shape = var_2513, x = linear_162_cast_fp16)[name = tensor<string, []>("op_2514_cast_fp16")];
            tensor<int32, [4]> var_2516 = const()[name = tensor<string, []>("op_2516"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2517_cast_fp16 = reshape(shape = var_2516, x = linear_163_cast_fp16)[name = tensor<string, []>("op_2517_cast_fp16")];
            tensor<int32, [4]> value_119_perm_0 = const()[name = tensor<string, []>("value_119_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_45_y_0_to_fp16 = const()[name = tensor<string, []>("mul_45_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_45_cast_fp16 = mul(x = var_2511_cast_fp16, y = mul_45_y_0_to_fp16)[name = tensor<string, []>("mul_45_cast_fp16")];
            tensor<bool, []> matmul_29_transpose_y_0 = const()[name = tensor<string, []>("matmul_29_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_29_transpose_x_0 = const()[name = tensor<string, []>("matmul_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_330_perm_0 = const()[name = tensor<string, []>("transpose_330_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_331_perm_0 = const()[name = tensor<string, []>("transpose_331_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_565 = transpose(perm = transpose_331_perm_0, x = var_2514_cast_fp16)[name = tensor<string, []>("transpose_565")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_566 = transpose(perm = transpose_330_perm_0, x = mul_45_cast_fp16)[name = tensor<string, []>("transpose_566")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_29_cast_fp16 = matmul(transpose_x = matmul_29_transpose_x_0, transpose_y = matmul_29_transpose_y_0, x = transpose_566, y = transpose_565)[name = tensor<string, []>("matmul_29_cast_fp16")];
            tensor<int32, []> softmax_29_axis_0 = const()[name = tensor<string, []>("softmax_29_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_29_cast_fp16 = softmax(axis = softmax_29_axis_0, x = matmul_29_cast_fp16)[name = tensor<string, []>("softmax_29_cast_fp16")];
            tensor<bool, []> hidden_states_347_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_347_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_347_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_347_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_564 = transpose(perm = value_119_perm_0, x = var_2517_cast_fp16)[name = tensor<string, []>("transpose_564")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_347_cast_fp16 = matmul(transpose_x = hidden_states_347_transpose_x_0, transpose_y = hidden_states_347_transpose_y_0, x = softmax_29_cast_fp16, y = transpose_564)[name = tensor<string, []>("hidden_states_347_cast_fp16")];
            tensor<int32, [4]> var_2520_perm_0 = const()[name = tensor<string, []>("op_2520_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2524 = const()[name = tensor<string, []>("op_2524"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_563 = transpose(perm = var_2520_perm_0, x = hidden_states_347_cast_fp16)[name = tensor<string, []>("transpose_563")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_349_cast_fp16 = reshape(shape = var_2524, x = transpose_563)[name = tensor<string, []>("hidden_states_349_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372006784))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373235648))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373235840)))];
            tensor<fp16, [1, 1024, 1280]> linear_164_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_349_cast_fp16)[name = tensor<string, []>("linear_164_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_445_cast_fp16 = add(x = linear_164_cast_fp16, y = input_437_cast_fp16)[name = tensor<string, []>("input_445_cast_fp16")];
            tensor<int32, [1]> input_447_axes_0 = const()[name = tensor<string, []>("input_447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373238464)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373241088)))];
            tensor<fp16, [1, 1024, 1280]> input_447_cast_fp16 = layer_norm(axes = input_447_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_0_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_0_norm3_weight_to_fp16, x = input_445_cast_fp16)[name = tensor<string, []>("input_447_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373243712))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383074176))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383074368))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383082112))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_165_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = tensor<string, []>("linear_165_cast_fp16")];
            tensor<int32, [2]> var_2546_split_sizes_0 = const()[name = tensor<string, []>("op_2546_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_2546_axis_0 = const()[name = tensor<string, []>("op_2546_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_2546_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_2546_cast_fp16_1 = split(axis = var_2546_axis_0, split_sizes = var_2546_split_sizes_0, x = linear_165_cast_fp16)[name = tensor<string, []>("op_2546_cast_fp16")];
            tensor<string, []> var_2548_mode_0 = const()[name = tensor<string, []>("op_2548_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_2548_cast_fp16 = gelu(mode = var_2548_mode_0, x = var_2546_cast_fp16_1)[name = tensor<string, []>("op_2548_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_449_cast_fp16 = mul(x = var_2546_cast_fp16_0, y = var_2548_cast_fp16)[name = tensor<string, []>("input_449_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383082304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387997568))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_0_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_0_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387997760)))];
            tensor<fp16, [1, 1024, 1280]> linear_166_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_0_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized, x = input_449_cast_fp16)[name = tensor<string, []>("linear_166_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_357_cast_fp16 = add(x = linear_166_cast_fp16, y = input_445_cast_fp16)[name = tensor<string, []>("hidden_states_357_cast_fp16")];
            tensor<int32, [1]> hidden_states_359_axes_0 = const()[name = tensor<string, []>("hidden_states_359_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388000384)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388003008)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_359_cast_fp16 = layer_norm(axes = hidden_states_359_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_1_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_1_norm1_weight_to_fp16, x = hidden_states_357_cast_fp16)[name = tensor<string, []>("hidden_states_359_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388005632))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389234496))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_167_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_359_cast_fp16)[name = tensor<string, []>("linear_167_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389234688))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390463552))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_168_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_359_cast_fp16)[name = tensor<string, []>("linear_168_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390463744))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391692608))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_169_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_359_cast_fp16)[name = tensor<string, []>("linear_169_cast_fp16")];
            tensor<int32, [4]> var_2583 = const()[name = tensor<string, []>("op_2583"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2584_cast_fp16 = reshape(shape = var_2583, x = linear_167_cast_fp16)[name = tensor<string, []>("op_2584_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = tensor<string, []>("op_2586"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_168_cast_fp16)[name = tensor<string, []>("op_2587_cast_fp16")];
            tensor<int32, [4]> var_2589 = const()[name = tensor<string, []>("op_2589"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2590_cast_fp16 = reshape(shape = var_2589, x = linear_169_cast_fp16)[name = tensor<string, []>("op_2590_cast_fp16")];
            tensor<int32, [4]> value_123_perm_0 = const()[name = tensor<string, []>("value_123_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_46_y_0_to_fp16 = const()[name = tensor<string, []>("mul_46_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_46_cast_fp16 = mul(x = var_2584_cast_fp16, y = mul_46_y_0_to_fp16)[name = tensor<string, []>("mul_46_cast_fp16")];
            tensor<bool, []> matmul_30_transpose_y_0 = const()[name = tensor<string, []>("matmul_30_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_30_transpose_x_0 = const()[name = tensor<string, []>("matmul_30_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_332_perm_0 = const()[name = tensor<string, []>("transpose_332_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_333_perm_0 = const()[name = tensor<string, []>("transpose_333_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_561 = transpose(perm = transpose_333_perm_0, x = var_2587_cast_fp16)[name = tensor<string, []>("transpose_561")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_562 = transpose(perm = transpose_332_perm_0, x = mul_46_cast_fp16)[name = tensor<string, []>("transpose_562")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_30_cast_fp16 = matmul(transpose_x = matmul_30_transpose_x_0, transpose_y = matmul_30_transpose_y_0, x = transpose_562, y = transpose_561)[name = tensor<string, []>("matmul_30_cast_fp16")];
            tensor<int32, []> softmax_30_axis_0 = const()[name = tensor<string, []>("softmax_30_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_30_cast_fp16 = softmax(axis = softmax_30_axis_0, x = matmul_30_cast_fp16)[name = tensor<string, []>("softmax_30_cast_fp16")];
            tensor<bool, []> hidden_states_361_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_361_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_361_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_361_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_560 = transpose(perm = value_123_perm_0, x = var_2590_cast_fp16)[name = tensor<string, []>("transpose_560")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_361_cast_fp16 = matmul(transpose_x = hidden_states_361_transpose_x_0, transpose_y = hidden_states_361_transpose_y_0, x = softmax_30_cast_fp16, y = transpose_560)[name = tensor<string, []>("hidden_states_361_cast_fp16")];
            tensor<int32, [4]> var_2593_perm_0 = const()[name = tensor<string, []>("op_2593_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2597 = const()[name = tensor<string, []>("op_2597"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_559 = transpose(perm = var_2593_perm_0, x = hidden_states_361_cast_fp16)[name = tensor<string, []>("transpose_559")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_363_cast_fp16 = reshape(shape = var_2597, x = transpose_559)[name = tensor<string, []>("hidden_states_363_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391692800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392921664))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392921856)))];
            tensor<fp16, [1, 1024, 1280]> linear_170_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_363_cast_fp16)[name = tensor<string, []>("linear_170_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_457_cast_fp16 = add(x = linear_170_cast_fp16, y = hidden_states_357_cast_fp16)[name = tensor<string, []>("input_457_cast_fp16")];
            tensor<int32, [1]> input_459_axes_0 = const()[name = tensor<string, []>("input_459_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392924480)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392927104)))];
            tensor<fp16, [1, 1024, 1280]> input_459_cast_fp16 = layer_norm(axes = input_459_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_1_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_1_norm2_weight_to_fp16, x = input_457_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392929728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394158592))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_171_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = tensor<string, []>("linear_171_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394158784))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396124928))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_172_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_172_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396125120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(398091264))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_173_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_173_cast_fp16")];
            tensor<int32, [4]> var_2629 = const()[name = tensor<string, []>("op_2629"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2630_cast_fp16 = reshape(shape = var_2629, x = linear_171_cast_fp16)[name = tensor<string, []>("op_2630_cast_fp16")];
            tensor<int32, [4]> var_2632 = const()[name = tensor<string, []>("op_2632"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2633_cast_fp16 = reshape(shape = var_2632, x = linear_172_cast_fp16)[name = tensor<string, []>("op_2633_cast_fp16")];
            tensor<int32, [4]> var_2635 = const()[name = tensor<string, []>("op_2635"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2636_cast_fp16 = reshape(shape = var_2635, x = linear_173_cast_fp16)[name = tensor<string, []>("op_2636_cast_fp16")];
            tensor<int32, [4]> value_127_perm_0 = const()[name = tensor<string, []>("value_127_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_47_y_0_to_fp16 = const()[name = tensor<string, []>("mul_47_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_47_cast_fp16 = mul(x = var_2630_cast_fp16, y = mul_47_y_0_to_fp16)[name = tensor<string, []>("mul_47_cast_fp16")];
            tensor<bool, []> matmul_31_transpose_y_0 = const()[name = tensor<string, []>("matmul_31_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_31_transpose_x_0 = const()[name = tensor<string, []>("matmul_31_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_334_perm_0 = const()[name = tensor<string, []>("transpose_334_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_335_perm_0 = const()[name = tensor<string, []>("transpose_335_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_557 = transpose(perm = transpose_335_perm_0, x = var_2633_cast_fp16)[name = tensor<string, []>("transpose_557")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_558 = transpose(perm = transpose_334_perm_0, x = mul_47_cast_fp16)[name = tensor<string, []>("transpose_558")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_31_cast_fp16 = matmul(transpose_x = matmul_31_transpose_x_0, transpose_y = matmul_31_transpose_y_0, x = transpose_558, y = transpose_557)[name = tensor<string, []>("matmul_31_cast_fp16")];
            tensor<int32, []> softmax_31_axis_0 = const()[name = tensor<string, []>("softmax_31_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_31_cast_fp16 = softmax(axis = softmax_31_axis_0, x = matmul_31_cast_fp16)[name = tensor<string, []>("softmax_31_cast_fp16")];
            tensor<bool, []> hidden_states_367_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_367_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_367_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_367_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_556 = transpose(perm = value_127_perm_0, x = var_2636_cast_fp16)[name = tensor<string, []>("transpose_556")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_367_cast_fp16 = matmul(transpose_x = hidden_states_367_transpose_x_0, transpose_y = hidden_states_367_transpose_y_0, x = softmax_31_cast_fp16, y = transpose_556)[name = tensor<string, []>("hidden_states_367_cast_fp16")];
            tensor<int32, [4]> var_2639_perm_0 = const()[name = tensor<string, []>("op_2639_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2643 = const()[name = tensor<string, []>("op_2643"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_555 = transpose(perm = var_2639_perm_0, x = hidden_states_367_cast_fp16)[name = tensor<string, []>("transpose_555")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_369_cast_fp16 = reshape(shape = var_2643, x = transpose_555)[name = tensor<string, []>("hidden_states_369_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(398091456))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399320320))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399320512)))];
            tensor<fp16, [1, 1024, 1280]> linear_174_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_369_cast_fp16)[name = tensor<string, []>("linear_174_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_465_cast_fp16 = add(x = linear_174_cast_fp16, y = input_457_cast_fp16)[name = tensor<string, []>("input_465_cast_fp16")];
            tensor<int32, [1]> input_467_axes_0 = const()[name = tensor<string, []>("input_467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399323136)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399325760)))];
            tensor<fp16, [1, 1024, 1280]> input_467_cast_fp16 = layer_norm(axes = input_467_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_1_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_1_norm3_weight_to_fp16, x = input_465_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399328384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409158848))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409159040))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409166784))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_175_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized, x = input_467_cast_fp16)[name = tensor<string, []>("linear_175_cast_fp16")];
            tensor<int32, [2]> var_2665_split_sizes_0 = const()[name = tensor<string, []>("op_2665_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_2665_axis_0 = const()[name = tensor<string, []>("op_2665_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_2665_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_2665_cast_fp16_1 = split(axis = var_2665_axis_0, split_sizes = var_2665_split_sizes_0, x = linear_175_cast_fp16)[name = tensor<string, []>("op_2665_cast_fp16")];
            tensor<string, []> var_2667_mode_0 = const()[name = tensor<string, []>("op_2667_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_2667_cast_fp16 = gelu(mode = var_2667_mode_0, x = var_2665_cast_fp16_1)[name = tensor<string, []>("op_2667_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_469_cast_fp16 = mul(x = var_2665_cast_fp16_0, y = var_2667_cast_fp16)[name = tensor<string, []>("input_469_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409166976))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414082240))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_1_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_1_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414082432)))];
            tensor<fp16, [1, 1024, 1280]> linear_176_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_1_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized, x = input_469_cast_fp16)[name = tensor<string, []>("linear_176_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_377_cast_fp16 = add(x = linear_176_cast_fp16, y = input_465_cast_fp16)[name = tensor<string, []>("hidden_states_377_cast_fp16")];
            tensor<int32, [1]> hidden_states_379_axes_0 = const()[name = tensor<string, []>("hidden_states_379_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414085056)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414087680)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_379_cast_fp16 = layer_norm(axes = hidden_states_379_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_2_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_2_norm1_weight_to_fp16, x = hidden_states_377_cast_fp16)[name = tensor<string, []>("hidden_states_379_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414090304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415319168))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_177_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_379_cast_fp16)[name = tensor<string, []>("linear_177_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415319360))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(416548224))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_178_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_379_cast_fp16)[name = tensor<string, []>("linear_178_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(416548416))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417777280))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_179_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_379_cast_fp16)[name = tensor<string, []>("linear_179_cast_fp16")];
            tensor<int32, [4]> var_2702 = const()[name = tensor<string, []>("op_2702"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2703_cast_fp16 = reshape(shape = var_2702, x = linear_177_cast_fp16)[name = tensor<string, []>("op_2703_cast_fp16")];
            tensor<int32, [4]> var_2705 = const()[name = tensor<string, []>("op_2705"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2706_cast_fp16 = reshape(shape = var_2705, x = linear_178_cast_fp16)[name = tensor<string, []>("op_2706_cast_fp16")];
            tensor<int32, [4]> var_2708 = const()[name = tensor<string, []>("op_2708"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2709_cast_fp16 = reshape(shape = var_2708, x = linear_179_cast_fp16)[name = tensor<string, []>("op_2709_cast_fp16")];
            tensor<int32, [4]> value_131_perm_0 = const()[name = tensor<string, []>("value_131_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_48_y_0_to_fp16 = const()[name = tensor<string, []>("mul_48_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_48_cast_fp16 = mul(x = var_2703_cast_fp16, y = mul_48_y_0_to_fp16)[name = tensor<string, []>("mul_48_cast_fp16")];
            tensor<bool, []> matmul_32_transpose_y_0 = const()[name = tensor<string, []>("matmul_32_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_32_transpose_x_0 = const()[name = tensor<string, []>("matmul_32_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_336_perm_0 = const()[name = tensor<string, []>("transpose_336_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_337_perm_0 = const()[name = tensor<string, []>("transpose_337_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_553 = transpose(perm = transpose_337_perm_0, x = var_2706_cast_fp16)[name = tensor<string, []>("transpose_553")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_554 = transpose(perm = transpose_336_perm_0, x = mul_48_cast_fp16)[name = tensor<string, []>("transpose_554")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_32_cast_fp16 = matmul(transpose_x = matmul_32_transpose_x_0, transpose_y = matmul_32_transpose_y_0, x = transpose_554, y = transpose_553)[name = tensor<string, []>("matmul_32_cast_fp16")];
            tensor<int32, []> softmax_32_axis_0 = const()[name = tensor<string, []>("softmax_32_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_32_cast_fp16 = softmax(axis = softmax_32_axis_0, x = matmul_32_cast_fp16)[name = tensor<string, []>("softmax_32_cast_fp16")];
            tensor<bool, []> hidden_states_381_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_381_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_381_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_381_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_552 = transpose(perm = value_131_perm_0, x = var_2709_cast_fp16)[name = tensor<string, []>("transpose_552")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_381_cast_fp16 = matmul(transpose_x = hidden_states_381_transpose_x_0, transpose_y = hidden_states_381_transpose_y_0, x = softmax_32_cast_fp16, y = transpose_552)[name = tensor<string, []>("hidden_states_381_cast_fp16")];
            tensor<int32, [4]> var_2712_perm_0 = const()[name = tensor<string, []>("op_2712_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2716 = const()[name = tensor<string, []>("op_2716"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_551 = transpose(perm = var_2712_perm_0, x = hidden_states_381_cast_fp16)[name = tensor<string, []>("transpose_551")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_383_cast_fp16 = reshape(shape = var_2716, x = transpose_551)[name = tensor<string, []>("hidden_states_383_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417777472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419006336))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419006528)))];
            tensor<fp16, [1, 1024, 1280]> linear_180_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_383_cast_fp16)[name = tensor<string, []>("linear_180_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_477_cast_fp16 = add(x = linear_180_cast_fp16, y = hidden_states_377_cast_fp16)[name = tensor<string, []>("input_477_cast_fp16")];
            tensor<int32, [1]> input_479_axes_0 = const()[name = tensor<string, []>("input_479_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419009152)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419011776)))];
            tensor<fp16, [1, 1024, 1280]> input_479_cast_fp16 = layer_norm(axes = input_479_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_2_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_2_norm2_weight_to_fp16, x = input_477_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419014400))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420243264))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_181_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized, x = input_479_cast_fp16)[name = tensor<string, []>("linear_181_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420243456))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422209600))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_182_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_182_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422209792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424175936))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_183_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_183_cast_fp16")];
            tensor<int32, [4]> var_2748 = const()[name = tensor<string, []>("op_2748"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2749_cast_fp16 = reshape(shape = var_2748, x = linear_181_cast_fp16)[name = tensor<string, []>("op_2749_cast_fp16")];
            tensor<int32, [4]> var_2751 = const()[name = tensor<string, []>("op_2751"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2752_cast_fp16 = reshape(shape = var_2751, x = linear_182_cast_fp16)[name = tensor<string, []>("op_2752_cast_fp16")];
            tensor<int32, [4]> var_2754 = const()[name = tensor<string, []>("op_2754"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2755_cast_fp16 = reshape(shape = var_2754, x = linear_183_cast_fp16)[name = tensor<string, []>("op_2755_cast_fp16")];
            tensor<int32, [4]> value_135_perm_0 = const()[name = tensor<string, []>("value_135_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_49_y_0_to_fp16 = const()[name = tensor<string, []>("mul_49_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_49_cast_fp16 = mul(x = var_2749_cast_fp16, y = mul_49_y_0_to_fp16)[name = tensor<string, []>("mul_49_cast_fp16")];
            tensor<bool, []> matmul_33_transpose_y_0 = const()[name = tensor<string, []>("matmul_33_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_33_transpose_x_0 = const()[name = tensor<string, []>("matmul_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_338_perm_0 = const()[name = tensor<string, []>("transpose_338_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_339_perm_0 = const()[name = tensor<string, []>("transpose_339_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_549 = transpose(perm = transpose_339_perm_0, x = var_2752_cast_fp16)[name = tensor<string, []>("transpose_549")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_550 = transpose(perm = transpose_338_perm_0, x = mul_49_cast_fp16)[name = tensor<string, []>("transpose_550")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_33_cast_fp16 = matmul(transpose_x = matmul_33_transpose_x_0, transpose_y = matmul_33_transpose_y_0, x = transpose_550, y = transpose_549)[name = tensor<string, []>("matmul_33_cast_fp16")];
            tensor<int32, []> softmax_33_axis_0 = const()[name = tensor<string, []>("softmax_33_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_33_cast_fp16 = softmax(axis = softmax_33_axis_0, x = matmul_33_cast_fp16)[name = tensor<string, []>("softmax_33_cast_fp16")];
            tensor<bool, []> hidden_states_387_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_387_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_387_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_387_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_548 = transpose(perm = value_135_perm_0, x = var_2755_cast_fp16)[name = tensor<string, []>("transpose_548")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_387_cast_fp16 = matmul(transpose_x = hidden_states_387_transpose_x_0, transpose_y = hidden_states_387_transpose_y_0, x = softmax_33_cast_fp16, y = transpose_548)[name = tensor<string, []>("hidden_states_387_cast_fp16")];
            tensor<int32, [4]> var_2758_perm_0 = const()[name = tensor<string, []>("op_2758_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2762 = const()[name = tensor<string, []>("op_2762"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_547 = transpose(perm = var_2758_perm_0, x = hidden_states_387_cast_fp16)[name = tensor<string, []>("transpose_547")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_389_cast_fp16 = reshape(shape = var_2762, x = transpose_547)[name = tensor<string, []>("hidden_states_389_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424176128))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425404992))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425405184)))];
            tensor<fp16, [1, 1024, 1280]> linear_184_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_389_cast_fp16)[name = tensor<string, []>("linear_184_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_485_cast_fp16 = add(x = linear_184_cast_fp16, y = input_477_cast_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
            tensor<int32, [1]> input_487_axes_0 = const()[name = tensor<string, []>("input_487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425407808)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425410432)))];
            tensor<fp16, [1, 1024, 1280]> input_487_cast_fp16 = layer_norm(axes = input_487_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_2_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_2_norm3_weight_to_fp16, x = input_485_cast_fp16)[name = tensor<string, []>("input_487_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425413056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435243520))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435243712))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435251456))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_185_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized, x = input_487_cast_fp16)[name = tensor<string, []>("linear_185_cast_fp16")];
            tensor<int32, [2]> var_2784_split_sizes_0 = const()[name = tensor<string, []>("op_2784_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_2784_axis_0 = const()[name = tensor<string, []>("op_2784_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_2784_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_2784_cast_fp16_1 = split(axis = var_2784_axis_0, split_sizes = var_2784_split_sizes_0, x = linear_185_cast_fp16)[name = tensor<string, []>("op_2784_cast_fp16")];
            tensor<string, []> var_2786_mode_0 = const()[name = tensor<string, []>("op_2786_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_2786_cast_fp16 = gelu(mode = var_2786_mode_0, x = var_2784_cast_fp16_1)[name = tensor<string, []>("op_2786_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_489_cast_fp16 = mul(x = var_2784_cast_fp16_0, y = var_2786_cast_fp16)[name = tensor<string, []>("input_489_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435251648))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440166912))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_2_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_2_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440167104)))];
            tensor<fp16, [1, 1024, 1280]> linear_186_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_2_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = tensor<string, []>("linear_186_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_397_cast_fp16 = add(x = linear_186_cast_fp16, y = input_485_cast_fp16)[name = tensor<string, []>("hidden_states_397_cast_fp16")];
            tensor<int32, [1]> hidden_states_399_axes_0 = const()[name = tensor<string, []>("hidden_states_399_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440169728)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440172352)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_399_cast_fp16 = layer_norm(axes = hidden_states_399_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_3_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_3_norm1_weight_to_fp16, x = hidden_states_397_cast_fp16)[name = tensor<string, []>("hidden_states_399_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440174976))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441403840))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_187_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_399_cast_fp16)[name = tensor<string, []>("linear_187_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441404032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442632896))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_188_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_399_cast_fp16)[name = tensor<string, []>("linear_188_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442633088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443861952))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_189_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_399_cast_fp16)[name = tensor<string, []>("linear_189_cast_fp16")];
            tensor<int32, [4]> var_2821 = const()[name = tensor<string, []>("op_2821"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2822_cast_fp16 = reshape(shape = var_2821, x = linear_187_cast_fp16)[name = tensor<string, []>("op_2822_cast_fp16")];
            tensor<int32, [4]> var_2824 = const()[name = tensor<string, []>("op_2824"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2825_cast_fp16 = reshape(shape = var_2824, x = linear_188_cast_fp16)[name = tensor<string, []>("op_2825_cast_fp16")];
            tensor<int32, [4]> var_2827 = const()[name = tensor<string, []>("op_2827"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2828_cast_fp16 = reshape(shape = var_2827, x = linear_189_cast_fp16)[name = tensor<string, []>("op_2828_cast_fp16")];
            tensor<int32, [4]> value_139_perm_0 = const()[name = tensor<string, []>("value_139_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_50_y_0_to_fp16 = const()[name = tensor<string, []>("mul_50_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_50_cast_fp16 = mul(x = var_2822_cast_fp16, y = mul_50_y_0_to_fp16)[name = tensor<string, []>("mul_50_cast_fp16")];
            tensor<bool, []> matmul_34_transpose_y_0 = const()[name = tensor<string, []>("matmul_34_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_34_transpose_x_0 = const()[name = tensor<string, []>("matmul_34_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_340_perm_0 = const()[name = tensor<string, []>("transpose_340_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_341_perm_0 = const()[name = tensor<string, []>("transpose_341_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_545 = transpose(perm = transpose_341_perm_0, x = var_2825_cast_fp16)[name = tensor<string, []>("transpose_545")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_546 = transpose(perm = transpose_340_perm_0, x = mul_50_cast_fp16)[name = tensor<string, []>("transpose_546")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_34_cast_fp16 = matmul(transpose_x = matmul_34_transpose_x_0, transpose_y = matmul_34_transpose_y_0, x = transpose_546, y = transpose_545)[name = tensor<string, []>("matmul_34_cast_fp16")];
            tensor<int32, []> softmax_34_axis_0 = const()[name = tensor<string, []>("softmax_34_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_34_cast_fp16 = softmax(axis = softmax_34_axis_0, x = matmul_34_cast_fp16)[name = tensor<string, []>("softmax_34_cast_fp16")];
            tensor<bool, []> hidden_states_401_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_401_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_401_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_401_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_544 = transpose(perm = value_139_perm_0, x = var_2828_cast_fp16)[name = tensor<string, []>("transpose_544")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_401_cast_fp16 = matmul(transpose_x = hidden_states_401_transpose_x_0, transpose_y = hidden_states_401_transpose_y_0, x = softmax_34_cast_fp16, y = transpose_544)[name = tensor<string, []>("hidden_states_401_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = tensor<string, []>("op_2831_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2835 = const()[name = tensor<string, []>("op_2835"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_543 = transpose(perm = var_2831_perm_0, x = hidden_states_401_cast_fp16)[name = tensor<string, []>("transpose_543")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_403_cast_fp16 = reshape(shape = var_2835, x = transpose_543)[name = tensor<string, []>("hidden_states_403_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443862144))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445091008))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445091200)))];
            tensor<fp16, [1, 1024, 1280]> linear_190_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_403_cast_fp16)[name = tensor<string, []>("linear_190_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_497_cast_fp16 = add(x = linear_190_cast_fp16, y = hidden_states_397_cast_fp16)[name = tensor<string, []>("input_497_cast_fp16")];
            tensor<int32, [1]> input_499_axes_0 = const()[name = tensor<string, []>("input_499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445093824)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445096448)))];
            tensor<fp16, [1, 1024, 1280]> input_499_cast_fp16 = layer_norm(axes = input_499_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_3_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_3_norm2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor<string, []>("input_499_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445099072))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446327936))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_191_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = tensor<string, []>("linear_191_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446328128))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448294272))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_192_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_192_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448294464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450260608))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_193_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_193_cast_fp16")];
            tensor<int32, [4]> var_2867 = const()[name = tensor<string, []>("op_2867"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2868_cast_fp16 = reshape(shape = var_2867, x = linear_191_cast_fp16)[name = tensor<string, []>("op_2868_cast_fp16")];
            tensor<int32, [4]> var_2870 = const()[name = tensor<string, []>("op_2870"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2871_cast_fp16 = reshape(shape = var_2870, x = linear_192_cast_fp16)[name = tensor<string, []>("op_2871_cast_fp16")];
            tensor<int32, [4]> var_2873 = const()[name = tensor<string, []>("op_2873"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2874_cast_fp16 = reshape(shape = var_2873, x = linear_193_cast_fp16)[name = tensor<string, []>("op_2874_cast_fp16")];
            tensor<int32, [4]> value_143_perm_0 = const()[name = tensor<string, []>("value_143_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_51_y_0_to_fp16 = const()[name = tensor<string, []>("mul_51_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_51_cast_fp16 = mul(x = var_2868_cast_fp16, y = mul_51_y_0_to_fp16)[name = tensor<string, []>("mul_51_cast_fp16")];
            tensor<bool, []> matmul_35_transpose_y_0 = const()[name = tensor<string, []>("matmul_35_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_35_transpose_x_0 = const()[name = tensor<string, []>("matmul_35_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_342_perm_0 = const()[name = tensor<string, []>("transpose_342_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_343_perm_0 = const()[name = tensor<string, []>("transpose_343_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_541 = transpose(perm = transpose_343_perm_0, x = var_2871_cast_fp16)[name = tensor<string, []>("transpose_541")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_542 = transpose(perm = transpose_342_perm_0, x = mul_51_cast_fp16)[name = tensor<string, []>("transpose_542")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_35_cast_fp16 = matmul(transpose_x = matmul_35_transpose_x_0, transpose_y = matmul_35_transpose_y_0, x = transpose_542, y = transpose_541)[name = tensor<string, []>("matmul_35_cast_fp16")];
            tensor<int32, []> softmax_35_axis_0 = const()[name = tensor<string, []>("softmax_35_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_35_cast_fp16 = softmax(axis = softmax_35_axis_0, x = matmul_35_cast_fp16)[name = tensor<string, []>("softmax_35_cast_fp16")];
            tensor<bool, []> hidden_states_407_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_407_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_407_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_407_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_540 = transpose(perm = value_143_perm_0, x = var_2874_cast_fp16)[name = tensor<string, []>("transpose_540")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_407_cast_fp16 = matmul(transpose_x = hidden_states_407_transpose_x_0, transpose_y = hidden_states_407_transpose_y_0, x = softmax_35_cast_fp16, y = transpose_540)[name = tensor<string, []>("hidden_states_407_cast_fp16")];
            tensor<int32, [4]> var_2877_perm_0 = const()[name = tensor<string, []>("op_2877_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2881 = const()[name = tensor<string, []>("op_2881"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_539 = transpose(perm = var_2877_perm_0, x = hidden_states_407_cast_fp16)[name = tensor<string, []>("transpose_539")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_409_cast_fp16 = reshape(shape = var_2881, x = transpose_539)[name = tensor<string, []>("hidden_states_409_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450260800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451489664))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451489856)))];
            tensor<fp16, [1, 1024, 1280]> linear_194_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_409_cast_fp16)[name = tensor<string, []>("linear_194_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_505_cast_fp16 = add(x = linear_194_cast_fp16, y = input_497_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
            tensor<int32, [1]> input_507_axes_0 = const()[name = tensor<string, []>("input_507_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451492480)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451495104)))];
            tensor<fp16, [1, 1024, 1280]> input_507_cast_fp16 = layer_norm(axes = input_507_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_3_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_3_norm3_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("input_507_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451497728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461328192))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461328384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461336128))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_195_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = tensor<string, []>("linear_195_cast_fp16")];
            tensor<int32, [2]> var_2903_split_sizes_0 = const()[name = tensor<string, []>("op_2903_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_2903_axis_0 = const()[name = tensor<string, []>("op_2903_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_2903_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_2903_cast_fp16_1 = split(axis = var_2903_axis_0, split_sizes = var_2903_split_sizes_0, x = linear_195_cast_fp16)[name = tensor<string, []>("op_2903_cast_fp16")];
            tensor<string, []> var_2905_mode_0 = const()[name = tensor<string, []>("op_2905_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_2905_cast_fp16 = gelu(mode = var_2905_mode_0, x = var_2903_cast_fp16_1)[name = tensor<string, []>("op_2905_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_509_cast_fp16 = mul(x = var_2903_cast_fp16_0, y = var_2905_cast_fp16)[name = tensor<string, []>("input_509_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461336320))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(466251584))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_3_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_3_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(466251776)))];
            tensor<fp16, [1, 1024, 1280]> linear_196_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_3_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized, x = input_509_cast_fp16)[name = tensor<string, []>("linear_196_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_417_cast_fp16 = add(x = linear_196_cast_fp16, y = input_505_cast_fp16)[name = tensor<string, []>("hidden_states_417_cast_fp16")];
            tensor<int32, [1]> hidden_states_419_axes_0 = const()[name = tensor<string, []>("hidden_states_419_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(466254400)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(466257024)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_419_cast_fp16 = layer_norm(axes = hidden_states_419_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_4_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_4_norm1_weight_to_fp16, x = hidden_states_417_cast_fp16)[name = tensor<string, []>("hidden_states_419_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(466259648))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467488512))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_197_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_419_cast_fp16)[name = tensor<string, []>("linear_197_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467488704))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468717568))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_198_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_419_cast_fp16)[name = tensor<string, []>("linear_198_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468717760))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(469946624))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_199_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_419_cast_fp16)[name = tensor<string, []>("linear_199_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = tensor<string, []>("op_2940"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2941_cast_fp16 = reshape(shape = var_2940, x = linear_197_cast_fp16)[name = tensor<string, []>("op_2941_cast_fp16")];
            tensor<int32, [4]> var_2943 = const()[name = tensor<string, []>("op_2943"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2944_cast_fp16 = reshape(shape = var_2943, x = linear_198_cast_fp16)[name = tensor<string, []>("op_2944_cast_fp16")];
            tensor<int32, [4]> var_2946 = const()[name = tensor<string, []>("op_2946"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2947_cast_fp16 = reshape(shape = var_2946, x = linear_199_cast_fp16)[name = tensor<string, []>("op_2947_cast_fp16")];
            tensor<int32, [4]> value_147_perm_0 = const()[name = tensor<string, []>("value_147_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_52_y_0_to_fp16 = const()[name = tensor<string, []>("mul_52_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_52_cast_fp16 = mul(x = var_2941_cast_fp16, y = mul_52_y_0_to_fp16)[name = tensor<string, []>("mul_52_cast_fp16")];
            tensor<bool, []> matmul_36_transpose_y_0 = const()[name = tensor<string, []>("matmul_36_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_36_transpose_x_0 = const()[name = tensor<string, []>("matmul_36_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_344_perm_0 = const()[name = tensor<string, []>("transpose_344_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_345_perm_0 = const()[name = tensor<string, []>("transpose_345_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_537 = transpose(perm = transpose_345_perm_0, x = var_2944_cast_fp16)[name = tensor<string, []>("transpose_537")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_538 = transpose(perm = transpose_344_perm_0, x = mul_52_cast_fp16)[name = tensor<string, []>("transpose_538")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_36_cast_fp16 = matmul(transpose_x = matmul_36_transpose_x_0, transpose_y = matmul_36_transpose_y_0, x = transpose_538, y = transpose_537)[name = tensor<string, []>("matmul_36_cast_fp16")];
            tensor<int32, []> softmax_36_axis_0 = const()[name = tensor<string, []>("softmax_36_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_36_cast_fp16 = softmax(axis = softmax_36_axis_0, x = matmul_36_cast_fp16)[name = tensor<string, []>("softmax_36_cast_fp16")];
            tensor<bool, []> hidden_states_421_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_421_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_421_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_421_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_536 = transpose(perm = value_147_perm_0, x = var_2947_cast_fp16)[name = tensor<string, []>("transpose_536")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_421_cast_fp16 = matmul(transpose_x = hidden_states_421_transpose_x_0, transpose_y = hidden_states_421_transpose_y_0, x = softmax_36_cast_fp16, y = transpose_536)[name = tensor<string, []>("hidden_states_421_cast_fp16")];
            tensor<int32, [4]> var_2950_perm_0 = const()[name = tensor<string, []>("op_2950_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2954 = const()[name = tensor<string, []>("op_2954"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_535 = transpose(perm = var_2950_perm_0, x = hidden_states_421_cast_fp16)[name = tensor<string, []>("transpose_535")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_423_cast_fp16 = reshape(shape = var_2954, x = transpose_535)[name = tensor<string, []>("hidden_states_423_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(469946816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(471175680))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(471175872)))];
            tensor<fp16, [1, 1024, 1280]> linear_200_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_423_cast_fp16)[name = tensor<string, []>("linear_200_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_517_cast_fp16 = add(x = linear_200_cast_fp16, y = hidden_states_417_cast_fp16)[name = tensor<string, []>("input_517_cast_fp16")];
            tensor<int32, [1]> input_519_axes_0 = const()[name = tensor<string, []>("input_519_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(471178496)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(471181120)))];
            tensor<fp16, [1, 1024, 1280]> input_519_cast_fp16 = layer_norm(axes = input_519_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_4_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_4_norm2_weight_to_fp16, x = input_517_cast_fp16)[name = tensor<string, []>("input_519_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(471183744))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(472412608))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_201_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized, x = input_519_cast_fp16)[name = tensor<string, []>("linear_201_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(472412800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474378944))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_202_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_202_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474379136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476345280))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_203_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_203_cast_fp16")];
            tensor<int32, [4]> var_2986 = const()[name = tensor<string, []>("op_2986"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_2987_cast_fp16 = reshape(shape = var_2986, x = linear_201_cast_fp16)[name = tensor<string, []>("op_2987_cast_fp16")];
            tensor<int32, [4]> var_2989 = const()[name = tensor<string, []>("op_2989"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2990_cast_fp16 = reshape(shape = var_2989, x = linear_202_cast_fp16)[name = tensor<string, []>("op_2990_cast_fp16")];
            tensor<int32, [4]> var_2992 = const()[name = tensor<string, []>("op_2992"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_2993_cast_fp16 = reshape(shape = var_2992, x = linear_203_cast_fp16)[name = tensor<string, []>("op_2993_cast_fp16")];
            tensor<int32, [4]> value_151_perm_0 = const()[name = tensor<string, []>("value_151_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_53_y_0_to_fp16 = const()[name = tensor<string, []>("mul_53_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_53_cast_fp16 = mul(x = var_2987_cast_fp16, y = mul_53_y_0_to_fp16)[name = tensor<string, []>("mul_53_cast_fp16")];
            tensor<bool, []> matmul_37_transpose_y_0 = const()[name = tensor<string, []>("matmul_37_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_37_transpose_x_0 = const()[name = tensor<string, []>("matmul_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_346_perm_0 = const()[name = tensor<string, []>("transpose_346_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_347_perm_0 = const()[name = tensor<string, []>("transpose_347_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_533 = transpose(perm = transpose_347_perm_0, x = var_2990_cast_fp16)[name = tensor<string, []>("transpose_533")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_534 = transpose(perm = transpose_346_perm_0, x = mul_53_cast_fp16)[name = tensor<string, []>("transpose_534")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_37_cast_fp16 = matmul(transpose_x = matmul_37_transpose_x_0, transpose_y = matmul_37_transpose_y_0, x = transpose_534, y = transpose_533)[name = tensor<string, []>("matmul_37_cast_fp16")];
            tensor<int32, []> softmax_37_axis_0 = const()[name = tensor<string, []>("softmax_37_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_37_cast_fp16 = softmax(axis = softmax_37_axis_0, x = matmul_37_cast_fp16)[name = tensor<string, []>("softmax_37_cast_fp16")];
            tensor<bool, []> hidden_states_427_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_427_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_427_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_427_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_532 = transpose(perm = value_151_perm_0, x = var_2993_cast_fp16)[name = tensor<string, []>("transpose_532")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_427_cast_fp16 = matmul(transpose_x = hidden_states_427_transpose_x_0, transpose_y = hidden_states_427_transpose_y_0, x = softmax_37_cast_fp16, y = transpose_532)[name = tensor<string, []>("hidden_states_427_cast_fp16")];
            tensor<int32, [4]> var_2996_perm_0 = const()[name = tensor<string, []>("op_2996_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3000 = const()[name = tensor<string, []>("op_3000"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_531 = transpose(perm = var_2996_perm_0, x = hidden_states_427_cast_fp16)[name = tensor<string, []>("transpose_531")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_429_cast_fp16 = reshape(shape = var_3000, x = transpose_531)[name = tensor<string, []>("hidden_states_429_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476345472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477574336))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477574528)))];
            tensor<fp16, [1, 1024, 1280]> linear_204_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_429_cast_fp16)[name = tensor<string, []>("linear_204_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_525_cast_fp16 = add(x = linear_204_cast_fp16, y = input_517_cast_fp16)[name = tensor<string, []>("input_525_cast_fp16")];
            tensor<int32, [1]> input_527_axes_0 = const()[name = tensor<string, []>("input_527_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477577152)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477579776)))];
            tensor<fp16, [1, 1024, 1280]> input_527_cast_fp16 = layer_norm(axes = input_527_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_4_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_4_norm3_weight_to_fp16, x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477582400))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487412864))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487413056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487420800))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_205_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized, x = input_527_cast_fp16)[name = tensor<string, []>("linear_205_cast_fp16")];
            tensor<int32, [2]> var_3022_split_sizes_0 = const()[name = tensor<string, []>("op_3022_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_3022_axis_0 = const()[name = tensor<string, []>("op_3022_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_3022_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_3022_cast_fp16_1 = split(axis = var_3022_axis_0, split_sizes = var_3022_split_sizes_0, x = linear_205_cast_fp16)[name = tensor<string, []>("op_3022_cast_fp16")];
            tensor<string, []> var_3024_mode_0 = const()[name = tensor<string, []>("op_3024_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_3024_cast_fp16 = gelu(mode = var_3024_mode_0, x = var_3022_cast_fp16_1)[name = tensor<string, []>("op_3024_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_529_cast_fp16 = mul(x = var_3022_cast_fp16_0, y = var_3024_cast_fp16)[name = tensor<string, []>("input_529_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487420992))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492336256))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_4_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_4_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492336448)))];
            tensor<fp16, [1, 1024, 1280]> linear_206_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_4_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized, x = input_529_cast_fp16)[name = tensor<string, []>("linear_206_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_437_cast_fp16 = add(x = linear_206_cast_fp16, y = input_525_cast_fp16)[name = tensor<string, []>("hidden_states_437_cast_fp16")];
            tensor<int32, [1]> hidden_states_439_axes_0 = const()[name = tensor<string, []>("hidden_states_439_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492339072)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492341696)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_439_cast_fp16 = layer_norm(axes = hidden_states_439_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_5_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_5_norm1_weight_to_fp16, x = hidden_states_437_cast_fp16)[name = tensor<string, []>("hidden_states_439_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492344320))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493573184))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_207_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_439_cast_fp16)[name = tensor<string, []>("linear_207_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493573376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494802240))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_208_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_439_cast_fp16)[name = tensor<string, []>("linear_208_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494802432))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496031296))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_209_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_439_cast_fp16)[name = tensor<string, []>("linear_209_cast_fp16")];
            tensor<int32, [4]> var_3059 = const()[name = tensor<string, []>("op_3059"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3060_cast_fp16 = reshape(shape = var_3059, x = linear_207_cast_fp16)[name = tensor<string, []>("op_3060_cast_fp16")];
            tensor<int32, [4]> var_3062 = const()[name = tensor<string, []>("op_3062"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3063_cast_fp16 = reshape(shape = var_3062, x = linear_208_cast_fp16)[name = tensor<string, []>("op_3063_cast_fp16")];
            tensor<int32, [4]> var_3065 = const()[name = tensor<string, []>("op_3065"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3066_cast_fp16 = reshape(shape = var_3065, x = linear_209_cast_fp16)[name = tensor<string, []>("op_3066_cast_fp16")];
            tensor<int32, [4]> value_155_perm_0 = const()[name = tensor<string, []>("value_155_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_54_y_0_to_fp16 = const()[name = tensor<string, []>("mul_54_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_54_cast_fp16 = mul(x = var_3060_cast_fp16, y = mul_54_y_0_to_fp16)[name = tensor<string, []>("mul_54_cast_fp16")];
            tensor<bool, []> matmul_38_transpose_y_0 = const()[name = tensor<string, []>("matmul_38_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_38_transpose_x_0 = const()[name = tensor<string, []>("matmul_38_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_348_perm_0 = const()[name = tensor<string, []>("transpose_348_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_349_perm_0 = const()[name = tensor<string, []>("transpose_349_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_529 = transpose(perm = transpose_349_perm_0, x = var_3063_cast_fp16)[name = tensor<string, []>("transpose_529")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_530 = transpose(perm = transpose_348_perm_0, x = mul_54_cast_fp16)[name = tensor<string, []>("transpose_530")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_38_cast_fp16 = matmul(transpose_x = matmul_38_transpose_x_0, transpose_y = matmul_38_transpose_y_0, x = transpose_530, y = transpose_529)[name = tensor<string, []>("matmul_38_cast_fp16")];
            tensor<int32, []> softmax_38_axis_0 = const()[name = tensor<string, []>("softmax_38_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_38_cast_fp16 = softmax(axis = softmax_38_axis_0, x = matmul_38_cast_fp16)[name = tensor<string, []>("softmax_38_cast_fp16")];
            tensor<bool, []> hidden_states_441_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_441_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_441_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_441_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_528 = transpose(perm = value_155_perm_0, x = var_3066_cast_fp16)[name = tensor<string, []>("transpose_528")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_441_cast_fp16 = matmul(transpose_x = hidden_states_441_transpose_x_0, transpose_y = hidden_states_441_transpose_y_0, x = softmax_38_cast_fp16, y = transpose_528)[name = tensor<string, []>("hidden_states_441_cast_fp16")];
            tensor<int32, [4]> var_3069_perm_0 = const()[name = tensor<string, []>("op_3069_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3073 = const()[name = tensor<string, []>("op_3073"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_527 = transpose(perm = var_3069_perm_0, x = hidden_states_441_cast_fp16)[name = tensor<string, []>("transpose_527")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_443_cast_fp16 = reshape(shape = var_3073, x = transpose_527)[name = tensor<string, []>("hidden_states_443_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496031488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497260352))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497260544)))];
            tensor<fp16, [1, 1024, 1280]> linear_210_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_443_cast_fp16)[name = tensor<string, []>("linear_210_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_537_cast_fp16 = add(x = linear_210_cast_fp16, y = hidden_states_437_cast_fp16)[name = tensor<string, []>("input_537_cast_fp16")];
            tensor<int32, [1]> input_539_axes_0 = const()[name = tensor<string, []>("input_539_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497263168)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497265792)))];
            tensor<fp16, [1, 1024, 1280]> input_539_cast_fp16 = layer_norm(axes = input_539_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_5_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_5_norm2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor<string, []>("input_539_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497268416))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498497280))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_211_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized, x = input_539_cast_fp16)[name = tensor<string, []>("linear_211_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498497472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500463616))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_212_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_212_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500463808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502429952))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_213_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_213_cast_fp16")];
            tensor<int32, [4]> var_3105 = const()[name = tensor<string, []>("op_3105"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3106_cast_fp16 = reshape(shape = var_3105, x = linear_211_cast_fp16)[name = tensor<string, []>("op_3106_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = tensor<string, []>("op_3108"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3109_cast_fp16 = reshape(shape = var_3108, x = linear_212_cast_fp16)[name = tensor<string, []>("op_3109_cast_fp16")];
            tensor<int32, [4]> var_3111 = const()[name = tensor<string, []>("op_3111"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3112_cast_fp16 = reshape(shape = var_3111, x = linear_213_cast_fp16)[name = tensor<string, []>("op_3112_cast_fp16")];
            tensor<int32, [4]> value_159_perm_0 = const()[name = tensor<string, []>("value_159_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_55_y_0_to_fp16 = const()[name = tensor<string, []>("mul_55_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_55_cast_fp16 = mul(x = var_3106_cast_fp16, y = mul_55_y_0_to_fp16)[name = tensor<string, []>("mul_55_cast_fp16")];
            tensor<bool, []> matmul_39_transpose_y_0 = const()[name = tensor<string, []>("matmul_39_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_39_transpose_x_0 = const()[name = tensor<string, []>("matmul_39_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_350_perm_0 = const()[name = tensor<string, []>("transpose_350_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_351_perm_0 = const()[name = tensor<string, []>("transpose_351_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_525 = transpose(perm = transpose_351_perm_0, x = var_3109_cast_fp16)[name = tensor<string, []>("transpose_525")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_526 = transpose(perm = transpose_350_perm_0, x = mul_55_cast_fp16)[name = tensor<string, []>("transpose_526")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_39_cast_fp16 = matmul(transpose_x = matmul_39_transpose_x_0, transpose_y = matmul_39_transpose_y_0, x = transpose_526, y = transpose_525)[name = tensor<string, []>("matmul_39_cast_fp16")];
            tensor<int32, []> softmax_39_axis_0 = const()[name = tensor<string, []>("softmax_39_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_39_cast_fp16 = softmax(axis = softmax_39_axis_0, x = matmul_39_cast_fp16)[name = tensor<string, []>("softmax_39_cast_fp16")];
            tensor<bool, []> hidden_states_447_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_447_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_447_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_447_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_524 = transpose(perm = value_159_perm_0, x = var_3112_cast_fp16)[name = tensor<string, []>("transpose_524")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_447_cast_fp16 = matmul(transpose_x = hidden_states_447_transpose_x_0, transpose_y = hidden_states_447_transpose_y_0, x = softmax_39_cast_fp16, y = transpose_524)[name = tensor<string, []>("hidden_states_447_cast_fp16")];
            tensor<int32, [4]> var_3115_perm_0 = const()[name = tensor<string, []>("op_3115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3119 = const()[name = tensor<string, []>("op_3119"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_523 = transpose(perm = var_3115_perm_0, x = hidden_states_447_cast_fp16)[name = tensor<string, []>("transpose_523")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_449_cast_fp16 = reshape(shape = var_3119, x = transpose_523)[name = tensor<string, []>("hidden_states_449_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502430144))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503659008))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503659200)))];
            tensor<fp16, [1, 1024, 1280]> linear_214_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_449_cast_fp16)[name = tensor<string, []>("linear_214_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_545_cast_fp16 = add(x = linear_214_cast_fp16, y = input_537_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
            tensor<int32, [1]> input_547_axes_0 = const()[name = tensor<string, []>("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503661824)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503664448)))];
            tensor<fp16, [1, 1024, 1280]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_5_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_5_norm3_weight_to_fp16, x = input_545_cast_fp16)[name = tensor<string, []>("input_547_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503667072))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513497536))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513497728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513505472))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_215_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = tensor<string, []>("linear_215_cast_fp16")];
            tensor<int32, [2]> var_3141_split_sizes_0 = const()[name = tensor<string, []>("op_3141_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_3141_axis_0 = const()[name = tensor<string, []>("op_3141_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_3141_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_3141_cast_fp16_1 = split(axis = var_3141_axis_0, split_sizes = var_3141_split_sizes_0, x = linear_215_cast_fp16)[name = tensor<string, []>("op_3141_cast_fp16")];
            tensor<string, []> var_3143_mode_0 = const()[name = tensor<string, []>("op_3143_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_3143_cast_fp16 = gelu(mode = var_3143_mode_0, x = var_3141_cast_fp16_1)[name = tensor<string, []>("op_3143_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_549_cast_fp16 = mul(x = var_3141_cast_fp16_0, y = var_3143_cast_fp16)[name = tensor<string, []>("input_549_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513505664))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518420928))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_5_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_5_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518421120)))];
            tensor<fp16, [1, 1024, 1280]> linear_216_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_5_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized, x = input_549_cast_fp16)[name = tensor<string, []>("linear_216_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_457_cast_fp16 = add(x = linear_216_cast_fp16, y = input_545_cast_fp16)[name = tensor<string, []>("hidden_states_457_cast_fp16")];
            tensor<int32, [1]> hidden_states_459_axes_0 = const()[name = tensor<string, []>("hidden_states_459_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518423744)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518426368)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_459_cast_fp16 = layer_norm(axes = hidden_states_459_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_6_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_6_norm1_weight_to_fp16, x = hidden_states_457_cast_fp16)[name = tensor<string, []>("hidden_states_459_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518428992))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(519657856))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_217_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_459_cast_fp16)[name = tensor<string, []>("linear_217_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(519658048))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520886912))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_218_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_459_cast_fp16)[name = tensor<string, []>("linear_218_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520887104))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522115968))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_219_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_459_cast_fp16)[name = tensor<string, []>("linear_219_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = tensor<string, []>("op_3178"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_217_cast_fp16)[name = tensor<string, []>("op_3179_cast_fp16")];
            tensor<int32, [4]> var_3181 = const()[name = tensor<string, []>("op_3181"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3182_cast_fp16 = reshape(shape = var_3181, x = linear_218_cast_fp16)[name = tensor<string, []>("op_3182_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = tensor<string, []>("op_3184"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_219_cast_fp16)[name = tensor<string, []>("op_3185_cast_fp16")];
            tensor<int32, [4]> value_163_perm_0 = const()[name = tensor<string, []>("value_163_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_56_y_0_to_fp16 = const()[name = tensor<string, []>("mul_56_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_56_cast_fp16 = mul(x = var_3179_cast_fp16, y = mul_56_y_0_to_fp16)[name = tensor<string, []>("mul_56_cast_fp16")];
            tensor<bool, []> matmul_40_transpose_y_0 = const()[name = tensor<string, []>("matmul_40_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_40_transpose_x_0 = const()[name = tensor<string, []>("matmul_40_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_352_perm_0 = const()[name = tensor<string, []>("transpose_352_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_353_perm_0 = const()[name = tensor<string, []>("transpose_353_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_521 = transpose(perm = transpose_353_perm_0, x = var_3182_cast_fp16)[name = tensor<string, []>("transpose_521")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_522 = transpose(perm = transpose_352_perm_0, x = mul_56_cast_fp16)[name = tensor<string, []>("transpose_522")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_40_cast_fp16 = matmul(transpose_x = matmul_40_transpose_x_0, transpose_y = matmul_40_transpose_y_0, x = transpose_522, y = transpose_521)[name = tensor<string, []>("matmul_40_cast_fp16")];
            tensor<int32, []> softmax_40_axis_0 = const()[name = tensor<string, []>("softmax_40_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_40_cast_fp16 = softmax(axis = softmax_40_axis_0, x = matmul_40_cast_fp16)[name = tensor<string, []>("softmax_40_cast_fp16")];
            tensor<bool, []> hidden_states_461_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_461_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_461_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_461_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_520 = transpose(perm = value_163_perm_0, x = var_3185_cast_fp16)[name = tensor<string, []>("transpose_520")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_461_cast_fp16 = matmul(transpose_x = hidden_states_461_transpose_x_0, transpose_y = hidden_states_461_transpose_y_0, x = softmax_40_cast_fp16, y = transpose_520)[name = tensor<string, []>("hidden_states_461_cast_fp16")];
            tensor<int32, [4]> var_3188_perm_0 = const()[name = tensor<string, []>("op_3188_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3192 = const()[name = tensor<string, []>("op_3192"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_519 = transpose(perm = var_3188_perm_0, x = hidden_states_461_cast_fp16)[name = tensor<string, []>("transpose_519")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_463_cast_fp16 = reshape(shape = var_3192, x = transpose_519)[name = tensor<string, []>("hidden_states_463_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522116160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(523345024))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(523345216)))];
            tensor<fp16, [1, 1024, 1280]> linear_220_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_463_cast_fp16)[name = tensor<string, []>("linear_220_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_557_cast_fp16 = add(x = linear_220_cast_fp16, y = hidden_states_457_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
            tensor<int32, [1]> input_559_axes_0 = const()[name = tensor<string, []>("input_559_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(523347840)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(523350464)))];
            tensor<fp16, [1, 1024, 1280]> input_559_cast_fp16 = layer_norm(axes = input_559_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_6_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_6_norm2_weight_to_fp16, x = input_557_cast_fp16)[name = tensor<string, []>("input_559_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(523353088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(524581952))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_221_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = tensor<string, []>("linear_221_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(524582144))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526548288))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_222_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_222_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526548480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528514624))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_223_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_223_cast_fp16")];
            tensor<int32, [4]> var_3224 = const()[name = tensor<string, []>("op_3224"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3225_cast_fp16 = reshape(shape = var_3224, x = linear_221_cast_fp16)[name = tensor<string, []>("op_3225_cast_fp16")];
            tensor<int32, [4]> var_3227 = const()[name = tensor<string, []>("op_3227"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3228_cast_fp16 = reshape(shape = var_3227, x = linear_222_cast_fp16)[name = tensor<string, []>("op_3228_cast_fp16")];
            tensor<int32, [4]> var_3230 = const()[name = tensor<string, []>("op_3230"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3231_cast_fp16 = reshape(shape = var_3230, x = linear_223_cast_fp16)[name = tensor<string, []>("op_3231_cast_fp16")];
            tensor<int32, [4]> value_167_perm_0 = const()[name = tensor<string, []>("value_167_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_57_y_0_to_fp16 = const()[name = tensor<string, []>("mul_57_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_57_cast_fp16 = mul(x = var_3225_cast_fp16, y = mul_57_y_0_to_fp16)[name = tensor<string, []>("mul_57_cast_fp16")];
            tensor<bool, []> matmul_41_transpose_y_0 = const()[name = tensor<string, []>("matmul_41_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_41_transpose_x_0 = const()[name = tensor<string, []>("matmul_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_354_perm_0 = const()[name = tensor<string, []>("transpose_354_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_355_perm_0 = const()[name = tensor<string, []>("transpose_355_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_517 = transpose(perm = transpose_355_perm_0, x = var_3228_cast_fp16)[name = tensor<string, []>("transpose_517")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_518 = transpose(perm = transpose_354_perm_0, x = mul_57_cast_fp16)[name = tensor<string, []>("transpose_518")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_41_cast_fp16 = matmul(transpose_x = matmul_41_transpose_x_0, transpose_y = matmul_41_transpose_y_0, x = transpose_518, y = transpose_517)[name = tensor<string, []>("matmul_41_cast_fp16")];
            tensor<int32, []> softmax_41_axis_0 = const()[name = tensor<string, []>("softmax_41_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_41_cast_fp16 = softmax(axis = softmax_41_axis_0, x = matmul_41_cast_fp16)[name = tensor<string, []>("softmax_41_cast_fp16")];
            tensor<bool, []> hidden_states_467_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_467_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_467_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_467_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_516 = transpose(perm = value_167_perm_0, x = var_3231_cast_fp16)[name = tensor<string, []>("transpose_516")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_467_cast_fp16 = matmul(transpose_x = hidden_states_467_transpose_x_0, transpose_y = hidden_states_467_transpose_y_0, x = softmax_41_cast_fp16, y = transpose_516)[name = tensor<string, []>("hidden_states_467_cast_fp16")];
            tensor<int32, [4]> var_3234_perm_0 = const()[name = tensor<string, []>("op_3234_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3238 = const()[name = tensor<string, []>("op_3238"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_515 = transpose(perm = var_3234_perm_0, x = hidden_states_467_cast_fp16)[name = tensor<string, []>("transpose_515")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_469_cast_fp16 = reshape(shape = var_3238, x = transpose_515)[name = tensor<string, []>("hidden_states_469_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528514816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529743680))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529743872)))];
            tensor<fp16, [1, 1024, 1280]> linear_224_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_469_cast_fp16)[name = tensor<string, []>("linear_224_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_565_cast_fp16 = add(x = linear_224_cast_fp16, y = input_557_cast_fp16)[name = tensor<string, []>("input_565_cast_fp16")];
            tensor<int32, [1]> input_567_axes_0 = const()[name = tensor<string, []>("input_567_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529746496)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529749120)))];
            tensor<fp16, [1, 1024, 1280]> input_567_cast_fp16 = layer_norm(axes = input_567_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_6_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_6_norm3_weight_to_fp16, x = input_565_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529751744))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539582208))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539582400))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539590144))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_225_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized, x = input_567_cast_fp16)[name = tensor<string, []>("linear_225_cast_fp16")];
            tensor<int32, [2]> var_3260_split_sizes_0 = const()[name = tensor<string, []>("op_3260_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_3260_axis_0 = const()[name = tensor<string, []>("op_3260_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_3260_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_3260_cast_fp16_1 = split(axis = var_3260_axis_0, split_sizes = var_3260_split_sizes_0, x = linear_225_cast_fp16)[name = tensor<string, []>("op_3260_cast_fp16")];
            tensor<string, []> var_3262_mode_0 = const()[name = tensor<string, []>("op_3262_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_3262_cast_fp16 = gelu(mode = var_3262_mode_0, x = var_3260_cast_fp16_1)[name = tensor<string, []>("op_3262_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_569_cast_fp16 = mul(x = var_3260_cast_fp16_0, y = var_3262_cast_fp16)[name = tensor<string, []>("input_569_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539590336))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(544505600))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_6_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_6_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(544505792)))];
            tensor<fp16, [1, 1024, 1280]> linear_226_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_6_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized, x = input_569_cast_fp16)[name = tensor<string, []>("linear_226_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_477_cast_fp16 = add(x = linear_226_cast_fp16, y = input_565_cast_fp16)[name = tensor<string, []>("hidden_states_477_cast_fp16")];
            tensor<int32, [1]> hidden_states_479_axes_0 = const()[name = tensor<string, []>("hidden_states_479_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(544508416)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(544511040)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_479_cast_fp16 = layer_norm(axes = hidden_states_479_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_7_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_7_norm1_weight_to_fp16, x = hidden_states_477_cast_fp16)[name = tensor<string, []>("hidden_states_479_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(544513664))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545742528))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_227_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_479_cast_fp16)[name = tensor<string, []>("linear_227_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545742720))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546971584))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_228_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_479_cast_fp16)[name = tensor<string, []>("linear_228_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546971776))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548200640))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_229_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_479_cast_fp16)[name = tensor<string, []>("linear_229_cast_fp16")];
            tensor<int32, [4]> var_3297 = const()[name = tensor<string, []>("op_3297"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3298_cast_fp16 = reshape(shape = var_3297, x = linear_227_cast_fp16)[name = tensor<string, []>("op_3298_cast_fp16")];
            tensor<int32, [4]> var_3300 = const()[name = tensor<string, []>("op_3300"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3301_cast_fp16 = reshape(shape = var_3300, x = linear_228_cast_fp16)[name = tensor<string, []>("op_3301_cast_fp16")];
            tensor<int32, [4]> var_3303 = const()[name = tensor<string, []>("op_3303"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3304_cast_fp16 = reshape(shape = var_3303, x = linear_229_cast_fp16)[name = tensor<string, []>("op_3304_cast_fp16")];
            tensor<int32, [4]> value_171_perm_0 = const()[name = tensor<string, []>("value_171_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_58_y_0_to_fp16 = const()[name = tensor<string, []>("mul_58_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_58_cast_fp16 = mul(x = var_3298_cast_fp16, y = mul_58_y_0_to_fp16)[name = tensor<string, []>("mul_58_cast_fp16")];
            tensor<bool, []> matmul_42_transpose_y_0 = const()[name = tensor<string, []>("matmul_42_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_42_transpose_x_0 = const()[name = tensor<string, []>("matmul_42_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_356_perm_0 = const()[name = tensor<string, []>("transpose_356_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_357_perm_0 = const()[name = tensor<string, []>("transpose_357_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_513 = transpose(perm = transpose_357_perm_0, x = var_3301_cast_fp16)[name = tensor<string, []>("transpose_513")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_514 = transpose(perm = transpose_356_perm_0, x = mul_58_cast_fp16)[name = tensor<string, []>("transpose_514")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_42_cast_fp16 = matmul(transpose_x = matmul_42_transpose_x_0, transpose_y = matmul_42_transpose_y_0, x = transpose_514, y = transpose_513)[name = tensor<string, []>("matmul_42_cast_fp16")];
            tensor<int32, []> softmax_42_axis_0 = const()[name = tensor<string, []>("softmax_42_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_42_cast_fp16 = softmax(axis = softmax_42_axis_0, x = matmul_42_cast_fp16)[name = tensor<string, []>("softmax_42_cast_fp16")];
            tensor<bool, []> hidden_states_481_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_481_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_481_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_481_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_512 = transpose(perm = value_171_perm_0, x = var_3304_cast_fp16)[name = tensor<string, []>("transpose_512")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_481_cast_fp16 = matmul(transpose_x = hidden_states_481_transpose_x_0, transpose_y = hidden_states_481_transpose_y_0, x = softmax_42_cast_fp16, y = transpose_512)[name = tensor<string, []>("hidden_states_481_cast_fp16")];
            tensor<int32, [4]> var_3307_perm_0 = const()[name = tensor<string, []>("op_3307_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3311 = const()[name = tensor<string, []>("op_3311"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_511 = transpose(perm = var_3307_perm_0, x = hidden_states_481_cast_fp16)[name = tensor<string, []>("transpose_511")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_483_cast_fp16 = reshape(shape = var_3311, x = transpose_511)[name = tensor<string, []>("hidden_states_483_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548200832))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549429696))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549429888)))];
            tensor<fp16, [1, 1024, 1280]> linear_230_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_483_cast_fp16)[name = tensor<string, []>("linear_230_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_577_cast_fp16 = add(x = linear_230_cast_fp16, y = hidden_states_477_cast_fp16)[name = tensor<string, []>("input_577_cast_fp16")];
            tensor<int32, [1]> input_579_axes_0 = const()[name = tensor<string, []>("input_579_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549432512)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549435136)))];
            tensor<fp16, [1, 1024, 1280]> input_579_cast_fp16 = layer_norm(axes = input_579_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_7_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_7_norm2_weight_to_fp16, x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549437760))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550666624))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_231_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized, x = input_579_cast_fp16)[name = tensor<string, []>("linear_231_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550666816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552632960))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_232_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_232_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552633152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554599296))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_233_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_233_cast_fp16")];
            tensor<int32, [4]> var_3343 = const()[name = tensor<string, []>("op_3343"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3344_cast_fp16 = reshape(shape = var_3343, x = linear_231_cast_fp16)[name = tensor<string, []>("op_3344_cast_fp16")];
            tensor<int32, [4]> var_3346 = const()[name = tensor<string, []>("op_3346"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3347_cast_fp16 = reshape(shape = var_3346, x = linear_232_cast_fp16)[name = tensor<string, []>("op_3347_cast_fp16")];
            tensor<int32, [4]> var_3349 = const()[name = tensor<string, []>("op_3349"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3350_cast_fp16 = reshape(shape = var_3349, x = linear_233_cast_fp16)[name = tensor<string, []>("op_3350_cast_fp16")];
            tensor<int32, [4]> value_175_perm_0 = const()[name = tensor<string, []>("value_175_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_59_y_0_to_fp16 = const()[name = tensor<string, []>("mul_59_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_59_cast_fp16 = mul(x = var_3344_cast_fp16, y = mul_59_y_0_to_fp16)[name = tensor<string, []>("mul_59_cast_fp16")];
            tensor<bool, []> matmul_43_transpose_y_0 = const()[name = tensor<string, []>("matmul_43_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_43_transpose_x_0 = const()[name = tensor<string, []>("matmul_43_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_358_perm_0 = const()[name = tensor<string, []>("transpose_358_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_359_perm_0 = const()[name = tensor<string, []>("transpose_359_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_509 = transpose(perm = transpose_359_perm_0, x = var_3347_cast_fp16)[name = tensor<string, []>("transpose_509")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_510 = transpose(perm = transpose_358_perm_0, x = mul_59_cast_fp16)[name = tensor<string, []>("transpose_510")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_43_cast_fp16 = matmul(transpose_x = matmul_43_transpose_x_0, transpose_y = matmul_43_transpose_y_0, x = transpose_510, y = transpose_509)[name = tensor<string, []>("matmul_43_cast_fp16")];
            tensor<int32, []> softmax_43_axis_0 = const()[name = tensor<string, []>("softmax_43_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_43_cast_fp16 = softmax(axis = softmax_43_axis_0, x = matmul_43_cast_fp16)[name = tensor<string, []>("softmax_43_cast_fp16")];
            tensor<bool, []> hidden_states_487_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_487_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_487_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_487_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_508 = transpose(perm = value_175_perm_0, x = var_3350_cast_fp16)[name = tensor<string, []>("transpose_508")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_487_cast_fp16 = matmul(transpose_x = hidden_states_487_transpose_x_0, transpose_y = hidden_states_487_transpose_y_0, x = softmax_43_cast_fp16, y = transpose_508)[name = tensor<string, []>("hidden_states_487_cast_fp16")];
            tensor<int32, [4]> var_3353_perm_0 = const()[name = tensor<string, []>("op_3353_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3357 = const()[name = tensor<string, []>("op_3357"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_507 = transpose(perm = var_3353_perm_0, x = hidden_states_487_cast_fp16)[name = tensor<string, []>("transpose_507")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_489_cast_fp16 = reshape(shape = var_3357, x = transpose_507)[name = tensor<string, []>("hidden_states_489_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554599488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555828352))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555828544)))];
            tensor<fp16, [1, 1024, 1280]> linear_234_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_489_cast_fp16)[name = tensor<string, []>("linear_234_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_585_cast_fp16 = add(x = linear_234_cast_fp16, y = input_577_cast_fp16)[name = tensor<string, []>("input_585_cast_fp16")];
            tensor<int32, [1]> input_587_axes_0 = const()[name = tensor<string, []>("input_587_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555831168)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555833792)))];
            tensor<fp16, [1, 1024, 1280]> input_587_cast_fp16 = layer_norm(axes = input_587_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_7_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_7_norm3_weight_to_fp16, x = input_585_cast_fp16)[name = tensor<string, []>("input_587_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555836416))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(565666880))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(565667072))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(565674816))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_235_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized, x = input_587_cast_fp16)[name = tensor<string, []>("linear_235_cast_fp16")];
            tensor<int32, [2]> var_3379_split_sizes_0 = const()[name = tensor<string, []>("op_3379_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_3379_axis_0 = const()[name = tensor<string, []>("op_3379_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_3379_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_3379_cast_fp16_1 = split(axis = var_3379_axis_0, split_sizes = var_3379_split_sizes_0, x = linear_235_cast_fp16)[name = tensor<string, []>("op_3379_cast_fp16")];
            tensor<string, []> var_3381_mode_0 = const()[name = tensor<string, []>("op_3381_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_3381_cast_fp16 = gelu(mode = var_3381_mode_0, x = var_3379_cast_fp16_1)[name = tensor<string, []>("op_3381_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_589_cast_fp16 = mul(x = var_3379_cast_fp16_0, y = var_3381_cast_fp16)[name = tensor<string, []>("input_589_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(565675008))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570590272))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_7_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_7_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570590464)))];
            tensor<fp16, [1, 1024, 1280]> linear_236_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_7_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = tensor<string, []>("linear_236_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_497_cast_fp16 = add(x = linear_236_cast_fp16, y = input_585_cast_fp16)[name = tensor<string, []>("hidden_states_497_cast_fp16")];
            tensor<int32, [1]> hidden_states_499_axes_0 = const()[name = tensor<string, []>("hidden_states_499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570593088)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570595712)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_499_cast_fp16 = layer_norm(axes = hidden_states_499_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_8_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_8_norm1_weight_to_fp16, x = hidden_states_497_cast_fp16)[name = tensor<string, []>("hidden_states_499_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570598336))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571827200))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_237_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_499_cast_fp16)[name = tensor<string, []>("linear_237_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571827392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(573056256))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_238_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_499_cast_fp16)[name = tensor<string, []>("linear_238_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(573056448))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(574285312))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_239_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_499_cast_fp16)[name = tensor<string, []>("linear_239_cast_fp16")];
            tensor<int32, [4]> var_3416 = const()[name = tensor<string, []>("op_3416"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3417_cast_fp16 = reshape(shape = var_3416, x = linear_237_cast_fp16)[name = tensor<string, []>("op_3417_cast_fp16")];
            tensor<int32, [4]> var_3419 = const()[name = tensor<string, []>("op_3419"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3420_cast_fp16 = reshape(shape = var_3419, x = linear_238_cast_fp16)[name = tensor<string, []>("op_3420_cast_fp16")];
            tensor<int32, [4]> var_3422 = const()[name = tensor<string, []>("op_3422"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3423_cast_fp16 = reshape(shape = var_3422, x = linear_239_cast_fp16)[name = tensor<string, []>("op_3423_cast_fp16")];
            tensor<int32, [4]> value_179_perm_0 = const()[name = tensor<string, []>("value_179_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_60_y_0_to_fp16 = const()[name = tensor<string, []>("mul_60_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_60_cast_fp16 = mul(x = var_3417_cast_fp16, y = mul_60_y_0_to_fp16)[name = tensor<string, []>("mul_60_cast_fp16")];
            tensor<bool, []> matmul_44_transpose_y_0 = const()[name = tensor<string, []>("matmul_44_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_44_transpose_x_0 = const()[name = tensor<string, []>("matmul_44_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_360_perm_0 = const()[name = tensor<string, []>("transpose_360_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_361_perm_0 = const()[name = tensor<string, []>("transpose_361_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_505 = transpose(perm = transpose_361_perm_0, x = var_3420_cast_fp16)[name = tensor<string, []>("transpose_505")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_506 = transpose(perm = transpose_360_perm_0, x = mul_60_cast_fp16)[name = tensor<string, []>("transpose_506")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_44_cast_fp16 = matmul(transpose_x = matmul_44_transpose_x_0, transpose_y = matmul_44_transpose_y_0, x = transpose_506, y = transpose_505)[name = tensor<string, []>("matmul_44_cast_fp16")];
            tensor<int32, []> softmax_44_axis_0 = const()[name = tensor<string, []>("softmax_44_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_44_cast_fp16 = softmax(axis = softmax_44_axis_0, x = matmul_44_cast_fp16)[name = tensor<string, []>("softmax_44_cast_fp16")];
            tensor<bool, []> hidden_states_501_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_501_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_501_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_501_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_504 = transpose(perm = value_179_perm_0, x = var_3423_cast_fp16)[name = tensor<string, []>("transpose_504")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_501_cast_fp16 = matmul(transpose_x = hidden_states_501_transpose_x_0, transpose_y = hidden_states_501_transpose_y_0, x = softmax_44_cast_fp16, y = transpose_504)[name = tensor<string, []>("hidden_states_501_cast_fp16")];
            tensor<int32, [4]> var_3426_perm_0 = const()[name = tensor<string, []>("op_3426_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3430 = const()[name = tensor<string, []>("op_3430"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_503 = transpose(perm = var_3426_perm_0, x = hidden_states_501_cast_fp16)[name = tensor<string, []>("transpose_503")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_503_cast_fp16 = reshape(shape = var_3430, x = transpose_503)[name = tensor<string, []>("hidden_states_503_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(574285504))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575514368))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575514560)))];
            tensor<fp16, [1, 1024, 1280]> linear_240_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_503_cast_fp16)[name = tensor<string, []>("linear_240_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_597_cast_fp16 = add(x = linear_240_cast_fp16, y = hidden_states_497_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
            tensor<int32, [1]> input_599_axes_0 = const()[name = tensor<string, []>("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575517184)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575519808)))];
            tensor<fp16, [1, 1024, 1280]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_8_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_8_norm2_weight_to_fp16, x = input_597_cast_fp16)[name = tensor<string, []>("input_599_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575522432))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(576751296))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_241_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = tensor<string, []>("linear_241_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(576751488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578717632))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_242_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_242_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578717824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(580683968))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_243_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_243_cast_fp16")];
            tensor<int32, [4]> var_3462 = const()[name = tensor<string, []>("op_3462"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3463_cast_fp16 = reshape(shape = var_3462, x = linear_241_cast_fp16)[name = tensor<string, []>("op_3463_cast_fp16")];
            tensor<int32, [4]> var_3465 = const()[name = tensor<string, []>("op_3465"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3466_cast_fp16 = reshape(shape = var_3465, x = linear_242_cast_fp16)[name = tensor<string, []>("op_3466_cast_fp16")];
            tensor<int32, [4]> var_3468 = const()[name = tensor<string, []>("op_3468"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3469_cast_fp16 = reshape(shape = var_3468, x = linear_243_cast_fp16)[name = tensor<string, []>("op_3469_cast_fp16")];
            tensor<int32, [4]> value_183_perm_0 = const()[name = tensor<string, []>("value_183_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_61_y_0_to_fp16 = const()[name = tensor<string, []>("mul_61_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_61_cast_fp16 = mul(x = var_3463_cast_fp16, y = mul_61_y_0_to_fp16)[name = tensor<string, []>("mul_61_cast_fp16")];
            tensor<bool, []> matmul_45_transpose_y_0 = const()[name = tensor<string, []>("matmul_45_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_45_transpose_x_0 = const()[name = tensor<string, []>("matmul_45_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_362_perm_0 = const()[name = tensor<string, []>("transpose_362_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_363_perm_0 = const()[name = tensor<string, []>("transpose_363_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_501 = transpose(perm = transpose_363_perm_0, x = var_3466_cast_fp16)[name = tensor<string, []>("transpose_501")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_502 = transpose(perm = transpose_362_perm_0, x = mul_61_cast_fp16)[name = tensor<string, []>("transpose_502")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_45_cast_fp16 = matmul(transpose_x = matmul_45_transpose_x_0, transpose_y = matmul_45_transpose_y_0, x = transpose_502, y = transpose_501)[name = tensor<string, []>("matmul_45_cast_fp16")];
            tensor<int32, []> softmax_45_axis_0 = const()[name = tensor<string, []>("softmax_45_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_45_cast_fp16 = softmax(axis = softmax_45_axis_0, x = matmul_45_cast_fp16)[name = tensor<string, []>("softmax_45_cast_fp16")];
            tensor<bool, []> hidden_states_507_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_507_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_507_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_507_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_500 = transpose(perm = value_183_perm_0, x = var_3469_cast_fp16)[name = tensor<string, []>("transpose_500")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_507_cast_fp16 = matmul(transpose_x = hidden_states_507_transpose_x_0, transpose_y = hidden_states_507_transpose_y_0, x = softmax_45_cast_fp16, y = transpose_500)[name = tensor<string, []>("hidden_states_507_cast_fp16")];
            tensor<int32, [4]> var_3472_perm_0 = const()[name = tensor<string, []>("op_3472_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3476 = const()[name = tensor<string, []>("op_3476"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_499 = transpose(perm = var_3472_perm_0, x = hidden_states_507_cast_fp16)[name = tensor<string, []>("transpose_499")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_509_cast_fp16 = reshape(shape = var_3476, x = transpose_499)[name = tensor<string, []>("hidden_states_509_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(580684160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581913024))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581913216)))];
            tensor<fp16, [1, 1024, 1280]> linear_244_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_509_cast_fp16)[name = tensor<string, []>("linear_244_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_605_cast_fp16 = add(x = linear_244_cast_fp16, y = input_597_cast_fp16)[name = tensor<string, []>("input_605_cast_fp16")];
            tensor<int32, [1]> input_607_axes_0 = const()[name = tensor<string, []>("input_607_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581915840)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581918464)))];
            tensor<fp16, [1, 1024, 1280]> input_607_cast_fp16 = layer_norm(axes = input_607_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_8_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_8_norm3_weight_to_fp16, x = input_605_cast_fp16)[name = tensor<string, []>("input_607_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581921088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591751552))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591751744))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591759488))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_245_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized, x = input_607_cast_fp16)[name = tensor<string, []>("linear_245_cast_fp16")];
            tensor<int32, [2]> var_3498_split_sizes_0 = const()[name = tensor<string, []>("op_3498_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_3498_axis_0 = const()[name = tensor<string, []>("op_3498_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_3498_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_3498_cast_fp16_1 = split(axis = var_3498_axis_0, split_sizes = var_3498_split_sizes_0, x = linear_245_cast_fp16)[name = tensor<string, []>("op_3498_cast_fp16")];
            tensor<string, []> var_3500_mode_0 = const()[name = tensor<string, []>("op_3500_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_3500_cast_fp16 = gelu(mode = var_3500_mode_0, x = var_3498_cast_fp16_1)[name = tensor<string, []>("op_3500_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_609_cast_fp16 = mul(x = var_3498_cast_fp16_0, y = var_3500_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591759680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(596674944))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_8_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_8_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(596675136)))];
            tensor<fp16, [1, 1024, 1280]> linear_246_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_8_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized, x = input_609_cast_fp16)[name = tensor<string, []>("linear_246_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_517_cast_fp16 = add(x = linear_246_cast_fp16, y = input_605_cast_fp16)[name = tensor<string, []>("hidden_states_517_cast_fp16")];
            tensor<int32, [1]> hidden_states_519_axes_0 = const()[name = tensor<string, []>("hidden_states_519_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_norm1_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(596677760)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_norm1_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(596680384)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_519_cast_fp16 = layer_norm(axes = hidden_states_519_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_9_norm1_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_9_norm1_weight_to_fp16, x = hidden_states_517_cast_fp16)[name = tensor<string, []>("hidden_states_519_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(596683008))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597911872))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_247_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_519_cast_fp16)[name = tensor<string, []>("linear_247_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597912064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(599140928))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_248_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_519_cast_fp16)[name = tensor<string, []>("linear_248_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(599141120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(600369984))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_249_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_519_cast_fp16)[name = tensor<string, []>("linear_249_cast_fp16")];
            tensor<int32, [4]> var_3535 = const()[name = tensor<string, []>("op_3535"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3536_cast_fp16 = reshape(shape = var_3535, x = linear_247_cast_fp16)[name = tensor<string, []>("op_3536_cast_fp16")];
            tensor<int32, [4]> var_3538 = const()[name = tensor<string, []>("op_3538"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3539_cast_fp16 = reshape(shape = var_3538, x = linear_248_cast_fp16)[name = tensor<string, []>("op_3539_cast_fp16")];
            tensor<int32, [4]> var_3541 = const()[name = tensor<string, []>("op_3541"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3542_cast_fp16 = reshape(shape = var_3541, x = linear_249_cast_fp16)[name = tensor<string, []>("op_3542_cast_fp16")];
            tensor<int32, [4]> value_187_perm_0 = const()[name = tensor<string, []>("value_187_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_62_y_0_to_fp16 = const()[name = tensor<string, []>("mul_62_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_62_cast_fp16 = mul(x = var_3536_cast_fp16, y = mul_62_y_0_to_fp16)[name = tensor<string, []>("mul_62_cast_fp16")];
            tensor<bool, []> matmul_46_transpose_y_0 = const()[name = tensor<string, []>("matmul_46_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_46_transpose_x_0 = const()[name = tensor<string, []>("matmul_46_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_364_perm_0 = const()[name = tensor<string, []>("transpose_364_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_365_perm_0 = const()[name = tensor<string, []>("transpose_365_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_497 = transpose(perm = transpose_365_perm_0, x = var_3539_cast_fp16)[name = tensor<string, []>("transpose_497")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_498 = transpose(perm = transpose_364_perm_0, x = mul_62_cast_fp16)[name = tensor<string, []>("transpose_498")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_46_cast_fp16 = matmul(transpose_x = matmul_46_transpose_x_0, transpose_y = matmul_46_transpose_y_0, x = transpose_498, y = transpose_497)[name = tensor<string, []>("matmul_46_cast_fp16")];
            tensor<int32, []> softmax_46_axis_0 = const()[name = tensor<string, []>("softmax_46_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_46_cast_fp16 = softmax(axis = softmax_46_axis_0, x = matmul_46_cast_fp16)[name = tensor<string, []>("softmax_46_cast_fp16")];
            tensor<bool, []> hidden_states_521_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_521_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_521_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_521_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_496 = transpose(perm = value_187_perm_0, x = var_3542_cast_fp16)[name = tensor<string, []>("transpose_496")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_521_cast_fp16 = matmul(transpose_x = hidden_states_521_transpose_x_0, transpose_y = hidden_states_521_transpose_y_0, x = softmax_46_cast_fp16, y = transpose_496)[name = tensor<string, []>("hidden_states_521_cast_fp16")];
            tensor<int32, [4]> var_3545_perm_0 = const()[name = tensor<string, []>("op_3545_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3549 = const()[name = tensor<string, []>("op_3549"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_495 = transpose(perm = var_3545_perm_0, x = hidden_states_521_cast_fp16)[name = tensor<string, []>("transpose_495")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_523_cast_fp16 = reshape(shape = var_3549, x = transpose_495)[name = tensor<string, []>("hidden_states_523_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(600370176))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(601599040))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(601599232)))];
            tensor<fp16, [1, 1024, 1280]> linear_250_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_523_cast_fp16)[name = tensor<string, []>("linear_250_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_617_cast_fp16 = add(x = linear_250_cast_fp16, y = hidden_states_517_cast_fp16)[name = tensor<string, []>("input_617_cast_fp16")];
            tensor<int32, [1]> input_619_axes_0 = const()[name = tensor<string, []>("input_619_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_norm2_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(601601856)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_norm2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(601604480)))];
            tensor<fp16, [1, 1024, 1280]> input_619_cast_fp16 = layer_norm(axes = input_619_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_9_norm2_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_9_norm2_weight_to_fp16, x = input_617_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(601607104))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602835968))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_251_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized, x = input_619_cast_fp16)[name = tensor<string, []>("linear_251_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602836160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604802304))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_252_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_252_cast_fp16")];
            tensor<fp16, [1280, 2048]> down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604802496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(606768640))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_253_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_253_cast_fp16")];
            tensor<int32, [4]> var_3581 = const()[name = tensor<string, []>("op_3581"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3582_cast_fp16 = reshape(shape = var_3581, x = linear_251_cast_fp16)[name = tensor<string, []>("op_3582_cast_fp16")];
            tensor<int32, [4]> var_3584 = const()[name = tensor<string, []>("op_3584"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3585_cast_fp16 = reshape(shape = var_3584, x = linear_252_cast_fp16)[name = tensor<string, []>("op_3585_cast_fp16")];
            tensor<int32, [4]> var_3587 = const()[name = tensor<string, []>("op_3587"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3588_cast_fp16 = reshape(shape = var_3587, x = linear_253_cast_fp16)[name = tensor<string, []>("op_3588_cast_fp16")];
            tensor<int32, [4]> value_191_perm_0 = const()[name = tensor<string, []>("value_191_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_63_y_0_to_fp16 = const()[name = tensor<string, []>("mul_63_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_63_cast_fp16 = mul(x = var_3582_cast_fp16, y = mul_63_y_0_to_fp16)[name = tensor<string, []>("mul_63_cast_fp16")];
            tensor<bool, []> matmul_47_transpose_y_0 = const()[name = tensor<string, []>("matmul_47_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_47_transpose_x_0 = const()[name = tensor<string, []>("matmul_47_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_366_perm_0 = const()[name = tensor<string, []>("transpose_366_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_367_perm_0 = const()[name = tensor<string, []>("transpose_367_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_493 = transpose(perm = transpose_367_perm_0, x = var_3585_cast_fp16)[name = tensor<string, []>("transpose_493")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_494 = transpose(perm = transpose_366_perm_0, x = mul_63_cast_fp16)[name = tensor<string, []>("transpose_494")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_47_cast_fp16 = matmul(transpose_x = matmul_47_transpose_x_0, transpose_y = matmul_47_transpose_y_0, x = transpose_494, y = transpose_493)[name = tensor<string, []>("matmul_47_cast_fp16")];
            tensor<int32, []> softmax_47_axis_0 = const()[name = tensor<string, []>("softmax_47_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_47_cast_fp16 = softmax(axis = softmax_47_axis_0, x = matmul_47_cast_fp16)[name = tensor<string, []>("softmax_47_cast_fp16")];
            tensor<bool, []> hidden_states_527_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_527_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_527_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_527_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_492 = transpose(perm = value_191_perm_0, x = var_3588_cast_fp16)[name = tensor<string, []>("transpose_492")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_527_cast_fp16 = matmul(transpose_x = hidden_states_527_transpose_x_0, transpose_y = hidden_states_527_transpose_y_0, x = softmax_47_cast_fp16, y = transpose_492)[name = tensor<string, []>("hidden_states_527_cast_fp16")];
            tensor<int32, [4]> var_3591_perm_0 = const()[name = tensor<string, []>("op_3591_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3595 = const()[name = tensor<string, []>("op_3595"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_491 = transpose(perm = var_3591_perm_0, x = hidden_states_527_cast_fp16)[name = tensor<string, []>("transpose_491")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_529_cast_fp16 = reshape(shape = var_3595, x = transpose_491)[name = tensor<string, []>("hidden_states_529_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(606768832))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(607997696))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(607997888)))];
            tensor<fp16, [1, 1024, 1280]> linear_254_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_out_0_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_529_cast_fp16)[name = tensor<string, []>("linear_254_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_625_cast_fp16 = add(x = linear_254_cast_fp16, y = input_617_cast_fp16)[name = tensor<string, []>("input_625_cast_fp16")];
            tensor<int32, [1]> input_627_axes_0 = const()[name = tensor<string, []>("input_627_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_norm3_weight_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(608000512)))];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_norm3_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(608003136)))];
            tensor<fp16, [1, 1024, 1280]> input_627_cast_fp16 = layer_norm(axes = input_627_axes_0, beta = down_blocks_2_attentions_1_transformer_blocks_9_norm3_bias_to_fp16, epsilon = var_1044_to_fp16, gamma = down_blocks_2_attentions_1_transformer_blocks_9_norm3_weight_to_fp16, x = input_625_cast_fp16)[name = tensor<string, []>("input_627_cast_fp16")];
            tensor<fp16, [10240, 1280]> down_blocks_2_attentions_1_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(608005760))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(617836224))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> down_blocks_2_attentions_1_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(617836416))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(617844160))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_255_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized, weight = down_blocks_2_attentions_1_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = tensor<string, []>("linear_255_cast_fp16")];
            tensor<int32, [2]> var_3617_split_sizes_0 = const()[name = tensor<string, []>("op_3617_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_3617_axis_0 = const()[name = tensor<string, []>("op_3617_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_3617_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_3617_cast_fp16_1 = split(axis = var_3617_axis_0, split_sizes = var_3617_split_sizes_0, x = linear_255_cast_fp16)[name = tensor<string, []>("op_3617_cast_fp16")];
            tensor<string, []> var_3619_mode_0 = const()[name = tensor<string, []>("op_3619_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_3619_cast_fp16 = gelu(mode = var_3619_mode_0, x = var_3617_cast_fp16_1)[name = tensor<string, []>("op_3619_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_629_cast_fp16 = mul(x = var_3617_cast_fp16_0, y = var_3619_cast_fp16)[name = tensor<string, []>("input_629_cast_fp16")];
            tensor<fp16, [1280, 5120]> down_blocks_2_attentions_1_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(617844352))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622759616))), name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_transformer_blocks_9_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_transformer_blocks_9_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622759808)))];
            tensor<fp16, [1, 1024, 1280]> linear_256_cast_fp16 = linear(bias = down_blocks_2_attentions_1_transformer_blocks_9_ff_net_2_bias_to_fp16, weight = down_blocks_2_attentions_1_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized, x = input_629_cast_fp16)[name = tensor<string, []>("linear_256_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_633_cast_fp16 = add(x = linear_256_cast_fp16, y = input_625_cast_fp16)[name = tensor<string, []>("input_633_cast_fp16")];
            tensor<fp16, [1280, 1280]> down_blocks_2_attentions_1_proj_out_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622762432))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623991296))), name = tensor<string, []>("down_blocks_2_attentions_1_proj_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> down_blocks_2_attentions_1_proj_out_bias_to_fp16 = const()[name = tensor<string, []>("down_blocks_2_attentions_1_proj_out_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623991488)))];
            tensor<fp16, [1, 1024, 1280]> linear_257_cast_fp16 = linear(bias = down_blocks_2_attentions_1_proj_out_bias_to_fp16, weight = down_blocks_2_attentions_1_proj_out_weight_to_fp16_palettized, x = input_633_cast_fp16)[name = tensor<string, []>("linear_257_cast_fp16")];
            tensor<int32, [4]> var_3629 = const()[name = tensor<string, []>("op_3629"), val = tensor<int32, [4]>([1, 32, 32, 1280])];
            tensor<fp16, [1, 32, 32, 1280]> var_3630_cast_fp16 = reshape(shape = var_3629, x = linear_257_cast_fp16)[name = tensor<string, []>("op_3630_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = tensor<string, []>("op_3631"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 1280, 32, 32]> transpose_490 = transpose(perm = var_3631, x = var_3630_cast_fp16)[name = tensor<string, []>("transpose_490")];
            tensor<fp16, [1, 1280, 32, 32]> input_635_cast_fp16 = add(x = transpose_490, y = var_2389_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
            tensor<int32, []> var_3651 = const()[name = tensor<string, []>("op_3651"), val = tensor<int32, []>(1)];
            tensor<int32, [5]> reshape_64_shape_0 = const()[name = tensor<string, []>("reshape_64_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_64_cast_fp16 = reshape(shape = reshape_64_shape_0, x = input_635_cast_fp16)[name = tensor<string, []>("reshape_64_cast_fp16")];
            tensor<int32, [3]> reduce_mean_48_axes_0 = const()[name = tensor<string, []>("reduce_mean_48_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_48_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_48_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_48_cast_fp16 = reduce_mean(axes = reduce_mean_48_axes_0, keep_dims = reduce_mean_48_keep_dims_0, x = reshape_64_cast_fp16)[name = tensor<string, []>("reduce_mean_48_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_32_cast_fp16 = sub(x = reshape_64_cast_fp16, y = reduce_mean_48_cast_fp16)[name = tensor<string, []>("sub_32_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_16_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor<string, []>("square_16_cast_fp16")];
            tensor<int32, [3]> reduce_mean_50_axes_0 = const()[name = tensor<string, []>("reduce_mean_50_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_50_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_50_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_50_cast_fp16 = reduce_mean(axes = reduce_mean_50_axes_0, keep_dims = reduce_mean_50_keep_dims_0, x = square_16_cast_fp16)[name = tensor<string, []>("reduce_mean_50_cast_fp16")];
            tensor<fp16, []> add_32_y_0_to_fp16 = const()[name = tensor<string, []>("add_32_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_32_cast_fp16 = add(x = reduce_mean_50_cast_fp16, y = add_32_y_0_to_fp16)[name = tensor<string, []>("add_32_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_16_cast_fp16 = sqrt(x = add_32_cast_fp16)[name = tensor<string, []>("sqrt_16_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_16_cast_fp16 = real_div(x = sub_32_cast_fp16, y = sqrt_16_cast_fp16)[name = tensor<string, []>("real_div_16_cast_fp16")];
            tensor<int32, [4]> reshape_65_shape_0 = const()[name = tensor<string, []>("reshape_65_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_65_cast_fp16 = reshape(shape = reshape_65_shape_0, x = real_div_16_cast_fp16)[name = tensor<string, []>("reshape_65_cast_fp16")];
            tensor<fp16, [1280]> add_33_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_33_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623994112)))];
            tensor<fp16, [1280]> add_33_beta_0_to_fp16 = const()[name = tensor<string, []>("add_33_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623996736)))];
            tensor<fp16, []> add_33_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_33_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_33_cast_fp16 = batch_norm(beta = add_33_beta_0_to_fp16, epsilon = add_33_epsilon_0_to_fp16, gamma = add_33_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_65_cast_fp16)[name = tensor<string, []>("add_33_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_639_cast_fp16 = silu(x = add_33_cast_fp16)[name = tensor<string, []>("input_639_cast_fp16")];
            tensor<int32, [2]> var_3673 = const()[name = tensor<string, []>("op_3673"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3675 = const()[name = tensor<string, []>("op_3675"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_541_pad_type_0 = const()[name = tensor<string, []>("hidden_states_541_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_541_pad_0 = const()[name = tensor<string, []>("hidden_states_541_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [1280, 1280, 3, 3]> mid_block_resnets_0_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [11059200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623999360))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(635058624))), name = tensor<string, []>("mid_block_resnets_0_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 3, 3])];
            tensor<fp16, [1280]> mid_block_resnets_0_conv1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_resnets_0_conv1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(635058816)))];
            tensor<fp16, [1, 1280, 32, 32]> hidden_states_541_cast_fp16 = conv(bias = mid_block_resnets_0_conv1_bias_to_fp16, dilations = var_3675, groups = var_3651, pad = hidden_states_541_pad_0, pad_type = hidden_states_541_pad_type_0, strides = var_3673, weight = mid_block_resnets_0_conv1_weight_to_fp16_palettized, x = input_639_cast_fp16)[name = tensor<string, []>("hidden_states_541_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_resnets_0_time_emb_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(635061440))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(636290304))), name = tensor<string, []>("mid_block_resnets_0_time_emb_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_resnets_0_time_emb_proj_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_resnets_0_time_emb_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(636290496)))];
            tensor<fp16, [1, 1280]> linear_258_cast_fp16 = linear(bias = mid_block_resnets_0_time_emb_proj_bias_to_fp16, weight = mid_block_resnets_0_time_emb_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_258_cast_fp16")];
            tensor<int32, [1]> var_3684_axes_0 = const()[name = tensor<string, []>("op_3684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1280, 1]> var_3684_cast_fp16 = expand_dims(axes = var_3684_axes_0, x = linear_258_cast_fp16)[name = tensor<string, []>("op_3684_cast_fp16")];
            tensor<int32, [1]> temb_13_axes_0 = const()[name = tensor<string, []>("temb_13_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 1280, 1, 1]> temb_13_cast_fp16 = expand_dims(axes = temb_13_axes_0, x = var_3684_cast_fp16)[name = tensor<string, []>("temb_13_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_643_cast_fp16 = add(x = hidden_states_541_cast_fp16, y = temb_13_cast_fp16)[name = tensor<string, []>("input_643_cast_fp16")];
            tensor<int32, [5]> reshape_68_shape_0 = const()[name = tensor<string, []>("reshape_68_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = input_643_cast_fp16)[name = tensor<string, []>("reshape_68_cast_fp16")];
            tensor<int32, [3]> reduce_mean_51_axes_0 = const()[name = tensor<string, []>("reduce_mean_51_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_51_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_51_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = reshape_68_cast_fp16)[name = tensor<string, []>("reduce_mean_51_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_34_cast_fp16 = sub(x = reshape_68_cast_fp16, y = reduce_mean_51_cast_fp16)[name = tensor<string, []>("sub_34_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_17_cast_fp16 = square(x = sub_34_cast_fp16)[name = tensor<string, []>("square_17_cast_fp16")];
            tensor<int32, [3]> reduce_mean_53_axes_0 = const()[name = tensor<string, []>("reduce_mean_53_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_53_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_53_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_17_cast_fp16)[name = tensor<string, []>("reduce_mean_53_cast_fp16")];
            tensor<fp16, []> add_34_y_0_to_fp16 = const()[name = tensor<string, []>("add_34_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_34_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = add_34_y_0_to_fp16)[name = tensor<string, []>("add_34_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_17_cast_fp16 = sqrt(x = add_34_cast_fp16)[name = tensor<string, []>("sqrt_17_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_17_cast_fp16 = real_div(x = sub_34_cast_fp16, y = sqrt_17_cast_fp16)[name = tensor<string, []>("real_div_17_cast_fp16")];
            tensor<int32, [4]> reshape_69_shape_0 = const()[name = tensor<string, []>("reshape_69_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_69_cast_fp16 = reshape(shape = reshape_69_shape_0, x = real_div_17_cast_fp16)[name = tensor<string, []>("reshape_69_cast_fp16")];
            tensor<fp16, [1280]> add_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_35_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(636293120)))];
            tensor<fp16, [1280]> add_35_beta_0_to_fp16 = const()[name = tensor<string, []>("add_35_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(636295744)))];
            tensor<fp16, []> add_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_35_cast_fp16 = batch_norm(beta = add_35_beta_0_to_fp16, epsilon = add_35_epsilon_0_to_fp16, gamma = add_35_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_69_cast_fp16)[name = tensor<string, []>("add_35_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_647_cast_fp16 = silu(x = add_35_cast_fp16)[name = tensor<string, []>("input_647_cast_fp16")];
            tensor<int32, [2]> var_3694 = const()[name = tensor<string, []>("op_3694"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3696 = const()[name = tensor<string, []>("op_3696"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_543_pad_type_0 = const()[name = tensor<string, []>("hidden_states_543_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_543_pad_0 = const()[name = tensor<string, []>("hidden_states_543_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [1280, 1280, 3, 3]> mid_block_resnets_0_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [11059200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(636298368))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647357632))), name = tensor<string, []>("mid_block_resnets_0_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 3, 3])];
            tensor<fp16, [1280]> mid_block_resnets_0_conv2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_resnets_0_conv2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647357824)))];
            tensor<fp16, [1, 1280, 32, 32]> hidden_states_543_cast_fp16 = conv(bias = mid_block_resnets_0_conv2_bias_to_fp16, dilations = var_3696, groups = var_3651, pad = hidden_states_543_pad_0, pad_type = hidden_states_543_pad_type_0, strides = var_3694, weight = mid_block_resnets_0_conv2_weight_to_fp16_palettized, x = input_647_cast_fp16)[name = tensor<string, []>("hidden_states_543_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> var_3699_cast_fp16 = add(x = input_635_cast_fp16, y = hidden_states_543_cast_fp16)[name = tensor<string, []>("op_3699_cast_fp16")];
            tensor<int32, [5]> reshape_72_shape_0 = const()[name = tensor<string, []>("reshape_72_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3699_cast_fp16)[name = tensor<string, []>("reshape_72_cast_fp16")];
            tensor<int32, [3]> reduce_mean_54_axes_0 = const()[name = tensor<string, []>("reduce_mean_54_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_54_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_54_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = reshape_72_cast_fp16)[name = tensor<string, []>("reduce_mean_54_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_36_cast_fp16 = sub(x = reshape_72_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor<string, []>("sub_36_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_18_cast_fp16 = square(x = sub_36_cast_fp16)[name = tensor<string, []>("square_18_cast_fp16")];
            tensor<int32, [3]> reduce_mean_56_axes_0 = const()[name = tensor<string, []>("reduce_mean_56_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_56_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_56_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = square_18_cast_fp16)[name = tensor<string, []>("reduce_mean_56_cast_fp16")];
            tensor<fp16, []> add_36_y_0_to_fp16 = const()[name = tensor<string, []>("add_36_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_36_cast_fp16 = add(x = reduce_mean_56_cast_fp16, y = add_36_y_0_to_fp16)[name = tensor<string, []>("add_36_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_18_cast_fp16 = sqrt(x = add_36_cast_fp16)[name = tensor<string, []>("sqrt_18_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_18_cast_fp16 = real_div(x = sub_36_cast_fp16, y = sqrt_18_cast_fp16)[name = tensor<string, []>("real_div_18_cast_fp16")];
            tensor<int32, [4]> reshape_73_shape_0 = const()[name = tensor<string, []>("reshape_73_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = real_div_18_cast_fp16)[name = tensor<string, []>("reshape_73_cast_fp16")];
            tensor<fp16, [1280]> add_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_37_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647360448)))];
            tensor<fp16, [1280]> add_37_beta_0_to_fp16 = const()[name = tensor<string, []>("add_37_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647363072)))];
            tensor<fp16, []> add_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_37_cast_fp16 = batch_norm(beta = add_37_beta_0_to_fp16, epsilon = add_37_epsilon_0_to_fp16, gamma = add_37_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_73_cast_fp16)[name = tensor<string, []>("add_37_cast_fp16")];
            tensor<int32, [4]> var_3737 = const()[name = tensor<string, []>("op_3737"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3741 = const()[name = tensor<string, []>("op_3741"), val = tensor<int32, [3]>([1, 1024, 1280])];
            tensor<fp16, [1, 32, 32, 1280]> transpose_489 = transpose(perm = var_3737, x = add_37_cast_fp16)[name = tensor<string, []>("transpose_489")];
            tensor<fp16, [1, 1024, 1280]> input_651_cast_fp16 = reshape(shape = var_3741, x = transpose_489)[name = tensor<string, []>("input_651_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_proj_in_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647365696))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648594560))), name = tensor<string, []>("mid_block_attentions_0_proj_in_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_proj_in_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_proj_in_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648594752)))];
            tensor<fp16, [1, 1024, 1280]> linear_259_cast_fp16 = linear(bias = mid_block_attentions_0_proj_in_bias_to_fp16, weight = mid_block_attentions_0_proj_in_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = tensor<string, []>("linear_259_cast_fp16")];
            tensor<int32, [1]> hidden_states_551_axes_0 = const()[name = tensor<string, []>("hidden_states_551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648597376)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648600000)))];
            tensor<fp16, []> var_3647_to_fp16 = const()[name = tensor<string, []>("op_3647_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1280]> hidden_states_551_cast_fp16 = layer_norm(axes = hidden_states_551_axes_0, beta = mid_block_attentions_0_transformer_blocks_0_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_0_norm1_weight_to_fp16, x = linear_259_cast_fp16)[name = tensor<string, []>("hidden_states_551_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648602624))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(649831488))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_260_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_551_cast_fp16)[name = tensor<string, []>("linear_260_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(649831680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(651060544))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_261_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_551_cast_fp16)[name = tensor<string, []>("linear_261_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(651060736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652289600))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_262_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_551_cast_fp16)[name = tensor<string, []>("linear_262_cast_fp16")];
            tensor<int32, [4]> var_3774 = const()[name = tensor<string, []>("op_3774"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3775_cast_fp16 = reshape(shape = var_3774, x = linear_260_cast_fp16)[name = tensor<string, []>("op_3775_cast_fp16")];
            tensor<int32, [4]> var_3777 = const()[name = tensor<string, []>("op_3777"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3778_cast_fp16 = reshape(shape = var_3777, x = linear_261_cast_fp16)[name = tensor<string, []>("op_3778_cast_fp16")];
            tensor<int32, [4]> var_3780 = const()[name = tensor<string, []>("op_3780"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3781_cast_fp16 = reshape(shape = var_3780, x = linear_262_cast_fp16)[name = tensor<string, []>("op_3781_cast_fp16")];
            tensor<int32, [4]> value_195_perm_0 = const()[name = tensor<string, []>("value_195_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_67_y_0_to_fp16 = const()[name = tensor<string, []>("mul_67_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_67_cast_fp16 = mul(x = var_3775_cast_fp16, y = mul_67_y_0_to_fp16)[name = tensor<string, []>("mul_67_cast_fp16")];
            tensor<bool, []> matmul_48_transpose_y_0 = const()[name = tensor<string, []>("matmul_48_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_48_transpose_x_0 = const()[name = tensor<string, []>("matmul_48_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_368_perm_0 = const()[name = tensor<string, []>("transpose_368_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_369_perm_0 = const()[name = tensor<string, []>("transpose_369_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_487 = transpose(perm = transpose_369_perm_0, x = var_3778_cast_fp16)[name = tensor<string, []>("transpose_487")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_488 = transpose(perm = transpose_368_perm_0, x = mul_67_cast_fp16)[name = tensor<string, []>("transpose_488")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_48_cast_fp16 = matmul(transpose_x = matmul_48_transpose_x_0, transpose_y = matmul_48_transpose_y_0, x = transpose_488, y = transpose_487)[name = tensor<string, []>("matmul_48_cast_fp16")];
            tensor<int32, []> softmax_48_axis_0 = const()[name = tensor<string, []>("softmax_48_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_48_cast_fp16 = softmax(axis = softmax_48_axis_0, x = matmul_48_cast_fp16)[name = tensor<string, []>("softmax_48_cast_fp16")];
            tensor<bool, []> hidden_states_553_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_553_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_553_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_553_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_486 = transpose(perm = value_195_perm_0, x = var_3781_cast_fp16)[name = tensor<string, []>("transpose_486")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_553_cast_fp16 = matmul(transpose_x = hidden_states_553_transpose_x_0, transpose_y = hidden_states_553_transpose_y_0, x = softmax_48_cast_fp16, y = transpose_486)[name = tensor<string, []>("hidden_states_553_cast_fp16")];
            tensor<int32, [4]> var_3784_perm_0 = const()[name = tensor<string, []>("op_3784_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3788 = const()[name = tensor<string, []>("op_3788"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_485 = transpose(perm = var_3784_perm_0, x = hidden_states_553_cast_fp16)[name = tensor<string, []>("transpose_485")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_555_cast_fp16 = reshape(shape = var_3788, x = transpose_485)[name = tensor<string, []>("hidden_states_555_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652289792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653518656))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653518848)))];
            tensor<fp16, [1, 1024, 1280]> linear_263_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_0_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_555_cast_fp16)[name = tensor<string, []>("linear_263_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_657_cast_fp16 = add(x = linear_263_cast_fp16, y = linear_259_cast_fp16)[name = tensor<string, []>("input_657_cast_fp16")];
            tensor<int32, [1]> input_659_axes_0 = const()[name = tensor<string, []>("input_659_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653521472)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653524096)))];
            tensor<fp16, [1, 1024, 1280]> input_659_cast_fp16 = layer_norm(axes = input_659_axes_0, beta = mid_block_attentions_0_transformer_blocks_0_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_0_norm2_weight_to_fp16, x = input_657_cast_fp16)[name = tensor<string, []>("input_659_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653526720))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(654755584))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_264_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_attn2_to_q_weight_to_fp16_palettized, x = input_659_cast_fp16)[name = tensor<string, []>("linear_264_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(654755776))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(656721920))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_265_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_265_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(656722112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658688256))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_266_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_266_cast_fp16")];
            tensor<int32, [4]> var_3820 = const()[name = tensor<string, []>("op_3820"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3821_cast_fp16 = reshape(shape = var_3820, x = linear_264_cast_fp16)[name = tensor<string, []>("op_3821_cast_fp16")];
            tensor<int32, [4]> var_3823 = const()[name = tensor<string, []>("op_3823"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3824_cast_fp16 = reshape(shape = var_3823, x = linear_265_cast_fp16)[name = tensor<string, []>("op_3824_cast_fp16")];
            tensor<int32, [4]> var_3826 = const()[name = tensor<string, []>("op_3826"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3827_cast_fp16 = reshape(shape = var_3826, x = linear_266_cast_fp16)[name = tensor<string, []>("op_3827_cast_fp16")];
            tensor<int32, [4]> value_199_perm_0 = const()[name = tensor<string, []>("value_199_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_68_y_0_to_fp16 = const()[name = tensor<string, []>("mul_68_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_68_cast_fp16 = mul(x = var_3821_cast_fp16, y = mul_68_y_0_to_fp16)[name = tensor<string, []>("mul_68_cast_fp16")];
            tensor<bool, []> matmul_49_transpose_y_0 = const()[name = tensor<string, []>("matmul_49_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_49_transpose_x_0 = const()[name = tensor<string, []>("matmul_49_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_370_perm_0 = const()[name = tensor<string, []>("transpose_370_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_371_perm_0 = const()[name = tensor<string, []>("transpose_371_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_483 = transpose(perm = transpose_371_perm_0, x = var_3824_cast_fp16)[name = tensor<string, []>("transpose_483")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_484 = transpose(perm = transpose_370_perm_0, x = mul_68_cast_fp16)[name = tensor<string, []>("transpose_484")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_49_cast_fp16 = matmul(transpose_x = matmul_49_transpose_x_0, transpose_y = matmul_49_transpose_y_0, x = transpose_484, y = transpose_483)[name = tensor<string, []>("matmul_49_cast_fp16")];
            tensor<int32, []> softmax_49_axis_0 = const()[name = tensor<string, []>("softmax_49_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_49_cast_fp16 = softmax(axis = softmax_49_axis_0, x = matmul_49_cast_fp16)[name = tensor<string, []>("softmax_49_cast_fp16")];
            tensor<bool, []> hidden_states_559_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_559_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_559_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_559_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_482 = transpose(perm = value_199_perm_0, x = var_3827_cast_fp16)[name = tensor<string, []>("transpose_482")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_559_cast_fp16 = matmul(transpose_x = hidden_states_559_transpose_x_0, transpose_y = hidden_states_559_transpose_y_0, x = softmax_49_cast_fp16, y = transpose_482)[name = tensor<string, []>("hidden_states_559_cast_fp16")];
            tensor<int32, [4]> var_3830_perm_0 = const()[name = tensor<string, []>("op_3830_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3834 = const()[name = tensor<string, []>("op_3834"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_481 = transpose(perm = var_3830_perm_0, x = hidden_states_559_cast_fp16)[name = tensor<string, []>("transpose_481")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_561_cast_fp16 = reshape(shape = var_3834, x = transpose_481)[name = tensor<string, []>("hidden_states_561_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658688448))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659917312))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659917504)))];
            tensor<fp16, [1, 1024, 1280]> linear_267_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_0_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_561_cast_fp16)[name = tensor<string, []>("linear_267_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_665_cast_fp16 = add(x = linear_267_cast_fp16, y = input_657_cast_fp16)[name = tensor<string, []>("input_665_cast_fp16")];
            tensor<int32, [1]> input_667_axes_0 = const()[name = tensor<string, []>("input_667_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659920128)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659922752)))];
            tensor<fp16, [1, 1024, 1280]> input_667_cast_fp16 = layer_norm(axes = input_667_axes_0, beta = mid_block_attentions_0_transformer_blocks_0_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_0_norm3_weight_to_fp16, x = input_665_cast_fp16)[name = tensor<string, []>("input_667_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659925376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(669755840))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(669756032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(669763776))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_268_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_0_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_0_ff_net_0_proj_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = tensor<string, []>("linear_268_cast_fp16")];
            tensor<int32, [2]> var_3856_split_sizes_0 = const()[name = tensor<string, []>("op_3856_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_3856_axis_0 = const()[name = tensor<string, []>("op_3856_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_3856_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_3856_cast_fp16_1 = split(axis = var_3856_axis_0, split_sizes = var_3856_split_sizes_0, x = linear_268_cast_fp16)[name = tensor<string, []>("op_3856_cast_fp16")];
            tensor<string, []> var_3858_mode_0 = const()[name = tensor<string, []>("op_3858_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_3858_cast_fp16 = gelu(mode = var_3858_mode_0, x = var_3856_cast_fp16_1)[name = tensor<string, []>("op_3858_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_669_cast_fp16 = mul(x = var_3856_cast_fp16_0, y = var_3858_cast_fp16)[name = tensor<string, []>("input_669_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(669763968))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(674679232))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(674679424)))];
            tensor<fp16, [1, 1024, 1280]> linear_269_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_0_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_0_ff_net_2_weight_to_fp16_palettized, x = input_669_cast_fp16)[name = tensor<string, []>("linear_269_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_569_cast_fp16 = add(x = linear_269_cast_fp16, y = input_665_cast_fp16)[name = tensor<string, []>("hidden_states_569_cast_fp16")];
            tensor<int32, [1]> hidden_states_571_axes_0 = const()[name = tensor<string, []>("hidden_states_571_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(674682048)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(674684672)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_571_cast_fp16 = layer_norm(axes = hidden_states_571_axes_0, beta = mid_block_attentions_0_transformer_blocks_1_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_1_norm1_weight_to_fp16, x = hidden_states_569_cast_fp16)[name = tensor<string, []>("hidden_states_571_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(674687296))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(675916160))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_270_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_571_cast_fp16)[name = tensor<string, []>("linear_270_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(675916352))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(677145216))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_271_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_571_cast_fp16)[name = tensor<string, []>("linear_271_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(677145408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678374272))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_272_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_571_cast_fp16)[name = tensor<string, []>("linear_272_cast_fp16")];
            tensor<int32, [4]> var_3893 = const()[name = tensor<string, []>("op_3893"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3894_cast_fp16 = reshape(shape = var_3893, x = linear_270_cast_fp16)[name = tensor<string, []>("op_3894_cast_fp16")];
            tensor<int32, [4]> var_3896 = const()[name = tensor<string, []>("op_3896"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3897_cast_fp16 = reshape(shape = var_3896, x = linear_271_cast_fp16)[name = tensor<string, []>("op_3897_cast_fp16")];
            tensor<int32, [4]> var_3899 = const()[name = tensor<string, []>("op_3899"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3900_cast_fp16 = reshape(shape = var_3899, x = linear_272_cast_fp16)[name = tensor<string, []>("op_3900_cast_fp16")];
            tensor<int32, [4]> value_203_perm_0 = const()[name = tensor<string, []>("value_203_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_69_y_0_to_fp16 = const()[name = tensor<string, []>("mul_69_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_69_cast_fp16 = mul(x = var_3894_cast_fp16, y = mul_69_y_0_to_fp16)[name = tensor<string, []>("mul_69_cast_fp16")];
            tensor<bool, []> matmul_50_transpose_y_0 = const()[name = tensor<string, []>("matmul_50_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_50_transpose_x_0 = const()[name = tensor<string, []>("matmul_50_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_372_perm_0 = const()[name = tensor<string, []>("transpose_372_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_373_perm_0 = const()[name = tensor<string, []>("transpose_373_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_479 = transpose(perm = transpose_373_perm_0, x = var_3897_cast_fp16)[name = tensor<string, []>("transpose_479")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_480 = transpose(perm = transpose_372_perm_0, x = mul_69_cast_fp16)[name = tensor<string, []>("transpose_480")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_50_cast_fp16 = matmul(transpose_x = matmul_50_transpose_x_0, transpose_y = matmul_50_transpose_y_0, x = transpose_480, y = transpose_479)[name = tensor<string, []>("matmul_50_cast_fp16")];
            tensor<int32, []> softmax_50_axis_0 = const()[name = tensor<string, []>("softmax_50_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_50_cast_fp16 = softmax(axis = softmax_50_axis_0, x = matmul_50_cast_fp16)[name = tensor<string, []>("softmax_50_cast_fp16")];
            tensor<bool, []> hidden_states_573_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_573_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_573_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_573_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_478 = transpose(perm = value_203_perm_0, x = var_3900_cast_fp16)[name = tensor<string, []>("transpose_478")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_573_cast_fp16 = matmul(transpose_x = hidden_states_573_transpose_x_0, transpose_y = hidden_states_573_transpose_y_0, x = softmax_50_cast_fp16, y = transpose_478)[name = tensor<string, []>("hidden_states_573_cast_fp16")];
            tensor<int32, [4]> var_3903_perm_0 = const()[name = tensor<string, []>("op_3903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3907 = const()[name = tensor<string, []>("op_3907"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_477 = transpose(perm = var_3903_perm_0, x = hidden_states_573_cast_fp16)[name = tensor<string, []>("transpose_477")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_575_cast_fp16 = reshape(shape = var_3907, x = transpose_477)[name = tensor<string, []>("hidden_states_575_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678374464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(679603328))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(679603520)))];
            tensor<fp16, [1, 1024, 1280]> linear_273_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_1_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_575_cast_fp16)[name = tensor<string, []>("linear_273_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_677_cast_fp16 = add(x = linear_273_cast_fp16, y = hidden_states_569_cast_fp16)[name = tensor<string, []>("input_677_cast_fp16")];
            tensor<int32, [1]> input_679_axes_0 = const()[name = tensor<string, []>("input_679_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(679606144)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(679608768)))];
            tensor<fp16, [1, 1024, 1280]> input_679_cast_fp16 = layer_norm(axes = input_679_axes_0, beta = mid_block_attentions_0_transformer_blocks_1_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_1_norm2_weight_to_fp16, x = input_677_cast_fp16)[name = tensor<string, []>("input_679_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(679611392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680840256))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_274_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_attn2_to_q_weight_to_fp16_palettized, x = input_679_cast_fp16)[name = tensor<string, []>("linear_274_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680840448))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(682806592))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_275_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_275_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(682806784))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(684772928))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_276_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_276_cast_fp16")];
            tensor<int32, [4]> var_3939 = const()[name = tensor<string, []>("op_3939"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_3940_cast_fp16 = reshape(shape = var_3939, x = linear_274_cast_fp16)[name = tensor<string, []>("op_3940_cast_fp16")];
            tensor<int32, [4]> var_3942 = const()[name = tensor<string, []>("op_3942"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3943_cast_fp16 = reshape(shape = var_3942, x = linear_275_cast_fp16)[name = tensor<string, []>("op_3943_cast_fp16")];
            tensor<int32, [4]> var_3945 = const()[name = tensor<string, []>("op_3945"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_3946_cast_fp16 = reshape(shape = var_3945, x = linear_276_cast_fp16)[name = tensor<string, []>("op_3946_cast_fp16")];
            tensor<int32, [4]> value_207_perm_0 = const()[name = tensor<string, []>("value_207_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_70_y_0_to_fp16 = const()[name = tensor<string, []>("mul_70_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_70_cast_fp16 = mul(x = var_3940_cast_fp16, y = mul_70_y_0_to_fp16)[name = tensor<string, []>("mul_70_cast_fp16")];
            tensor<bool, []> matmul_51_transpose_y_0 = const()[name = tensor<string, []>("matmul_51_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_51_transpose_x_0 = const()[name = tensor<string, []>("matmul_51_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_374_perm_0 = const()[name = tensor<string, []>("transpose_374_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_375_perm_0 = const()[name = tensor<string, []>("transpose_375_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_475 = transpose(perm = transpose_375_perm_0, x = var_3943_cast_fp16)[name = tensor<string, []>("transpose_475")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_476 = transpose(perm = transpose_374_perm_0, x = mul_70_cast_fp16)[name = tensor<string, []>("transpose_476")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_51_cast_fp16 = matmul(transpose_x = matmul_51_transpose_x_0, transpose_y = matmul_51_transpose_y_0, x = transpose_476, y = transpose_475)[name = tensor<string, []>("matmul_51_cast_fp16")];
            tensor<int32, []> softmax_51_axis_0 = const()[name = tensor<string, []>("softmax_51_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_51_cast_fp16 = softmax(axis = softmax_51_axis_0, x = matmul_51_cast_fp16)[name = tensor<string, []>("softmax_51_cast_fp16")];
            tensor<bool, []> hidden_states_579_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_579_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_579_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_579_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_474 = transpose(perm = value_207_perm_0, x = var_3946_cast_fp16)[name = tensor<string, []>("transpose_474")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_579_cast_fp16 = matmul(transpose_x = hidden_states_579_transpose_x_0, transpose_y = hidden_states_579_transpose_y_0, x = softmax_51_cast_fp16, y = transpose_474)[name = tensor<string, []>("hidden_states_579_cast_fp16")];
            tensor<int32, [4]> var_3949_perm_0 = const()[name = tensor<string, []>("op_3949_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3953 = const()[name = tensor<string, []>("op_3953"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_473 = transpose(perm = var_3949_perm_0, x = hidden_states_579_cast_fp16)[name = tensor<string, []>("transpose_473")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_581_cast_fp16 = reshape(shape = var_3953, x = transpose_473)[name = tensor<string, []>("hidden_states_581_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(684773120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686001984))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686002176)))];
            tensor<fp16, [1, 1024, 1280]> linear_277_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_1_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_581_cast_fp16)[name = tensor<string, []>("linear_277_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_685_cast_fp16 = add(x = linear_277_cast_fp16, y = input_677_cast_fp16)[name = tensor<string, []>("input_685_cast_fp16")];
            tensor<int32, [1]> input_687_axes_0 = const()[name = tensor<string, []>("input_687_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686004800)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686007424)))];
            tensor<fp16, [1, 1024, 1280]> input_687_cast_fp16 = layer_norm(axes = input_687_axes_0, beta = mid_block_attentions_0_transformer_blocks_1_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_1_norm3_weight_to_fp16, x = input_685_cast_fp16)[name = tensor<string, []>("input_687_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686010048))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695840512))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695840704))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695848448))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_278_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_1_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_1_ff_net_0_proj_weight_to_fp16_palettized, x = input_687_cast_fp16)[name = tensor<string, []>("linear_278_cast_fp16")];
            tensor<int32, [2]> var_3975_split_sizes_0 = const()[name = tensor<string, []>("op_3975_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_3975_axis_0 = const()[name = tensor<string, []>("op_3975_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_3975_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_3975_cast_fp16_1 = split(axis = var_3975_axis_0, split_sizes = var_3975_split_sizes_0, x = linear_278_cast_fp16)[name = tensor<string, []>("op_3975_cast_fp16")];
            tensor<string, []> var_3977_mode_0 = const()[name = tensor<string, []>("op_3977_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_3977_cast_fp16 = gelu(mode = var_3977_mode_0, x = var_3975_cast_fp16_1)[name = tensor<string, []>("op_3977_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_689_cast_fp16 = mul(x = var_3975_cast_fp16_0, y = var_3977_cast_fp16)[name = tensor<string, []>("input_689_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695848640))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(700763904))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(700764096)))];
            tensor<fp16, [1, 1024, 1280]> linear_279_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_1_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_1_ff_net_2_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = tensor<string, []>("linear_279_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_589_cast_fp16 = add(x = linear_279_cast_fp16, y = input_685_cast_fp16)[name = tensor<string, []>("hidden_states_589_cast_fp16")];
            tensor<int32, [1]> hidden_states_591_axes_0 = const()[name = tensor<string, []>("hidden_states_591_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(700766720)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(700769344)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_591_cast_fp16 = layer_norm(axes = hidden_states_591_axes_0, beta = mid_block_attentions_0_transformer_blocks_2_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_2_norm1_weight_to_fp16, x = hidden_states_589_cast_fp16)[name = tensor<string, []>("hidden_states_591_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(700771968))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(702000832))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_280_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_591_cast_fp16)[name = tensor<string, []>("linear_280_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(702001024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(703229888))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_281_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_591_cast_fp16)[name = tensor<string, []>("linear_281_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(703230080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(704458944))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_282_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_591_cast_fp16)[name = tensor<string, []>("linear_282_cast_fp16")];
            tensor<int32, [4]> var_4012 = const()[name = tensor<string, []>("op_4012"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4013_cast_fp16 = reshape(shape = var_4012, x = linear_280_cast_fp16)[name = tensor<string, []>("op_4013_cast_fp16")];
            tensor<int32, [4]> var_4015 = const()[name = tensor<string, []>("op_4015"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4016_cast_fp16 = reshape(shape = var_4015, x = linear_281_cast_fp16)[name = tensor<string, []>("op_4016_cast_fp16")];
            tensor<int32, [4]> var_4018 = const()[name = tensor<string, []>("op_4018"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4019_cast_fp16 = reshape(shape = var_4018, x = linear_282_cast_fp16)[name = tensor<string, []>("op_4019_cast_fp16")];
            tensor<int32, [4]> value_211_perm_0 = const()[name = tensor<string, []>("value_211_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_71_y_0_to_fp16 = const()[name = tensor<string, []>("mul_71_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_71_cast_fp16 = mul(x = var_4013_cast_fp16, y = mul_71_y_0_to_fp16)[name = tensor<string, []>("mul_71_cast_fp16")];
            tensor<bool, []> matmul_52_transpose_y_0 = const()[name = tensor<string, []>("matmul_52_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_52_transpose_x_0 = const()[name = tensor<string, []>("matmul_52_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_376_perm_0 = const()[name = tensor<string, []>("transpose_376_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_377_perm_0 = const()[name = tensor<string, []>("transpose_377_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_471 = transpose(perm = transpose_377_perm_0, x = var_4016_cast_fp16)[name = tensor<string, []>("transpose_471")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_472 = transpose(perm = transpose_376_perm_0, x = mul_71_cast_fp16)[name = tensor<string, []>("transpose_472")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_52_cast_fp16 = matmul(transpose_x = matmul_52_transpose_x_0, transpose_y = matmul_52_transpose_y_0, x = transpose_472, y = transpose_471)[name = tensor<string, []>("matmul_52_cast_fp16")];
            tensor<int32, []> softmax_52_axis_0 = const()[name = tensor<string, []>("softmax_52_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_52_cast_fp16 = softmax(axis = softmax_52_axis_0, x = matmul_52_cast_fp16)[name = tensor<string, []>("softmax_52_cast_fp16")];
            tensor<bool, []> hidden_states_593_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_593_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_593_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_593_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_470 = transpose(perm = value_211_perm_0, x = var_4019_cast_fp16)[name = tensor<string, []>("transpose_470")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_593_cast_fp16 = matmul(transpose_x = hidden_states_593_transpose_x_0, transpose_y = hidden_states_593_transpose_y_0, x = softmax_52_cast_fp16, y = transpose_470)[name = tensor<string, []>("hidden_states_593_cast_fp16")];
            tensor<int32, [4]> var_4022_perm_0 = const()[name = tensor<string, []>("op_4022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4026 = const()[name = tensor<string, []>("op_4026"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_469 = transpose(perm = var_4022_perm_0, x = hidden_states_593_cast_fp16)[name = tensor<string, []>("transpose_469")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_595_cast_fp16 = reshape(shape = var_4026, x = transpose_469)[name = tensor<string, []>("hidden_states_595_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(704459136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(705688000))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(705688192)))];
            tensor<fp16, [1, 1024, 1280]> linear_283_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_2_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_595_cast_fp16)[name = tensor<string, []>("linear_283_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_697_cast_fp16 = add(x = linear_283_cast_fp16, y = hidden_states_589_cast_fp16)[name = tensor<string, []>("input_697_cast_fp16")];
            tensor<int32, [1]> input_699_axes_0 = const()[name = tensor<string, []>("input_699_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(705690816)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(705693440)))];
            tensor<fp16, [1, 1024, 1280]> input_699_cast_fp16 = layer_norm(axes = input_699_axes_0, beta = mid_block_attentions_0_transformer_blocks_2_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_2_norm2_weight_to_fp16, x = input_697_cast_fp16)[name = tensor<string, []>("input_699_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(705696064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(706924928))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_284_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_attn2_to_q_weight_to_fp16_palettized, x = input_699_cast_fp16)[name = tensor<string, []>("linear_284_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(706925120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(708891264))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_285_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_285_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(708891456))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(710857600))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_286_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_286_cast_fp16")];
            tensor<int32, [4]> var_4058 = const()[name = tensor<string, []>("op_4058"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4059_cast_fp16 = reshape(shape = var_4058, x = linear_284_cast_fp16)[name = tensor<string, []>("op_4059_cast_fp16")];
            tensor<int32, [4]> var_4061 = const()[name = tensor<string, []>("op_4061"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4062_cast_fp16 = reshape(shape = var_4061, x = linear_285_cast_fp16)[name = tensor<string, []>("op_4062_cast_fp16")];
            tensor<int32, [4]> var_4064 = const()[name = tensor<string, []>("op_4064"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4065_cast_fp16 = reshape(shape = var_4064, x = linear_286_cast_fp16)[name = tensor<string, []>("op_4065_cast_fp16")];
            tensor<int32, [4]> value_215_perm_0 = const()[name = tensor<string, []>("value_215_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_72_y_0_to_fp16 = const()[name = tensor<string, []>("mul_72_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_72_cast_fp16 = mul(x = var_4059_cast_fp16, y = mul_72_y_0_to_fp16)[name = tensor<string, []>("mul_72_cast_fp16")];
            tensor<bool, []> matmul_53_transpose_y_0 = const()[name = tensor<string, []>("matmul_53_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_53_transpose_x_0 = const()[name = tensor<string, []>("matmul_53_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_378_perm_0 = const()[name = tensor<string, []>("transpose_378_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_379_perm_0 = const()[name = tensor<string, []>("transpose_379_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_467 = transpose(perm = transpose_379_perm_0, x = var_4062_cast_fp16)[name = tensor<string, []>("transpose_467")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_468 = transpose(perm = transpose_378_perm_0, x = mul_72_cast_fp16)[name = tensor<string, []>("transpose_468")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_53_cast_fp16 = matmul(transpose_x = matmul_53_transpose_x_0, transpose_y = matmul_53_transpose_y_0, x = transpose_468, y = transpose_467)[name = tensor<string, []>("matmul_53_cast_fp16")];
            tensor<int32, []> softmax_53_axis_0 = const()[name = tensor<string, []>("softmax_53_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_53_cast_fp16 = softmax(axis = softmax_53_axis_0, x = matmul_53_cast_fp16)[name = tensor<string, []>("softmax_53_cast_fp16")];
            tensor<bool, []> hidden_states_599_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_599_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_599_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_599_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_466 = transpose(perm = value_215_perm_0, x = var_4065_cast_fp16)[name = tensor<string, []>("transpose_466")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_599_cast_fp16 = matmul(transpose_x = hidden_states_599_transpose_x_0, transpose_y = hidden_states_599_transpose_y_0, x = softmax_53_cast_fp16, y = transpose_466)[name = tensor<string, []>("hidden_states_599_cast_fp16")];
            tensor<int32, [4]> var_4068_perm_0 = const()[name = tensor<string, []>("op_4068_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4072 = const()[name = tensor<string, []>("op_4072"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_465 = transpose(perm = var_4068_perm_0, x = hidden_states_599_cast_fp16)[name = tensor<string, []>("transpose_465")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_601_cast_fp16 = reshape(shape = var_4072, x = transpose_465)[name = tensor<string, []>("hidden_states_601_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(710857792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712086656))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712086848)))];
            tensor<fp16, [1, 1024, 1280]> linear_287_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_2_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_601_cast_fp16)[name = tensor<string, []>("linear_287_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_705_cast_fp16 = add(x = linear_287_cast_fp16, y = input_697_cast_fp16)[name = tensor<string, []>("input_705_cast_fp16")];
            tensor<int32, [1]> input_707_axes_0 = const()[name = tensor<string, []>("input_707_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712089472)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712092096)))];
            tensor<fp16, [1, 1024, 1280]> input_707_cast_fp16 = layer_norm(axes = input_707_axes_0, beta = mid_block_attentions_0_transformer_blocks_2_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_2_norm3_weight_to_fp16, x = input_705_cast_fp16)[name = tensor<string, []>("input_707_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712094720))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721925184))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721925376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721933120))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_288_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_2_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_2_ff_net_0_proj_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = tensor<string, []>("linear_288_cast_fp16")];
            tensor<int32, [2]> var_4094_split_sizes_0 = const()[name = tensor<string, []>("op_4094_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_4094_axis_0 = const()[name = tensor<string, []>("op_4094_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_4094_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_4094_cast_fp16_1 = split(axis = var_4094_axis_0, split_sizes = var_4094_split_sizes_0, x = linear_288_cast_fp16)[name = tensor<string, []>("op_4094_cast_fp16")];
            tensor<string, []> var_4096_mode_0 = const()[name = tensor<string, []>("op_4096_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_4096_cast_fp16 = gelu(mode = var_4096_mode_0, x = var_4094_cast_fp16_1)[name = tensor<string, []>("op_4096_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_709_cast_fp16 = mul(x = var_4094_cast_fp16_0, y = var_4096_cast_fp16)[name = tensor<string, []>("input_709_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721933312))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726848576))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_2_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_2_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726848768)))];
            tensor<fp16, [1, 1024, 1280]> linear_289_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_2_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_2_ff_net_2_weight_to_fp16_palettized, x = input_709_cast_fp16)[name = tensor<string, []>("linear_289_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_609_cast_fp16 = add(x = linear_289_cast_fp16, y = input_705_cast_fp16)[name = tensor<string, []>("hidden_states_609_cast_fp16")];
            tensor<int32, [1]> hidden_states_611_axes_0 = const()[name = tensor<string, []>("hidden_states_611_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726851392)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726854016)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_611_cast_fp16 = layer_norm(axes = hidden_states_611_axes_0, beta = mid_block_attentions_0_transformer_blocks_3_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_3_norm1_weight_to_fp16, x = hidden_states_609_cast_fp16)[name = tensor<string, []>("hidden_states_611_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726856640))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(728085504))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_290_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_611_cast_fp16)[name = tensor<string, []>("linear_290_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(728085696))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(729314560))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_291_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_611_cast_fp16)[name = tensor<string, []>("linear_291_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(729314752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(730543616))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_292_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_611_cast_fp16)[name = tensor<string, []>("linear_292_cast_fp16")];
            tensor<int32, [4]> var_4131 = const()[name = tensor<string, []>("op_4131"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4132_cast_fp16 = reshape(shape = var_4131, x = linear_290_cast_fp16)[name = tensor<string, []>("op_4132_cast_fp16")];
            tensor<int32, [4]> var_4134 = const()[name = tensor<string, []>("op_4134"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4135_cast_fp16 = reshape(shape = var_4134, x = linear_291_cast_fp16)[name = tensor<string, []>("op_4135_cast_fp16")];
            tensor<int32, [4]> var_4137 = const()[name = tensor<string, []>("op_4137"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4138_cast_fp16 = reshape(shape = var_4137, x = linear_292_cast_fp16)[name = tensor<string, []>("op_4138_cast_fp16")];
            tensor<int32, [4]> value_219_perm_0 = const()[name = tensor<string, []>("value_219_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_73_y_0_to_fp16 = const()[name = tensor<string, []>("mul_73_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_73_cast_fp16 = mul(x = var_4132_cast_fp16, y = mul_73_y_0_to_fp16)[name = tensor<string, []>("mul_73_cast_fp16")];
            tensor<bool, []> matmul_54_transpose_y_0 = const()[name = tensor<string, []>("matmul_54_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_54_transpose_x_0 = const()[name = tensor<string, []>("matmul_54_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_380_perm_0 = const()[name = tensor<string, []>("transpose_380_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_381_perm_0 = const()[name = tensor<string, []>("transpose_381_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_463 = transpose(perm = transpose_381_perm_0, x = var_4135_cast_fp16)[name = tensor<string, []>("transpose_463")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_464 = transpose(perm = transpose_380_perm_0, x = mul_73_cast_fp16)[name = tensor<string, []>("transpose_464")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_54_cast_fp16 = matmul(transpose_x = matmul_54_transpose_x_0, transpose_y = matmul_54_transpose_y_0, x = transpose_464, y = transpose_463)[name = tensor<string, []>("matmul_54_cast_fp16")];
            tensor<int32, []> softmax_54_axis_0 = const()[name = tensor<string, []>("softmax_54_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_54_cast_fp16 = softmax(axis = softmax_54_axis_0, x = matmul_54_cast_fp16)[name = tensor<string, []>("softmax_54_cast_fp16")];
            tensor<bool, []> hidden_states_613_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_613_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_613_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_613_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_462 = transpose(perm = value_219_perm_0, x = var_4138_cast_fp16)[name = tensor<string, []>("transpose_462")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_613_cast_fp16 = matmul(transpose_x = hidden_states_613_transpose_x_0, transpose_y = hidden_states_613_transpose_y_0, x = softmax_54_cast_fp16, y = transpose_462)[name = tensor<string, []>("hidden_states_613_cast_fp16")];
            tensor<int32, [4]> var_4141_perm_0 = const()[name = tensor<string, []>("op_4141_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4145 = const()[name = tensor<string, []>("op_4145"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_461 = transpose(perm = var_4141_perm_0, x = hidden_states_613_cast_fp16)[name = tensor<string, []>("transpose_461")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_615_cast_fp16 = reshape(shape = var_4145, x = transpose_461)[name = tensor<string, []>("hidden_states_615_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(730543808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(731772672))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(731772864)))];
            tensor<fp16, [1, 1024, 1280]> linear_293_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_3_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_615_cast_fp16)[name = tensor<string, []>("linear_293_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_717_cast_fp16 = add(x = linear_293_cast_fp16, y = hidden_states_609_cast_fp16)[name = tensor<string, []>("input_717_cast_fp16")];
            tensor<int32, [1]> input_719_axes_0 = const()[name = tensor<string, []>("input_719_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(731775488)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(731778112)))];
            tensor<fp16, [1, 1024, 1280]> input_719_cast_fp16 = layer_norm(axes = input_719_axes_0, beta = mid_block_attentions_0_transformer_blocks_3_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_3_norm2_weight_to_fp16, x = input_717_cast_fp16)[name = tensor<string, []>("input_719_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(731780736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(733009600))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_294_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_attn2_to_q_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = tensor<string, []>("linear_294_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(733009792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734975936))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_295_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_295_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734976128))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736942272))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_296_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_296_cast_fp16")];
            tensor<int32, [4]> var_4177 = const()[name = tensor<string, []>("op_4177"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4178_cast_fp16 = reshape(shape = var_4177, x = linear_294_cast_fp16)[name = tensor<string, []>("op_4178_cast_fp16")];
            tensor<int32, [4]> var_4180 = const()[name = tensor<string, []>("op_4180"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4181_cast_fp16 = reshape(shape = var_4180, x = linear_295_cast_fp16)[name = tensor<string, []>("op_4181_cast_fp16")];
            tensor<int32, [4]> var_4183 = const()[name = tensor<string, []>("op_4183"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4184_cast_fp16 = reshape(shape = var_4183, x = linear_296_cast_fp16)[name = tensor<string, []>("op_4184_cast_fp16")];
            tensor<int32, [4]> value_223_perm_0 = const()[name = tensor<string, []>("value_223_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_74_y_0_to_fp16 = const()[name = tensor<string, []>("mul_74_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_74_cast_fp16 = mul(x = var_4178_cast_fp16, y = mul_74_y_0_to_fp16)[name = tensor<string, []>("mul_74_cast_fp16")];
            tensor<bool, []> matmul_55_transpose_y_0 = const()[name = tensor<string, []>("matmul_55_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_55_transpose_x_0 = const()[name = tensor<string, []>("matmul_55_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_382_perm_0 = const()[name = tensor<string, []>("transpose_382_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_383_perm_0 = const()[name = tensor<string, []>("transpose_383_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_459 = transpose(perm = transpose_383_perm_0, x = var_4181_cast_fp16)[name = tensor<string, []>("transpose_459")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_460 = transpose(perm = transpose_382_perm_0, x = mul_74_cast_fp16)[name = tensor<string, []>("transpose_460")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_55_cast_fp16 = matmul(transpose_x = matmul_55_transpose_x_0, transpose_y = matmul_55_transpose_y_0, x = transpose_460, y = transpose_459)[name = tensor<string, []>("matmul_55_cast_fp16")];
            tensor<int32, []> softmax_55_axis_0 = const()[name = tensor<string, []>("softmax_55_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_55_cast_fp16 = softmax(axis = softmax_55_axis_0, x = matmul_55_cast_fp16)[name = tensor<string, []>("softmax_55_cast_fp16")];
            tensor<bool, []> hidden_states_619_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_619_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_619_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_619_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_458 = transpose(perm = value_223_perm_0, x = var_4184_cast_fp16)[name = tensor<string, []>("transpose_458")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_619_cast_fp16 = matmul(transpose_x = hidden_states_619_transpose_x_0, transpose_y = hidden_states_619_transpose_y_0, x = softmax_55_cast_fp16, y = transpose_458)[name = tensor<string, []>("hidden_states_619_cast_fp16")];
            tensor<int32, [4]> var_4187_perm_0 = const()[name = tensor<string, []>("op_4187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4191 = const()[name = tensor<string, []>("op_4191"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_457 = transpose(perm = var_4187_perm_0, x = hidden_states_619_cast_fp16)[name = tensor<string, []>("transpose_457")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_621_cast_fp16 = reshape(shape = var_4191, x = transpose_457)[name = tensor<string, []>("hidden_states_621_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736942464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738171328))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738171520)))];
            tensor<fp16, [1, 1024, 1280]> linear_297_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_3_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_621_cast_fp16)[name = tensor<string, []>("linear_297_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_725_cast_fp16 = add(x = linear_297_cast_fp16, y = input_717_cast_fp16)[name = tensor<string, []>("input_725_cast_fp16")];
            tensor<int32, [1]> input_727_axes_0 = const()[name = tensor<string, []>("input_727_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738174144)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738176768)))];
            tensor<fp16, [1, 1024, 1280]> input_727_cast_fp16 = layer_norm(axes = input_727_axes_0, beta = mid_block_attentions_0_transformer_blocks_3_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_3_norm3_weight_to_fp16, x = input_725_cast_fp16)[name = tensor<string, []>("input_727_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738179392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748009856))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748010048))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748017792))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_298_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_3_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_3_ff_net_0_proj_weight_to_fp16_palettized, x = input_727_cast_fp16)[name = tensor<string, []>("linear_298_cast_fp16")];
            tensor<int32, [2]> var_4213_split_sizes_0 = const()[name = tensor<string, []>("op_4213_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_4213_axis_0 = const()[name = tensor<string, []>("op_4213_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_4213_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_4213_cast_fp16_1 = split(axis = var_4213_axis_0, split_sizes = var_4213_split_sizes_0, x = linear_298_cast_fp16)[name = tensor<string, []>("op_4213_cast_fp16")];
            tensor<string, []> var_4215_mode_0 = const()[name = tensor<string, []>("op_4215_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_4215_cast_fp16 = gelu(mode = var_4215_mode_0, x = var_4213_cast_fp16_1)[name = tensor<string, []>("op_4215_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_729_cast_fp16 = mul(x = var_4213_cast_fp16_0, y = var_4215_cast_fp16)[name = tensor<string, []>("input_729_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748017984))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(752933248))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_3_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_3_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(752933440)))];
            tensor<fp16, [1, 1024, 1280]> linear_299_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_3_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_3_ff_net_2_weight_to_fp16_palettized, x = input_729_cast_fp16)[name = tensor<string, []>("linear_299_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_629_cast_fp16 = add(x = linear_299_cast_fp16, y = input_725_cast_fp16)[name = tensor<string, []>("hidden_states_629_cast_fp16")];
            tensor<int32, [1]> hidden_states_631_axes_0 = const()[name = tensor<string, []>("hidden_states_631_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(752936064)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(752938688)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_631_cast_fp16 = layer_norm(axes = hidden_states_631_axes_0, beta = mid_block_attentions_0_transformer_blocks_4_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_4_norm1_weight_to_fp16, x = hidden_states_629_cast_fp16)[name = tensor<string, []>("hidden_states_631_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(752941312))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754170176))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_300_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_631_cast_fp16)[name = tensor<string, []>("linear_300_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754170368))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(755399232))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_301_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_631_cast_fp16)[name = tensor<string, []>("linear_301_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(755399424))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756628288))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_302_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_631_cast_fp16)[name = tensor<string, []>("linear_302_cast_fp16")];
            tensor<int32, [4]> var_4250 = const()[name = tensor<string, []>("op_4250"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4251_cast_fp16 = reshape(shape = var_4250, x = linear_300_cast_fp16)[name = tensor<string, []>("op_4251_cast_fp16")];
            tensor<int32, [4]> var_4253 = const()[name = tensor<string, []>("op_4253"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4254_cast_fp16 = reshape(shape = var_4253, x = linear_301_cast_fp16)[name = tensor<string, []>("op_4254_cast_fp16")];
            tensor<int32, [4]> var_4256 = const()[name = tensor<string, []>("op_4256"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4257_cast_fp16 = reshape(shape = var_4256, x = linear_302_cast_fp16)[name = tensor<string, []>("op_4257_cast_fp16")];
            tensor<int32, [4]> value_227_perm_0 = const()[name = tensor<string, []>("value_227_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_75_y_0_to_fp16 = const()[name = tensor<string, []>("mul_75_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_75_cast_fp16 = mul(x = var_4251_cast_fp16, y = mul_75_y_0_to_fp16)[name = tensor<string, []>("mul_75_cast_fp16")];
            tensor<bool, []> matmul_56_transpose_y_0 = const()[name = tensor<string, []>("matmul_56_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_56_transpose_x_0 = const()[name = tensor<string, []>("matmul_56_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_384_perm_0 = const()[name = tensor<string, []>("transpose_384_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_385_perm_0 = const()[name = tensor<string, []>("transpose_385_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_455 = transpose(perm = transpose_385_perm_0, x = var_4254_cast_fp16)[name = tensor<string, []>("transpose_455")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_456 = transpose(perm = transpose_384_perm_0, x = mul_75_cast_fp16)[name = tensor<string, []>("transpose_456")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_56_cast_fp16 = matmul(transpose_x = matmul_56_transpose_x_0, transpose_y = matmul_56_transpose_y_0, x = transpose_456, y = transpose_455)[name = tensor<string, []>("matmul_56_cast_fp16")];
            tensor<int32, []> softmax_56_axis_0 = const()[name = tensor<string, []>("softmax_56_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_56_cast_fp16 = softmax(axis = softmax_56_axis_0, x = matmul_56_cast_fp16)[name = tensor<string, []>("softmax_56_cast_fp16")];
            tensor<bool, []> hidden_states_633_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_633_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_633_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_633_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_454 = transpose(perm = value_227_perm_0, x = var_4257_cast_fp16)[name = tensor<string, []>("transpose_454")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_633_cast_fp16 = matmul(transpose_x = hidden_states_633_transpose_x_0, transpose_y = hidden_states_633_transpose_y_0, x = softmax_56_cast_fp16, y = transpose_454)[name = tensor<string, []>("hidden_states_633_cast_fp16")];
            tensor<int32, [4]> var_4260_perm_0 = const()[name = tensor<string, []>("op_4260_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4264 = const()[name = tensor<string, []>("op_4264"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_453 = transpose(perm = var_4260_perm_0, x = hidden_states_633_cast_fp16)[name = tensor<string, []>("transpose_453")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_635_cast_fp16 = reshape(shape = var_4264, x = transpose_453)[name = tensor<string, []>("hidden_states_635_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756628480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757857344))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757857536)))];
            tensor<fp16, [1, 1024, 1280]> linear_303_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_4_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_635_cast_fp16)[name = tensor<string, []>("linear_303_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_737_cast_fp16 = add(x = linear_303_cast_fp16, y = hidden_states_629_cast_fp16)[name = tensor<string, []>("input_737_cast_fp16")];
            tensor<int32, [1]> input_739_axes_0 = const()[name = tensor<string, []>("input_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757860160)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757862784)))];
            tensor<fp16, [1, 1024, 1280]> input_739_cast_fp16 = layer_norm(axes = input_739_axes_0, beta = mid_block_attentions_0_transformer_blocks_4_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_4_norm2_weight_to_fp16, x = input_737_cast_fp16)[name = tensor<string, []>("input_739_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757865408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(759094272))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_304_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_attn2_to_q_weight_to_fp16_palettized, x = input_739_cast_fp16)[name = tensor<string, []>("linear_304_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(759094464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(761060608))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_305_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_305_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(761060800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(763026944))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_306_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_306_cast_fp16")];
            tensor<int32, [4]> var_4296 = const()[name = tensor<string, []>("op_4296"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4297_cast_fp16 = reshape(shape = var_4296, x = linear_304_cast_fp16)[name = tensor<string, []>("op_4297_cast_fp16")];
            tensor<int32, [4]> var_4299 = const()[name = tensor<string, []>("op_4299"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4300_cast_fp16 = reshape(shape = var_4299, x = linear_305_cast_fp16)[name = tensor<string, []>("op_4300_cast_fp16")];
            tensor<int32, [4]> var_4302 = const()[name = tensor<string, []>("op_4302"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4303_cast_fp16 = reshape(shape = var_4302, x = linear_306_cast_fp16)[name = tensor<string, []>("op_4303_cast_fp16")];
            tensor<int32, [4]> value_231_perm_0 = const()[name = tensor<string, []>("value_231_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_76_y_0_to_fp16 = const()[name = tensor<string, []>("mul_76_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_76_cast_fp16 = mul(x = var_4297_cast_fp16, y = mul_76_y_0_to_fp16)[name = tensor<string, []>("mul_76_cast_fp16")];
            tensor<bool, []> matmul_57_transpose_y_0 = const()[name = tensor<string, []>("matmul_57_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_57_transpose_x_0 = const()[name = tensor<string, []>("matmul_57_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_386_perm_0 = const()[name = tensor<string, []>("transpose_386_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_387_perm_0 = const()[name = tensor<string, []>("transpose_387_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_451 = transpose(perm = transpose_387_perm_0, x = var_4300_cast_fp16)[name = tensor<string, []>("transpose_451")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_452 = transpose(perm = transpose_386_perm_0, x = mul_76_cast_fp16)[name = tensor<string, []>("transpose_452")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_57_cast_fp16 = matmul(transpose_x = matmul_57_transpose_x_0, transpose_y = matmul_57_transpose_y_0, x = transpose_452, y = transpose_451)[name = tensor<string, []>("matmul_57_cast_fp16")];
            tensor<int32, []> softmax_57_axis_0 = const()[name = tensor<string, []>("softmax_57_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_57_cast_fp16 = softmax(axis = softmax_57_axis_0, x = matmul_57_cast_fp16)[name = tensor<string, []>("softmax_57_cast_fp16")];
            tensor<bool, []> hidden_states_639_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_639_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_639_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_639_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_450 = transpose(perm = value_231_perm_0, x = var_4303_cast_fp16)[name = tensor<string, []>("transpose_450")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_639_cast_fp16 = matmul(transpose_x = hidden_states_639_transpose_x_0, transpose_y = hidden_states_639_transpose_y_0, x = softmax_57_cast_fp16, y = transpose_450)[name = tensor<string, []>("hidden_states_639_cast_fp16")];
            tensor<int32, [4]> var_4306_perm_0 = const()[name = tensor<string, []>("op_4306_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4310 = const()[name = tensor<string, []>("op_4310"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_449 = transpose(perm = var_4306_perm_0, x = hidden_states_639_cast_fp16)[name = tensor<string, []>("transpose_449")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_641_cast_fp16 = reshape(shape = var_4310, x = transpose_449)[name = tensor<string, []>("hidden_states_641_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(763027136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(764256000))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(764256192)))];
            tensor<fp16, [1, 1024, 1280]> linear_307_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_4_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_641_cast_fp16)[name = tensor<string, []>("linear_307_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_745_cast_fp16 = add(x = linear_307_cast_fp16, y = input_737_cast_fp16)[name = tensor<string, []>("input_745_cast_fp16")];
            tensor<int32, [1]> input_747_axes_0 = const()[name = tensor<string, []>("input_747_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(764258816)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(764261440)))];
            tensor<fp16, [1, 1024, 1280]> input_747_cast_fp16 = layer_norm(axes = input_747_axes_0, beta = mid_block_attentions_0_transformer_blocks_4_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_4_norm3_weight_to_fp16, x = input_745_cast_fp16)[name = tensor<string, []>("input_747_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(764264064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774094528))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774094720))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774102464))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_308_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_4_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_4_ff_net_0_proj_weight_to_fp16_palettized, x = input_747_cast_fp16)[name = tensor<string, []>("linear_308_cast_fp16")];
            tensor<int32, [2]> var_4332_split_sizes_0 = const()[name = tensor<string, []>("op_4332_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_4332_axis_0 = const()[name = tensor<string, []>("op_4332_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_4332_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_4332_cast_fp16_1 = split(axis = var_4332_axis_0, split_sizes = var_4332_split_sizes_0, x = linear_308_cast_fp16)[name = tensor<string, []>("op_4332_cast_fp16")];
            tensor<string, []> var_4334_mode_0 = const()[name = tensor<string, []>("op_4334_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_4334_cast_fp16 = gelu(mode = var_4334_mode_0, x = var_4332_cast_fp16_1)[name = tensor<string, []>("op_4334_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_749_cast_fp16 = mul(x = var_4332_cast_fp16_0, y = var_4334_cast_fp16)[name = tensor<string, []>("input_749_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774102656))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779017920))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_4_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_4_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779018112)))];
            tensor<fp16, [1, 1024, 1280]> linear_309_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_4_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_4_ff_net_2_weight_to_fp16_palettized, x = input_749_cast_fp16)[name = tensor<string, []>("linear_309_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_649_cast_fp16 = add(x = linear_309_cast_fp16, y = input_745_cast_fp16)[name = tensor<string, []>("hidden_states_649_cast_fp16")];
            tensor<int32, [1]> hidden_states_651_axes_0 = const()[name = tensor<string, []>("hidden_states_651_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779020736)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779023360)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_651_cast_fp16 = layer_norm(axes = hidden_states_651_axes_0, beta = mid_block_attentions_0_transformer_blocks_5_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_5_norm1_weight_to_fp16, x = hidden_states_649_cast_fp16)[name = tensor<string, []>("hidden_states_651_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779025984))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(780254848))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_310_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_651_cast_fp16)[name = tensor<string, []>("linear_310_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(780255040))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781483904))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_311_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_651_cast_fp16)[name = tensor<string, []>("linear_311_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781484096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(782712960))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_312_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_651_cast_fp16)[name = tensor<string, []>("linear_312_cast_fp16")];
            tensor<int32, [4]> var_4369 = const()[name = tensor<string, []>("op_4369"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4370_cast_fp16 = reshape(shape = var_4369, x = linear_310_cast_fp16)[name = tensor<string, []>("op_4370_cast_fp16")];
            tensor<int32, [4]> var_4372 = const()[name = tensor<string, []>("op_4372"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4373_cast_fp16 = reshape(shape = var_4372, x = linear_311_cast_fp16)[name = tensor<string, []>("op_4373_cast_fp16")];
            tensor<int32, [4]> var_4375 = const()[name = tensor<string, []>("op_4375"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4376_cast_fp16 = reshape(shape = var_4375, x = linear_312_cast_fp16)[name = tensor<string, []>("op_4376_cast_fp16")];
            tensor<int32, [4]> value_235_perm_0 = const()[name = tensor<string, []>("value_235_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_77_y_0_to_fp16 = const()[name = tensor<string, []>("mul_77_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_77_cast_fp16 = mul(x = var_4370_cast_fp16, y = mul_77_y_0_to_fp16)[name = tensor<string, []>("mul_77_cast_fp16")];
            tensor<bool, []> matmul_58_transpose_y_0 = const()[name = tensor<string, []>("matmul_58_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_58_transpose_x_0 = const()[name = tensor<string, []>("matmul_58_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_388_perm_0 = const()[name = tensor<string, []>("transpose_388_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_389_perm_0 = const()[name = tensor<string, []>("transpose_389_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_447 = transpose(perm = transpose_389_perm_0, x = var_4373_cast_fp16)[name = tensor<string, []>("transpose_447")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_448 = transpose(perm = transpose_388_perm_0, x = mul_77_cast_fp16)[name = tensor<string, []>("transpose_448")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_58_cast_fp16 = matmul(transpose_x = matmul_58_transpose_x_0, transpose_y = matmul_58_transpose_y_0, x = transpose_448, y = transpose_447)[name = tensor<string, []>("matmul_58_cast_fp16")];
            tensor<int32, []> softmax_58_axis_0 = const()[name = tensor<string, []>("softmax_58_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_58_cast_fp16 = softmax(axis = softmax_58_axis_0, x = matmul_58_cast_fp16)[name = tensor<string, []>("softmax_58_cast_fp16")];
            tensor<bool, []> hidden_states_653_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_653_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_653_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_653_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_446 = transpose(perm = value_235_perm_0, x = var_4376_cast_fp16)[name = tensor<string, []>("transpose_446")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_653_cast_fp16 = matmul(transpose_x = hidden_states_653_transpose_x_0, transpose_y = hidden_states_653_transpose_y_0, x = softmax_58_cast_fp16, y = transpose_446)[name = tensor<string, []>("hidden_states_653_cast_fp16")];
            tensor<int32, [4]> var_4379_perm_0 = const()[name = tensor<string, []>("op_4379_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4383 = const()[name = tensor<string, []>("op_4383"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_445 = transpose(perm = var_4379_perm_0, x = hidden_states_653_cast_fp16)[name = tensor<string, []>("transpose_445")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_655_cast_fp16 = reshape(shape = var_4383, x = transpose_445)[name = tensor<string, []>("hidden_states_655_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(782713152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(783942016))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(783942208)))];
            tensor<fp16, [1, 1024, 1280]> linear_313_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_5_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_655_cast_fp16)[name = tensor<string, []>("linear_313_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_757_cast_fp16 = add(x = linear_313_cast_fp16, y = hidden_states_649_cast_fp16)[name = tensor<string, []>("input_757_cast_fp16")];
            tensor<int32, [1]> input_759_axes_0 = const()[name = tensor<string, []>("input_759_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(783944832)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(783947456)))];
            tensor<fp16, [1, 1024, 1280]> input_759_cast_fp16 = layer_norm(axes = input_759_axes_0, beta = mid_block_attentions_0_transformer_blocks_5_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_5_norm2_weight_to_fp16, x = input_757_cast_fp16)[name = tensor<string, []>("input_759_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(783950080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785178944))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_314_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_attn2_to_q_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = tensor<string, []>("linear_314_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785179136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(787145280))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_315_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_315_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(787145472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(789111616))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_316_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_316_cast_fp16")];
            tensor<int32, [4]> var_4415 = const()[name = tensor<string, []>("op_4415"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4416_cast_fp16 = reshape(shape = var_4415, x = linear_314_cast_fp16)[name = tensor<string, []>("op_4416_cast_fp16")];
            tensor<int32, [4]> var_4418 = const()[name = tensor<string, []>("op_4418"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4419_cast_fp16 = reshape(shape = var_4418, x = linear_315_cast_fp16)[name = tensor<string, []>("op_4419_cast_fp16")];
            tensor<int32, [4]> var_4421 = const()[name = tensor<string, []>("op_4421"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4422_cast_fp16 = reshape(shape = var_4421, x = linear_316_cast_fp16)[name = tensor<string, []>("op_4422_cast_fp16")];
            tensor<int32, [4]> value_239_perm_0 = const()[name = tensor<string, []>("value_239_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_78_y_0_to_fp16 = const()[name = tensor<string, []>("mul_78_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_78_cast_fp16 = mul(x = var_4416_cast_fp16, y = mul_78_y_0_to_fp16)[name = tensor<string, []>("mul_78_cast_fp16")];
            tensor<bool, []> matmul_59_transpose_y_0 = const()[name = tensor<string, []>("matmul_59_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_59_transpose_x_0 = const()[name = tensor<string, []>("matmul_59_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_390_perm_0 = const()[name = tensor<string, []>("transpose_390_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_391_perm_0 = const()[name = tensor<string, []>("transpose_391_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_443 = transpose(perm = transpose_391_perm_0, x = var_4419_cast_fp16)[name = tensor<string, []>("transpose_443")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_444 = transpose(perm = transpose_390_perm_0, x = mul_78_cast_fp16)[name = tensor<string, []>("transpose_444")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_59_cast_fp16 = matmul(transpose_x = matmul_59_transpose_x_0, transpose_y = matmul_59_transpose_y_0, x = transpose_444, y = transpose_443)[name = tensor<string, []>("matmul_59_cast_fp16")];
            tensor<int32, []> softmax_59_axis_0 = const()[name = tensor<string, []>("softmax_59_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_59_cast_fp16 = softmax(axis = softmax_59_axis_0, x = matmul_59_cast_fp16)[name = tensor<string, []>("softmax_59_cast_fp16")];
            tensor<bool, []> hidden_states_659_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_659_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_659_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_659_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_442 = transpose(perm = value_239_perm_0, x = var_4422_cast_fp16)[name = tensor<string, []>("transpose_442")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_659_cast_fp16 = matmul(transpose_x = hidden_states_659_transpose_x_0, transpose_y = hidden_states_659_transpose_y_0, x = softmax_59_cast_fp16, y = transpose_442)[name = tensor<string, []>("hidden_states_659_cast_fp16")];
            tensor<int32, [4]> var_4425_perm_0 = const()[name = tensor<string, []>("op_4425_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4429 = const()[name = tensor<string, []>("op_4429"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_441 = transpose(perm = var_4425_perm_0, x = hidden_states_659_cast_fp16)[name = tensor<string, []>("transpose_441")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_661_cast_fp16 = reshape(shape = var_4429, x = transpose_441)[name = tensor<string, []>("hidden_states_661_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(789111808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(790340672))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(790340864)))];
            tensor<fp16, [1, 1024, 1280]> linear_317_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_5_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_661_cast_fp16)[name = tensor<string, []>("linear_317_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_765_cast_fp16 = add(x = linear_317_cast_fp16, y = input_757_cast_fp16)[name = tensor<string, []>("input_765_cast_fp16")];
            tensor<int32, [1]> input_767_axes_0 = const()[name = tensor<string, []>("input_767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(790343488)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(790346112)))];
            tensor<fp16, [1, 1024, 1280]> input_767_cast_fp16 = layer_norm(axes = input_767_axes_0, beta = mid_block_attentions_0_transformer_blocks_5_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_5_norm3_weight_to_fp16, x = input_765_cast_fp16)[name = tensor<string, []>("input_767_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(790348736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800179200))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800179392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800187136))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_318_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_5_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_5_ff_net_0_proj_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = tensor<string, []>("linear_318_cast_fp16")];
            tensor<int32, [2]> var_4451_split_sizes_0 = const()[name = tensor<string, []>("op_4451_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_4451_axis_0 = const()[name = tensor<string, []>("op_4451_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_4451_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_4451_cast_fp16_1 = split(axis = var_4451_axis_0, split_sizes = var_4451_split_sizes_0, x = linear_318_cast_fp16)[name = tensor<string, []>("op_4451_cast_fp16")];
            tensor<string, []> var_4453_mode_0 = const()[name = tensor<string, []>("op_4453_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_4453_cast_fp16 = gelu(mode = var_4453_mode_0, x = var_4451_cast_fp16_1)[name = tensor<string, []>("op_4453_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_769_cast_fp16 = mul(x = var_4451_cast_fp16_0, y = var_4453_cast_fp16)[name = tensor<string, []>("input_769_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800187328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805102592))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_5_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_5_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805102784)))];
            tensor<fp16, [1, 1024, 1280]> linear_319_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_5_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_5_ff_net_2_weight_to_fp16_palettized, x = input_769_cast_fp16)[name = tensor<string, []>("linear_319_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_669_cast_fp16 = add(x = linear_319_cast_fp16, y = input_765_cast_fp16)[name = tensor<string, []>("hidden_states_669_cast_fp16")];
            tensor<int32, [1]> hidden_states_671_axes_0 = const()[name = tensor<string, []>("hidden_states_671_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805105408)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805108032)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_671_cast_fp16 = layer_norm(axes = hidden_states_671_axes_0, beta = mid_block_attentions_0_transformer_blocks_6_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_6_norm1_weight_to_fp16, x = hidden_states_669_cast_fp16)[name = tensor<string, []>("hidden_states_671_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805110656))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(806339520))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_320_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_671_cast_fp16)[name = tensor<string, []>("linear_320_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(806339712))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(807568576))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_321_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_671_cast_fp16)[name = tensor<string, []>("linear_321_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(807568768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(808797632))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_322_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_671_cast_fp16)[name = tensor<string, []>("linear_322_cast_fp16")];
            tensor<int32, [4]> var_4488 = const()[name = tensor<string, []>("op_4488"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4489_cast_fp16 = reshape(shape = var_4488, x = linear_320_cast_fp16)[name = tensor<string, []>("op_4489_cast_fp16")];
            tensor<int32, [4]> var_4491 = const()[name = tensor<string, []>("op_4491"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4492_cast_fp16 = reshape(shape = var_4491, x = linear_321_cast_fp16)[name = tensor<string, []>("op_4492_cast_fp16")];
            tensor<int32, [4]> var_4494 = const()[name = tensor<string, []>("op_4494"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4495_cast_fp16 = reshape(shape = var_4494, x = linear_322_cast_fp16)[name = tensor<string, []>("op_4495_cast_fp16")];
            tensor<int32, [4]> value_243_perm_0 = const()[name = tensor<string, []>("value_243_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_79_y_0_to_fp16 = const()[name = tensor<string, []>("mul_79_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_79_cast_fp16 = mul(x = var_4489_cast_fp16, y = mul_79_y_0_to_fp16)[name = tensor<string, []>("mul_79_cast_fp16")];
            tensor<bool, []> matmul_60_transpose_y_0 = const()[name = tensor<string, []>("matmul_60_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_60_transpose_x_0 = const()[name = tensor<string, []>("matmul_60_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_392_perm_0 = const()[name = tensor<string, []>("transpose_392_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_393_perm_0 = const()[name = tensor<string, []>("transpose_393_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_439 = transpose(perm = transpose_393_perm_0, x = var_4492_cast_fp16)[name = tensor<string, []>("transpose_439")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_440 = transpose(perm = transpose_392_perm_0, x = mul_79_cast_fp16)[name = tensor<string, []>("transpose_440")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_60_cast_fp16 = matmul(transpose_x = matmul_60_transpose_x_0, transpose_y = matmul_60_transpose_y_0, x = transpose_440, y = transpose_439)[name = tensor<string, []>("matmul_60_cast_fp16")];
            tensor<int32, []> softmax_60_axis_0 = const()[name = tensor<string, []>("softmax_60_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_60_cast_fp16 = softmax(axis = softmax_60_axis_0, x = matmul_60_cast_fp16)[name = tensor<string, []>("softmax_60_cast_fp16")];
            tensor<bool, []> hidden_states_673_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_673_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_673_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_673_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_438 = transpose(perm = value_243_perm_0, x = var_4495_cast_fp16)[name = tensor<string, []>("transpose_438")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_673_cast_fp16 = matmul(transpose_x = hidden_states_673_transpose_x_0, transpose_y = hidden_states_673_transpose_y_0, x = softmax_60_cast_fp16, y = transpose_438)[name = tensor<string, []>("hidden_states_673_cast_fp16")];
            tensor<int32, [4]> var_4498_perm_0 = const()[name = tensor<string, []>("op_4498_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4502 = const()[name = tensor<string, []>("op_4502"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_437 = transpose(perm = var_4498_perm_0, x = hidden_states_673_cast_fp16)[name = tensor<string, []>("transpose_437")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_675_cast_fp16 = reshape(shape = var_4502, x = transpose_437)[name = tensor<string, []>("hidden_states_675_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(808797824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(810026688))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(810026880)))];
            tensor<fp16, [1, 1024, 1280]> linear_323_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_6_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_675_cast_fp16)[name = tensor<string, []>("linear_323_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_777_cast_fp16 = add(x = linear_323_cast_fp16, y = hidden_states_669_cast_fp16)[name = tensor<string, []>("input_777_cast_fp16")];
            tensor<int32, [1]> input_779_axes_0 = const()[name = tensor<string, []>("input_779_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(810029504)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(810032128)))];
            tensor<fp16, [1, 1024, 1280]> input_779_cast_fp16 = layer_norm(axes = input_779_axes_0, beta = mid_block_attentions_0_transformer_blocks_6_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_6_norm2_weight_to_fp16, x = input_777_cast_fp16)[name = tensor<string, []>("input_779_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(810034752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(811263616))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_324_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_attn2_to_q_weight_to_fp16_palettized, x = input_779_cast_fp16)[name = tensor<string, []>("linear_324_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(811263808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(813229952))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_325_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_325_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(813230144))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815196288))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_326_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_326_cast_fp16")];
            tensor<int32, [4]> var_4534 = const()[name = tensor<string, []>("op_4534"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4535_cast_fp16 = reshape(shape = var_4534, x = linear_324_cast_fp16)[name = tensor<string, []>("op_4535_cast_fp16")];
            tensor<int32, [4]> var_4537 = const()[name = tensor<string, []>("op_4537"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4538_cast_fp16 = reshape(shape = var_4537, x = linear_325_cast_fp16)[name = tensor<string, []>("op_4538_cast_fp16")];
            tensor<int32, [4]> var_4540 = const()[name = tensor<string, []>("op_4540"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4541_cast_fp16 = reshape(shape = var_4540, x = linear_326_cast_fp16)[name = tensor<string, []>("op_4541_cast_fp16")];
            tensor<int32, [4]> value_247_perm_0 = const()[name = tensor<string, []>("value_247_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_80_y_0_to_fp16 = const()[name = tensor<string, []>("mul_80_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_80_cast_fp16 = mul(x = var_4535_cast_fp16, y = mul_80_y_0_to_fp16)[name = tensor<string, []>("mul_80_cast_fp16")];
            tensor<bool, []> matmul_61_transpose_y_0 = const()[name = tensor<string, []>("matmul_61_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_61_transpose_x_0 = const()[name = tensor<string, []>("matmul_61_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_394_perm_0 = const()[name = tensor<string, []>("transpose_394_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_395_perm_0 = const()[name = tensor<string, []>("transpose_395_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_435 = transpose(perm = transpose_395_perm_0, x = var_4538_cast_fp16)[name = tensor<string, []>("transpose_435")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_436 = transpose(perm = transpose_394_perm_0, x = mul_80_cast_fp16)[name = tensor<string, []>("transpose_436")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_61_cast_fp16 = matmul(transpose_x = matmul_61_transpose_x_0, transpose_y = matmul_61_transpose_y_0, x = transpose_436, y = transpose_435)[name = tensor<string, []>("matmul_61_cast_fp16")];
            tensor<int32, []> softmax_61_axis_0 = const()[name = tensor<string, []>("softmax_61_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_61_cast_fp16 = softmax(axis = softmax_61_axis_0, x = matmul_61_cast_fp16)[name = tensor<string, []>("softmax_61_cast_fp16")];
            tensor<bool, []> hidden_states_679_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_679_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_679_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_679_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_434 = transpose(perm = value_247_perm_0, x = var_4541_cast_fp16)[name = tensor<string, []>("transpose_434")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_679_cast_fp16 = matmul(transpose_x = hidden_states_679_transpose_x_0, transpose_y = hidden_states_679_transpose_y_0, x = softmax_61_cast_fp16, y = transpose_434)[name = tensor<string, []>("hidden_states_679_cast_fp16")];
            tensor<int32, [4]> var_4544_perm_0 = const()[name = tensor<string, []>("op_4544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4548 = const()[name = tensor<string, []>("op_4548"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_433 = transpose(perm = var_4544_perm_0, x = hidden_states_679_cast_fp16)[name = tensor<string, []>("transpose_433")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_681_cast_fp16 = reshape(shape = var_4548, x = transpose_433)[name = tensor<string, []>("hidden_states_681_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815196480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(816425344))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(816425536)))];
            tensor<fp16, [1, 1024, 1280]> linear_327_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_6_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_681_cast_fp16)[name = tensor<string, []>("linear_327_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_785_cast_fp16 = add(x = linear_327_cast_fp16, y = input_777_cast_fp16)[name = tensor<string, []>("input_785_cast_fp16")];
            tensor<int32, [1]> input_787_axes_0 = const()[name = tensor<string, []>("input_787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(816428160)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(816430784)))];
            tensor<fp16, [1, 1024, 1280]> input_787_cast_fp16 = layer_norm(axes = input_787_axes_0, beta = mid_block_attentions_0_transformer_blocks_6_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_6_norm3_weight_to_fp16, x = input_785_cast_fp16)[name = tensor<string, []>("input_787_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(816433408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(826263872))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(826264064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(826271808))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_328_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_6_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_6_ff_net_0_proj_weight_to_fp16_palettized, x = input_787_cast_fp16)[name = tensor<string, []>("linear_328_cast_fp16")];
            tensor<int32, [2]> var_4570_split_sizes_0 = const()[name = tensor<string, []>("op_4570_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_4570_axis_0 = const()[name = tensor<string, []>("op_4570_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_4570_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_4570_cast_fp16_1 = split(axis = var_4570_axis_0, split_sizes = var_4570_split_sizes_0, x = linear_328_cast_fp16)[name = tensor<string, []>("op_4570_cast_fp16")];
            tensor<string, []> var_4572_mode_0 = const()[name = tensor<string, []>("op_4572_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_4572_cast_fp16 = gelu(mode = var_4572_mode_0, x = var_4570_cast_fp16_1)[name = tensor<string, []>("op_4572_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_789_cast_fp16 = mul(x = var_4570_cast_fp16_0, y = var_4572_cast_fp16)[name = tensor<string, []>("input_789_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(826272000))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(831187264))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_6_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_6_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(831187456)))];
            tensor<fp16, [1, 1024, 1280]> linear_329_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_6_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_6_ff_net_2_weight_to_fp16_palettized, x = input_789_cast_fp16)[name = tensor<string, []>("linear_329_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_689_cast_fp16 = add(x = linear_329_cast_fp16, y = input_785_cast_fp16)[name = tensor<string, []>("hidden_states_689_cast_fp16")];
            tensor<int32, [1]> hidden_states_691_axes_0 = const()[name = tensor<string, []>("hidden_states_691_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(831190080)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(831192704)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_691_cast_fp16 = layer_norm(axes = hidden_states_691_axes_0, beta = mid_block_attentions_0_transformer_blocks_7_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_7_norm1_weight_to_fp16, x = hidden_states_689_cast_fp16)[name = tensor<string, []>("hidden_states_691_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(831195328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832424192))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_330_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_691_cast_fp16)[name = tensor<string, []>("linear_330_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832424384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(833653248))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_331_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_691_cast_fp16)[name = tensor<string, []>("linear_331_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(833653440))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(834882304))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_332_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_691_cast_fp16)[name = tensor<string, []>("linear_332_cast_fp16")];
            tensor<int32, [4]> var_4607 = const()[name = tensor<string, []>("op_4607"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4608_cast_fp16 = reshape(shape = var_4607, x = linear_330_cast_fp16)[name = tensor<string, []>("op_4608_cast_fp16")];
            tensor<int32, [4]> var_4610 = const()[name = tensor<string, []>("op_4610"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4611_cast_fp16 = reshape(shape = var_4610, x = linear_331_cast_fp16)[name = tensor<string, []>("op_4611_cast_fp16")];
            tensor<int32, [4]> var_4613 = const()[name = tensor<string, []>("op_4613"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4614_cast_fp16 = reshape(shape = var_4613, x = linear_332_cast_fp16)[name = tensor<string, []>("op_4614_cast_fp16")];
            tensor<int32, [4]> value_251_perm_0 = const()[name = tensor<string, []>("value_251_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_81_y_0_to_fp16 = const()[name = tensor<string, []>("mul_81_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_81_cast_fp16 = mul(x = var_4608_cast_fp16, y = mul_81_y_0_to_fp16)[name = tensor<string, []>("mul_81_cast_fp16")];
            tensor<bool, []> matmul_62_transpose_y_0 = const()[name = tensor<string, []>("matmul_62_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_62_transpose_x_0 = const()[name = tensor<string, []>("matmul_62_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_396_perm_0 = const()[name = tensor<string, []>("transpose_396_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_397_perm_0 = const()[name = tensor<string, []>("transpose_397_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_431 = transpose(perm = transpose_397_perm_0, x = var_4611_cast_fp16)[name = tensor<string, []>("transpose_431")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_432 = transpose(perm = transpose_396_perm_0, x = mul_81_cast_fp16)[name = tensor<string, []>("transpose_432")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_62_cast_fp16 = matmul(transpose_x = matmul_62_transpose_x_0, transpose_y = matmul_62_transpose_y_0, x = transpose_432, y = transpose_431)[name = tensor<string, []>("matmul_62_cast_fp16")];
            tensor<int32, []> softmax_62_axis_0 = const()[name = tensor<string, []>("softmax_62_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_62_cast_fp16 = softmax(axis = softmax_62_axis_0, x = matmul_62_cast_fp16)[name = tensor<string, []>("softmax_62_cast_fp16")];
            tensor<bool, []> hidden_states_693_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_693_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_693_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_693_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_430 = transpose(perm = value_251_perm_0, x = var_4614_cast_fp16)[name = tensor<string, []>("transpose_430")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_693_cast_fp16 = matmul(transpose_x = hidden_states_693_transpose_x_0, transpose_y = hidden_states_693_transpose_y_0, x = softmax_62_cast_fp16, y = transpose_430)[name = tensor<string, []>("hidden_states_693_cast_fp16")];
            tensor<int32, [4]> var_4617_perm_0 = const()[name = tensor<string, []>("op_4617_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4621 = const()[name = tensor<string, []>("op_4621"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_429 = transpose(perm = var_4617_perm_0, x = hidden_states_693_cast_fp16)[name = tensor<string, []>("transpose_429")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_695_cast_fp16 = reshape(shape = var_4621, x = transpose_429)[name = tensor<string, []>("hidden_states_695_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(834882496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(836111360))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(836111552)))];
            tensor<fp16, [1, 1024, 1280]> linear_333_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_7_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_695_cast_fp16)[name = tensor<string, []>("linear_333_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_797_cast_fp16 = add(x = linear_333_cast_fp16, y = hidden_states_689_cast_fp16)[name = tensor<string, []>("input_797_cast_fp16")];
            tensor<int32, [1]> input_799_axes_0 = const()[name = tensor<string, []>("input_799_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(836114176)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(836116800)))];
            tensor<fp16, [1, 1024, 1280]> input_799_cast_fp16 = layer_norm(axes = input_799_axes_0, beta = mid_block_attentions_0_transformer_blocks_7_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_7_norm2_weight_to_fp16, x = input_797_cast_fp16)[name = tensor<string, []>("input_799_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(836119424))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(837348288))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_334_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_attn2_to_q_weight_to_fp16_palettized, x = input_799_cast_fp16)[name = tensor<string, []>("linear_334_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(837348480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(839314624))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_335_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_335_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(839314816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841280960))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_336_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_336_cast_fp16")];
            tensor<int32, [4]> var_4653 = const()[name = tensor<string, []>("op_4653"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4654_cast_fp16 = reshape(shape = var_4653, x = linear_334_cast_fp16)[name = tensor<string, []>("op_4654_cast_fp16")];
            tensor<int32, [4]> var_4656 = const()[name = tensor<string, []>("op_4656"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4657_cast_fp16 = reshape(shape = var_4656, x = linear_335_cast_fp16)[name = tensor<string, []>("op_4657_cast_fp16")];
            tensor<int32, [4]> var_4659 = const()[name = tensor<string, []>("op_4659"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4660_cast_fp16 = reshape(shape = var_4659, x = linear_336_cast_fp16)[name = tensor<string, []>("op_4660_cast_fp16")];
            tensor<int32, [4]> value_255_perm_0 = const()[name = tensor<string, []>("value_255_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_82_y_0_to_fp16 = const()[name = tensor<string, []>("mul_82_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_82_cast_fp16 = mul(x = var_4654_cast_fp16, y = mul_82_y_0_to_fp16)[name = tensor<string, []>("mul_82_cast_fp16")];
            tensor<bool, []> matmul_63_transpose_y_0 = const()[name = tensor<string, []>("matmul_63_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_63_transpose_x_0 = const()[name = tensor<string, []>("matmul_63_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_398_perm_0 = const()[name = tensor<string, []>("transpose_398_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_399_perm_0 = const()[name = tensor<string, []>("transpose_399_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_427 = transpose(perm = transpose_399_perm_0, x = var_4657_cast_fp16)[name = tensor<string, []>("transpose_427")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_428 = transpose(perm = transpose_398_perm_0, x = mul_82_cast_fp16)[name = tensor<string, []>("transpose_428")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_63_cast_fp16 = matmul(transpose_x = matmul_63_transpose_x_0, transpose_y = matmul_63_transpose_y_0, x = transpose_428, y = transpose_427)[name = tensor<string, []>("matmul_63_cast_fp16")];
            tensor<int32, []> softmax_63_axis_0 = const()[name = tensor<string, []>("softmax_63_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_63_cast_fp16 = softmax(axis = softmax_63_axis_0, x = matmul_63_cast_fp16)[name = tensor<string, []>("softmax_63_cast_fp16")];
            tensor<bool, []> hidden_states_699_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_699_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_699_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_699_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_426 = transpose(perm = value_255_perm_0, x = var_4660_cast_fp16)[name = tensor<string, []>("transpose_426")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_699_cast_fp16 = matmul(transpose_x = hidden_states_699_transpose_x_0, transpose_y = hidden_states_699_transpose_y_0, x = softmax_63_cast_fp16, y = transpose_426)[name = tensor<string, []>("hidden_states_699_cast_fp16")];
            tensor<int32, [4]> var_4663_perm_0 = const()[name = tensor<string, []>("op_4663_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4667 = const()[name = tensor<string, []>("op_4667"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_425 = transpose(perm = var_4663_perm_0, x = hidden_states_699_cast_fp16)[name = tensor<string, []>("transpose_425")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_701_cast_fp16 = reshape(shape = var_4667, x = transpose_425)[name = tensor<string, []>("hidden_states_701_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841281152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842510016))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842510208)))];
            tensor<fp16, [1, 1024, 1280]> linear_337_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_7_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_701_cast_fp16)[name = tensor<string, []>("linear_337_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_805_cast_fp16 = add(x = linear_337_cast_fp16, y = input_797_cast_fp16)[name = tensor<string, []>("input_805_cast_fp16")];
            tensor<int32, [1]> input_807_axes_0 = const()[name = tensor<string, []>("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842512832)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842515456)))];
            tensor<fp16, [1, 1024, 1280]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = mid_block_attentions_0_transformer_blocks_7_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_7_norm3_weight_to_fp16, x = input_805_cast_fp16)[name = tensor<string, []>("input_807_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842518080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(852348544))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(852348736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(852356480))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_338_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_7_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_7_ff_net_0_proj_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = tensor<string, []>("linear_338_cast_fp16")];
            tensor<int32, [2]> var_4689_split_sizes_0 = const()[name = tensor<string, []>("op_4689_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_4689_axis_0 = const()[name = tensor<string, []>("op_4689_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_4689_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_4689_cast_fp16_1 = split(axis = var_4689_axis_0, split_sizes = var_4689_split_sizes_0, x = linear_338_cast_fp16)[name = tensor<string, []>("op_4689_cast_fp16")];
            tensor<string, []> var_4691_mode_0 = const()[name = tensor<string, []>("op_4691_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_4691_cast_fp16 = gelu(mode = var_4691_mode_0, x = var_4689_cast_fp16_1)[name = tensor<string, []>("op_4691_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_809_cast_fp16 = mul(x = var_4689_cast_fp16_0, y = var_4691_cast_fp16)[name = tensor<string, []>("input_809_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(852356672))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(857271936))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_7_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_7_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(857272128)))];
            tensor<fp16, [1, 1024, 1280]> linear_339_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_7_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_7_ff_net_2_weight_to_fp16_palettized, x = input_809_cast_fp16)[name = tensor<string, []>("linear_339_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_709_cast_fp16 = add(x = linear_339_cast_fp16, y = input_805_cast_fp16)[name = tensor<string, []>("hidden_states_709_cast_fp16")];
            tensor<int32, [1]> hidden_states_711_axes_0 = const()[name = tensor<string, []>("hidden_states_711_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(857274752)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(857277376)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_711_cast_fp16 = layer_norm(axes = hidden_states_711_axes_0, beta = mid_block_attentions_0_transformer_blocks_8_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_8_norm1_weight_to_fp16, x = hidden_states_709_cast_fp16)[name = tensor<string, []>("hidden_states_711_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(857280000))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(858508864))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_340_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_711_cast_fp16)[name = tensor<string, []>("linear_340_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(858509056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(859737920))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_341_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_711_cast_fp16)[name = tensor<string, []>("linear_341_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(859738112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(860966976))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_342_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_711_cast_fp16)[name = tensor<string, []>("linear_342_cast_fp16")];
            tensor<int32, [4]> var_4726 = const()[name = tensor<string, []>("op_4726"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4727_cast_fp16 = reshape(shape = var_4726, x = linear_340_cast_fp16)[name = tensor<string, []>("op_4727_cast_fp16")];
            tensor<int32, [4]> var_4729 = const()[name = tensor<string, []>("op_4729"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4730_cast_fp16 = reshape(shape = var_4729, x = linear_341_cast_fp16)[name = tensor<string, []>("op_4730_cast_fp16")];
            tensor<int32, [4]> var_4732 = const()[name = tensor<string, []>("op_4732"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4733_cast_fp16 = reshape(shape = var_4732, x = linear_342_cast_fp16)[name = tensor<string, []>("op_4733_cast_fp16")];
            tensor<int32, [4]> value_259_perm_0 = const()[name = tensor<string, []>("value_259_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_83_y_0_to_fp16 = const()[name = tensor<string, []>("mul_83_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_83_cast_fp16 = mul(x = var_4727_cast_fp16, y = mul_83_y_0_to_fp16)[name = tensor<string, []>("mul_83_cast_fp16")];
            tensor<bool, []> matmul_64_transpose_y_0 = const()[name = tensor<string, []>("matmul_64_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_64_transpose_x_0 = const()[name = tensor<string, []>("matmul_64_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_400_perm_0 = const()[name = tensor<string, []>("transpose_400_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_401_perm_0 = const()[name = tensor<string, []>("transpose_401_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_423 = transpose(perm = transpose_401_perm_0, x = var_4730_cast_fp16)[name = tensor<string, []>("transpose_423")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_424 = transpose(perm = transpose_400_perm_0, x = mul_83_cast_fp16)[name = tensor<string, []>("transpose_424")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_64_cast_fp16 = matmul(transpose_x = matmul_64_transpose_x_0, transpose_y = matmul_64_transpose_y_0, x = transpose_424, y = transpose_423)[name = tensor<string, []>("matmul_64_cast_fp16")];
            tensor<int32, []> softmax_64_axis_0 = const()[name = tensor<string, []>("softmax_64_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_64_cast_fp16 = softmax(axis = softmax_64_axis_0, x = matmul_64_cast_fp16)[name = tensor<string, []>("softmax_64_cast_fp16")];
            tensor<bool, []> hidden_states_713_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_713_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_713_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_713_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_422 = transpose(perm = value_259_perm_0, x = var_4733_cast_fp16)[name = tensor<string, []>("transpose_422")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_713_cast_fp16 = matmul(transpose_x = hidden_states_713_transpose_x_0, transpose_y = hidden_states_713_transpose_y_0, x = softmax_64_cast_fp16, y = transpose_422)[name = tensor<string, []>("hidden_states_713_cast_fp16")];
            tensor<int32, [4]> var_4736_perm_0 = const()[name = tensor<string, []>("op_4736_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4740 = const()[name = tensor<string, []>("op_4740"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_421 = transpose(perm = var_4736_perm_0, x = hidden_states_713_cast_fp16)[name = tensor<string, []>("transpose_421")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_715_cast_fp16 = reshape(shape = var_4740, x = transpose_421)[name = tensor<string, []>("hidden_states_715_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(860967168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862196032))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862196224)))];
            tensor<fp16, [1, 1024, 1280]> linear_343_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_8_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_715_cast_fp16)[name = tensor<string, []>("linear_343_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_817_cast_fp16 = add(x = linear_343_cast_fp16, y = hidden_states_709_cast_fp16)[name = tensor<string, []>("input_817_cast_fp16")];
            tensor<int32, [1]> input_819_axes_0 = const()[name = tensor<string, []>("input_819_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862198848)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862201472)))];
            tensor<fp16, [1, 1024, 1280]> input_819_cast_fp16 = layer_norm(axes = input_819_axes_0, beta = mid_block_attentions_0_transformer_blocks_8_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_8_norm2_weight_to_fp16, x = input_817_cast_fp16)[name = tensor<string, []>("input_819_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862204096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(863432960))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_344_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_attn2_to_q_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = tensor<string, []>("linear_344_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(863433152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865399296))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_345_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_345_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865399488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(867365632))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_346_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_346_cast_fp16")];
            tensor<int32, [4]> var_4772 = const()[name = tensor<string, []>("op_4772"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4773_cast_fp16 = reshape(shape = var_4772, x = linear_344_cast_fp16)[name = tensor<string, []>("op_4773_cast_fp16")];
            tensor<int32, [4]> var_4775 = const()[name = tensor<string, []>("op_4775"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4776_cast_fp16 = reshape(shape = var_4775, x = linear_345_cast_fp16)[name = tensor<string, []>("op_4776_cast_fp16")];
            tensor<int32, [4]> var_4778 = const()[name = tensor<string, []>("op_4778"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4779_cast_fp16 = reshape(shape = var_4778, x = linear_346_cast_fp16)[name = tensor<string, []>("op_4779_cast_fp16")];
            tensor<int32, [4]> value_263_perm_0 = const()[name = tensor<string, []>("value_263_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_84_y_0_to_fp16 = const()[name = tensor<string, []>("mul_84_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_84_cast_fp16 = mul(x = var_4773_cast_fp16, y = mul_84_y_0_to_fp16)[name = tensor<string, []>("mul_84_cast_fp16")];
            tensor<bool, []> matmul_65_transpose_y_0 = const()[name = tensor<string, []>("matmul_65_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_65_transpose_x_0 = const()[name = tensor<string, []>("matmul_65_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_402_perm_0 = const()[name = tensor<string, []>("transpose_402_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_403_perm_0 = const()[name = tensor<string, []>("transpose_403_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_419 = transpose(perm = transpose_403_perm_0, x = var_4776_cast_fp16)[name = tensor<string, []>("transpose_419")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_420 = transpose(perm = transpose_402_perm_0, x = mul_84_cast_fp16)[name = tensor<string, []>("transpose_420")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_65_cast_fp16 = matmul(transpose_x = matmul_65_transpose_x_0, transpose_y = matmul_65_transpose_y_0, x = transpose_420, y = transpose_419)[name = tensor<string, []>("matmul_65_cast_fp16")];
            tensor<int32, []> softmax_65_axis_0 = const()[name = tensor<string, []>("softmax_65_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_65_cast_fp16 = softmax(axis = softmax_65_axis_0, x = matmul_65_cast_fp16)[name = tensor<string, []>("softmax_65_cast_fp16")];
            tensor<bool, []> hidden_states_719_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_719_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_719_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_719_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_418 = transpose(perm = value_263_perm_0, x = var_4779_cast_fp16)[name = tensor<string, []>("transpose_418")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_719_cast_fp16 = matmul(transpose_x = hidden_states_719_transpose_x_0, transpose_y = hidden_states_719_transpose_y_0, x = softmax_65_cast_fp16, y = transpose_418)[name = tensor<string, []>("hidden_states_719_cast_fp16")];
            tensor<int32, [4]> var_4782_perm_0 = const()[name = tensor<string, []>("op_4782_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4786 = const()[name = tensor<string, []>("op_4786"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_417 = transpose(perm = var_4782_perm_0, x = hidden_states_719_cast_fp16)[name = tensor<string, []>("transpose_417")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_721_cast_fp16 = reshape(shape = var_4786, x = transpose_417)[name = tensor<string, []>("hidden_states_721_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(867365824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(868594688))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(868594880)))];
            tensor<fp16, [1, 1024, 1280]> linear_347_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_8_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_721_cast_fp16)[name = tensor<string, []>("linear_347_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_825_cast_fp16 = add(x = linear_347_cast_fp16, y = input_817_cast_fp16)[name = tensor<string, []>("input_825_cast_fp16")];
            tensor<int32, [1]> input_827_axes_0 = const()[name = tensor<string, []>("input_827_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(868597504)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(868600128)))];
            tensor<fp16, [1, 1024, 1280]> input_827_cast_fp16 = layer_norm(axes = input_827_axes_0, beta = mid_block_attentions_0_transformer_blocks_8_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_8_norm3_weight_to_fp16, x = input_825_cast_fp16)[name = tensor<string, []>("input_827_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(868602752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(878433216))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(878433408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(878441152))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_348_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_8_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_8_ff_net_0_proj_weight_to_fp16_palettized, x = input_827_cast_fp16)[name = tensor<string, []>("linear_348_cast_fp16")];
            tensor<int32, [2]> var_4808_split_sizes_0 = const()[name = tensor<string, []>("op_4808_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_4808_axis_0 = const()[name = tensor<string, []>("op_4808_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_4808_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_4808_cast_fp16_1 = split(axis = var_4808_axis_0, split_sizes = var_4808_split_sizes_0, x = linear_348_cast_fp16)[name = tensor<string, []>("op_4808_cast_fp16")];
            tensor<string, []> var_4810_mode_0 = const()[name = tensor<string, []>("op_4810_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_4810_cast_fp16 = gelu(mode = var_4810_mode_0, x = var_4808_cast_fp16_1)[name = tensor<string, []>("op_4810_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_829_cast_fp16 = mul(x = var_4808_cast_fp16_0, y = var_4810_cast_fp16)[name = tensor<string, []>("input_829_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(878441344))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883356608))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_8_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_8_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883356800)))];
            tensor<fp16, [1, 1024, 1280]> linear_349_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_8_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_8_ff_net_2_weight_to_fp16_palettized, x = input_829_cast_fp16)[name = tensor<string, []>("linear_349_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_729_cast_fp16 = add(x = linear_349_cast_fp16, y = input_825_cast_fp16)[name = tensor<string, []>("hidden_states_729_cast_fp16")];
            tensor<int32, [1]> hidden_states_731_axes_0 = const()[name = tensor<string, []>("hidden_states_731_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_norm1_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_norm1_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883359424)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_norm1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_norm1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883362048)))];
            tensor<fp16, [1, 1024, 1280]> hidden_states_731_cast_fp16 = layer_norm(axes = hidden_states_731_axes_0, beta = mid_block_attentions_0_transformer_blocks_9_norm1_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_9_norm1_weight_to_fp16, x = hidden_states_729_cast_fp16)[name = tensor<string, []>("hidden_states_731_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883364672))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(884593536))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_350_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_attn1_to_q_weight_to_fp16_palettized, x = hidden_states_731_cast_fp16)[name = tensor<string, []>("linear_350_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(884593728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(885822592))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_351_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_attn1_to_k_weight_to_fp16_palettized, x = hidden_states_731_cast_fp16)[name = tensor<string, []>("linear_351_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(885822784))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(887051648))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_352_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_attn1_to_v_weight_to_fp16_palettized, x = hidden_states_731_cast_fp16)[name = tensor<string, []>("linear_352_cast_fp16")];
            tensor<int32, [4]> var_4845 = const()[name = tensor<string, []>("op_4845"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4846_cast_fp16 = reshape(shape = var_4845, x = linear_350_cast_fp16)[name = tensor<string, []>("op_4846_cast_fp16")];
            tensor<int32, [4]> var_4848 = const()[name = tensor<string, []>("op_4848"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4849_cast_fp16 = reshape(shape = var_4848, x = linear_351_cast_fp16)[name = tensor<string, []>("op_4849_cast_fp16")];
            tensor<int32, [4]> var_4851 = const()[name = tensor<string, []>("op_4851"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4852_cast_fp16 = reshape(shape = var_4851, x = linear_352_cast_fp16)[name = tensor<string, []>("op_4852_cast_fp16")];
            tensor<int32, [4]> value_267_perm_0 = const()[name = tensor<string, []>("value_267_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_85_y_0_to_fp16 = const()[name = tensor<string, []>("mul_85_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_85_cast_fp16 = mul(x = var_4846_cast_fp16, y = mul_85_y_0_to_fp16)[name = tensor<string, []>("mul_85_cast_fp16")];
            tensor<bool, []> matmul_66_transpose_y_0 = const()[name = tensor<string, []>("matmul_66_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_66_transpose_x_0 = const()[name = tensor<string, []>("matmul_66_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_404_perm_0 = const()[name = tensor<string, []>("transpose_404_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_405_perm_0 = const()[name = tensor<string, []>("transpose_405_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 1024, 64]> transpose_415 = transpose(perm = transpose_405_perm_0, x = var_4849_cast_fp16)[name = tensor<string, []>("transpose_415")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_416 = transpose(perm = transpose_404_perm_0, x = mul_85_cast_fp16)[name = tensor<string, []>("transpose_416")];
            tensor<fp16, [1, 20, 1024, 1024]> matmul_66_cast_fp16 = matmul(transpose_x = matmul_66_transpose_x_0, transpose_y = matmul_66_transpose_y_0, x = transpose_416, y = transpose_415)[name = tensor<string, []>("matmul_66_cast_fp16")];
            tensor<int32, []> softmax_66_axis_0 = const()[name = tensor<string, []>("softmax_66_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 1024]> softmax_66_cast_fp16 = softmax(axis = softmax_66_axis_0, x = matmul_66_cast_fp16)[name = tensor<string, []>("softmax_66_cast_fp16")];
            tensor<bool, []> hidden_states_733_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_733_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_733_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_733_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1024, 64]> transpose_414 = transpose(perm = value_267_perm_0, x = var_4852_cast_fp16)[name = tensor<string, []>("transpose_414")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_733_cast_fp16 = matmul(transpose_x = hidden_states_733_transpose_x_0, transpose_y = hidden_states_733_transpose_y_0, x = softmax_66_cast_fp16, y = transpose_414)[name = tensor<string, []>("hidden_states_733_cast_fp16")];
            tensor<int32, [4]> var_4855_perm_0 = const()[name = tensor<string, []>("op_4855_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4859 = const()[name = tensor<string, []>("op_4859"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_413 = transpose(perm = var_4855_perm_0, x = hidden_states_733_cast_fp16)[name = tensor<string, []>("transpose_413")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_735_cast_fp16 = reshape(shape = var_4859, x = transpose_413)[name = tensor<string, []>("hidden_states_735_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(887051840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888280704))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_attn1_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn1_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888280896)))];
            tensor<fp16, [1, 1024, 1280]> linear_353_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_9_attn1_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_attn1_to_out_0_weight_to_fp16_palettized, x = hidden_states_735_cast_fp16)[name = tensor<string, []>("linear_353_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_837_cast_fp16 = add(x = linear_353_cast_fp16, y = hidden_states_729_cast_fp16)[name = tensor<string, []>("input_837_cast_fp16")];
            tensor<int32, [1]> input_839_axes_0 = const()[name = tensor<string, []>("input_839_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_norm2_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_norm2_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888283520)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_norm2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_norm2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888286144)))];
            tensor<fp16, [1, 1024, 1280]> input_839_cast_fp16 = layer_norm(axes = input_839_axes_0, beta = mid_block_attentions_0_transformer_blocks_9_norm2_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_9_norm2_weight_to_fp16, x = input_837_cast_fp16)[name = tensor<string, []>("input_839_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888288768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(889517632))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1, 1024, 1280]> linear_354_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_attn2_to_q_weight_to_fp16_palettized, x = input_839_cast_fp16)[name = tensor<string, []>("linear_354_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(889517824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(891483968))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_355_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_attn2_to_k_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_355_cast_fp16")];
            tensor<fp16, [1280, 2048]> mid_block_attentions_0_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1966080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(891484160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893450304))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 2048])];
            tensor<fp16, [1, 77, 1280]> linear_356_cast_fp16 = linear(bias = add_23_mean_0_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_attn2_to_v_weight_to_fp16_palettized, x = encoder_hidden_states)[name = tensor<string, []>("linear_356_cast_fp16")];
            tensor<int32, [4]> var_4891 = const()[name = tensor<string, []>("op_4891"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 1024, 20, 64]> var_4892_cast_fp16 = reshape(shape = var_4891, x = linear_354_cast_fp16)[name = tensor<string, []>("op_4892_cast_fp16")];
            tensor<int32, [4]> var_4894 = const()[name = tensor<string, []>("op_4894"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4895_cast_fp16 = reshape(shape = var_4894, x = linear_355_cast_fp16)[name = tensor<string, []>("op_4895_cast_fp16")];
            tensor<int32, [4]> var_4897 = const()[name = tensor<string, []>("op_4897"), val = tensor<int32, [4]>([1, -1, 20, 64])];
            tensor<fp16, [1, 77, 20, 64]> var_4898_cast_fp16 = reshape(shape = var_4897, x = linear_356_cast_fp16)[name = tensor<string, []>("op_4898_cast_fp16")];
            tensor<int32, [4]> value_perm_0 = const()[name = tensor<string, []>("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, []> mul_86_y_0_to_fp16 = const()[name = tensor<string, []>("mul_86_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1024, 20, 64]> mul_86_cast_fp16 = mul(x = var_4892_cast_fp16, y = mul_86_y_0_to_fp16)[name = tensor<string, []>("mul_86_cast_fp16")];
            tensor<bool, []> matmul_67_transpose_y_0 = const()[name = tensor<string, []>("matmul_67_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_67_transpose_x_0 = const()[name = tensor<string, []>("matmul_67_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_406_perm_0 = const()[name = tensor<string, []>("transpose_406_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_407_perm_0 = const()[name = tensor<string, []>("transpose_407_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 20, 77, 64]> transpose_411 = transpose(perm = transpose_407_perm_0, x = var_4895_cast_fp16)[name = tensor<string, []>("transpose_411")];
            tensor<fp16, [1, 20, 1024, 64]> transpose_412 = transpose(perm = transpose_406_perm_0, x = mul_86_cast_fp16)[name = tensor<string, []>("transpose_412")];
            tensor<fp16, [1, 20, 1024, 77]> matmul_67_cast_fp16 = matmul(transpose_x = matmul_67_transpose_x_0, transpose_y = matmul_67_transpose_y_0, x = transpose_412, y = transpose_411)[name = tensor<string, []>("matmul_67_cast_fp16")];
            tensor<int32, []> softmax_67_axis_0 = const()[name = tensor<string, []>("softmax_67_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 20, 1024, 77]> softmax_67_cast_fp16 = softmax(axis = softmax_67_axis_0, x = matmul_67_cast_fp16)[name = tensor<string, []>("softmax_67_cast_fp16")];
            tensor<bool, []> hidden_states_739_transpose_x_0 = const()[name = tensor<string, []>("hidden_states_739_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> hidden_states_739_transpose_y_0 = const()[name = tensor<string, []>("hidden_states_739_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 77, 64]> transpose_410 = transpose(perm = value_perm_0, x = var_4898_cast_fp16)[name = tensor<string, []>("transpose_410")];
            tensor<fp16, [1, 20, 1024, 64]> hidden_states_739_cast_fp16 = matmul(transpose_x = hidden_states_739_transpose_x_0, transpose_y = hidden_states_739_transpose_y_0, x = softmax_67_cast_fp16, y = transpose_410)[name = tensor<string, []>("hidden_states_739_cast_fp16")];
            tensor<int32, [4]> var_4901_perm_0 = const()[name = tensor<string, []>("op_4901_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4905 = const()[name = tensor<string, []>("op_4905"), val = tensor<int32, [3]>([1, -1, 1280])];
            tensor<fp16, [1, 1024, 20, 64]> transpose_409 = transpose(perm = var_4901_perm_0, x = hidden_states_739_cast_fp16)[name = tensor<string, []>("transpose_409")];
            tensor<fp16, [1, 1024, 1280]> hidden_states_741_cast_fp16 = reshape(shape = var_4905, x = transpose_409)[name = tensor<string, []>("hidden_states_741_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893450496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(894679360))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_attn2_to_out_0_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_attn2_to_out_0_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(894679552)))];
            tensor<fp16, [1, 1024, 1280]> linear_357_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_9_attn2_to_out_0_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_attn2_to_out_0_weight_to_fp16_palettized, x = hidden_states_741_cast_fp16)[name = tensor<string, []>("linear_357_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_845_cast_fp16 = add(x = linear_357_cast_fp16, y = input_837_cast_fp16)[name = tensor<string, []>("input_845_cast_fp16")];
            tensor<int32, [1]> input_847_axes_0 = const()[name = tensor<string, []>("input_847_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_norm3_weight_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_norm3_weight_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(894682176)))];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_norm3_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_norm3_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(894684800)))];
            tensor<fp16, [1, 1024, 1280]> input_847_cast_fp16 = layer_norm(axes = input_847_axes_0, beta = mid_block_attentions_0_transformer_blocks_9_norm3_bias_to_fp16, epsilon = var_3647_to_fp16, gamma = mid_block_attentions_0_transformer_blocks_9_norm3_weight_to_fp16, x = input_845_cast_fp16)[name = tensor<string, []>("input_847_cast_fp16")];
            tensor<fp16, [10240, 1280]> mid_block_attentions_0_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [9830400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(894687424))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(904517888))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10240, 1280])];
            tensor<fp16, [10240]> mid_block_attentions_0_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(904518080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(904525824))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized"), shape = tensor<uint32, [1]>([10240])];
            tensor<fp16, [1, 1024, 10240]> linear_358_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_9_ff_net_0_proj_bias_to_fp16_palettized, weight = mid_block_attentions_0_transformer_blocks_9_ff_net_0_proj_weight_to_fp16_palettized, x = input_847_cast_fp16)[name = tensor<string, []>("linear_358_cast_fp16")];
            tensor<int32, [2]> var_4927_split_sizes_0 = const()[name = tensor<string, []>("op_4927_split_sizes_0"), val = tensor<int32, [2]>([5120, 5120])];
            tensor<int32, []> var_4927_axis_0 = const()[name = tensor<string, []>("op_4927_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1024, 5120]> var_4927_cast_fp16_0, tensor<fp16, [1, 1024, 5120]> var_4927_cast_fp16_1 = split(axis = var_4927_axis_0, split_sizes = var_4927_split_sizes_0, x = linear_358_cast_fp16)[name = tensor<string, []>("op_4927_cast_fp16")];
            tensor<string, []> var_4929_mode_0 = const()[name = tensor<string, []>("op_4929_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 1024, 5120]> var_4929_cast_fp16 = gelu(mode = var_4929_mode_0, x = var_4927_cast_fp16_1)[name = tensor<string, []>("op_4929_cast_fp16")];
            tensor<fp16, [1, 1024, 5120]> input_849_cast_fp16 = mul(x = var_4927_cast_fp16_0, y = var_4929_cast_fp16)[name = tensor<string, []>("input_849_cast_fp16")];
            tensor<fp16, [1280, 5120]> mid_block_attentions_0_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(904526016))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(909441280))), name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 5120])];
            tensor<fp16, [1280]> mid_block_attentions_0_transformer_blocks_9_ff_net_2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_transformer_blocks_9_ff_net_2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(909441472)))];
            tensor<fp16, [1, 1024, 1280]> linear_359_cast_fp16 = linear(bias = mid_block_attentions_0_transformer_blocks_9_ff_net_2_bias_to_fp16, weight = mid_block_attentions_0_transformer_blocks_9_ff_net_2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = tensor<string, []>("linear_359_cast_fp16")];
            tensor<fp16, [1, 1024, 1280]> input_853_cast_fp16 = add(x = linear_359_cast_fp16, y = input_845_cast_fp16)[name = tensor<string, []>("input_853_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_attentions_0_proj_out_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(909444096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(910672960))), name = tensor<string, []>("mid_block_attentions_0_proj_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_attentions_0_proj_out_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_attentions_0_proj_out_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(910673152)))];
            tensor<fp16, [1, 1024, 1280]> linear_360_cast_fp16 = linear(bias = mid_block_attentions_0_proj_out_bias_to_fp16, weight = mid_block_attentions_0_proj_out_weight_to_fp16_palettized, x = input_853_cast_fp16)[name = tensor<string, []>("linear_360_cast_fp16")];
            tensor<int32, [4]> var_4939 = const()[name = tensor<string, []>("op_4939"), val = tensor<int32, [4]>([1, 32, 32, 1280])];
            tensor<fp16, [1, 32, 32, 1280]> var_4940_cast_fp16 = reshape(shape = var_4939, x = linear_360_cast_fp16)[name = tensor<string, []>("op_4940_cast_fp16")];
            tensor<int32, [4]> var_4941 = const()[name = tensor<string, []>("op_4941"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 1280, 32, 32]> transpose_408 = transpose(perm = var_4941, x = var_4940_cast_fp16)[name = tensor<string, []>("transpose_408")];
            tensor<fp16, [1, 1280, 32, 32]> input_855_cast_fp16 = add(x = transpose_408, y = var_3699_cast_fp16)[name = tensor<string, []>("input_855_cast_fp16")];
            tensor<int32, [5]> reshape_76_shape_0 = const()[name = tensor<string, []>("reshape_76_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_76_cast_fp16 = reshape(shape = reshape_76_shape_0, x = input_855_cast_fp16)[name = tensor<string, []>("reshape_76_cast_fp16")];
            tensor<int32, [3]> reduce_mean_57_axes_0 = const()[name = tensor<string, []>("reduce_mean_57_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_57_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_57_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = reshape_76_cast_fp16)[name = tensor<string, []>("reduce_mean_57_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_38_cast_fp16 = sub(x = reshape_76_cast_fp16, y = reduce_mean_57_cast_fp16)[name = tensor<string, []>("sub_38_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_19_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor<string, []>("square_19_cast_fp16")];
            tensor<int32, [3]> reduce_mean_59_axes_0 = const()[name = tensor<string, []>("reduce_mean_59_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_59_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_59_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_19_cast_fp16)[name = tensor<string, []>("reduce_mean_59_cast_fp16")];
            tensor<fp16, []> add_38_y_0_to_fp16 = const()[name = tensor<string, []>("add_38_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_38_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = add_38_y_0_to_fp16)[name = tensor<string, []>("add_38_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_19_cast_fp16 = sqrt(x = add_38_cast_fp16)[name = tensor<string, []>("sqrt_19_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_19_cast_fp16 = real_div(x = sub_38_cast_fp16, y = sqrt_19_cast_fp16)[name = tensor<string, []>("real_div_19_cast_fp16")];
            tensor<int32, [4]> reshape_77_shape_0 = const()[name = tensor<string, []>("reshape_77_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = real_div_19_cast_fp16)[name = tensor<string, []>("reshape_77_cast_fp16")];
            tensor<fp16, [1280]> add_39_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_39_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(910675776)))];
            tensor<fp16, [1280]> add_39_beta_0_to_fp16 = const()[name = tensor<string, []>("add_39_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(910678400)))];
            tensor<fp16, []> add_39_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_39_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_39_cast_fp16 = batch_norm(beta = add_39_beta_0_to_fp16, epsilon = add_39_epsilon_0_to_fp16, gamma = add_39_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_77_cast_fp16)[name = tensor<string, []>("add_39_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_859_cast_fp16 = silu(x = add_39_cast_fp16)[name = tensor<string, []>("input_859_cast_fp16")];
            tensor<int32, [2]> var_4956 = const()[name = tensor<string, []>("op_4956"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4958 = const()[name = tensor<string, []>("op_4958"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_753_pad_type_0 = const()[name = tensor<string, []>("hidden_states_753_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_753_pad_0 = const()[name = tensor<string, []>("hidden_states_753_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [1280, 1280, 3, 3]> mid_block_resnets_1_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [11059200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(910681024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(921740288))), name = tensor<string, []>("mid_block_resnets_1_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 3, 3])];
            tensor<fp16, [1280]> mid_block_resnets_1_conv1_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_resnets_1_conv1_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(921740480)))];
            tensor<fp16, [1, 1280, 32, 32]> hidden_states_753_cast_fp16 = conv(bias = mid_block_resnets_1_conv1_bias_to_fp16, dilations = var_4958, groups = var_3651, pad = hidden_states_753_pad_0, pad_type = hidden_states_753_pad_type_0, strides = var_4956, weight = mid_block_resnets_1_conv1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = tensor<string, []>("hidden_states_753_cast_fp16")];
            tensor<fp16, [1280, 1280]> mid_block_resnets_1_time_emb_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(921743104))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(922971968))), name = tensor<string, []>("mid_block_resnets_1_time_emb_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1280, 1280])];
            tensor<fp16, [1280]> mid_block_resnets_1_time_emb_proj_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_resnets_1_time_emb_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(922972160)))];
            tensor<fp16, [1, 1280]> linear_361_cast_fp16 = linear(bias = mid_block_resnets_1_time_emb_proj_bias_to_fp16, weight = mid_block_resnets_1_time_emb_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_361_cast_fp16")];
            tensor<int32, [1]> var_4967_axes_0 = const()[name = tensor<string, []>("op_4967_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1280, 1]> var_4967_cast_fp16 = expand_dims(axes = var_4967_axes_0, x = linear_361_cast_fp16)[name = tensor<string, []>("op_4967_cast_fp16")];
            tensor<int32, [1]> temb_axes_0 = const()[name = tensor<string, []>("temb_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 1280, 1, 1]> temb_cast_fp16 = expand_dims(axes = temb_axes_0, x = var_4967_cast_fp16)[name = tensor<string, []>("temb_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_863_cast_fp16 = add(x = hidden_states_753_cast_fp16, y = temb_cast_fp16)[name = tensor<string, []>("input_863_cast_fp16")];
            tensor<int32, [5]> reshape_80_shape_0 = const()[name = tensor<string, []>("reshape_80_shape_0"), val = tensor<int32, [5]>([1, 32, 40, 32, 32])];
            tensor<fp16, [1, 32, 40, 32, 32]> reshape_80_cast_fp16 = reshape(shape = reshape_80_shape_0, x = input_863_cast_fp16)[name = tensor<string, []>("reshape_80_cast_fp16")];
            tensor<int32, [3]> reduce_mean_60_axes_0 = const()[name = tensor<string, []>("reduce_mean_60_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_60_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_60_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_60_cast_fp16 = reduce_mean(axes = reduce_mean_60_axes_0, keep_dims = reduce_mean_60_keep_dims_0, x = reshape_80_cast_fp16)[name = tensor<string, []>("reduce_mean_60_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> sub_40_cast_fp16 = sub(x = reshape_80_cast_fp16, y = reduce_mean_60_cast_fp16)[name = tensor<string, []>("sub_40_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> square_20_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor<string, []>("square_20_cast_fp16")];
            tensor<int32, [3]> reduce_mean_62_axes_0 = const()[name = tensor<string, []>("reduce_mean_62_axes_0"), val = tensor<int32, [3]>([2, 3, 4])];
            tensor<bool, []> reduce_mean_62_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_62_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 32, 1, 1, 1]> reduce_mean_62_cast_fp16 = reduce_mean(axes = reduce_mean_62_axes_0, keep_dims = reduce_mean_62_keep_dims_0, x = square_20_cast_fp16)[name = tensor<string, []>("reduce_mean_62_cast_fp16")];
            tensor<fp16, []> add_40_y_0_to_fp16 = const()[name = tensor<string, []>("add_40_y_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 32, 1, 1, 1]> add_40_cast_fp16 = add(x = reduce_mean_62_cast_fp16, y = add_40_y_0_to_fp16)[name = tensor<string, []>("add_40_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1, 1]> sqrt_20_cast_fp16 = sqrt(x = add_40_cast_fp16)[name = tensor<string, []>("sqrt_20_cast_fp16")];
            tensor<fp16, [1, 32, 40, 32, 32]> real_div_20_cast_fp16 = real_div(x = sub_40_cast_fp16, y = sqrt_20_cast_fp16)[name = tensor<string, []>("real_div_20_cast_fp16")];
            tensor<int32, [4]> reshape_81_shape_0 = const()[name = tensor<string, []>("reshape_81_shape_0"), val = tensor<int32, [4]>([1, 1280, 32, 32])];
            tensor<fp16, [1, 1280, 32, 32]> reshape_81_cast_fp16 = reshape(shape = reshape_81_shape_0, x = real_div_20_cast_fp16)[name = tensor<string, []>("reshape_81_cast_fp16")];
            tensor<fp16, [1280]> add_41_gamma_0_to_fp16 = const()[name = tensor<string, []>("add_41_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(922974784)))];
            tensor<fp16, [1280]> add_41_beta_0_to_fp16 = const()[name = tensor<string, []>("add_41_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(922977408)))];
            tensor<fp16, []> add_41_epsilon_0_to_fp16 = const()[name = tensor<string, []>("add_41_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 32, 32]> add_41_cast_fp16 = batch_norm(beta = add_41_beta_0_to_fp16, epsilon = add_41_epsilon_0_to_fp16, gamma = add_41_gamma_0_to_fp16, mean = add_23_mean_0_to_fp16, variance = add_23_variance_0_to_fp16, x = reshape_81_cast_fp16)[name = tensor<string, []>("add_41_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> input_867_cast_fp16 = silu(x = add_41_cast_fp16)[name = tensor<string, []>("input_867_cast_fp16")];
            tensor<int32, [2]> var_4977 = const()[name = tensor<string, []>("op_4977"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4979 = const()[name = tensor<string, []>("op_4979"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_pad_type_0 = const()[name = tensor<string, []>("hidden_states_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = tensor<string, []>("hidden_states_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp16, [1280, 1280, 3, 3]> mid_block_resnets_1_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [11059200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(922980032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934039296))), name = tensor<string, []>("mid_block_resnets_1_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 3, 3])];
            tensor<fp16, [1280]> mid_block_resnets_1_conv2_bias_to_fp16 = const()[name = tensor<string, []>("mid_block_resnets_1_conv2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934039488)))];
            tensor<fp16, [1, 1280, 32, 32]> hidden_states_cast_fp16 = conv(bias = mid_block_resnets_1_conv2_bias_to_fp16, dilations = var_4979, groups = var_3651, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = var_4977, weight = mid_block_resnets_1_conv2_weight_to_fp16_palettized, x = input_867_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
            tensor<fp16, [1, 1280, 32, 32]> var_4982_cast_fp16 = add(x = input_855_cast_fp16, y = hidden_states_cast_fp16)[name = tensor<string, []>("op_4982_cast_fp16")];
            tensor<int32, []> var_4988 = const()[name = tensor<string, []>("op_4988"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_4991 = const()[name = tensor<string, []>("op_4991"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4993 = const()[name = tensor<string, []>("op_4993"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_5_pad_type_0 = const()[name = tensor<string, []>("sample_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_5_pad_0 = const()[name = tensor<string, []>("sample_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [320, 320, 1, 1]> controlnet_down_blocks_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [76800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934042112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934118976))), name = tensor<string, []>("controlnet_down_blocks_0_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 1, 1])];
            tensor<fp16, [320]> controlnet_down_blocks_0_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_0_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934119168)))];
            tensor<fp16, [1, 320, 128, 128]> sample_5_cast_fp16 = conv(bias = controlnet_down_blocks_0_bias_to_fp16, dilations = var_4993, groups = var_4988, pad = sample_5_pad_0, pad_type = sample_5_pad_type_0, strides = var_4991, weight = controlnet_down_blocks_0_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = tensor<string, []>("sample_5_cast_fp16")];
            tensor<int32, []> var_4999 = const()[name = tensor<string, []>("op_4999"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5002 = const()[name = tensor<string, []>("op_5002"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5004 = const()[name = tensor<string, []>("op_5004"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_7_pad_type_0 = const()[name = tensor<string, []>("sample_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_7_pad_0 = const()[name = tensor<string, []>("sample_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [320, 320, 1, 1]> controlnet_down_blocks_1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [76800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934119872))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934196736))), name = tensor<string, []>("controlnet_down_blocks_1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 1, 1])];
            tensor<fp16, [320]> controlnet_down_blocks_1_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_1_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934196928)))];
            tensor<fp16, [1, 320, 128, 128]> sample_7_cast_fp16 = conv(bias = controlnet_down_blocks_1_bias_to_fp16, dilations = var_5004, groups = var_4999, pad = sample_7_pad_0, pad_type = sample_7_pad_type_0, strides = var_5002, weight = controlnet_down_blocks_1_weight_to_fp16_palettized, x = var_302_cast_fp16)[name = tensor<string, []>("sample_7_cast_fp16")];
            tensor<int32, []> var_5010 = const()[name = tensor<string, []>("op_5010"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5013 = const()[name = tensor<string, []>("op_5013"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5015 = const()[name = tensor<string, []>("op_5015"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_9_pad_type_0 = const()[name = tensor<string, []>("sample_9_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_9_pad_0 = const()[name = tensor<string, []>("sample_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [320, 320, 1, 1]> controlnet_down_blocks_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [76800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934197632))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934274496))), name = tensor<string, []>("controlnet_down_blocks_2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 1, 1])];
            tensor<fp16, [320]> controlnet_down_blocks_2_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_2_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934274688)))];
            tensor<fp16, [1, 320, 128, 128]> sample_9_cast_fp16 = conv(bias = controlnet_down_blocks_2_bias_to_fp16, dilations = var_5015, groups = var_5010, pad = sample_9_pad_0, pad_type = sample_9_pad_type_0, strides = var_5013, weight = controlnet_down_blocks_2_weight_to_fp16_palettized, x = var_342_cast_fp16)[name = tensor<string, []>("sample_9_cast_fp16")];
            tensor<int32, []> var_5021 = const()[name = tensor<string, []>("op_5021"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5024 = const()[name = tensor<string, []>("op_5024"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5026 = const()[name = tensor<string, []>("op_5026"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_11_pad_type_0 = const()[name = tensor<string, []>("sample_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_11_pad_0 = const()[name = tensor<string, []>("sample_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [320, 320, 1, 1]> controlnet_down_blocks_3_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [76800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934275392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934352256))), name = tensor<string, []>("controlnet_down_blocks_3_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([320, 320, 1, 1])];
            tensor<fp16, [320]> controlnet_down_blocks_3_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_3_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934352448)))];
            tensor<fp16, [1, 320, 64, 64]> sample_11_cast_fp16 = conv(bias = controlnet_down_blocks_3_bias_to_fp16, dilations = var_5026, groups = var_5021, pad = sample_11_pad_0, pad_type = sample_11_pad_type_0, strides = var_5024, weight = controlnet_down_blocks_3_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = tensor<string, []>("sample_11_cast_fp16")];
            tensor<int32, []> var_5032 = const()[name = tensor<string, []>("op_5032"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5035 = const()[name = tensor<string, []>("op_5035"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5037 = const()[name = tensor<string, []>("op_5037"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_13_pad_type_0 = const()[name = tensor<string, []>("sample_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_13_pad_0 = const()[name = tensor<string, []>("sample_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [640, 640, 1, 1]> controlnet_down_blocks_4_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934353152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934660416))), name = tensor<string, []>("controlnet_down_blocks_4_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 640, 1, 1])];
            tensor<fp16, [640]> controlnet_down_blocks_4_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_4_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934660608)))];
            tensor<fp16, [1, 640, 64, 64]> sample_13_cast_fp16 = conv(bias = controlnet_down_blocks_4_bias_to_fp16, dilations = var_5037, groups = var_5032, pad = sample_13_pad_0, pad_type = sample_13_pad_type_0, strides = var_5035, weight = controlnet_down_blocks_4_weight_to_fp16_palettized, x = input_135_cast_fp16)[name = tensor<string, []>("sample_13_cast_fp16")];
            tensor<int32, []> var_5043 = const()[name = tensor<string, []>("op_5043"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5046 = const()[name = tensor<string, []>("op_5046"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5048 = const()[name = tensor<string, []>("op_5048"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_15_pad_type_0 = const()[name = tensor<string, []>("sample_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_15_pad_0 = const()[name = tensor<string, []>("sample_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [640, 640, 1, 1]> controlnet_down_blocks_5_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934661952))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934969216))), name = tensor<string, []>("controlnet_down_blocks_5_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 640, 1, 1])];
            tensor<fp16, [640]> controlnet_down_blocks_5_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_5_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934969408)))];
            tensor<fp16, [1, 640, 64, 64]> sample_15_cast_fp16 = conv(bias = controlnet_down_blocks_5_bias_to_fp16, dilations = var_5048, groups = var_5043, pad = sample_15_pad_0, pad_type = sample_15_pad_type_0, strides = var_5046, weight = controlnet_down_blocks_5_weight_to_fp16_palettized, x = hidden_states_115_cast_fp16)[name = tensor<string, []>("sample_15_cast_fp16")];
            tensor<int32, []> var_5054 = const()[name = tensor<string, []>("op_5054"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5057 = const()[name = tensor<string, []>("op_5057"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5059 = const()[name = tensor<string, []>("op_5059"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_17_pad_type_0 = const()[name = tensor<string, []>("sample_17_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_17_pad_0 = const()[name = tensor<string, []>("sample_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [640, 640, 1, 1]> controlnet_down_blocks_6_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [307200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(934970752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(935278016))), name = tensor<string, []>("controlnet_down_blocks_6_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([640, 640, 1, 1])];
            tensor<fp16, [640]> controlnet_down_blocks_6_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_6_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(935278208)))];
            tensor<fp16, [1, 640, 32, 32]> sample_17_cast_fp16 = conv(bias = controlnet_down_blocks_6_bias_to_fp16, dilations = var_5059, groups = var_5054, pad = sample_17_pad_0, pad_type = sample_17_pad_type_0, strides = var_5057, weight = controlnet_down_blocks_6_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = tensor<string, []>("sample_17_cast_fp16")];
            tensor<int32, []> var_5065 = const()[name = tensor<string, []>("op_5065"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5068 = const()[name = tensor<string, []>("op_5068"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5070 = const()[name = tensor<string, []>("op_5070"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_19_pad_type_0 = const()[name = tensor<string, []>("sample_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_19_pad_0 = const()[name = tensor<string, []>("sample_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> controlnet_down_blocks_7_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(935279552))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(936508416))), name = tensor<string, []>("controlnet_down_blocks_7_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> controlnet_down_blocks_7_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_7_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(936508608)))];
            tensor<fp16, [1, 1280, 32, 32]> sample_19_cast_fp16 = conv(bias = controlnet_down_blocks_7_bias_to_fp16, dilations = var_5070, groups = var_5065, pad = sample_19_pad_0, pad_type = sample_19_pad_type_0, strides = var_5068, weight = controlnet_down_blocks_7_weight_to_fp16_palettized, x = input_415_cast_fp16)[name = tensor<string, []>("sample_19_cast_fp16")];
            tensor<int32, []> var_5076 = const()[name = tensor<string, []>("op_5076"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5079 = const()[name = tensor<string, []>("op_5079"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5081 = const()[name = tensor<string, []>("op_5081"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> sample_pad_type_0 = const()[name = tensor<string, []>("sample_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> sample_pad_0 = const()[name = tensor<string, []>("sample_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> controlnet_down_blocks_8_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(936511232))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(937740096))), name = tensor<string, []>("controlnet_down_blocks_8_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> controlnet_down_blocks_8_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_down_blocks_8_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(937740288)))];
            tensor<fp16, [1, 1280, 32, 32]> sample_cast_fp16 = conv(bias = controlnet_down_blocks_8_bias_to_fp16, dilations = var_5081, groups = var_5076, pad = sample_pad_0, pad_type = sample_pad_type_0, strides = var_5079, weight = controlnet_down_blocks_8_weight_to_fp16_palettized, x = input_635_cast_fp16)[name = tensor<string, []>("sample_cast_fp16")];
            tensor<int32, []> var_5087 = const()[name = tensor<string, []>("op_5087"), val = tensor<int32, []>(1)];
            tensor<int32, [2]> var_5090 = const()[name = tensor<string, []>("op_5090"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5092 = const()[name = tensor<string, []>("op_5092"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> mid_block_res_sample_pad_type_0 = const()[name = tensor<string, []>("mid_block_res_sample_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> mid_block_res_sample_pad_0 = const()[name = tensor<string, []>("mid_block_res_sample_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> controlnet_mid_block_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(937742912))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(938971776))), name = tensor<string, []>("controlnet_mid_block_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> controlnet_mid_block_bias_to_fp16 = const()[name = tensor<string, []>("controlnet_mid_block_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(938971968)))];
            tensor<fp16, [1, 1280, 32, 32]> mid_block_res_sample_cast_fp16 = conv(bias = controlnet_mid_block_bias_to_fp16, dilations = var_5092, groups = var_5087, pad = mid_block_res_sample_pad_0, pad_type = mid_block_res_sample_pad_type_0, strides = var_5090, weight = controlnet_mid_block_weight_to_fp16_palettized, x = var_4982_cast_fp16)[name = tensor<string, []>("mid_block_res_sample_cast_fp16")];
            tensor<fp16, [1, 320, 128, 128]> var_5095_cast_fp16 = mul(x = sample_5_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5095_cast_fp16")];
            tensor<string, []> var_5095_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5095_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 320, 128, 128]> var_5096_cast_fp16 = mul(x = sample_7_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5096_cast_fp16")];
            tensor<string, []> var_5096_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5096_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 320, 128, 128]> var_5097_cast_fp16 = mul(x = sample_9_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5097_cast_fp16")];
            tensor<string, []> var_5097_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5097_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 320, 64, 64]> var_5098_cast_fp16 = mul(x = sample_11_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5098_cast_fp16")];
            tensor<string, []> var_5098_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5098_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 640, 64, 64]> var_5099_cast_fp16 = mul(x = sample_13_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5099_cast_fp16")];
            tensor<string, []> var_5099_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5099_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 640, 64, 64]> var_5100_cast_fp16 = mul(x = sample_15_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5100_cast_fp16")];
            tensor<string, []> var_5100_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5100_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 640, 32, 32]> var_5101_cast_fp16 = mul(x = sample_17_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5101_cast_fp16")];
            tensor<string, []> var_5101_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5101_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 1280, 32, 32]> var_5102_cast_fp16 = mul(x = sample_19_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5102_cast_fp16")];
            tensor<string, []> var_5102_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5102_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 1280, 32, 32]> var_5103_cast_fp16 = mul(x = sample_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5103_cast_fp16")];
            tensor<string, []> var_5103_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5103_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 1280, 32, 32]> var_5104_cast_fp16 = mul(x = mid_block_res_sample_cast_fp16, y = conditioning_scale)[name = tensor<string, []>("op_5104_cast_fp16")];
            tensor<string, []> var_5104_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5104_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 1280, 32, 32]> additional_residual_9 = cast(dtype = var_5104_cast_fp16_to_fp32_dtype_0, x = var_5104_cast_fp16)[name = tensor<string, []>("cast_0")];
            tensor<fp32, [1, 1280, 32, 32]> additional_residual_8 = cast(dtype = var_5103_cast_fp16_to_fp32_dtype_0, x = var_5103_cast_fp16)[name = tensor<string, []>("cast_1")];
            tensor<fp32, [1, 1280, 32, 32]> additional_residual_7 = cast(dtype = var_5102_cast_fp16_to_fp32_dtype_0, x = var_5102_cast_fp16)[name = tensor<string, []>("cast_2")];
            tensor<fp32, [1, 640, 32, 32]> additional_residual_6 = cast(dtype = var_5101_cast_fp16_to_fp32_dtype_0, x = var_5101_cast_fp16)[name = tensor<string, []>("cast_3")];
            tensor<fp32, [1, 640, 64, 64]> additional_residual_5 = cast(dtype = var_5100_cast_fp16_to_fp32_dtype_0, x = var_5100_cast_fp16)[name = tensor<string, []>("cast_4")];
            tensor<fp32, [1, 640, 64, 64]> additional_residual_4 = cast(dtype = var_5099_cast_fp16_to_fp32_dtype_0, x = var_5099_cast_fp16)[name = tensor<string, []>("cast_5")];
            tensor<fp32, [1, 320, 64, 64]> additional_residual_3 = cast(dtype = var_5098_cast_fp16_to_fp32_dtype_0, x = var_5098_cast_fp16)[name = tensor<string, []>("cast_6")];
            tensor<fp32, [1, 320, 128, 128]> additional_residual_2 = cast(dtype = var_5097_cast_fp16_to_fp32_dtype_0, x = var_5097_cast_fp16)[name = tensor<string, []>("cast_7")];
            tensor<fp32, [1, 320, 128, 128]> additional_residual_1 = cast(dtype = var_5096_cast_fp16_to_fp32_dtype_0, x = var_5096_cast_fp16)[name = tensor<string, []>("cast_8")];
            tensor<fp32, [1, 320, 128, 128]> additional_residual_0 = cast(dtype = var_5095_cast_fp16_to_fp32_dtype_0, x = var_5095_cast_fp16)[name = tensor<string, []>("cast_9")];
        } -> (additional_residual_0, additional_residual_1, additional_residual_2, additional_residual_3, additional_residual_4, additional_residual_5, additional_residual_6, additional_residual_7, additional_residual_8, additional_residual_9);
}