program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.4.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
{
func main<ios16>(tensor<fp16, [1, 128, 1, 3000]> melspectrogram_features) {
tensor<string, []> var_106_pad_type_0 = const()[name = tensor<string, []>("op_106_pad_type_0"), val = tensor<string, []>("custom")];
tensor<int32, [4]> var_106_pad_0 = const()[name = tensor<string, []>("op_106_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
tensor<int32, [2]> var_106_strides_0 = const()[name = tensor<string, []>("op_106_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [2]> var_106_dilations_0 = const()[name = tensor<string, []>("op_106_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> var_106_groups_0 = const()[name = tensor<string, []>("op_106_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 128, 1, 3]> var_81_to_fp16 = const()[name = tensor<string, []>("op_81_to_fp16"), val = tensor<fp16, [1280, 128, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
tensor<fp16, [1280]> var_87_to_fp16 = const()[name = tensor<string, []>("op_87_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(983168)))];
tensor<fp16, [1, 1280, 1, 3000]> var_106_cast_fp16 = conv(bias = var_87_to_fp16, dilations = var_106_dilations_0, groups = var_106_groups_0, pad = var_106_pad_0, pad_type = var_106_pad_type_0, strides = var_106_strides_0, weight = var_81_to_fp16, x = melspectrogram_features)[name = tensor<string, []>("op_106_cast_fp16")];
tensor<string, []> hidden_states_1_mode_0 = const()[name = tensor<string, []>("hidden_states_1_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 1280, 1, 3000]> hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_106_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
tensor<string, []> var_146_pad_type_0 = const()[name = tensor<string, []>("op_146_pad_type_0"), val = tensor<string, []>("custom")];
tensor<int32, [4]> var_146_pad_0 = const()[name = tensor<string, []>("op_146_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
tensor<int32, [2]> var_146_strides_0 = const()[name = tensor<string, []>("op_146_strides_0"), val = tensor<int32, [2]>([2, 2])];
tensor<int32, [2]> var_146_dilations_0 = const()[name = tensor<string, []>("op_146_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> var_146_groups_0 = const()[name = tensor<string, []>("op_146_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 3]> var_121_to_fp16 = const()[name = tensor<string, []>("op_121_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(985792)))];
tensor<fp16, [1280]> var_127_to_fp16 = const()[name = tensor<string, []>("op_127_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10816256)))];
tensor<fp16, [1, 1280, 1, 1500]> var_146_cast_fp16 = conv(bias = var_127_to_fp16, dilations = var_146_dilations_0, groups = var_146_groups_0, pad = var_146_pad_0, pad_type = var_146_pad_type_0, strides = var_146_strides_0, weight = var_121_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_146_cast_fp16")];
tensor<string, []> hidden_states_3_mode_0 = const()[name = tensor<string, []>("hidden_states_3_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_146_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> var_164_to_fp16 = const()[name = tensor<string, []>("op_164_to_fp16"), val = tensor<fp16, [1, 1280, 1, 1500]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10818880)))];
tensor<fp16, [1, 1280, 1, 1500]> inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_164_to_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
tensor<int32, []> var_178 = const()[name = tensor<string, []>("op_178"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_197_to_fp16 = const()[name = tensor<string, []>("op_197_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_197_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
tensor<fp16, [1280]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14658944)))];
tensor<fp16, [1280]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14661568)))];
tensor<fp16, [1280]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14664192)))];
tensor<fp16, [1280]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14666816)))];
tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_1_strides_0 = const()[name = tensor<string, []>("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_1_dilations_0 = const()[name = tensor<string, []>("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_1_groups_0 = const()[name = tensor<string, []>("query_1_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14669440)))];
tensor<fp16, [1280]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17946304)))];
tensor<fp16, [1, 1280, 1, 1500]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
tensor<string, []> key_1_pad_type_0 = const()[name = tensor<string, []>("key_1_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_1_strides_0 = const()[name = tensor<string, []>("key_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_1_pad_0 = const()[name = tensor<string, []>("key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_1_dilations_0 = const()[name = tensor<string, []>("key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_1_groups_0 = const()[name = tensor<string, []>("key_1_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17948928)))];
tensor<fp16, [1, 1280, 1, 1500]> key_1_cast_fp16 = conv(dilations = key_1_dilations_0, groups = key_1_groups_0, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
tensor<string, []> value_1_pad_type_0 = const()[name = tensor<string, []>("value_1_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_1_strides_0 = const()[name = tensor<string, []>("value_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_1_pad_0 = const()[name = tensor<string, []>("value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_1_dilations_0 = const()[name = tensor<string, []>("value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_1_groups_0 = const()[name = tensor<string, []>("value_1_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21225792)))];
tensor<fp16, [1280]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24502656)))];
tensor<fp16, [1, 1280, 1, 1500]> value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = value_1_dilations_0, groups = value_1_groups_0, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
tensor<int32, [4]> var_232 = const()[name = tensor<string, []>("op_232"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_1_cast_fp16 = reshape(shape = var_232, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")];
tensor<fp16, []> var_234_to_fp16 = const()[name = tensor<string, []>("op_234_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_235_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_234_to_fp16)[name = tensor<string, []>("op_235_cast_fp16")];
tensor<int32, [4]> var_236 = const()[name = tensor<string, []>("op_236"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_237_cast_fp16 = reshape(shape = var_236, x = key_1_cast_fp16)[name = tensor<string, []>("op_237_cast_fp16")];
tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_235_cast_fp16, y = var_237_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_240_cast_fp16 = softmax(axis = var_178, x = mh_w_1_cast_fp16)[name = tensor<string, []>("op_240_cast_fp16")];
tensor<int32, [4]> var_241 = const()[name = tensor<string, []>("op_241"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_242_cast_fp16 = reshape(shape = var_241, x = value_1_cast_fp16)[name = tensor<string, []>("op_242_cast_fp16")];
tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_242_cast_fp16, y = var_240_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
tensor<int32, [4]> var_245 = const()[name = tensor<string, []>("op_245"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_1_cast_fp16 = reshape(shape = var_245, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
tensor<string, []> obj_3_pad_type_0 = const()[name = tensor<string, []>("obj_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_3_strides_0 = const()[name = tensor<string, []>("obj_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_3_pad_0 = const()[name = tensor<string, []>("obj_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_3_dilations_0 = const()[name = tensor<string, []>("obj_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_3_groups_0 = const()[name = tensor<string, []>("obj_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24505280)))];
tensor<fp16, [1280]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27782144)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_3_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_3_dilations_0, groups = obj_3_groups_0, pad = obj_3_pad_0, pad_type = obj_3_pad_type_0, strides = obj_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_3_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_263_to_fp16 = const()[name = tensor<string, []>("op_263_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_263_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
tensor<fp16, [1280]> input_3_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_3_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27784768)))];
tensor<fp16, [1280]> input_3_beta_0_to_fp16 = const()[name = tensor<string, []>("input_3_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27787392)))];
tensor<fp16, []> input_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_3_cast_fp16 = batch_norm(beta = input_3_beta_0_to_fp16, epsilon = input_3_epsilon_0_to_fp16, gamma = input_3_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_5_strides_0 = const()[name = tensor<string, []>("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_5_dilations_0 = const()[name = tensor<string, []>("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_5_groups_0 = const()[name = tensor<string, []>("input_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27790016)))];
tensor<fp16, [5120]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40897280)))];
tensor<fp16, [1, 5120, 1, 1500]> input_5_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
tensor<string, []> input_7_mode_0 = const()[name = tensor<string, []>("input_7_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = tensor<string, []>("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = tensor<string, []>("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_5_groups_0 = const()[name = tensor<string, []>("hidden_states_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40907584)))];
tensor<fp16, [1280]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54014848)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_5_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
tensor<int32, []> var_296 = const()[name = tensor<string, []>("op_296"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_315_to_fp16 = const()[name = tensor<string, []>("op_315_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_315_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
tensor<fp16, [1280]> obj_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54017472)))];
tensor<fp16, [1280]> obj_5_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54020096)))];
tensor<fp16, []> obj_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("obj_5_cast_fp16")];
tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_3_strides_0 = const()[name = tensor<string, []>("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_3_dilations_0 = const()[name = tensor<string, []>("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_3_groups_0 = const()[name = tensor<string, []>("query_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54022720)))];
tensor<fp16, [1280]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57299584)))];
tensor<fp16, [1, 1280, 1, 1500]> query_3_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_3_strides_0 = const()[name = tensor<string, []>("key_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_3_dilations_0 = const()[name = tensor<string, []>("key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_3_groups_0 = const()[name = tensor<string, []>("key_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57302208)))];
tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_3_strides_0 = const()[name = tensor<string, []>("value_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_3_dilations_0 = const()[name = tensor<string, []>("value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_3_groups_0 = const()[name = tensor<string, []>("value_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60579072)))];
tensor<fp16, [1280]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63855936)))];
tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")];
tensor<int32, [4]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_3_cast_fp16 = reshape(shape = var_350, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")];
tensor<fp16, []> var_352_to_fp16 = const()[name = tensor<string, []>("op_352_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_353_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_352_to_fp16)[name = tensor<string, []>("op_353_cast_fp16")];
tensor<int32, [4]> var_354 = const()[name = tensor<string, []>("op_354"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_355_cast_fp16 = reshape(shape = var_354, x = key_3_cast_fp16)[name = tensor<string, []>("op_355_cast_fp16")];
tensor<bool, []> mh_w_3_transpose_x_0 = const()[name = tensor<string, []>("mh_w_3_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_3_transpose_y_0 = const()[name = tensor<string, []>("mh_w_3_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_3_cast_fp16 = matmul(transpose_x = mh_w_3_transpose_x_0, transpose_y = mh_w_3_transpose_y_0, x = var_353_cast_fp16, y = var_355_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_358_cast_fp16 = softmax(axis = var_296, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_358_cast_fp16")];
tensor<int32, [4]> var_359 = const()[name = tensor<string, []>("op_359"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_360_cast_fp16 = reshape(shape = var_359, x = value_3_cast_fp16)[name = tensor<string, []>("op_360_cast_fp16")];
tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_360_cast_fp16, y = var_358_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
tensor<int32, [4]> var_363 = const()[name = tensor<string, []>("op_363"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_9_cast_fp16 = reshape(shape = var_363, x = attn_3_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_7_strides_0 = const()[name = tensor<string, []>("obj_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_7_dilations_0 = const()[name = tensor<string, []>("obj_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_7_groups_0 = const()[name = tensor<string, []>("obj_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63858560)))];
tensor<fp16, [1280]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67135424)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_7_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_381_to_fp16 = const()[name = tensor<string, []>("op_381_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_381_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
tensor<fp16, [1280]> input_11_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_11_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67138048)))];
tensor<fp16, [1280]> input_11_beta_0_to_fp16 = const()[name = tensor<string, []>("input_11_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67140672)))];
tensor<fp16, []> input_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_11_cast_fp16 = batch_norm(beta = input_11_beta_0_to_fp16, epsilon = input_11_epsilon_0_to_fp16, gamma = input_11_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67143296)))];
tensor<fp16, [5120]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80250560)))];
tensor<fp16, [1, 5120, 1, 1500]> input_13_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
tensor<string, []> input_15_mode_0 = const()[name = tensor<string, []>("input_15_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = tensor<string, []>("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = tensor<string, []>("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_7_groups_0 = const()[name = tensor<string, []>("hidden_states_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80260864)))];
tensor<fp16, [1280]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93368128)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_7_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
tensor<int32, []> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_433_to_fp16 = const()[name = tensor<string, []>("op_433_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_433_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
tensor<fp16, [1280]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93370752)))];
tensor<fp16, [1280]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93373376)))];
tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_5_strides_0 = const()[name = tensor<string, []>("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_5_dilations_0 = const()[name = tensor<string, []>("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_5_groups_0 = const()[name = tensor<string, []>("query_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93376000)))];
tensor<fp16, [1280]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96652864)))];
tensor<fp16, [1, 1280, 1, 1500]> query_5_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
tensor<string, []> key_5_pad_type_0 = const()[name = tensor<string, []>("key_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_5_strides_0 = const()[name = tensor<string, []>("key_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_5_pad_0 = const()[name = tensor<string, []>("key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_5_dilations_0 = const()[name = tensor<string, []>("key_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_5_groups_0 = const()[name = tensor<string, []>("key_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96655488)))];
tensor<fp16, [1, 1280, 1, 1500]> key_5_cast_fp16 = conv(dilations = key_5_dilations_0, groups = key_5_groups_0, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
tensor<string, []> value_5_pad_type_0 = const()[name = tensor<string, []>("value_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_5_strides_0 = const()[name = tensor<string, []>("value_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_5_pad_0 = const()[name = tensor<string, []>("value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_5_dilations_0 = const()[name = tensor<string, []>("value_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_5_groups_0 = const()[name = tensor<string, []>("value_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99932352)))];
tensor<fp16, [1280]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103209216)))];
tensor<fp16, [1, 1280, 1, 1500]> value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = value_5_dilations_0, groups = value_5_groups_0, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
tensor<int32, [4]> var_468 = const()[name = tensor<string, []>("op_468"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_5_cast_fp16 = reshape(shape = var_468, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")];
tensor<fp16, []> var_470_to_fp16 = const()[name = tensor<string, []>("op_470_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_471_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_470_to_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
tensor<int32, [4]> var_472 = const()[name = tensor<string, []>("op_472"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_473_cast_fp16 = reshape(shape = var_472, x = key_5_cast_fp16)[name = tensor<string, []>("op_473_cast_fp16")];
tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_471_cast_fp16, y = var_473_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_476_cast_fp16 = softmax(axis = var_414, x = mh_w_5_cast_fp16)[name = tensor<string, []>("op_476_cast_fp16")];
tensor<int32, [4]> var_477 = const()[name = tensor<string, []>("op_477"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_478_cast_fp16 = reshape(shape = var_477, x = value_5_cast_fp16)[name = tensor<string, []>("op_478_cast_fp16")];
tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_478_cast_fp16, y = var_476_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
tensor<int32, [4]> var_481 = const()[name = tensor<string, []>("op_481"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_17_cast_fp16 = reshape(shape = var_481, x = attn_5_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_11_strides_0 = const()[name = tensor<string, []>("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_11_dilations_0 = const()[name = tensor<string, []>("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_11_groups_0 = const()[name = tensor<string, []>("obj_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103211840)))];
tensor<fp16, [1280]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106488704)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_11_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_499_to_fp16 = const()[name = tensor<string, []>("op_499_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_499_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
tensor<fp16, [1280]> input_19_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_19_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106491328)))];
tensor<fp16, [1280]> input_19_beta_0_to_fp16 = const()[name = tensor<string, []>("input_19_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106493952)))];
tensor<fp16, []> input_19_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_19_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_19_cast_fp16 = batch_norm(beta = input_19_beta_0_to_fp16, epsilon = input_19_epsilon_0_to_fp16, gamma = input_19_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
tensor<string, []> input_21_pad_type_0 = const()[name = tensor<string, []>("input_21_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_21_strides_0 = const()[name = tensor<string, []>("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_21_pad_0 = const()[name = tensor<string, []>("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_21_dilations_0 = const()[name = tensor<string, []>("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_21_groups_0 = const()[name = tensor<string, []>("input_21_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106496576)))];
tensor<fp16, [5120]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119603840)))];
tensor<fp16, [1, 5120, 1, 1500]> input_21_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
tensor<string, []> input_23_mode_0 = const()[name = tensor<string, []>("input_23_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = tensor<string, []>("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = tensor<string, []>("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_9_groups_0 = const()[name = tensor<string, []>("hidden_states_9_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119614144)))];
tensor<fp16, [1280]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132721408)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_9_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
tensor<int32, []> var_532 = const()[name = tensor<string, []>("op_532"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_551_to_fp16 = const()[name = tensor<string, []>("op_551_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_551_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
tensor<fp16, [1280]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132724032)))];
tensor<fp16, [1280]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132726656)))];
tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
tensor<string, []> query_7_pad_type_0 = const()[name = tensor<string, []>("query_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_7_strides_0 = const()[name = tensor<string, []>("query_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_7_pad_0 = const()[name = tensor<string, []>("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_7_dilations_0 = const()[name = tensor<string, []>("query_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_7_groups_0 = const()[name = tensor<string, []>("query_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132729280)))];
tensor<fp16, [1280]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136006144)))];
tensor<fp16, [1, 1280, 1, 1500]> query_7_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_7_strides_0 = const()[name = tensor<string, []>("key_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_7_pad_0 = const()[name = tensor<string, []>("key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_7_dilations_0 = const()[name = tensor<string, []>("key_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_7_groups_0 = const()[name = tensor<string, []>("key_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136008768)))];
tensor<fp16, [1, 1280, 1, 1500]> key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
tensor<string, []> value_7_pad_type_0 = const()[name = tensor<string, []>("value_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_7_strides_0 = const()[name = tensor<string, []>("value_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_7_pad_0 = const()[name = tensor<string, []>("value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_7_dilations_0 = const()[name = tensor<string, []>("value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_7_groups_0 = const()[name = tensor<string, []>("value_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139285632)))];
tensor<fp16, [1280]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142562496)))];
tensor<fp16, [1, 1280, 1, 1500]> value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")];
tensor<int32, [4]> var_586 = const()[name = tensor<string, []>("op_586"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_7_cast_fp16 = reshape(shape = var_586, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")];
tensor<fp16, []> var_588_to_fp16 = const()[name = tensor<string, []>("op_588_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_589_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_588_to_fp16)[name = tensor<string, []>("op_589_cast_fp16")];
tensor<int32, [4]> var_590 = const()[name = tensor<string, []>("op_590"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_591_cast_fp16 = reshape(shape = var_590, x = key_7_cast_fp16)[name = tensor<string, []>("op_591_cast_fp16")];
tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_589_cast_fp16, y = var_591_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_594_cast_fp16 = softmax(axis = var_532, x = mh_w_7_cast_fp16)[name = tensor<string, []>("op_594_cast_fp16")];
tensor<int32, [4]> var_595 = const()[name = tensor<string, []>("op_595"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_596_cast_fp16 = reshape(shape = var_595, x = value_7_cast_fp16)[name = tensor<string, []>("op_596_cast_fp16")];
tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_596_cast_fp16, y = var_594_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
tensor<int32, [4]> var_599 = const()[name = tensor<string, []>("op_599"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_25_cast_fp16 = reshape(shape = var_599, x = attn_7_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
tensor<string, []> obj_15_pad_type_0 = const()[name = tensor<string, []>("obj_15_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_15_strides_0 = const()[name = tensor<string, []>("obj_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_15_pad_0 = const()[name = tensor<string, []>("obj_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_15_dilations_0 = const()[name = tensor<string, []>("obj_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_15_groups_0 = const()[name = tensor<string, []>("obj_15_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142565120)))];
tensor<fp16, [1280]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145841984)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_15_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_15_dilations_0, groups = obj_15_groups_0, pad = obj_15_pad_0, pad_type = obj_15_pad_type_0, strides = obj_15_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_15_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_617_to_fp16 = const()[name = tensor<string, []>("op_617_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_617_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
tensor<fp16, [1280]> input_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_27_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145844608)))];
tensor<fp16, [1280]> input_27_beta_0_to_fp16 = const()[name = tensor<string, []>("input_27_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145847232)))];
tensor<fp16, []> input_27_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_27_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_27_cast_fp16 = batch_norm(beta = input_27_beta_0_to_fp16, epsilon = input_27_epsilon_0_to_fp16, gamma = input_27_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145849856)))];
tensor<fp16, [5120]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158957120)))];
tensor<fp16, [1, 5120, 1, 1500]> input_29_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
tensor<string, []> input_31_mode_0 = const()[name = tensor<string, []>("input_31_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
tensor<string, []> hidden_states_11_pad_type_0 = const()[name = tensor<string, []>("hidden_states_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_11_strides_0 = const()[name = tensor<string, []>("hidden_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = tensor<string, []>("hidden_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_11_dilations_0 = const()[name = tensor<string, []>("hidden_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_11_groups_0 = const()[name = tensor<string, []>("hidden_states_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158967424)))];
tensor<fp16, [1280]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172074688)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_11_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("hidden_states_11_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
tensor<int32, []> var_650 = const()[name = tensor<string, []>("op_650"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_669_to_fp16 = const()[name = tensor<string, []>("op_669_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_669_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
tensor<fp16, [1280]> obj_17_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_17_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172077312)))];
tensor<fp16, [1280]> obj_17_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_17_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172079936)))];
tensor<fp16, []> obj_17_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_17_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_17_cast_fp16 = batch_norm(beta = obj_17_beta_0_to_fp16, epsilon = obj_17_epsilon_0_to_fp16, gamma = obj_17_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("obj_17_cast_fp16")];
tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_9_strides_0 = const()[name = tensor<string, []>("query_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_9_dilations_0 = const()[name = tensor<string, []>("query_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_9_groups_0 = const()[name = tensor<string, []>("query_9_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172082560)))];
tensor<fp16, [1280]> layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175359424)))];
tensor<fp16, [1, 1280, 1, 1500]> query_9_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
tensor<string, []> key_9_pad_type_0 = const()[name = tensor<string, []>("key_9_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_9_strides_0 = const()[name = tensor<string, []>("key_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_9_pad_0 = const()[name = tensor<string, []>("key_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_9_dilations_0 = const()[name = tensor<string, []>("key_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_9_groups_0 = const()[name = tensor<string, []>("key_9_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175362048)))];
tensor<fp16, [1, 1280, 1, 1500]> key_9_cast_fp16 = conv(dilations = key_9_dilations_0, groups = key_9_groups_0, pad = key_9_pad_0, pad_type = key_9_pad_type_0, strides = key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
tensor<string, []> value_9_pad_type_0 = const()[name = tensor<string, []>("value_9_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_9_strides_0 = const()[name = tensor<string, []>("value_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_9_pad_0 = const()[name = tensor<string, []>("value_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_9_dilations_0 = const()[name = tensor<string, []>("value_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_9_groups_0 = const()[name = tensor<string, []>("value_9_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178638912)))];
tensor<fp16, [1280]> layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181915776)))];
tensor<fp16, [1, 1280, 1, 1500]> value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = value_9_dilations_0, groups = value_9_groups_0, pad = value_9_pad_0, pad_type = value_9_pad_type_0, strides = value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
tensor<int32, [4]> var_704 = const()[name = tensor<string, []>("op_704"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_9_cast_fp16 = reshape(shape = var_704, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")];
tensor<fp16, []> var_706_to_fp16 = const()[name = tensor<string, []>("op_706_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_707_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_706_to_fp16)[name = tensor<string, []>("op_707_cast_fp16")];
tensor<int32, [4]> var_708 = const()[name = tensor<string, []>("op_708"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_709_cast_fp16 = reshape(shape = var_708, x = key_9_cast_fp16)[name = tensor<string, []>("op_709_cast_fp16")];
tensor<bool, []> mh_w_9_transpose_x_0 = const()[name = tensor<string, []>("mh_w_9_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_9_transpose_y_0 = const()[name = tensor<string, []>("mh_w_9_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_707_cast_fp16, y = var_709_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_712_cast_fp16 = softmax(axis = var_650, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_712_cast_fp16")];
tensor<int32, [4]> var_713 = const()[name = tensor<string, []>("op_713"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_714_cast_fp16 = reshape(shape = var_713, x = value_9_cast_fp16)[name = tensor<string, []>("op_714_cast_fp16")];
tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_714_cast_fp16, y = var_712_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
tensor<int32, [4]> var_717 = const()[name = tensor<string, []>("op_717"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_33_cast_fp16 = reshape(shape = var_717, x = attn_9_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
tensor<string, []> obj_19_pad_type_0 = const()[name = tensor<string, []>("obj_19_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_19_strides_0 = const()[name = tensor<string, []>("obj_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_19_pad_0 = const()[name = tensor<string, []>("obj_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_19_dilations_0 = const()[name = tensor<string, []>("obj_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_19_groups_0 = const()[name = tensor<string, []>("obj_19_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181918400)))];
tensor<fp16, [1280]> layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185195264)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_19_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_19_dilations_0, groups = obj_19_groups_0, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = obj_19_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = obj_19_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_735_to_fp16 = const()[name = tensor<string, []>("op_735_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_735_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
tensor<fp16, [1280]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185197888)))];
tensor<fp16, [1280]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185200512)))];
tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_37_strides_0 = const()[name = tensor<string, []>("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_37_dilations_0 = const()[name = tensor<string, []>("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_37_groups_0 = const()[name = tensor<string, []>("input_37_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_4_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185203136)))];
tensor<fp16, [5120]> layers_4_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198310400)))];
tensor<fp16, [1, 5120, 1, 1500]> input_37_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_4_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
tensor<string, []> input_39_mode_0 = const()[name = tensor<string, []>("input_39_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
tensor<string, []> hidden_states_13_pad_type_0 = const()[name = tensor<string, []>("hidden_states_13_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = tensor<string, []>("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = tensor<string, []>("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = tensor<string, []>("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_13_groups_0 = const()[name = tensor<string, []>("hidden_states_13_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_4_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198320704)))];
tensor<fp16, [1280]> layers_4_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211427968)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_13_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
tensor<int32, []> var_768 = const()[name = tensor<string, []>("op_768"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_787_to_fp16 = const()[name = tensor<string, []>("op_787_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_787_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
tensor<fp16, [1280]> obj_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_21_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211430592)))];
tensor<fp16, [1280]> obj_21_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_21_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211433216)))];
tensor<fp16, []> obj_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
tensor<string, []> query_11_pad_type_0 = const()[name = tensor<string, []>("query_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_11_strides_0 = const()[name = tensor<string, []>("query_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_11_pad_0 = const()[name = tensor<string, []>("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_11_dilations_0 = const()[name = tensor<string, []>("query_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_11_groups_0 = const()[name = tensor<string, []>("query_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211435840)))];
tensor<fp16, [1280]> layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214712704)))];
tensor<fp16, [1, 1280, 1, 1500]> query_11_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_11_strides_0 = const()[name = tensor<string, []>("key_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_11_dilations_0 = const()[name = tensor<string, []>("key_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_11_groups_0 = const()[name = tensor<string, []>("key_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214715328)))];
tensor<fp16, [1, 1280, 1, 1500]> key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
tensor<string, []> value_11_pad_type_0 = const()[name = tensor<string, []>("value_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_11_strides_0 = const()[name = tensor<string, []>("value_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_11_pad_0 = const()[name = tensor<string, []>("value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_11_dilations_0 = const()[name = tensor<string, []>("value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_11_groups_0 = const()[name = tensor<string, []>("value_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217992192)))];
tensor<fp16, [1280]> layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221269056)))];
tensor<fp16, [1, 1280, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")];
tensor<int32, [4]> var_822 = const()[name = tensor<string, []>("op_822"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_11_cast_fp16 = reshape(shape = var_822, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")];
tensor<fp16, []> var_824_to_fp16 = const()[name = tensor<string, []>("op_824_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_825_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_824_to_fp16)[name = tensor<string, []>("op_825_cast_fp16")];
tensor<int32, [4]> var_826 = const()[name = tensor<string, []>("op_826"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_827_cast_fp16 = reshape(shape = var_826, x = key_11_cast_fp16)[name = tensor<string, []>("op_827_cast_fp16")];
tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_825_cast_fp16, y = var_827_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_830_cast_fp16 = softmax(axis = var_768, x = mh_w_11_cast_fp16)[name = tensor<string, []>("op_830_cast_fp16")];
tensor<int32, [4]> var_831 = const()[name = tensor<string, []>("op_831"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_832_cast_fp16 = reshape(shape = var_831, x = value_11_cast_fp16)[name = tensor<string, []>("op_832_cast_fp16")];
tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_832_cast_fp16, y = var_830_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
tensor<int32, [4]> var_835 = const()[name = tensor<string, []>("op_835"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_41_cast_fp16 = reshape(shape = var_835, x = attn_11_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
tensor<string, []> obj_23_pad_type_0 = const()[name = tensor<string, []>("obj_23_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_23_strides_0 = const()[name = tensor<string, []>("obj_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_23_pad_0 = const()[name = tensor<string, []>("obj_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_23_dilations_0 = const()[name = tensor<string, []>("obj_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_23_groups_0 = const()[name = tensor<string, []>("obj_23_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221271680)))];
tensor<fp16, [1280]> layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224548544)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_23_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_23_dilations_0, groups = obj_23_groups_0, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = obj_23_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_23_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_853_to_fp16 = const()[name = tensor<string, []>("op_853_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_853_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
tensor<fp16, [1280]> input_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_43_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224551168)))];
tensor<fp16, [1280]> input_43_beta_0_to_fp16 = const()[name = tensor<string, []>("input_43_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224553792)))];
tensor<fp16, []> input_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_43_cast_fp16 = batch_norm(beta = input_43_beta_0_to_fp16, epsilon = input_43_epsilon_0_to_fp16, gamma = input_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
tensor<string, []> input_45_pad_type_0 = const()[name = tensor<string, []>("input_45_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_45_strides_0 = const()[name = tensor<string, []>("input_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_45_pad_0 = const()[name = tensor<string, []>("input_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_45_dilations_0 = const()[name = tensor<string, []>("input_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_45_groups_0 = const()[name = tensor<string, []>("input_45_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_5_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224556416)))];
tensor<fp16, [5120]> layers_5_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237663680)))];
tensor<fp16, [1, 5120, 1, 1500]> input_45_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
tensor<string, []> input_47_mode_0 = const()[name = tensor<string, []>("input_47_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
tensor<string, []> hidden_states_15_pad_type_0 = const()[name = tensor<string, []>("hidden_states_15_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_15_strides_0 = const()[name = tensor<string, []>("hidden_states_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = tensor<string, []>("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_15_dilations_0 = const()[name = tensor<string, []>("hidden_states_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_15_groups_0 = const()[name = tensor<string, []>("hidden_states_15_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_5_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237673984)))];
tensor<fp16, [1280]> layers_5_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250781248)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_15_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("hidden_states_15_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor<string, []>("inputs_25_cast_fp16")];
tensor<int32, []> var_886 = const()[name = tensor<string, []>("op_886"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_25_axes_0 = const()[name = tensor<string, []>("out_25_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_905_to_fp16 = const()[name = tensor<string, []>("op_905_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_905_to_fp16, x = inputs_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
tensor<fp16, [1280]> obj_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250783872)))];
tensor<fp16, [1280]> obj_25_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250786496)))];
tensor<fp16, []> obj_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_13_strides_0 = const()[name = tensor<string, []>("query_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_13_dilations_0 = const()[name = tensor<string, []>("query_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_13_groups_0 = const()[name = tensor<string, []>("query_13_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250789120)))];
tensor<fp16, [1280]> layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254065984)))];
tensor<fp16, [1, 1280, 1, 1500]> query_13_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
tensor<string, []> key_13_pad_type_0 = const()[name = tensor<string, []>("key_13_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_13_strides_0 = const()[name = tensor<string, []>("key_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_13_pad_0 = const()[name = tensor<string, []>("key_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_13_dilations_0 = const()[name = tensor<string, []>("key_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_13_groups_0 = const()[name = tensor<string, []>("key_13_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254068608)))];
tensor<fp16, [1, 1280, 1, 1500]> key_13_cast_fp16 = conv(dilations = key_13_dilations_0, groups = key_13_groups_0, pad = key_13_pad_0, pad_type = key_13_pad_type_0, strides = key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
tensor<string, []> value_13_pad_type_0 = const()[name = tensor<string, []>("value_13_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_13_strides_0 = const()[name = tensor<string, []>("value_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_13_pad_0 = const()[name = tensor<string, []>("value_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_13_dilations_0 = const()[name = tensor<string, []>("value_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_13_groups_0 = const()[name = tensor<string, []>("value_13_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257345472)))];
tensor<fp16, [1280]> layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260622336)))];
tensor<fp16, [1, 1280, 1, 1500]> value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = value_13_dilations_0, groups = value_13_groups_0, pad = value_13_pad_0, pad_type = value_13_pad_type_0, strides = value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
tensor<int32, [4]> var_940 = const()[name = tensor<string, []>("op_940"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_13_cast_fp16 = reshape(shape = var_940, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")];
tensor<fp16, []> var_942_to_fp16 = const()[name = tensor<string, []>("op_942_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_943_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_942_to_fp16)[name = tensor<string, []>("op_943_cast_fp16")];
tensor<int32, [4]> var_944 = const()[name = tensor<string, []>("op_944"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_945_cast_fp16 = reshape(shape = var_944, x = key_13_cast_fp16)[name = tensor<string, []>("op_945_cast_fp16")];
tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_943_cast_fp16, y = var_945_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_948_cast_fp16 = softmax(axis = var_886, x = mh_w_13_cast_fp16)[name = tensor<string, []>("op_948_cast_fp16")];
tensor<int32, [4]> var_949 = const()[name = tensor<string, []>("op_949"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_950_cast_fp16 = reshape(shape = var_949, x = value_13_cast_fp16)[name = tensor<string, []>("op_950_cast_fp16")];
tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_950_cast_fp16, y = var_948_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
tensor<int32, [4]> var_953 = const()[name = tensor<string, []>("op_953"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_49_cast_fp16 = reshape(shape = var_953, x = attn_13_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
tensor<string, []> obj_27_pad_type_0 = const()[name = tensor<string, []>("obj_27_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_27_strides_0 = const()[name = tensor<string, []>("obj_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_27_pad_0 = const()[name = tensor<string, []>("obj_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_27_dilations_0 = const()[name = tensor<string, []>("obj_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_27_groups_0 = const()[name = tensor<string, []>("obj_27_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260624960)))];
tensor<fp16, [1280]> layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263901824)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_27_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_27_dilations_0, groups = obj_27_groups_0, pad = obj_27_pad_0, pad_type = obj_27_pad_type_0, strides = obj_27_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("inputs_27_cast_fp16")];
tensor<int32, [1]> out_27_axes_0 = const()[name = tensor<string, []>("out_27_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_971_to_fp16 = const()[name = tensor<string, []>("op_971_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_971_to_fp16, x = inputs_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
tensor<fp16, [1280]> input_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_51_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263904448)))];
tensor<fp16, [1280]> input_51_beta_0_to_fp16 = const()[name = tensor<string, []>("input_51_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263907072)))];
tensor<fp16, []> input_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
tensor<string, []> input_53_pad_type_0 = const()[name = tensor<string, []>("input_53_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_53_strides_0 = const()[name = tensor<string, []>("input_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_53_pad_0 = const()[name = tensor<string, []>("input_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_53_dilations_0 = const()[name = tensor<string, []>("input_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_53_groups_0 = const()[name = tensor<string, []>("input_53_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_6_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263909696)))];
tensor<fp16, [5120]> layers_6_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277016960)))];
tensor<fp16, [1, 5120, 1, 1500]> input_53_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
tensor<string, []> input_55_mode_0 = const()[name = tensor<string, []>("input_55_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = input_53_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
tensor<string, []> hidden_states_17_pad_type_0 = const()[name = tensor<string, []>("hidden_states_17_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = tensor<string, []>("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = tensor<string, []>("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = tensor<string, []>("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_17_groups_0 = const()[name = tensor<string, []>("hidden_states_17_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_6_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277027264)))];
tensor<fp16, [1280]> layers_6_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290134528)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_17_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("hidden_states_17_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor<string, []>("inputs_29_cast_fp16")];
tensor<int32, []> var_1004 = const()[name = tensor<string, []>("op_1004"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_29_axes_0 = const()[name = tensor<string, []>("out_29_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1023_to_fp16 = const()[name = tensor<string, []>("op_1023_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1023_to_fp16, x = inputs_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
tensor<fp16, [1280]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290137152)))];
tensor<fp16, [1280]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290139776)))];
tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
tensor<string, []> query_15_pad_type_0 = const()[name = tensor<string, []>("query_15_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_15_strides_0 = const()[name = tensor<string, []>("query_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_15_pad_0 = const()[name = tensor<string, []>("query_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_15_dilations_0 = const()[name = tensor<string, []>("query_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_15_groups_0 = const()[name = tensor<string, []>("query_15_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290142400)))];
tensor<fp16, [1280]> layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293419264)))];
tensor<fp16, [1, 1280, 1, 1500]> query_15_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
tensor<string, []> key_15_pad_type_0 = const()[name = tensor<string, []>("key_15_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_15_strides_0 = const()[name = tensor<string, []>("key_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_15_pad_0 = const()[name = tensor<string, []>("key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_15_dilations_0 = const()[name = tensor<string, []>("key_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_15_groups_0 = const()[name = tensor<string, []>("key_15_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293421888)))];
tensor<fp16, [1, 1280, 1, 1500]> key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
tensor<string, []> value_15_pad_type_0 = const()[name = tensor<string, []>("value_15_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_15_strides_0 = const()[name = tensor<string, []>("value_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_15_pad_0 = const()[name = tensor<string, []>("value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_15_dilations_0 = const()[name = tensor<string, []>("value_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_15_groups_0 = const()[name = tensor<string, []>("value_15_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296698752)))];
tensor<fp16, [1280]> layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299975616)))];
tensor<fp16, [1, 1280, 1, 1500]> value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("value_15_cast_fp16")];
tensor<int32, [4]> var_1058 = const()[name = tensor<string, []>("op_1058"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_15_cast_fp16 = reshape(shape = var_1058, x = query_15_cast_fp16)[name = tensor<string, []>("mh_q_15_cast_fp16")];
tensor<fp16, []> var_1060_to_fp16 = const()[name = tensor<string, []>("op_1060_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_1061_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_1060_to_fp16)[name = tensor<string, []>("op_1061_cast_fp16")];
tensor<int32, [4]> var_1062 = const()[name = tensor<string, []>("op_1062"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1063_cast_fp16 = reshape(shape = var_1062, x = key_15_cast_fp16)[name = tensor<string, []>("op_1063_cast_fp16")];
tensor<bool, []> mh_w_15_transpose_x_0 = const()[name = tensor<string, []>("mh_w_15_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_15_transpose_y_0 = const()[name = tensor<string, []>("mh_w_15_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_15_cast_fp16 = matmul(transpose_x = mh_w_15_transpose_x_0, transpose_y = mh_w_15_transpose_y_0, x = var_1061_cast_fp16, y = var_1063_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_1066_cast_fp16 = softmax(axis = var_1004, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_1066_cast_fp16")];
tensor<int32, [4]> var_1067 = const()[name = tensor<string, []>("op_1067"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1068_cast_fp16 = reshape(shape = var_1067, x = value_15_cast_fp16)[name = tensor<string, []>("op_1068_cast_fp16")];
tensor<bool, []> attn_15_transpose_x_0 = const()[name = tensor<string, []>("attn_15_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_15_transpose_y_0 = const()[name = tensor<string, []>("attn_15_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_1068_cast_fp16, y = var_1066_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
tensor<int32, [4]> var_1071 = const()[name = tensor<string, []>("op_1071"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_57_cast_fp16 = reshape(shape = var_1071, x = attn_15_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
tensor<string, []> obj_31_pad_type_0 = const()[name = tensor<string, []>("obj_31_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_31_strides_0 = const()[name = tensor<string, []>("obj_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_31_pad_0 = const()[name = tensor<string, []>("obj_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_31_dilations_0 = const()[name = tensor<string, []>("obj_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_31_groups_0 = const()[name = tensor<string, []>("obj_31_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299978240)))];
tensor<fp16, [1280]> layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303255104)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_31_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_31_dilations_0, groups = obj_31_groups_0, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = obj_31_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("obj_31_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = obj_31_cast_fp16)[name = tensor<string, []>("inputs_31_cast_fp16")];
tensor<int32, [1]> out_31_axes_0 = const()[name = tensor<string, []>("out_31_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1089_to_fp16 = const()[name = tensor<string, []>("op_1089_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1089_to_fp16, x = inputs_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
tensor<fp16, [1280]> input_59_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_59_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303257728)))];
tensor<fp16, [1280]> input_59_beta_0_to_fp16 = const()[name = tensor<string, []>("input_59_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303260352)))];
tensor<fp16, []> input_59_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_59_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_59_cast_fp16 = batch_norm(beta = input_59_beta_0_to_fp16, epsilon = input_59_epsilon_0_to_fp16, gamma = input_59_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
tensor<string, []> input_61_pad_type_0 = const()[name = tensor<string, []>("input_61_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_61_strides_0 = const()[name = tensor<string, []>("input_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_61_pad_0 = const()[name = tensor<string, []>("input_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_61_dilations_0 = const()[name = tensor<string, []>("input_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_61_groups_0 = const()[name = tensor<string, []>("input_61_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_7_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303262976)))];
tensor<fp16, [5120]> layers_7_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316370240)))];
tensor<fp16, [1, 5120, 1, 1500]> input_61_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_61_dilations_0, groups = input_61_groups_0, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = input_61_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
tensor<string, []> input_63_mode_0 = const()[name = tensor<string, []>("input_63_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_63_cast_fp16 = gelu(mode = input_63_mode_0, x = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
tensor<string, []> hidden_states_19_pad_type_0 = const()[name = tensor<string, []>("hidden_states_19_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_19_strides_0 = const()[name = tensor<string, []>("hidden_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_19_pad_0 = const()[name = tensor<string, []>("hidden_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_19_dilations_0 = const()[name = tensor<string, []>("hidden_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_19_groups_0 = const()[name = tensor<string, []>("hidden_states_19_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_7_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316380544)))];
tensor<fp16, [1280]> layers_7_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329487808)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_19_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("hidden_states_19_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")];
tensor<int32, []> var_1122 = const()[name = tensor<string, []>("op_1122"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_33_axes_0 = const()[name = tensor<string, []>("out_33_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1141_to_fp16 = const()[name = tensor<string, []>("op_1141_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1141_to_fp16, x = inputs_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
tensor<fp16, [1280]> obj_33_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329490432)))];
tensor<fp16, [1280]> obj_33_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_33_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329493056)))];
tensor<fp16, []> obj_33_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_33_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor<string, []>("obj_33_cast_fp16")];
tensor<string, []> query_17_pad_type_0 = const()[name = tensor<string, []>("query_17_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_17_strides_0 = const()[name = tensor<string, []>("query_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_17_pad_0 = const()[name = tensor<string, []>("query_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_17_dilations_0 = const()[name = tensor<string, []>("query_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_17_groups_0 = const()[name = tensor<string, []>("query_17_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329495680)))];
tensor<fp16, [1280]> layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332772544)))];
tensor<fp16, [1, 1280, 1, 1500]> query_17_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
tensor<string, []> key_17_pad_type_0 = const()[name = tensor<string, []>("key_17_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_17_strides_0 = const()[name = tensor<string, []>("key_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_17_pad_0 = const()[name = tensor<string, []>("key_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_17_dilations_0 = const()[name = tensor<string, []>("key_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_17_groups_0 = const()[name = tensor<string, []>("key_17_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332775168)))];
tensor<fp16, [1, 1280, 1, 1500]> key_17_cast_fp16 = conv(dilations = key_17_dilations_0, groups = key_17_groups_0, pad = key_17_pad_0, pad_type = key_17_pad_type_0, strides = key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
tensor<string, []> value_17_pad_type_0 = const()[name = tensor<string, []>("value_17_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_17_strides_0 = const()[name = tensor<string, []>("value_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_17_pad_0 = const()[name = tensor<string, []>("value_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_17_dilations_0 = const()[name = tensor<string, []>("value_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_17_groups_0 = const()[name = tensor<string, []>("value_17_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336052032)))];
tensor<fp16, [1280]> layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339328896)))];
tensor<fp16, [1, 1280, 1, 1500]> value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = value_17_dilations_0, groups = value_17_groups_0, pad = value_17_pad_0, pad_type = value_17_pad_type_0, strides = value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")];
tensor<int32, [4]> var_1176 = const()[name = tensor<string, []>("op_1176"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_17_cast_fp16 = reshape(shape = var_1176, x = query_17_cast_fp16)[name = tensor<string, []>("mh_q_17_cast_fp16")];
tensor<fp16, []> var_1178_to_fp16 = const()[name = tensor<string, []>("op_1178_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_1179_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1178_to_fp16)[name = tensor<string, []>("op_1179_cast_fp16")];
tensor<int32, [4]> var_1180 = const()[name = tensor<string, []>("op_1180"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1181_cast_fp16 = reshape(shape = var_1180, x = key_17_cast_fp16)[name = tensor<string, []>("op_1181_cast_fp16")];
tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1179_cast_fp16, y = var_1181_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_1184_cast_fp16 = softmax(axis = var_1122, x = mh_w_17_cast_fp16)[name = tensor<string, []>("op_1184_cast_fp16")];
tensor<int32, [4]> var_1185 = const()[name = tensor<string, []>("op_1185"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1186_cast_fp16 = reshape(shape = var_1185, x = value_17_cast_fp16)[name = tensor<string, []>("op_1186_cast_fp16")];
tensor<bool, []> attn_17_transpose_x_0 = const()[name = tensor<string, []>("attn_17_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_17_transpose_y_0 = const()[name = tensor<string, []>("attn_17_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1186_cast_fp16, y = var_1184_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
tensor<int32, [4]> var_1189 = const()[name = tensor<string, []>("op_1189"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_65_cast_fp16 = reshape(shape = var_1189, x = attn_17_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_35_strides_0 = const()[name = tensor<string, []>("obj_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_35_dilations_0 = const()[name = tensor<string, []>("obj_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_35_groups_0 = const()[name = tensor<string, []>("obj_35_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339331520)))];
tensor<fp16, [1280]> layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342608384)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_35_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_35_cast_fp16")];
tensor<int32, [1]> out_35_axes_0 = const()[name = tensor<string, []>("out_35_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1207_to_fp16 = const()[name = tensor<string, []>("op_1207_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1207_to_fp16, x = inputs_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
tensor<fp16, [1280]> input_67_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_67_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342611008)))];
tensor<fp16, [1280]> input_67_beta_0_to_fp16 = const()[name = tensor<string, []>("input_67_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342613632)))];
tensor<fp16, []> input_67_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_67_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_67_cast_fp16 = batch_norm(beta = input_67_beta_0_to_fp16, epsilon = input_67_epsilon_0_to_fp16, gamma = input_67_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
tensor<string, []> input_69_pad_type_0 = const()[name = tensor<string, []>("input_69_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_69_strides_0 = const()[name = tensor<string, []>("input_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_69_pad_0 = const()[name = tensor<string, []>("input_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_69_dilations_0 = const()[name = tensor<string, []>("input_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_69_groups_0 = const()[name = tensor<string, []>("input_69_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_8_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342616256)))];
tensor<fp16, [5120]> layers_8_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355723520)))];
tensor<fp16, [1, 5120, 1, 1500]> input_69_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
tensor<string, []> input_71_mode_0 = const()[name = tensor<string, []>("input_71_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
tensor<string, []> hidden_states_21_pad_type_0 = const()[name = tensor<string, []>("hidden_states_21_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_21_strides_0 = const()[name = tensor<string, []>("hidden_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_21_pad_0 = const()[name = tensor<string, []>("hidden_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_21_dilations_0 = const()[name = tensor<string, []>("hidden_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_21_groups_0 = const()[name = tensor<string, []>("hidden_states_21_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_8_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355733824)))];
tensor<fp16, [1280]> layers_8_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368841088)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_21_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("hidden_states_21_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor<string, []>("inputs_37_cast_fp16")];
tensor<int32, []> var_1240 = const()[name = tensor<string, []>("op_1240"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_37_axes_0 = const()[name = tensor<string, []>("out_37_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1259_to_fp16 = const()[name = tensor<string, []>("op_1259_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1259_to_fp16, x = inputs_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
tensor<fp16, [1280]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368843712)))];
tensor<fp16, [1280]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368846336)))];
tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
tensor<string, []> query_19_pad_type_0 = const()[name = tensor<string, []>("query_19_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_19_strides_0 = const()[name = tensor<string, []>("query_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_19_pad_0 = const()[name = tensor<string, []>("query_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_19_dilations_0 = const()[name = tensor<string, []>("query_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_19_groups_0 = const()[name = tensor<string, []>("query_19_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368848960)))];
tensor<fp16, [1280]> layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372125824)))];
tensor<fp16, [1, 1280, 1, 1500]> query_19_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
tensor<string, []> key_19_pad_type_0 = const()[name = tensor<string, []>("key_19_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_19_strides_0 = const()[name = tensor<string, []>("key_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_19_pad_0 = const()[name = tensor<string, []>("key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_19_dilations_0 = const()[name = tensor<string, []>("key_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_19_groups_0 = const()[name = tensor<string, []>("key_19_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372128448)))];
tensor<fp16, [1, 1280, 1, 1500]> key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
tensor<string, []> value_19_pad_type_0 = const()[name = tensor<string, []>("value_19_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_19_strides_0 = const()[name = tensor<string, []>("value_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_19_pad_0 = const()[name = tensor<string, []>("value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_19_dilations_0 = const()[name = tensor<string, []>("value_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_19_groups_0 = const()[name = tensor<string, []>("value_19_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375405312)))];
tensor<fp16, [1280]> layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378682176)))];
tensor<fp16, [1, 1280, 1, 1500]> value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("value_19_cast_fp16")];
tensor<int32, [4]> var_1294 = const()[name = tensor<string, []>("op_1294"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_19_cast_fp16 = reshape(shape = var_1294, x = query_19_cast_fp16)[name = tensor<string, []>("mh_q_19_cast_fp16")];
tensor<fp16, []> var_1296_to_fp16 = const()[name = tensor<string, []>("op_1296_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_1297_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1296_to_fp16)[name = tensor<string, []>("op_1297_cast_fp16")];
tensor<int32, [4]> var_1298 = const()[name = tensor<string, []>("op_1298"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1299_cast_fp16 = reshape(shape = var_1298, x = key_19_cast_fp16)[name = tensor<string, []>("op_1299_cast_fp16")];
tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1297_cast_fp16, y = var_1299_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_1302_cast_fp16 = softmax(axis = var_1240, x = mh_w_19_cast_fp16)[name = tensor<string, []>("op_1302_cast_fp16")];
tensor<int32, [4]> var_1303 = const()[name = tensor<string, []>("op_1303"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1304_cast_fp16 = reshape(shape = var_1303, x = value_19_cast_fp16)[name = tensor<string, []>("op_1304_cast_fp16")];
tensor<bool, []> attn_19_transpose_x_0 = const()[name = tensor<string, []>("attn_19_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_19_transpose_y_0 = const()[name = tensor<string, []>("attn_19_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1304_cast_fp16, y = var_1302_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
tensor<int32, [4]> var_1307 = const()[name = tensor<string, []>("op_1307"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_73_cast_fp16 = reshape(shape = var_1307, x = attn_19_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
tensor<string, []> obj_39_pad_type_0 = const()[name = tensor<string, []>("obj_39_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_39_strides_0 = const()[name = tensor<string, []>("obj_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_39_pad_0 = const()[name = tensor<string, []>("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_39_dilations_0 = const()[name = tensor<string, []>("obj_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_39_groups_0 = const()[name = tensor<string, []>("obj_39_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378684800)))];
tensor<fp16, [1280]> layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381961664)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_39_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_39_cast_fp16")];
tensor<int32, [1]> out_39_axes_0 = const()[name = tensor<string, []>("out_39_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1325_to_fp16 = const()[name = tensor<string, []>("op_1325_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1325_to_fp16, x = inputs_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
tensor<fp16, [1280]> input_75_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_75_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381964288)))];
tensor<fp16, [1280]> input_75_beta_0_to_fp16 = const()[name = tensor<string, []>("input_75_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381966912)))];
tensor<fp16, []> input_75_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_75_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
tensor<string, []> input_77_pad_type_0 = const()[name = tensor<string, []>("input_77_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_77_strides_0 = const()[name = tensor<string, []>("input_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_77_pad_0 = const()[name = tensor<string, []>("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_77_dilations_0 = const()[name = tensor<string, []>("input_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_77_groups_0 = const()[name = tensor<string, []>("input_77_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_9_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381969536)))];
tensor<fp16, [5120]> layers_9_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395076800)))];
tensor<fp16, [1, 5120, 1, 1500]> input_77_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
tensor<string, []> input_79_mode_0 = const()[name = tensor<string, []>("input_79_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
tensor<string, []> hidden_states_23_pad_type_0 = const()[name = tensor<string, []>("hidden_states_23_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = tensor<string, []>("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = tensor<string, []>("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = tensor<string, []>("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_23_groups_0 = const()[name = tensor<string, []>("hidden_states_23_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_9_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395087104)))];
tensor<fp16, [1280]> layers_9_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408194368)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_23_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("hidden_states_23_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor<string, []>("inputs_41_cast_fp16")];
tensor<int32, []> var_1358 = const()[name = tensor<string, []>("op_1358"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_41_axes_0 = const()[name = tensor<string, []>("out_41_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1377_to_fp16 = const()[name = tensor<string, []>("op_1377_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1377_to_fp16, x = inputs_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
tensor<fp16, [1280]> obj_41_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_41_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408196992)))];
tensor<fp16, [1280]> obj_41_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_41_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408199616)))];
tensor<fp16, []> obj_41_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_41_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_41_cast_fp16 = batch_norm(beta = obj_41_beta_0_to_fp16, epsilon = obj_41_epsilon_0_to_fp16, gamma = obj_41_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
tensor<string, []> query_21_pad_type_0 = const()[name = tensor<string, []>("query_21_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_21_strides_0 = const()[name = tensor<string, []>("query_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_21_pad_0 = const()[name = tensor<string, []>("query_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_21_dilations_0 = const()[name = tensor<string, []>("query_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_21_groups_0 = const()[name = tensor<string, []>("query_21_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408202240)))];
tensor<fp16, [1280]> layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411479104)))];
tensor<fp16, [1, 1280, 1, 1500]> query_21_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
tensor<string, []> key_21_pad_type_0 = const()[name = tensor<string, []>("key_21_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_21_strides_0 = const()[name = tensor<string, []>("key_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_21_pad_0 = const()[name = tensor<string, []>("key_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_21_dilations_0 = const()[name = tensor<string, []>("key_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_21_groups_0 = const()[name = tensor<string, []>("key_21_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411481728)))];
tensor<fp16, [1, 1280, 1, 1500]> key_21_cast_fp16 = conv(dilations = key_21_dilations_0, groups = key_21_groups_0, pad = key_21_pad_0, pad_type = key_21_pad_type_0, strides = key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
tensor<string, []> value_21_pad_type_0 = const()[name = tensor<string, []>("value_21_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_21_strides_0 = const()[name = tensor<string, []>("value_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_21_pad_0 = const()[name = tensor<string, []>("value_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_21_dilations_0 = const()[name = tensor<string, []>("value_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_21_groups_0 = const()[name = tensor<string, []>("value_21_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414758592)))];
tensor<fp16, [1280]> layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418035456)))];
tensor<fp16, [1, 1280, 1, 1500]> value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = value_21_dilations_0, groups = value_21_groups_0, pad = value_21_pad_0, pad_type = value_21_pad_type_0, strides = value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")];
tensor<int32, [4]> var_1412 = const()[name = tensor<string, []>("op_1412"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_21_cast_fp16 = reshape(shape = var_1412, x = query_21_cast_fp16)[name = tensor<string, []>("mh_q_21_cast_fp16")];
tensor<fp16, []> var_1414_to_fp16 = const()[name = tensor<string, []>("op_1414_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_1415_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1414_to_fp16)[name = tensor<string, []>("op_1415_cast_fp16")];
tensor<int32, [4]> var_1416 = const()[name = tensor<string, []>("op_1416"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1417_cast_fp16 = reshape(shape = var_1416, x = key_21_cast_fp16)[name = tensor<string, []>("op_1417_cast_fp16")];
tensor<bool, []> mh_w_21_transpose_x_0 = const()[name = tensor<string, []>("mh_w_21_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_21_transpose_y_0 = const()[name = tensor<string, []>("mh_w_21_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_1415_cast_fp16, y = var_1417_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_1420_cast_fp16 = softmax(axis = var_1358, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_1420_cast_fp16")];
tensor<int32, [4]> var_1421 = const()[name = tensor<string, []>("op_1421"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1422_cast_fp16 = reshape(shape = var_1421, x = value_21_cast_fp16)[name = tensor<string, []>("op_1422_cast_fp16")];
tensor<bool, []> attn_21_transpose_x_0 = const()[name = tensor<string, []>("attn_21_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_21_transpose_y_0 = const()[name = tensor<string, []>("attn_21_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1422_cast_fp16, y = var_1420_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
tensor<int32, [4]> var_1425 = const()[name = tensor<string, []>("op_1425"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_81_cast_fp16 = reshape(shape = var_1425, x = attn_21_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
tensor<string, []> obj_43_pad_type_0 = const()[name = tensor<string, []>("obj_43_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_43_strides_0 = const()[name = tensor<string, []>("obj_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_43_pad_0 = const()[name = tensor<string, []>("obj_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_43_dilations_0 = const()[name = tensor<string, []>("obj_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_43_groups_0 = const()[name = tensor<string, []>("obj_43_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418038080)))];
tensor<fp16, [1280]> layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421314944)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_43_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_43_dilations_0, groups = obj_43_groups_0, pad = obj_43_pad_0, pad_type = obj_43_pad_type_0, strides = obj_43_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = obj_43_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")];
tensor<int32, [1]> out_43_axes_0 = const()[name = tensor<string, []>("out_43_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1443_to_fp16 = const()[name = tensor<string, []>("op_1443_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1443_to_fp16, x = inputs_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
tensor<fp16, [1280]> input_83_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_83_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421317568)))];
tensor<fp16, [1280]> input_83_beta_0_to_fp16 = const()[name = tensor<string, []>("input_83_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421320192)))];
tensor<fp16, []> input_83_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_83_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_83_cast_fp16 = batch_norm(beta = input_83_beta_0_to_fp16, epsilon = input_83_epsilon_0_to_fp16, gamma = input_83_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
tensor<string, []> input_85_pad_type_0 = const()[name = tensor<string, []>("input_85_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_85_strides_0 = const()[name = tensor<string, []>("input_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_85_pad_0 = const()[name = tensor<string, []>("input_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_85_dilations_0 = const()[name = tensor<string, []>("input_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_85_groups_0 = const()[name = tensor<string, []>("input_85_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_10_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421322816)))];
tensor<fp16, [5120]> layers_10_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434430080)))];
tensor<fp16, [1, 5120, 1, 1500]> input_85_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_85_dilations_0, groups = input_85_groups_0, pad = input_85_pad_0, pad_type = input_85_pad_type_0, strides = input_85_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
tensor<string, []> input_87_mode_0 = const()[name = tensor<string, []>("input_87_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
tensor<string, []> hidden_states_25_pad_type_0 = const()[name = tensor<string, []>("hidden_states_25_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_25_strides_0 = const()[name = tensor<string, []>("hidden_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_25_pad_0 = const()[name = tensor<string, []>("hidden_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_25_dilations_0 = const()[name = tensor<string, []>("hidden_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_25_groups_0 = const()[name = tensor<string, []>("hidden_states_25_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_10_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434440384)))];
tensor<fp16, [1280]> layers_10_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447547648)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_25_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("hidden_states_25_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor<string, []>("inputs_45_cast_fp16")];
tensor<int32, []> var_1476 = const()[name = tensor<string, []>("op_1476"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_45_axes_0 = const()[name = tensor<string, []>("out_45_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1495_to_fp16 = const()[name = tensor<string, []>("op_1495_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1495_to_fp16, x = inputs_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
tensor<fp16, [1280]> obj_45_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_45_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447550272)))];
tensor<fp16, [1280]> obj_45_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_45_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447552896)))];
tensor<fp16, []> obj_45_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_45_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor<string, []>("obj_45_cast_fp16")];
tensor<string, []> query_23_pad_type_0 = const()[name = tensor<string, []>("query_23_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_23_strides_0 = const()[name = tensor<string, []>("query_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_23_pad_0 = const()[name = tensor<string, []>("query_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_23_dilations_0 = const()[name = tensor<string, []>("query_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_23_groups_0 = const()[name = tensor<string, []>("query_23_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447555520)))];
tensor<fp16, [1280]> layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450832384)))];
tensor<fp16, [1, 1280, 1, 1500]> query_23_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor<string, []>("query_23_cast_fp16")];
tensor<string, []> key_23_pad_type_0 = const()[name = tensor<string, []>("key_23_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_23_strides_0 = const()[name = tensor<string, []>("key_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_23_pad_0 = const()[name = tensor<string, []>("key_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_23_dilations_0 = const()[name = tensor<string, []>("key_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_23_groups_0 = const()[name = tensor<string, []>("key_23_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450835008)))];
tensor<fp16, [1, 1280, 1, 1500]> key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor<string, []>("key_23_cast_fp16")];
tensor<string, []> value_23_pad_type_0 = const()[name = tensor<string, []>("value_23_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_23_strides_0 = const()[name = tensor<string, []>("value_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_23_pad_0 = const()[name = tensor<string, []>("value_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_23_dilations_0 = const()[name = tensor<string, []>("value_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_23_groups_0 = const()[name = tensor<string, []>("value_23_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454111872)))];
tensor<fp16, [1280]> layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457388736)))];
tensor<fp16, [1, 1280, 1, 1500]> value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor<string, []>("value_23_cast_fp16")];
tensor<int32, [4]> var_1530 = const()[name = tensor<string, []>("op_1530"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_23_cast_fp16 = reshape(shape = var_1530, x = query_23_cast_fp16)[name = tensor<string, []>("mh_q_23_cast_fp16")];
tensor<fp16, []> var_1532_to_fp16 = const()[name = tensor<string, []>("op_1532_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_1533_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1532_to_fp16)[name = tensor<string, []>("op_1533_cast_fp16")];
tensor<int32, [4]> var_1534 = const()[name = tensor<string, []>("op_1534"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1535_cast_fp16 = reshape(shape = var_1534, x = key_23_cast_fp16)[name = tensor<string, []>("op_1535_cast_fp16")];
tensor<bool, []> mh_w_23_transpose_x_0 = const()[name = tensor<string, []>("mh_w_23_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_23_transpose_y_0 = const()[name = tensor<string, []>("mh_w_23_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_1533_cast_fp16, y = var_1535_cast_fp16)[name = tensor<string, []>("mh_w_23_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_1538_cast_fp16 = softmax(axis = var_1476, x = mh_w_23_cast_fp16)[name = tensor<string, []>("op_1538_cast_fp16")];
tensor<int32, [4]> var_1539 = const()[name = tensor<string, []>("op_1539"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1540_cast_fp16 = reshape(shape = var_1539, x = value_23_cast_fp16)[name = tensor<string, []>("op_1540_cast_fp16")];
tensor<bool, []> attn_23_transpose_x_0 = const()[name = tensor<string, []>("attn_23_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_23_transpose_y_0 = const()[name = tensor<string, []>("attn_23_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1540_cast_fp16, y = var_1538_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")];
tensor<int32, [4]> var_1543 = const()[name = tensor<string, []>("op_1543"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_89_cast_fp16 = reshape(shape = var_1543, x = attn_23_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
tensor<string, []> obj_47_pad_type_0 = const()[name = tensor<string, []>("obj_47_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_47_strides_0 = const()[name = tensor<string, []>("obj_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_47_pad_0 = const()[name = tensor<string, []>("obj_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_47_dilations_0 = const()[name = tensor<string, []>("obj_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_47_groups_0 = const()[name = tensor<string, []>("obj_47_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457391360)))];
tensor<fp16, [1280]> layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460668224)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_47_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_47_dilations_0, groups = obj_47_groups_0, pad = obj_47_pad_0, pad_type = obj_47_pad_type_0, strides = obj_47_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("obj_47_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_47_cast_fp16)[name = tensor<string, []>("inputs_47_cast_fp16")];
tensor<int32, [1]> out_47_axes_0 = const()[name = tensor<string, []>("out_47_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1561_to_fp16 = const()[name = tensor<string, []>("op_1561_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1561_to_fp16, x = inputs_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")];
tensor<fp16, [1280]> input_91_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_91_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460670848)))];
tensor<fp16, [1280]> input_91_beta_0_to_fp16 = const()[name = tensor<string, []>("input_91_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460673472)))];
tensor<fp16, []> input_91_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_91_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_91_cast_fp16 = batch_norm(beta = input_91_beta_0_to_fp16, epsilon = input_91_epsilon_0_to_fp16, gamma = input_91_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
tensor<string, []> input_93_pad_type_0 = const()[name = tensor<string, []>("input_93_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_93_strides_0 = const()[name = tensor<string, []>("input_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_93_pad_0 = const()[name = tensor<string, []>("input_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_93_dilations_0 = const()[name = tensor<string, []>("input_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_93_groups_0 = const()[name = tensor<string, []>("input_93_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_11_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460676096)))];
tensor<fp16, [5120]> layers_11_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(473783360)))];
tensor<fp16, [1, 5120, 1, 1500]> input_93_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
tensor<string, []> input_95_mode_0 = const()[name = tensor<string, []>("input_95_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = input_93_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
tensor<string, []> hidden_states_27_pad_type_0 = const()[name = tensor<string, []>("hidden_states_27_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = tensor<string, []>("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = tensor<string, []>("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = tensor<string, []>("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_27_groups_0 = const()[name = tensor<string, []>("hidden_states_27_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_11_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(473793664)))];
tensor<fp16, [1280]> layers_11_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(486900928)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_27_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("hidden_states_27_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor<string, []>("inputs_49_cast_fp16")];
tensor<int32, []> var_1594 = const()[name = tensor<string, []>("op_1594"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_49_axes_0 = const()[name = tensor<string, []>("out_49_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1613_to_fp16 = const()[name = tensor<string, []>("op_1613_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1613_to_fp16, x = inputs_49_cast_fp16)[name = tensor<string, []>("out_49_cast_fp16")];
tensor<fp16, [1280]> obj_49_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_49_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(486903552)))];
tensor<fp16, [1280]> obj_49_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_49_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(486906176)))];
tensor<fp16, []> obj_49_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_49_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
tensor<string, []> query_25_pad_type_0 = const()[name = tensor<string, []>("query_25_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_25_strides_0 = const()[name = tensor<string, []>("query_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_25_pad_0 = const()[name = tensor<string, []>("query_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_25_dilations_0 = const()[name = tensor<string, []>("query_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_25_groups_0 = const()[name = tensor<string, []>("query_25_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(486908800)))];
tensor<fp16, [1280]> layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490185664)))];
tensor<fp16, [1, 1280, 1, 1500]> query_25_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")];
tensor<string, []> key_25_pad_type_0 = const()[name = tensor<string, []>("key_25_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_25_strides_0 = const()[name = tensor<string, []>("key_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_25_pad_0 = const()[name = tensor<string, []>("key_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_25_dilations_0 = const()[name = tensor<string, []>("key_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_25_groups_0 = const()[name = tensor<string, []>("key_25_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490188288)))];
tensor<fp16, [1, 1280, 1, 1500]> key_25_cast_fp16 = conv(dilations = key_25_dilations_0, groups = key_25_groups_0, pad = key_25_pad_0, pad_type = key_25_pad_type_0, strides = key_25_strides_0, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
tensor<string, []> value_25_pad_type_0 = const()[name = tensor<string, []>("value_25_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_25_strides_0 = const()[name = tensor<string, []>("value_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_25_pad_0 = const()[name = tensor<string, []>("value_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_25_dilations_0 = const()[name = tensor<string, []>("value_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_25_groups_0 = const()[name = tensor<string, []>("value_25_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493465152)))];
tensor<fp16, [1280]> layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496742016)))];
tensor<fp16, [1, 1280, 1, 1500]> value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = value_25_dilations_0, groups = value_25_groups_0, pad = value_25_pad_0, pad_type = value_25_pad_type_0, strides = value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor<string, []>("value_25_cast_fp16")];
tensor<int32, [4]> var_1648 = const()[name = tensor<string, []>("op_1648"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_25_cast_fp16 = reshape(shape = var_1648, x = query_25_cast_fp16)[name = tensor<string, []>("mh_q_25_cast_fp16")];
tensor<fp16, []> var_1650_to_fp16 = const()[name = tensor<string, []>("op_1650_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_1651_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1650_to_fp16)[name = tensor<string, []>("op_1651_cast_fp16")];
tensor<int32, [4]> var_1652 = const()[name = tensor<string, []>("op_1652"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1653_cast_fp16 = reshape(shape = var_1652, x = key_25_cast_fp16)[name = tensor<string, []>("op_1653_cast_fp16")];
tensor<bool, []> mh_w_25_transpose_x_0 = const()[name = tensor<string, []>("mh_w_25_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_25_transpose_y_0 = const()[name = tensor<string, []>("mh_w_25_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1651_cast_fp16, y = var_1653_cast_fp16)[name = tensor<string, []>("mh_w_25_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_1656_cast_fp16 = softmax(axis = var_1594, x = mh_w_25_cast_fp16)[name = tensor<string, []>("op_1656_cast_fp16")];
tensor<int32, [4]> var_1657 = const()[name = tensor<string, []>("op_1657"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1658_cast_fp16 = reshape(shape = var_1657, x = value_25_cast_fp16)[name = tensor<string, []>("op_1658_cast_fp16")];
tensor<bool, []> attn_25_transpose_x_0 = const()[name = tensor<string, []>("attn_25_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_25_transpose_y_0 = const()[name = tensor<string, []>("attn_25_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1658_cast_fp16, y = var_1656_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
tensor<int32, [4]> var_1661 = const()[name = tensor<string, []>("op_1661"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_97_cast_fp16 = reshape(shape = var_1661, x = attn_25_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
tensor<string, []> obj_51_pad_type_0 = const()[name = tensor<string, []>("obj_51_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_51_strides_0 = const()[name = tensor<string, []>("obj_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_51_pad_0 = const()[name = tensor<string, []>("obj_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_51_dilations_0 = const()[name = tensor<string, []>("obj_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_51_groups_0 = const()[name = tensor<string, []>("obj_51_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496744640)))];
tensor<fp16, [1280]> layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500021504)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_51_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = obj_51_dilations_0, groups = obj_51_groups_0, pad = obj_51_pad_0, pad_type = obj_51_pad_type_0, strides = obj_51_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_51_cast_fp16)[name = tensor<string, []>("inputs_51_cast_fp16")];
tensor<int32, [1]> out_51_axes_0 = const()[name = tensor<string, []>("out_51_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1679_to_fp16 = const()[name = tensor<string, []>("op_1679_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1679_to_fp16, x = inputs_51_cast_fp16)[name = tensor<string, []>("out_51_cast_fp16")];
tensor<fp16, [1280]> input_99_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_99_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500024128)))];
tensor<fp16, [1280]> input_99_beta_0_to_fp16 = const()[name = tensor<string, []>("input_99_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500026752)))];
tensor<fp16, []> input_99_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_99_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_99_cast_fp16 = batch_norm(beta = input_99_beta_0_to_fp16, epsilon = input_99_epsilon_0_to_fp16, gamma = input_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
tensor<string, []> input_101_pad_type_0 = const()[name = tensor<string, []>("input_101_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_101_strides_0 = const()[name = tensor<string, []>("input_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_101_pad_0 = const()[name = tensor<string, []>("input_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_101_dilations_0 = const()[name = tensor<string, []>("input_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_101_groups_0 = const()[name = tensor<string, []>("input_101_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_12_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500029376)))];
tensor<fp16, [5120]> layers_12_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513136640)))];
tensor<fp16, [1, 5120, 1, 1500]> input_101_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = layers_12_fc1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
tensor<string, []> input_103_mode_0 = const()[name = tensor<string, []>("input_103_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
tensor<string, []> hidden_states_29_pad_type_0 = const()[name = tensor<string, []>("hidden_states_29_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_29_strides_0 = const()[name = tensor<string, []>("hidden_states_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_29_pad_0 = const()[name = tensor<string, []>("hidden_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_29_dilations_0 = const()[name = tensor<string, []>("hidden_states_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_29_groups_0 = const()[name = tensor<string, []>("hidden_states_29_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_12_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513146944)))];
tensor<fp16, [1280]> layers_12_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526254208)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_29_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = layers_12_fc2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("hidden_states_29_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor<string, []>("inputs_53_cast_fp16")];
tensor<int32, []> var_1712 = const()[name = tensor<string, []>("op_1712"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_53_axes_0 = const()[name = tensor<string, []>("out_53_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1731_to_fp16 = const()[name = tensor<string, []>("op_1731_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1731_to_fp16, x = inputs_53_cast_fp16)[name = tensor<string, []>("out_53_cast_fp16")];
tensor<fp16, [1280]> obj_53_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_53_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526256832)))];
tensor<fp16, [1280]> obj_53_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_53_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526259456)))];
tensor<fp16, []> obj_53_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_53_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_53_cast_fp16 = batch_norm(beta = obj_53_beta_0_to_fp16, epsilon = obj_53_epsilon_0_to_fp16, gamma = obj_53_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
tensor<string, []> query_27_pad_type_0 = const()[name = tensor<string, []>("query_27_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_27_strides_0 = const()[name = tensor<string, []>("query_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_27_pad_0 = const()[name = tensor<string, []>("query_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_27_dilations_0 = const()[name = tensor<string, []>("query_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_27_groups_0 = const()[name = tensor<string, []>("query_27_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526262080)))];
tensor<fp16, [1280]> layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529538944)))];
tensor<fp16, [1, 1280, 1, 1500]> query_27_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor<string, []>("query_27_cast_fp16")];
tensor<string, []> key_27_pad_type_0 = const()[name = tensor<string, []>("key_27_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_27_strides_0 = const()[name = tensor<string, []>("key_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_27_pad_0 = const()[name = tensor<string, []>("key_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_27_dilations_0 = const()[name = tensor<string, []>("key_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_27_groups_0 = const()[name = tensor<string, []>("key_27_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529541568)))];
tensor<fp16, [1, 1280, 1, 1500]> key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor<string, []>("key_27_cast_fp16")];
tensor<string, []> value_27_pad_type_0 = const()[name = tensor<string, []>("value_27_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_27_strides_0 = const()[name = tensor<string, []>("value_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_27_pad_0 = const()[name = tensor<string, []>("value_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_27_dilations_0 = const()[name = tensor<string, []>("value_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_27_groups_0 = const()[name = tensor<string, []>("value_27_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532818432)))];
tensor<fp16, [1280]> layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536095296)))];
tensor<fp16, [1, 1280, 1, 1500]> value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor<string, []>("value_27_cast_fp16")];
tensor<int32, [4]> var_1766 = const()[name = tensor<string, []>("op_1766"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_27_cast_fp16 = reshape(shape = var_1766, x = query_27_cast_fp16)[name = tensor<string, []>("mh_q_27_cast_fp16")];
tensor<fp16, []> var_1768_to_fp16 = const()[name = tensor<string, []>("op_1768_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_1769_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1768_to_fp16)[name = tensor<string, []>("op_1769_cast_fp16")];
tensor<int32, [4]> var_1770 = const()[name = tensor<string, []>("op_1770"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1771_cast_fp16 = reshape(shape = var_1770, x = key_27_cast_fp16)[name = tensor<string, []>("op_1771_cast_fp16")];
tensor<bool, []> mh_w_27_transpose_x_0 = const()[name = tensor<string, []>("mh_w_27_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_27_transpose_y_0 = const()[name = tensor<string, []>("mh_w_27_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_27_cast_fp16 = matmul(transpose_x = mh_w_27_transpose_x_0, transpose_y = mh_w_27_transpose_y_0, x = var_1769_cast_fp16, y = var_1771_cast_fp16)[name = tensor<string, []>("mh_w_27_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_1774_cast_fp16 = softmax(axis = var_1712, x = mh_w_27_cast_fp16)[name = tensor<string, []>("op_1774_cast_fp16")];
tensor<int32, [4]> var_1775 = const()[name = tensor<string, []>("op_1775"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1776_cast_fp16 = reshape(shape = var_1775, x = value_27_cast_fp16)[name = tensor<string, []>("op_1776_cast_fp16")];
tensor<bool, []> attn_27_transpose_x_0 = const()[name = tensor<string, []>("attn_27_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_27_transpose_y_0 = const()[name = tensor<string, []>("attn_27_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1776_cast_fp16, y = var_1774_cast_fp16)[name = tensor<string, []>("attn_27_cast_fp16")];
tensor<int32, [4]> var_1779 = const()[name = tensor<string, []>("op_1779"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_105_cast_fp16 = reshape(shape = var_1779, x = attn_27_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
tensor<string, []> obj_55_pad_type_0 = const()[name = tensor<string, []>("obj_55_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_55_strides_0 = const()[name = tensor<string, []>("obj_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_55_pad_0 = const()[name = tensor<string, []>("obj_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_55_dilations_0 = const()[name = tensor<string, []>("obj_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_55_groups_0 = const()[name = tensor<string, []>("obj_55_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536097920)))];
tensor<fp16, [1280]> layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539374784)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_55_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = obj_55_dilations_0, groups = obj_55_groups_0, pad = obj_55_pad_0, pad_type = obj_55_pad_type_0, strides = obj_55_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("inputs_55_cast_fp16")];
tensor<int32, [1]> out_55_axes_0 = const()[name = tensor<string, []>("out_55_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1797_to_fp16 = const()[name = tensor<string, []>("op_1797_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_1797_to_fp16, x = inputs_55_cast_fp16)[name = tensor<string, []>("out_55_cast_fp16")];
tensor<fp16, [1280]> input_107_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_107_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539377408)))];
tensor<fp16, [1280]> input_107_beta_0_to_fp16 = const()[name = tensor<string, []>("input_107_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539380032)))];
tensor<fp16, []> input_107_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_107_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_107_cast_fp16 = batch_norm(beta = input_107_beta_0_to_fp16, epsilon = input_107_epsilon_0_to_fp16, gamma = input_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
tensor<string, []> input_109_pad_type_0 = const()[name = tensor<string, []>("input_109_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_109_strides_0 = const()[name = tensor<string, []>("input_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_109_pad_0 = const()[name = tensor<string, []>("input_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_109_dilations_0 = const()[name = tensor<string, []>("input_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_109_groups_0 = const()[name = tensor<string, []>("input_109_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_13_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539382656)))];
tensor<fp16, [5120]> layers_13_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552489920)))];
tensor<fp16, [1, 5120, 1, 1500]> input_109_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = layers_13_fc1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
tensor<string, []> input_111_mode_0 = const()[name = tensor<string, []>("input_111_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
tensor<string, []> hidden_states_31_pad_type_0 = const()[name = tensor<string, []>("hidden_states_31_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_31_strides_0 = const()[name = tensor<string, []>("hidden_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_31_pad_0 = const()[name = tensor<string, []>("hidden_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_31_dilations_0 = const()[name = tensor<string, []>("hidden_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_31_groups_0 = const()[name = tensor<string, []>("hidden_states_31_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_13_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552500224)))];
tensor<fp16, [1280]> layers_13_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(565607488)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_31_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = hidden_states_31_dilations_0, groups = hidden_states_31_groups_0, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = hidden_states_31_strides_0, weight = layers_13_fc2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("hidden_states_31_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor<string, []>("inputs_57_cast_fp16")];
tensor<int32, []> var_1830 = const()[name = tensor<string, []>("op_1830"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_57_axes_0 = const()[name = tensor<string, []>("out_57_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1849_to_fp16 = const()[name = tensor<string, []>("op_1849_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_1849_to_fp16, x = inputs_57_cast_fp16)[name = tensor<string, []>("out_57_cast_fp16")];
tensor<fp16, [1280]> obj_57_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_57_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(565610112)))];
tensor<fp16, [1280]> obj_57_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_57_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(565612736)))];
tensor<fp16, []> obj_57_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_57_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor<string, []>("obj_57_cast_fp16")];
tensor<string, []> query_29_pad_type_0 = const()[name = tensor<string, []>("query_29_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_29_strides_0 = const()[name = tensor<string, []>("query_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_29_pad_0 = const()[name = tensor<string, []>("query_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_29_dilations_0 = const()[name = tensor<string, []>("query_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_29_groups_0 = const()[name = tensor<string, []>("query_29_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(565615360)))];
tensor<fp16, [1280]> layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(568892224)))];
tensor<fp16, [1, 1280, 1, 1500]> query_29_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")];
tensor<string, []> key_29_pad_type_0 = const()[name = tensor<string, []>("key_29_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_29_strides_0 = const()[name = tensor<string, []>("key_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_29_pad_0 = const()[name = tensor<string, []>("key_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_29_dilations_0 = const()[name = tensor<string, []>("key_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_29_groups_0 = const()[name = tensor<string, []>("key_29_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(568894848)))];
tensor<fp16, [1, 1280, 1, 1500]> key_29_cast_fp16 = conv(dilations = key_29_dilations_0, groups = key_29_groups_0, pad = key_29_pad_0, pad_type = key_29_pad_type_0, strides = key_29_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
tensor<string, []> value_29_pad_type_0 = const()[name = tensor<string, []>("value_29_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_29_strides_0 = const()[name = tensor<string, []>("value_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_29_pad_0 = const()[name = tensor<string, []>("value_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_29_dilations_0 = const()[name = tensor<string, []>("value_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_29_groups_0 = const()[name = tensor<string, []>("value_29_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(572171712)))];
tensor<fp16, [1280]> layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575448576)))];
tensor<fp16, [1, 1280, 1, 1500]> value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = value_29_dilations_0, groups = value_29_groups_0, pad = value_29_pad_0, pad_type = value_29_pad_type_0, strides = value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("value_29_cast_fp16")];
tensor<int32, [4]> var_1884 = const()[name = tensor<string, []>("op_1884"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_29_cast_fp16 = reshape(shape = var_1884, x = query_29_cast_fp16)[name = tensor<string, []>("mh_q_29_cast_fp16")];
tensor<fp16, []> var_1886_to_fp16 = const()[name = tensor<string, []>("op_1886_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_1887_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1886_to_fp16)[name = tensor<string, []>("op_1887_cast_fp16")];
tensor<int32, [4]> var_1888 = const()[name = tensor<string, []>("op_1888"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1889_cast_fp16 = reshape(shape = var_1888, x = key_29_cast_fp16)[name = tensor<string, []>("op_1889_cast_fp16")];
tensor<bool, []> mh_w_29_transpose_x_0 = const()[name = tensor<string, []>("mh_w_29_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_29_transpose_y_0 = const()[name = tensor<string, []>("mh_w_29_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1887_cast_fp16, y = var_1889_cast_fp16)[name = tensor<string, []>("mh_w_29_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_1892_cast_fp16 = softmax(axis = var_1830, x = mh_w_29_cast_fp16)[name = tensor<string, []>("op_1892_cast_fp16")];
tensor<int32, [4]> var_1893 = const()[name = tensor<string, []>("op_1893"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_1894_cast_fp16 = reshape(shape = var_1893, x = value_29_cast_fp16)[name = tensor<string, []>("op_1894_cast_fp16")];
tensor<bool, []> attn_29_transpose_x_0 = const()[name = tensor<string, []>("attn_29_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_29_transpose_y_0 = const()[name = tensor<string, []>("attn_29_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1894_cast_fp16, y = var_1892_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
tensor<int32, [4]> var_1897 = const()[name = tensor<string, []>("op_1897"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_113_cast_fp16 = reshape(shape = var_1897, x = attn_29_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
tensor<string, []> obj_59_pad_type_0 = const()[name = tensor<string, []>("obj_59_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_59_strides_0 = const()[name = tensor<string, []>("obj_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_59_pad_0 = const()[name = tensor<string, []>("obj_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_59_dilations_0 = const()[name = tensor<string, []>("obj_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_59_groups_0 = const()[name = tensor<string, []>("obj_59_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575451200)))];
tensor<fp16, [1280]> layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578728064)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_59_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = obj_59_dilations_0, groups = obj_59_groups_0, pad = obj_59_pad_0, pad_type = obj_59_pad_type_0, strides = obj_59_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("obj_59_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_59_cast_fp16)[name = tensor<string, []>("inputs_59_cast_fp16")];
tensor<int32, [1]> out_59_axes_0 = const()[name = tensor<string, []>("out_59_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1915_to_fp16 = const()[name = tensor<string, []>("op_1915_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_1915_to_fp16, x = inputs_59_cast_fp16)[name = tensor<string, []>("out_59_cast_fp16")];
tensor<fp16, [1280]> input_115_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_115_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578730688)))];
tensor<fp16, [1280]> input_115_beta_0_to_fp16 = const()[name = tensor<string, []>("input_115_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578733312)))];
tensor<fp16, []> input_115_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_115_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
tensor<string, []> input_117_pad_type_0 = const()[name = tensor<string, []>("input_117_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_117_strides_0 = const()[name = tensor<string, []>("input_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_117_pad_0 = const()[name = tensor<string, []>("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_117_dilations_0 = const()[name = tensor<string, []>("input_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_117_groups_0 = const()[name = tensor<string, []>("input_117_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_14_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578735936)))];
tensor<fp16, [5120]> layers_14_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591843200)))];
tensor<fp16, [1, 5120, 1, 1500]> input_117_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_14_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
tensor<string, []> input_119_mode_0 = const()[name = tensor<string, []>("input_119_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
tensor<string, []> hidden_states_33_pad_type_0 = const()[name = tensor<string, []>("hidden_states_33_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = tensor<string, []>("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = tensor<string, []>("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = tensor<string, []>("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_33_groups_0 = const()[name = tensor<string, []>("hidden_states_33_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_14_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591853504)))];
tensor<fp16, [1280]> layers_14_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604960768)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_33_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_14_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("hidden_states_33_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor<string, []>("inputs_61_cast_fp16")];
tensor<int32, []> var_1948 = const()[name = tensor<string, []>("op_1948"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_61_axes_0 = const()[name = tensor<string, []>("out_61_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_1967_to_fp16 = const()[name = tensor<string, []>("op_1967_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_1967_to_fp16, x = inputs_61_cast_fp16)[name = tensor<string, []>("out_61_cast_fp16")];
tensor<fp16, [1280]> obj_61_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_61_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604963392)))];
tensor<fp16, [1280]> obj_61_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_61_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604966016)))];
tensor<fp16, []> obj_61_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_61_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor<string, []>("obj_61_cast_fp16")];
tensor<string, []> query_31_pad_type_0 = const()[name = tensor<string, []>("query_31_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_31_strides_0 = const()[name = tensor<string, []>("query_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_31_pad_0 = const()[name = tensor<string, []>("query_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_31_dilations_0 = const()[name = tensor<string, []>("query_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_31_groups_0 = const()[name = tensor<string, []>("query_31_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604968640)))];
tensor<fp16, [1280]> layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(608245504)))];
tensor<fp16, [1, 1280, 1, 1500]> query_31_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor<string, []>("query_31_cast_fp16")];
tensor<string, []> key_31_pad_type_0 = const()[name = tensor<string, []>("key_31_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_31_strides_0 = const()[name = tensor<string, []>("key_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_31_pad_0 = const()[name = tensor<string, []>("key_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_31_dilations_0 = const()[name = tensor<string, []>("key_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_31_groups_0 = const()[name = tensor<string, []>("key_31_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(608248128)))];
tensor<fp16, [1, 1280, 1, 1500]> key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor<string, []>("key_31_cast_fp16")];
tensor<string, []> value_31_pad_type_0 = const()[name = tensor<string, []>("value_31_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_31_strides_0 = const()[name = tensor<string, []>("value_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_31_pad_0 = const()[name = tensor<string, []>("value_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_31_dilations_0 = const()[name = tensor<string, []>("value_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_31_groups_0 = const()[name = tensor<string, []>("value_31_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(611524992)))];
tensor<fp16, [1280]> layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(614801856)))];
tensor<fp16, [1, 1280, 1, 1500]> value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor<string, []>("value_31_cast_fp16")];
tensor<int32, [4]> var_2002 = const()[name = tensor<string, []>("op_2002"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_31_cast_fp16 = reshape(shape = var_2002, x = query_31_cast_fp16)[name = tensor<string, []>("mh_q_31_cast_fp16")];
tensor<fp16, []> var_2004_to_fp16 = const()[name = tensor<string, []>("op_2004_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2005_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_2004_to_fp16)[name = tensor<string, []>("op_2005_cast_fp16")];
tensor<int32, [4]> var_2006 = const()[name = tensor<string, []>("op_2006"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2007_cast_fp16 = reshape(shape = var_2006, x = key_31_cast_fp16)[name = tensor<string, []>("op_2007_cast_fp16")];
tensor<bool, []> mh_w_31_transpose_x_0 = const()[name = tensor<string, []>("mh_w_31_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_31_transpose_y_0 = const()[name = tensor<string, []>("mh_w_31_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_2005_cast_fp16, y = var_2007_cast_fp16)[name = tensor<string, []>("mh_w_31_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2010_cast_fp16 = softmax(axis = var_1948, x = mh_w_31_cast_fp16)[name = tensor<string, []>("op_2010_cast_fp16")];
tensor<int32, [4]> var_2011 = const()[name = tensor<string, []>("op_2011"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2012_cast_fp16 = reshape(shape = var_2011, x = value_31_cast_fp16)[name = tensor<string, []>("op_2012_cast_fp16")];
tensor<bool, []> attn_31_transpose_x_0 = const()[name = tensor<string, []>("attn_31_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_31_transpose_y_0 = const()[name = tensor<string, []>("attn_31_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_2012_cast_fp16, y = var_2010_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")];
tensor<int32, [4]> var_2015 = const()[name = tensor<string, []>("op_2015"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_121_cast_fp16 = reshape(shape = var_2015, x = attn_31_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
tensor<string, []> obj_63_pad_type_0 = const()[name = tensor<string, []>("obj_63_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_63_strides_0 = const()[name = tensor<string, []>("obj_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_63_pad_0 = const()[name = tensor<string, []>("obj_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_63_dilations_0 = const()[name = tensor<string, []>("obj_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_63_groups_0 = const()[name = tensor<string, []>("obj_63_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(614804480)))];
tensor<fp16, [1280]> layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(618081344)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_63_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = obj_63_dilations_0, groups = obj_63_groups_0, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = obj_63_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("obj_63_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_63_cast_fp16)[name = tensor<string, []>("inputs_63_cast_fp16")];
tensor<int32, [1]> out_63_axes_0 = const()[name = tensor<string, []>("out_63_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2033_to_fp16 = const()[name = tensor<string, []>("op_2033_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2033_to_fp16, x = inputs_63_cast_fp16)[name = tensor<string, []>("out_63_cast_fp16")];
tensor<fp16, [1280]> input_123_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_123_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(618083968)))];
tensor<fp16, [1280]> input_123_beta_0_to_fp16 = const()[name = tensor<string, []>("input_123_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(618086592)))];
tensor<fp16, []> input_123_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_123_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_123_cast_fp16 = batch_norm(beta = input_123_beta_0_to_fp16, epsilon = input_123_epsilon_0_to_fp16, gamma = input_123_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
tensor<string, []> input_125_pad_type_0 = const()[name = tensor<string, []>("input_125_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_125_strides_0 = const()[name = tensor<string, []>("input_125_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_125_pad_0 = const()[name = tensor<string, []>("input_125_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_125_dilations_0 = const()[name = tensor<string, []>("input_125_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_125_groups_0 = const()[name = tensor<string, []>("input_125_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_15_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(618089216)))];
tensor<fp16, [5120]> layers_15_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(631196480)))];
tensor<fp16, [1, 5120, 1, 1500]> input_125_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = input_125_dilations_0, groups = input_125_groups_0, pad = input_125_pad_0, pad_type = input_125_pad_type_0, strides = input_125_strides_0, weight = layers_15_fc1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
tensor<string, []> input_127_mode_0 = const()[name = tensor<string, []>("input_127_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = input_125_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
tensor<string, []> hidden_states_35_pad_type_0 = const()[name = tensor<string, []>("hidden_states_35_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_35_strides_0 = const()[name = tensor<string, []>("hidden_states_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_35_pad_0 = const()[name = tensor<string, []>("hidden_states_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_35_dilations_0 = const()[name = tensor<string, []>("hidden_states_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_35_groups_0 = const()[name = tensor<string, []>("hidden_states_35_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_15_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(631206784)))];
tensor<fp16, [1280]> layers_15_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644314048)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_35_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = layers_15_fc2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("hidden_states_35_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor<string, []>("inputs_65_cast_fp16")];
tensor<int32, []> var_2066 = const()[name = tensor<string, []>("op_2066"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_65_axes_0 = const()[name = tensor<string, []>("out_65_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2085_to_fp16 = const()[name = tensor<string, []>("op_2085_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2085_to_fp16, x = inputs_65_cast_fp16)[name = tensor<string, []>("out_65_cast_fp16")];
tensor<fp16, [1280]> obj_65_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_65_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644316672)))];
tensor<fp16, [1280]> obj_65_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_65_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644319296)))];
tensor<fp16, []> obj_65_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_65_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor<string, []>("obj_65_cast_fp16")];
tensor<string, []> query_33_pad_type_0 = const()[name = tensor<string, []>("query_33_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_33_strides_0 = const()[name = tensor<string, []>("query_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_33_pad_0 = const()[name = tensor<string, []>("query_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_33_dilations_0 = const()[name = tensor<string, []>("query_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_33_groups_0 = const()[name = tensor<string, []>("query_33_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644321920)))];
tensor<fp16, [1280]> layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647598784)))];
tensor<fp16, [1, 1280, 1, 1500]> query_33_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")];
tensor<string, []> key_33_pad_type_0 = const()[name = tensor<string, []>("key_33_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_33_strides_0 = const()[name = tensor<string, []>("key_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_33_pad_0 = const()[name = tensor<string, []>("key_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_33_dilations_0 = const()[name = tensor<string, []>("key_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_33_groups_0 = const()[name = tensor<string, []>("key_33_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647601408)))];
tensor<fp16, [1, 1280, 1, 1500]> key_33_cast_fp16 = conv(dilations = key_33_dilations_0, groups = key_33_groups_0, pad = key_33_pad_0, pad_type = key_33_pad_type_0, strides = key_33_strides_0, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor<string, []>("key_33_cast_fp16")];
tensor<string, []> value_33_pad_type_0 = const()[name = tensor<string, []>("value_33_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_33_strides_0 = const()[name = tensor<string, []>("value_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_33_pad_0 = const()[name = tensor<string, []>("value_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_33_dilations_0 = const()[name = tensor<string, []>("value_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_33_groups_0 = const()[name = tensor<string, []>("value_33_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(650878272)))];
tensor<fp16, [1280]> layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(654155136)))];
tensor<fp16, [1, 1280, 1, 1500]> value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = value_33_dilations_0, groups = value_33_groups_0, pad = value_33_pad_0, pad_type = value_33_pad_type_0, strides = value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor<string, []>("value_33_cast_fp16")];
tensor<int32, [4]> var_2120 = const()[name = tensor<string, []>("op_2120"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_33_cast_fp16 = reshape(shape = var_2120, x = query_33_cast_fp16)[name = tensor<string, []>("mh_q_33_cast_fp16")];
tensor<fp16, []> var_2122_to_fp16 = const()[name = tensor<string, []>("op_2122_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2123_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_2122_to_fp16)[name = tensor<string, []>("op_2123_cast_fp16")];
tensor<int32, [4]> var_2124 = const()[name = tensor<string, []>("op_2124"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2125_cast_fp16 = reshape(shape = var_2124, x = key_33_cast_fp16)[name = tensor<string, []>("op_2125_cast_fp16")];
tensor<bool, []> mh_w_33_transpose_x_0 = const()[name = tensor<string, []>("mh_w_33_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_33_transpose_y_0 = const()[name = tensor<string, []>("mh_w_33_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_2123_cast_fp16, y = var_2125_cast_fp16)[name = tensor<string, []>("mh_w_33_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2128_cast_fp16 = softmax(axis = var_2066, x = mh_w_33_cast_fp16)[name = tensor<string, []>("op_2128_cast_fp16")];
tensor<int32, [4]> var_2129 = const()[name = tensor<string, []>("op_2129"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2130_cast_fp16 = reshape(shape = var_2129, x = value_33_cast_fp16)[name = tensor<string, []>("op_2130_cast_fp16")];
tensor<bool, []> attn_33_transpose_x_0 = const()[name = tensor<string, []>("attn_33_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_33_transpose_y_0 = const()[name = tensor<string, []>("attn_33_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_2130_cast_fp16, y = var_2128_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
tensor<int32, [4]> var_2133 = const()[name = tensor<string, []>("op_2133"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_129_cast_fp16 = reshape(shape = var_2133, x = attn_33_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
tensor<string, []> obj_67_pad_type_0 = const()[name = tensor<string, []>("obj_67_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_67_strides_0 = const()[name = tensor<string, []>("obj_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_67_pad_0 = const()[name = tensor<string, []>("obj_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_67_dilations_0 = const()[name = tensor<string, []>("obj_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_67_groups_0 = const()[name = tensor<string, []>("obj_67_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(654157760)))];
tensor<fp16, [1280]> layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(657434624)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_67_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("obj_67_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = obj_67_cast_fp16)[name = tensor<string, []>("inputs_67_cast_fp16")];
tensor<int32, [1]> out_67_axes_0 = const()[name = tensor<string, []>("out_67_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2151_to_fp16 = const()[name = tensor<string, []>("op_2151_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2151_to_fp16, x = inputs_67_cast_fp16)[name = tensor<string, []>("out_67_cast_fp16")];
tensor<fp16, [1280]> input_131_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_131_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(657437248)))];
tensor<fp16, [1280]> input_131_beta_0_to_fp16 = const()[name = tensor<string, []>("input_131_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(657439872)))];
tensor<fp16, []> input_131_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_131_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_131_cast_fp16 = batch_norm(beta = input_131_beta_0_to_fp16, epsilon = input_131_epsilon_0_to_fp16, gamma = input_131_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
tensor<string, []> input_133_pad_type_0 = const()[name = tensor<string, []>("input_133_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_133_strides_0 = const()[name = tensor<string, []>("input_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_133_pad_0 = const()[name = tensor<string, []>("input_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_133_dilations_0 = const()[name = tensor<string, []>("input_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_133_groups_0 = const()[name = tensor<string, []>("input_133_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_16_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(657442496)))];
tensor<fp16, [5120]> layers_16_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(670549760)))];
tensor<fp16, [1, 5120, 1, 1500]> input_133_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = input_133_dilations_0, groups = input_133_groups_0, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = input_133_strides_0, weight = layers_16_fc1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
tensor<string, []> input_135_mode_0 = const()[name = tensor<string, []>("input_135_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_135_cast_fp16 = gelu(mode = input_135_mode_0, x = input_133_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
tensor<string, []> hidden_states_37_pad_type_0 = const()[name = tensor<string, []>("hidden_states_37_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = tensor<string, []>("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = tensor<string, []>("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = tensor<string, []>("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_37_groups_0 = const()[name = tensor<string, []>("hidden_states_37_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_16_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(670560064)))];
tensor<fp16, [1280]> layers_16_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683667328)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_37_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_16_fc2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("hidden_states_37_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor<string, []>("inputs_69_cast_fp16")];
tensor<int32, []> var_2184 = const()[name = tensor<string, []>("op_2184"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_69_axes_0 = const()[name = tensor<string, []>("out_69_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2203_to_fp16 = const()[name = tensor<string, []>("op_2203_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2203_to_fp16, x = inputs_69_cast_fp16)[name = tensor<string, []>("out_69_cast_fp16")];
tensor<fp16, [1280]> obj_69_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_69_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683669952)))];
tensor<fp16, [1280]> obj_69_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_69_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683672576)))];
tensor<fp16, []> obj_69_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_69_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor<string, []>("obj_69_cast_fp16")];
tensor<string, []> query_35_pad_type_0 = const()[name = tensor<string, []>("query_35_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_35_strides_0 = const()[name = tensor<string, []>("query_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_35_pad_0 = const()[name = tensor<string, []>("query_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_35_dilations_0 = const()[name = tensor<string, []>("query_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_35_groups_0 = const()[name = tensor<string, []>("query_35_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683675200)))];
tensor<fp16, [1280]> layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686952064)))];
tensor<fp16, [1, 1280, 1, 1500]> query_35_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor<string, []>("query_35_cast_fp16")];
tensor<string, []> key_35_pad_type_0 = const()[name = tensor<string, []>("key_35_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_35_strides_0 = const()[name = tensor<string, []>("key_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_35_pad_0 = const()[name = tensor<string, []>("key_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_35_dilations_0 = const()[name = tensor<string, []>("key_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_35_groups_0 = const()[name = tensor<string, []>("key_35_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686954688)))];
tensor<fp16, [1, 1280, 1, 1500]> key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor<string, []>("key_35_cast_fp16")];
tensor<string, []> value_35_pad_type_0 = const()[name = tensor<string, []>("value_35_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_35_strides_0 = const()[name = tensor<string, []>("value_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_35_pad_0 = const()[name = tensor<string, []>("value_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_35_dilations_0 = const()[name = tensor<string, []>("value_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_35_groups_0 = const()[name = tensor<string, []>("value_35_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(690231552)))];
tensor<fp16, [1280]> layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693508416)))];
tensor<fp16, [1, 1280, 1, 1500]> value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor<string, []>("value_35_cast_fp16")];
tensor<int32, [4]> var_2238 = const()[name = tensor<string, []>("op_2238"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_35_cast_fp16 = reshape(shape = var_2238, x = query_35_cast_fp16)[name = tensor<string, []>("mh_q_35_cast_fp16")];
tensor<fp16, []> var_2240_to_fp16 = const()[name = tensor<string, []>("op_2240_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2241_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_2240_to_fp16)[name = tensor<string, []>("op_2241_cast_fp16")];
tensor<int32, [4]> var_2242 = const()[name = tensor<string, []>("op_2242"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2243_cast_fp16 = reshape(shape = var_2242, x = key_35_cast_fp16)[name = tensor<string, []>("op_2243_cast_fp16")];
tensor<bool, []> mh_w_35_transpose_x_0 = const()[name = tensor<string, []>("mh_w_35_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_35_transpose_y_0 = const()[name = tensor<string, []>("mh_w_35_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_2241_cast_fp16, y = var_2243_cast_fp16)[name = tensor<string, []>("mh_w_35_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2246_cast_fp16 = softmax(axis = var_2184, x = mh_w_35_cast_fp16)[name = tensor<string, []>("op_2246_cast_fp16")];
tensor<int32, [4]> var_2247 = const()[name = tensor<string, []>("op_2247"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2248_cast_fp16 = reshape(shape = var_2247, x = value_35_cast_fp16)[name = tensor<string, []>("op_2248_cast_fp16")];
tensor<bool, []> attn_35_transpose_x_0 = const()[name = tensor<string, []>("attn_35_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_35_transpose_y_0 = const()[name = tensor<string, []>("attn_35_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2248_cast_fp16, y = var_2246_cast_fp16)[name = tensor<string, []>("attn_35_cast_fp16")];
tensor<int32, [4]> var_2251 = const()[name = tensor<string, []>("op_2251"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_137_cast_fp16 = reshape(shape = var_2251, x = attn_35_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
tensor<string, []> obj_71_pad_type_0 = const()[name = tensor<string, []>("obj_71_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_71_strides_0 = const()[name = tensor<string, []>("obj_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_71_pad_0 = const()[name = tensor<string, []>("obj_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_71_dilations_0 = const()[name = tensor<string, []>("obj_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_71_groups_0 = const()[name = tensor<string, []>("obj_71_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693511040)))];
tensor<fp16, [1280]> layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696787904)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_71_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = obj_71_dilations_0, groups = obj_71_groups_0, pad = obj_71_pad_0, pad_type = obj_71_pad_type_0, strides = obj_71_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("obj_71_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_71_cast_fp16)[name = tensor<string, []>("inputs_71_cast_fp16")];
tensor<int32, [1]> out_71_axes_0 = const()[name = tensor<string, []>("out_71_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2269_to_fp16 = const()[name = tensor<string, []>("op_2269_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2269_to_fp16, x = inputs_71_cast_fp16)[name = tensor<string, []>("out_71_cast_fp16")];
tensor<fp16, [1280]> input_139_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_139_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696790528)))];
tensor<fp16, [1280]> input_139_beta_0_to_fp16 = const()[name = tensor<string, []>("input_139_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696793152)))];
tensor<fp16, []> input_139_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_139_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_139_cast_fp16 = batch_norm(beta = input_139_beta_0_to_fp16, epsilon = input_139_epsilon_0_to_fp16, gamma = input_139_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
tensor<string, []> input_141_pad_type_0 = const()[name = tensor<string, []>("input_141_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_141_strides_0 = const()[name = tensor<string, []>("input_141_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_141_pad_0 = const()[name = tensor<string, []>("input_141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_141_dilations_0 = const()[name = tensor<string, []>("input_141_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_141_groups_0 = const()[name = tensor<string, []>("input_141_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_17_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696795776)))];
tensor<fp16, [5120]> layers_17_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(709903040)))];
tensor<fp16, [1, 5120, 1, 1500]> input_141_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = input_141_dilations_0, groups = input_141_groups_0, pad = input_141_pad_0, pad_type = input_141_pad_type_0, strides = input_141_strides_0, weight = layers_17_fc1_weight_to_fp16, x = input_139_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
tensor<string, []> input_143_mode_0 = const()[name = tensor<string, []>("input_143_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = input_141_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
tensor<string, []> hidden_states_39_pad_type_0 = const()[name = tensor<string, []>("hidden_states_39_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_39_strides_0 = const()[name = tensor<string, []>("hidden_states_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_39_pad_0 = const()[name = tensor<string, []>("hidden_states_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_39_dilations_0 = const()[name = tensor<string, []>("hidden_states_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_39_groups_0 = const()[name = tensor<string, []>("hidden_states_39_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_17_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(709913344)))];
tensor<fp16, [1280]> layers_17_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(723020608)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_39_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = layers_17_fc2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("hidden_states_39_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor<string, []>("inputs_73_cast_fp16")];
tensor<int32, []> var_2302 = const()[name = tensor<string, []>("op_2302"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_73_axes_0 = const()[name = tensor<string, []>("out_73_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2321_to_fp16 = const()[name = tensor<string, []>("op_2321_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2321_to_fp16, x = inputs_73_cast_fp16)[name = tensor<string, []>("out_73_cast_fp16")];
tensor<fp16, [1280]> obj_73_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_73_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(723023232)))];
tensor<fp16, [1280]> obj_73_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_73_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(723025856)))];
tensor<fp16, []> obj_73_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_73_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor<string, []>("obj_73_cast_fp16")];
tensor<string, []> query_37_pad_type_0 = const()[name = tensor<string, []>("query_37_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_37_strides_0 = const()[name = tensor<string, []>("query_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_37_pad_0 = const()[name = tensor<string, []>("query_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_37_dilations_0 = const()[name = tensor<string, []>("query_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_37_groups_0 = const()[name = tensor<string, []>("query_37_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(723028480)))];
tensor<fp16, [1280]> layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726305344)))];
tensor<fp16, [1, 1280, 1, 1500]> query_37_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")];
tensor<string, []> key_37_pad_type_0 = const()[name = tensor<string, []>("key_37_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_37_strides_0 = const()[name = tensor<string, []>("key_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_37_pad_0 = const()[name = tensor<string, []>("key_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_37_dilations_0 = const()[name = tensor<string, []>("key_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_37_groups_0 = const()[name = tensor<string, []>("key_37_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726307968)))];
tensor<fp16, [1, 1280, 1, 1500]> key_37_cast_fp16 = conv(dilations = key_37_dilations_0, groups = key_37_groups_0, pad = key_37_pad_0, pad_type = key_37_pad_type_0, strides = key_37_strides_0, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor<string, []>("key_37_cast_fp16")];
tensor<string, []> value_37_pad_type_0 = const()[name = tensor<string, []>("value_37_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_37_strides_0 = const()[name = tensor<string, []>("value_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_37_pad_0 = const()[name = tensor<string, []>("value_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_37_dilations_0 = const()[name = tensor<string, []>("value_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_37_groups_0 = const()[name = tensor<string, []>("value_37_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(729584832)))];
tensor<fp16, [1280]> layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(732861696)))];
tensor<fp16, [1, 1280, 1, 1500]> value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = value_37_dilations_0, groups = value_37_groups_0, pad = value_37_pad_0, pad_type = value_37_pad_type_0, strides = value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor<string, []>("value_37_cast_fp16")];
tensor<int32, [4]> var_2356 = const()[name = tensor<string, []>("op_2356"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_37_cast_fp16 = reshape(shape = var_2356, x = query_37_cast_fp16)[name = tensor<string, []>("mh_q_37_cast_fp16")];
tensor<fp16, []> var_2358_to_fp16 = const()[name = tensor<string, []>("op_2358_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2359_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2358_to_fp16)[name = tensor<string, []>("op_2359_cast_fp16")];
tensor<int32, [4]> var_2360 = const()[name = tensor<string, []>("op_2360"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2361_cast_fp16 = reshape(shape = var_2360, x = key_37_cast_fp16)[name = tensor<string, []>("op_2361_cast_fp16")];
tensor<bool, []> mh_w_37_transpose_x_0 = const()[name = tensor<string, []>("mh_w_37_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_37_transpose_y_0 = const()[name = tensor<string, []>("mh_w_37_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2359_cast_fp16, y = var_2361_cast_fp16)[name = tensor<string, []>("mh_w_37_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2364_cast_fp16 = softmax(axis = var_2302, x = mh_w_37_cast_fp16)[name = tensor<string, []>("op_2364_cast_fp16")];
tensor<int32, [4]> var_2365 = const()[name = tensor<string, []>("op_2365"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2366_cast_fp16 = reshape(shape = var_2365, x = value_37_cast_fp16)[name = tensor<string, []>("op_2366_cast_fp16")];
tensor<bool, []> attn_37_transpose_x_0 = const()[name = tensor<string, []>("attn_37_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_37_transpose_y_0 = const()[name = tensor<string, []>("attn_37_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2366_cast_fp16, y = var_2364_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
tensor<int32, [4]> var_2369 = const()[name = tensor<string, []>("op_2369"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_145_cast_fp16 = reshape(shape = var_2369, x = attn_37_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
tensor<string, []> obj_75_pad_type_0 = const()[name = tensor<string, []>("obj_75_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_75_strides_0 = const()[name = tensor<string, []>("obj_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_75_pad_0 = const()[name = tensor<string, []>("obj_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_75_dilations_0 = const()[name = tensor<string, []>("obj_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_75_groups_0 = const()[name = tensor<string, []>("obj_75_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(732864320)))];
tensor<fp16, [1280]> layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736141184)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_75_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = obj_75_dilations_0, groups = obj_75_groups_0, pad = obj_75_pad_0, pad_type = obj_75_pad_type_0, strides = obj_75_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("obj_75_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_75_cast_fp16)[name = tensor<string, []>("inputs_75_cast_fp16")];
tensor<int32, [1]> out_75_axes_0 = const()[name = tensor<string, []>("out_75_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2387_to_fp16 = const()[name = tensor<string, []>("op_2387_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2387_to_fp16, x = inputs_75_cast_fp16)[name = tensor<string, []>("out_75_cast_fp16")];
tensor<fp16, [1280]> input_147_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_147_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736143808)))];
tensor<fp16, [1280]> input_147_beta_0_to_fp16 = const()[name = tensor<string, []>("input_147_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736146432)))];
tensor<fp16, []> input_147_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_147_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_147_cast_fp16 = batch_norm(beta = input_147_beta_0_to_fp16, epsilon = input_147_epsilon_0_to_fp16, gamma = input_147_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
tensor<string, []> input_149_pad_type_0 = const()[name = tensor<string, []>("input_149_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_149_strides_0 = const()[name = tensor<string, []>("input_149_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_149_pad_0 = const()[name = tensor<string, []>("input_149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_149_dilations_0 = const()[name = tensor<string, []>("input_149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_149_groups_0 = const()[name = tensor<string, []>("input_149_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_18_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736149056)))];
tensor<fp16, [5120]> layers_18_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(749256320)))];
tensor<fp16, [1, 5120, 1, 1500]> input_149_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = layers_18_fc1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
tensor<string, []> input_151_mode_0 = const()[name = tensor<string, []>("input_151_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
tensor<string, []> hidden_states_41_pad_type_0 = const()[name = tensor<string, []>("hidden_states_41_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_41_strides_0 = const()[name = tensor<string, []>("hidden_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_41_pad_0 = const()[name = tensor<string, []>("hidden_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_41_dilations_0 = const()[name = tensor<string, []>("hidden_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_41_groups_0 = const()[name = tensor<string, []>("hidden_states_41_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_18_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(749266624)))];
tensor<fp16, [1280]> layers_18_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(762373888)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_41_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = hidden_states_41_dilations_0, groups = hidden_states_41_groups_0, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = hidden_states_41_strides_0, weight = layers_18_fc2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("hidden_states_41_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor<string, []>("inputs_77_cast_fp16")];
tensor<int32, []> var_2420 = const()[name = tensor<string, []>("op_2420"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_77_axes_0 = const()[name = tensor<string, []>("out_77_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2439_to_fp16 = const()[name = tensor<string, []>("op_2439_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2439_to_fp16, x = inputs_77_cast_fp16)[name = tensor<string, []>("out_77_cast_fp16")];
tensor<fp16, [1280]> obj_77_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_77_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(762376512)))];
tensor<fp16, [1280]> obj_77_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_77_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(762379136)))];
tensor<fp16, []> obj_77_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_77_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_77_cast_fp16 = batch_norm(beta = obj_77_beta_0_to_fp16, epsilon = obj_77_epsilon_0_to_fp16, gamma = obj_77_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor<string, []>("obj_77_cast_fp16")];
tensor<string, []> query_39_pad_type_0 = const()[name = tensor<string, []>("query_39_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_39_strides_0 = const()[name = tensor<string, []>("query_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_39_pad_0 = const()[name = tensor<string, []>("query_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_39_dilations_0 = const()[name = tensor<string, []>("query_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_39_groups_0 = const()[name = tensor<string, []>("query_39_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(762381760)))];
tensor<fp16, [1280]> layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765658624)))];
tensor<fp16, [1, 1280, 1, 1500]> query_39_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor<string, []>("query_39_cast_fp16")];
tensor<string, []> key_39_pad_type_0 = const()[name = tensor<string, []>("key_39_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_39_strides_0 = const()[name = tensor<string, []>("key_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_39_pad_0 = const()[name = tensor<string, []>("key_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_39_dilations_0 = const()[name = tensor<string, []>("key_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_39_groups_0 = const()[name = tensor<string, []>("key_39_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765661248)))];
tensor<fp16, [1, 1280, 1, 1500]> key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor<string, []>("key_39_cast_fp16")];
tensor<string, []> value_39_pad_type_0 = const()[name = tensor<string, []>("value_39_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_39_strides_0 = const()[name = tensor<string, []>("value_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_39_pad_0 = const()[name = tensor<string, []>("value_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_39_dilations_0 = const()[name = tensor<string, []>("value_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_39_groups_0 = const()[name = tensor<string, []>("value_39_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(768938112)))];
tensor<fp16, [1280]> layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(772214976)))];
tensor<fp16, [1, 1280, 1, 1500]> value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor<string, []>("value_39_cast_fp16")];
tensor<int32, [4]> var_2474 = const()[name = tensor<string, []>("op_2474"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_39_cast_fp16 = reshape(shape = var_2474, x = query_39_cast_fp16)[name = tensor<string, []>("mh_q_39_cast_fp16")];
tensor<fp16, []> var_2476_to_fp16 = const()[name = tensor<string, []>("op_2476_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2477_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2476_to_fp16)[name = tensor<string, []>("op_2477_cast_fp16")];
tensor<int32, [4]> var_2478 = const()[name = tensor<string, []>("op_2478"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2479_cast_fp16 = reshape(shape = var_2478, x = key_39_cast_fp16)[name = tensor<string, []>("op_2479_cast_fp16")];
tensor<bool, []> mh_w_39_transpose_x_0 = const()[name = tensor<string, []>("mh_w_39_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_39_transpose_y_0 = const()[name = tensor<string, []>("mh_w_39_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_39_cast_fp16 = matmul(transpose_x = mh_w_39_transpose_x_0, transpose_y = mh_w_39_transpose_y_0, x = var_2477_cast_fp16, y = var_2479_cast_fp16)[name = tensor<string, []>("mh_w_39_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2482_cast_fp16 = softmax(axis = var_2420, x = mh_w_39_cast_fp16)[name = tensor<string, []>("op_2482_cast_fp16")];
tensor<int32, [4]> var_2483 = const()[name = tensor<string, []>("op_2483"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2484_cast_fp16 = reshape(shape = var_2483, x = value_39_cast_fp16)[name = tensor<string, []>("op_2484_cast_fp16")];
tensor<bool, []> attn_39_transpose_x_0 = const()[name = tensor<string, []>("attn_39_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_39_transpose_y_0 = const()[name = tensor<string, []>("attn_39_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2484_cast_fp16, y = var_2482_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")];
tensor<int32, [4]> var_2487 = const()[name = tensor<string, []>("op_2487"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_153_cast_fp16 = reshape(shape = var_2487, x = attn_39_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
tensor<string, []> obj_79_pad_type_0 = const()[name = tensor<string, []>("obj_79_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_79_strides_0 = const()[name = tensor<string, []>("obj_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_79_pad_0 = const()[name = tensor<string, []>("obj_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_79_dilations_0 = const()[name = tensor<string, []>("obj_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_79_groups_0 = const()[name = tensor<string, []>("obj_79_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(772217600)))];
tensor<fp16, [1280]> layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(775494464)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_79_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = obj_79_dilations_0, groups = obj_79_groups_0, pad = obj_79_pad_0, pad_type = obj_79_pad_type_0, strides = obj_79_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("obj_79_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = obj_79_cast_fp16)[name = tensor<string, []>("inputs_79_cast_fp16")];
tensor<int32, [1]> out_79_axes_0 = const()[name = tensor<string, []>("out_79_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2505_to_fp16 = const()[name = tensor<string, []>("op_2505_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_2505_to_fp16, x = inputs_79_cast_fp16)[name = tensor<string, []>("out_79_cast_fp16")];
tensor<fp16, [1280]> input_155_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_155_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(775497088)))];
tensor<fp16, [1280]> input_155_beta_0_to_fp16 = const()[name = tensor<string, []>("input_155_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(775499712)))];
tensor<fp16, []> input_155_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_155_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
tensor<string, []> input_157_pad_type_0 = const()[name = tensor<string, []>("input_157_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_157_strides_0 = const()[name = tensor<string, []>("input_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_157_pad_0 = const()[name = tensor<string, []>("input_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_157_dilations_0 = const()[name = tensor<string, []>("input_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_157_groups_0 = const()[name = tensor<string, []>("input_157_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_19_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(775502336)))];
tensor<fp16, [5120]> layers_19_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(788609600)))];
tensor<fp16, [1, 5120, 1, 1500]> input_157_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = layers_19_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
tensor<string, []> input_159_mode_0 = const()[name = tensor<string, []>("input_159_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
tensor<string, []> hidden_states_43_pad_type_0 = const()[name = tensor<string, []>("hidden_states_43_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = tensor<string, []>("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = tensor<string, []>("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = tensor<string, []>("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_43_groups_0 = const()[name = tensor<string, []>("hidden_states_43_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_19_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(788619904)))];
tensor<fp16, [1280]> layers_19_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(801727168)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_43_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_19_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("hidden_states_43_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor<string, []>("inputs_81_cast_fp16")];
tensor<int32, []> var_2538 = const()[name = tensor<string, []>("op_2538"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_81_axes_0 = const()[name = tensor<string, []>("out_81_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2557_to_fp16 = const()[name = tensor<string, []>("op_2557_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_2557_to_fp16, x = inputs_81_cast_fp16)[name = tensor<string, []>("out_81_cast_fp16")];
tensor<fp16, [1280]> obj_81_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_81_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(801729792)))];
tensor<fp16, [1280]> obj_81_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_81_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(801732416)))];
tensor<fp16, []> obj_81_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_81_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor<string, []>("obj_81_cast_fp16")];
tensor<string, []> query_41_pad_type_0 = const()[name = tensor<string, []>("query_41_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_41_strides_0 = const()[name = tensor<string, []>("query_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_41_pad_0 = const()[name = tensor<string, []>("query_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_41_dilations_0 = const()[name = tensor<string, []>("query_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_41_groups_0 = const()[name = tensor<string, []>("query_41_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(801735040)))];
tensor<fp16, [1280]> layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805011904)))];
tensor<fp16, [1, 1280, 1, 1500]> query_41_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")];
tensor<string, []> key_41_pad_type_0 = const()[name = tensor<string, []>("key_41_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_41_strides_0 = const()[name = tensor<string, []>("key_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_41_pad_0 = const()[name = tensor<string, []>("key_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_41_dilations_0 = const()[name = tensor<string, []>("key_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_41_groups_0 = const()[name = tensor<string, []>("key_41_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805014528)))];
tensor<fp16, [1, 1280, 1, 1500]> key_41_cast_fp16 = conv(dilations = key_41_dilations_0, groups = key_41_groups_0, pad = key_41_pad_0, pad_type = key_41_pad_type_0, strides = key_41_strides_0, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor<string, []>("key_41_cast_fp16")];
tensor<string, []> value_41_pad_type_0 = const()[name = tensor<string, []>("value_41_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_41_strides_0 = const()[name = tensor<string, []>("value_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_41_pad_0 = const()[name = tensor<string, []>("value_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_41_dilations_0 = const()[name = tensor<string, []>("value_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_41_groups_0 = const()[name = tensor<string, []>("value_41_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(808291392)))];
tensor<fp16, [1280]> layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(811568256)))];
tensor<fp16, [1, 1280, 1, 1500]> value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = value_41_dilations_0, groups = value_41_groups_0, pad = value_41_pad_0, pad_type = value_41_pad_type_0, strides = value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor<string, []>("value_41_cast_fp16")];
tensor<int32, [4]> var_2592 = const()[name = tensor<string, []>("op_2592"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_41_cast_fp16 = reshape(shape = var_2592, x = query_41_cast_fp16)[name = tensor<string, []>("mh_q_41_cast_fp16")];
tensor<fp16, []> var_2594_to_fp16 = const()[name = tensor<string, []>("op_2594_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2595_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2594_to_fp16)[name = tensor<string, []>("op_2595_cast_fp16")];
tensor<int32, [4]> var_2596 = const()[name = tensor<string, []>("op_2596"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2597_cast_fp16 = reshape(shape = var_2596, x = key_41_cast_fp16)[name = tensor<string, []>("op_2597_cast_fp16")];
tensor<bool, []> mh_w_41_transpose_x_0 = const()[name = tensor<string, []>("mh_w_41_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_41_transpose_y_0 = const()[name = tensor<string, []>("mh_w_41_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_2595_cast_fp16, y = var_2597_cast_fp16)[name = tensor<string, []>("mh_w_41_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2600_cast_fp16 = softmax(axis = var_2538, x = mh_w_41_cast_fp16)[name = tensor<string, []>("op_2600_cast_fp16")];
tensor<int32, [4]> var_2601 = const()[name = tensor<string, []>("op_2601"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2602_cast_fp16 = reshape(shape = var_2601, x = value_41_cast_fp16)[name = tensor<string, []>("op_2602_cast_fp16")];
tensor<bool, []> attn_41_transpose_x_0 = const()[name = tensor<string, []>("attn_41_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_41_transpose_y_0 = const()[name = tensor<string, []>("attn_41_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2602_cast_fp16, y = var_2600_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
tensor<int32, [4]> var_2605 = const()[name = tensor<string, []>("op_2605"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_161_cast_fp16 = reshape(shape = var_2605, x = attn_41_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
tensor<string, []> obj_83_pad_type_0 = const()[name = tensor<string, []>("obj_83_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_83_strides_0 = const()[name = tensor<string, []>("obj_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_83_pad_0 = const()[name = tensor<string, []>("obj_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_83_dilations_0 = const()[name = tensor<string, []>("obj_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_83_groups_0 = const()[name = tensor<string, []>("obj_83_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(811570880)))];
tensor<fp16, [1280]> layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(814847744)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_83_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = obj_83_dilations_0, groups = obj_83_groups_0, pad = obj_83_pad_0, pad_type = obj_83_pad_type_0, strides = obj_83_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("obj_83_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_83_cast_fp16)[name = tensor<string, []>("inputs_83_cast_fp16")];
tensor<int32, [1]> out_83_axes_0 = const()[name = tensor<string, []>("out_83_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2623_to_fp16 = const()[name = tensor<string, []>("op_2623_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_2623_to_fp16, x = inputs_83_cast_fp16)[name = tensor<string, []>("out_83_cast_fp16")];
tensor<fp16, [1280]> input_163_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_163_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(814850368)))];
tensor<fp16, [1280]> input_163_beta_0_to_fp16 = const()[name = tensor<string, []>("input_163_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(814852992)))];
tensor<fp16, []> input_163_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_163_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_163_cast_fp16 = batch_norm(beta = input_163_beta_0_to_fp16, epsilon = input_163_epsilon_0_to_fp16, gamma = input_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
tensor<string, []> input_165_pad_type_0 = const()[name = tensor<string, []>("input_165_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_165_strides_0 = const()[name = tensor<string, []>("input_165_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_165_pad_0 = const()[name = tensor<string, []>("input_165_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_165_dilations_0 = const()[name = tensor<string, []>("input_165_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_165_groups_0 = const()[name = tensor<string, []>("input_165_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_20_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(814855616)))];
tensor<fp16, [5120]> layers_20_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(827962880)))];
tensor<fp16, [1, 5120, 1, 1500]> input_165_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = input_165_dilations_0, groups = input_165_groups_0, pad = input_165_pad_0, pad_type = input_165_pad_type_0, strides = input_165_strides_0, weight = layers_20_fc1_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
tensor<string, []> input_167_mode_0 = const()[name = tensor<string, []>("input_167_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
tensor<string, []> hidden_states_45_pad_type_0 = const()[name = tensor<string, []>("hidden_states_45_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_45_strides_0 = const()[name = tensor<string, []>("hidden_states_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_45_pad_0 = const()[name = tensor<string, []>("hidden_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_45_dilations_0 = const()[name = tensor<string, []>("hidden_states_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_45_groups_0 = const()[name = tensor<string, []>("hidden_states_45_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_20_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(827973184)))];
tensor<fp16, [1280]> layers_20_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841080448)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_45_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = layers_20_fc2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("hidden_states_45_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor<string, []>("inputs_85_cast_fp16")];
tensor<int32, []> var_2656 = const()[name = tensor<string, []>("op_2656"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_85_axes_0 = const()[name = tensor<string, []>("out_85_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2675_to_fp16 = const()[name = tensor<string, []>("op_2675_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_2675_to_fp16, x = inputs_85_cast_fp16)[name = tensor<string, []>("out_85_cast_fp16")];
tensor<fp16, [1280]> obj_85_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_85_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841083072)))];
tensor<fp16, [1280]> obj_85_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_85_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841085696)))];
tensor<fp16, []> obj_85_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_85_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor<string, []>("obj_85_cast_fp16")];
tensor<string, []> query_43_pad_type_0 = const()[name = tensor<string, []>("query_43_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_43_strides_0 = const()[name = tensor<string, []>("query_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_43_pad_0 = const()[name = tensor<string, []>("query_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_43_dilations_0 = const()[name = tensor<string, []>("query_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_43_groups_0 = const()[name = tensor<string, []>("query_43_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841088320)))];
tensor<fp16, [1280]> layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(844365184)))];
tensor<fp16, [1, 1280, 1, 1500]> query_43_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("query_43_cast_fp16")];
tensor<string, []> key_43_pad_type_0 = const()[name = tensor<string, []>("key_43_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_43_strides_0 = const()[name = tensor<string, []>("key_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_43_pad_0 = const()[name = tensor<string, []>("key_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_43_dilations_0 = const()[name = tensor<string, []>("key_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_43_groups_0 = const()[name = tensor<string, []>("key_43_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(844367808)))];
tensor<fp16, [1, 1280, 1, 1500]> key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("key_43_cast_fp16")];
tensor<string, []> value_43_pad_type_0 = const()[name = tensor<string, []>("value_43_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_43_strides_0 = const()[name = tensor<string, []>("value_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_43_pad_0 = const()[name = tensor<string, []>("value_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_43_dilations_0 = const()[name = tensor<string, []>("value_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_43_groups_0 = const()[name = tensor<string, []>("value_43_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(847644672)))];
tensor<fp16, [1280]> layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(850921536)))];
tensor<fp16, [1, 1280, 1, 1500]> value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("value_43_cast_fp16")];
tensor<int32, [4]> var_2710 = const()[name = tensor<string, []>("op_2710"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_43_cast_fp16 = reshape(shape = var_2710, x = query_43_cast_fp16)[name = tensor<string, []>("mh_q_43_cast_fp16")];
tensor<fp16, []> var_2712_to_fp16 = const()[name = tensor<string, []>("op_2712_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2713_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2712_to_fp16)[name = tensor<string, []>("op_2713_cast_fp16")];
tensor<int32, [4]> var_2714 = const()[name = tensor<string, []>("op_2714"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2715_cast_fp16 = reshape(shape = var_2714, x = key_43_cast_fp16)[name = tensor<string, []>("op_2715_cast_fp16")];
tensor<bool, []> mh_w_43_transpose_x_0 = const()[name = tensor<string, []>("mh_w_43_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_43_transpose_y_0 = const()[name = tensor<string, []>("mh_w_43_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_2713_cast_fp16, y = var_2715_cast_fp16)[name = tensor<string, []>("mh_w_43_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2718_cast_fp16 = softmax(axis = var_2656, x = mh_w_43_cast_fp16)[name = tensor<string, []>("op_2718_cast_fp16")];
tensor<int32, [4]> var_2719 = const()[name = tensor<string, []>("op_2719"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2720_cast_fp16 = reshape(shape = var_2719, x = value_43_cast_fp16)[name = tensor<string, []>("op_2720_cast_fp16")];
tensor<bool, []> attn_43_transpose_x_0 = const()[name = tensor<string, []>("attn_43_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_43_transpose_y_0 = const()[name = tensor<string, []>("attn_43_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2720_cast_fp16, y = var_2718_cast_fp16)[name = tensor<string, []>("attn_43_cast_fp16")];
tensor<int32, [4]> var_2723 = const()[name = tensor<string, []>("op_2723"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_169_cast_fp16 = reshape(shape = var_2723, x = attn_43_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
tensor<string, []> obj_87_pad_type_0 = const()[name = tensor<string, []>("obj_87_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_87_strides_0 = const()[name = tensor<string, []>("obj_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_87_pad_0 = const()[name = tensor<string, []>("obj_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_87_dilations_0 = const()[name = tensor<string, []>("obj_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_87_groups_0 = const()[name = tensor<string, []>("obj_87_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(850924160)))];
tensor<fp16, [1280]> layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854201024)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_87_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = obj_87_dilations_0, groups = obj_87_groups_0, pad = obj_87_pad_0, pad_type = obj_87_pad_type_0, strides = obj_87_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("obj_87_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_87_cast_fp16)[name = tensor<string, []>("inputs_87_cast_fp16")];
tensor<int32, [1]> out_87_axes_0 = const()[name = tensor<string, []>("out_87_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2741_to_fp16 = const()[name = tensor<string, []>("op_2741_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_2741_to_fp16, x = inputs_87_cast_fp16)[name = tensor<string, []>("out_87_cast_fp16")];
tensor<fp16, [1280]> input_171_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_171_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854203648)))];
tensor<fp16, [1280]> input_171_beta_0_to_fp16 = const()[name = tensor<string, []>("input_171_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854206272)))];
tensor<fp16, []> input_171_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_171_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_171_cast_fp16 = batch_norm(beta = input_171_beta_0_to_fp16, epsilon = input_171_epsilon_0_to_fp16, gamma = input_171_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
tensor<string, []> input_173_pad_type_0 = const()[name = tensor<string, []>("input_173_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_173_strides_0 = const()[name = tensor<string, []>("input_173_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_173_pad_0 = const()[name = tensor<string, []>("input_173_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_173_dilations_0 = const()[name = tensor<string, []>("input_173_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_173_groups_0 = const()[name = tensor<string, []>("input_173_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_21_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854208896)))];
tensor<fp16, [5120]> layers_21_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(867316160)))];
tensor<fp16, [1, 5120, 1, 1500]> input_173_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = input_173_dilations_0, groups = input_173_groups_0, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = input_173_strides_0, weight = layers_21_fc1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
tensor<string, []> input_175_mode_0 = const()[name = tensor<string, []>("input_175_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
tensor<string, []> hidden_states_47_pad_type_0 = const()[name = tensor<string, []>("hidden_states_47_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = tensor<string, []>("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = tensor<string, []>("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = tensor<string, []>("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_47_groups_0 = const()[name = tensor<string, []>("hidden_states_47_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_21_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(867326464)))];
tensor<fp16, [1280]> layers_21_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(880433728)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_47_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_21_fc2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("hidden_states_47_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor<string, []>("inputs_89_cast_fp16")];
tensor<int32, []> var_2774 = const()[name = tensor<string, []>("op_2774"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_89_axes_0 = const()[name = tensor<string, []>("out_89_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2793_to_fp16 = const()[name = tensor<string, []>("op_2793_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_2793_to_fp16, x = inputs_89_cast_fp16)[name = tensor<string, []>("out_89_cast_fp16")];
tensor<fp16, [1280]> obj_89_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_89_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(880436352)))];
tensor<fp16, [1280]> obj_89_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_89_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(880438976)))];
tensor<fp16, []> obj_89_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_89_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_89_cast_fp16 = batch_norm(beta = obj_89_beta_0_to_fp16, epsilon = obj_89_epsilon_0_to_fp16, gamma = obj_89_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor<string, []>("obj_89_cast_fp16")];
tensor<string, []> query_45_pad_type_0 = const()[name = tensor<string, []>("query_45_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_45_strides_0 = const()[name = tensor<string, []>("query_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_45_pad_0 = const()[name = tensor<string, []>("query_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_45_dilations_0 = const()[name = tensor<string, []>("query_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_45_groups_0 = const()[name = tensor<string, []>("query_45_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(880441600)))];
tensor<fp16, [1280]> layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883718464)))];
tensor<fp16, [1, 1280, 1, 1500]> query_45_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")];
tensor<string, []> key_45_pad_type_0 = const()[name = tensor<string, []>("key_45_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_45_strides_0 = const()[name = tensor<string, []>("key_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_45_pad_0 = const()[name = tensor<string, []>("key_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_45_dilations_0 = const()[name = tensor<string, []>("key_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_45_groups_0 = const()[name = tensor<string, []>("key_45_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883721088)))];
tensor<fp16, [1, 1280, 1, 1500]> key_45_cast_fp16 = conv(dilations = key_45_dilations_0, groups = key_45_groups_0, pad = key_45_pad_0, pad_type = key_45_pad_type_0, strides = key_45_strides_0, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor<string, []>("key_45_cast_fp16")];
tensor<string, []> value_45_pad_type_0 = const()[name = tensor<string, []>("value_45_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_45_strides_0 = const()[name = tensor<string, []>("value_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_45_pad_0 = const()[name = tensor<string, []>("value_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_45_dilations_0 = const()[name = tensor<string, []>("value_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_45_groups_0 = const()[name = tensor<string, []>("value_45_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(886997952)))];
tensor<fp16, [1280]> layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(890274816)))];
tensor<fp16, [1, 1280, 1, 1500]> value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = value_45_dilations_0, groups = value_45_groups_0, pad = value_45_pad_0, pad_type = value_45_pad_type_0, strides = value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor<string, []>("value_45_cast_fp16")];
tensor<int32, [4]> var_2828 = const()[name = tensor<string, []>("op_2828"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_45_cast_fp16 = reshape(shape = var_2828, x = query_45_cast_fp16)[name = tensor<string, []>("mh_q_45_cast_fp16")];
tensor<fp16, []> var_2830_to_fp16 = const()[name = tensor<string, []>("op_2830_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2831_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2830_to_fp16)[name = tensor<string, []>("op_2831_cast_fp16")];
tensor<int32, [4]> var_2832 = const()[name = tensor<string, []>("op_2832"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2833_cast_fp16 = reshape(shape = var_2832, x = key_45_cast_fp16)[name = tensor<string, []>("op_2833_cast_fp16")];
tensor<bool, []> mh_w_45_transpose_x_0 = const()[name = tensor<string, []>("mh_w_45_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_45_transpose_y_0 = const()[name = tensor<string, []>("mh_w_45_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_2831_cast_fp16, y = var_2833_cast_fp16)[name = tensor<string, []>("mh_w_45_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2836_cast_fp16 = softmax(axis = var_2774, x = mh_w_45_cast_fp16)[name = tensor<string, []>("op_2836_cast_fp16")];
tensor<int32, [4]> var_2837 = const()[name = tensor<string, []>("op_2837"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2838_cast_fp16 = reshape(shape = var_2837, x = value_45_cast_fp16)[name = tensor<string, []>("op_2838_cast_fp16")];
tensor<bool, []> attn_45_transpose_x_0 = const()[name = tensor<string, []>("attn_45_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_45_transpose_y_0 = const()[name = tensor<string, []>("attn_45_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2838_cast_fp16, y = var_2836_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
tensor<int32, [4]> var_2841 = const()[name = tensor<string, []>("op_2841"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_177_cast_fp16 = reshape(shape = var_2841, x = attn_45_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
tensor<string, []> obj_91_pad_type_0 = const()[name = tensor<string, []>("obj_91_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_91_strides_0 = const()[name = tensor<string, []>("obj_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_91_pad_0 = const()[name = tensor<string, []>("obj_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_91_dilations_0 = const()[name = tensor<string, []>("obj_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_91_groups_0 = const()[name = tensor<string, []>("obj_91_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(890277440)))];
tensor<fp16, [1280]> layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893554304)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_91_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("obj_91_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = obj_91_cast_fp16)[name = tensor<string, []>("inputs_91_cast_fp16")];
tensor<int32, [1]> out_91_axes_0 = const()[name = tensor<string, []>("out_91_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2859_to_fp16 = const()[name = tensor<string, []>("op_2859_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_2859_to_fp16, x = inputs_91_cast_fp16)[name = tensor<string, []>("out_91_cast_fp16")];
tensor<fp16, [1280]> input_179_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_179_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893556928)))];
tensor<fp16, [1280]> input_179_beta_0_to_fp16 = const()[name = tensor<string, []>("input_179_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893559552)))];
tensor<fp16, []> input_179_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_179_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_179_cast_fp16 = batch_norm(beta = input_179_beta_0_to_fp16, epsilon = input_179_epsilon_0_to_fp16, gamma = input_179_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
tensor<string, []> input_181_pad_type_0 = const()[name = tensor<string, []>("input_181_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_181_strides_0 = const()[name = tensor<string, []>("input_181_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_181_pad_0 = const()[name = tensor<string, []>("input_181_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_181_dilations_0 = const()[name = tensor<string, []>("input_181_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_181_groups_0 = const()[name = tensor<string, []>("input_181_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_22_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893562176)))];
tensor<fp16, [5120]> layers_22_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(906669440)))];
tensor<fp16, [1, 5120, 1, 1500]> input_181_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = input_181_dilations_0, groups = input_181_groups_0, pad = input_181_pad_0, pad_type = input_181_pad_type_0, strides = input_181_strides_0, weight = layers_22_fc1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
tensor<string, []> input_183_mode_0 = const()[name = tensor<string, []>("input_183_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_183_cast_fp16 = gelu(mode = input_183_mode_0, x = input_181_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
tensor<string, []> hidden_states_49_pad_type_0 = const()[name = tensor<string, []>("hidden_states_49_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_49_strides_0 = const()[name = tensor<string, []>("hidden_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_49_pad_0 = const()[name = tensor<string, []>("hidden_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_49_dilations_0 = const()[name = tensor<string, []>("hidden_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_49_groups_0 = const()[name = tensor<string, []>("hidden_states_49_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_22_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(906679744)))];
tensor<fp16, [1280]> layers_22_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(919787008)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_49_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = hidden_states_49_dilations_0, groups = hidden_states_49_groups_0, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = hidden_states_49_strides_0, weight = layers_22_fc2_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("hidden_states_49_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor<string, []>("inputs_93_cast_fp16")];
tensor<int32, []> var_2892 = const()[name = tensor<string, []>("op_2892"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_93_axes_0 = const()[name = tensor<string, []>("out_93_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2911_to_fp16 = const()[name = tensor<string, []>("op_2911_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_2911_to_fp16, x = inputs_93_cast_fp16)[name = tensor<string, []>("out_93_cast_fp16")];
tensor<fp16, [1280]> obj_93_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_93_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(919789632)))];
tensor<fp16, [1280]> obj_93_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_93_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(919792256)))];
tensor<fp16, []> obj_93_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_93_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor<string, []>("obj_93_cast_fp16")];
tensor<string, []> query_47_pad_type_0 = const()[name = tensor<string, []>("query_47_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_47_strides_0 = const()[name = tensor<string, []>("query_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_47_pad_0 = const()[name = tensor<string, []>("query_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_47_dilations_0 = const()[name = tensor<string, []>("query_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_47_groups_0 = const()[name = tensor<string, []>("query_47_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(919794880)))];
tensor<fp16, [1280]> layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(923071744)))];
tensor<fp16, [1, 1280, 1, 1500]> query_47_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = query_47_dilations_0, groups = query_47_groups_0, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = query_47_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor<string, []>("query_47_cast_fp16")];
tensor<string, []> key_47_pad_type_0 = const()[name = tensor<string, []>("key_47_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_47_strides_0 = const()[name = tensor<string, []>("key_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_47_pad_0 = const()[name = tensor<string, []>("key_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_47_dilations_0 = const()[name = tensor<string, []>("key_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_47_groups_0 = const()[name = tensor<string, []>("key_47_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(923074368)))];
tensor<fp16, [1, 1280, 1, 1500]> key_47_cast_fp16 = conv(dilations = key_47_dilations_0, groups = key_47_groups_0, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = key_47_strides_0, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor<string, []>("key_47_cast_fp16")];
tensor<string, []> value_47_pad_type_0 = const()[name = tensor<string, []>("value_47_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_47_strides_0 = const()[name = tensor<string, []>("value_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_47_pad_0 = const()[name = tensor<string, []>("value_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_47_dilations_0 = const()[name = tensor<string, []>("value_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_47_groups_0 = const()[name = tensor<string, []>("value_47_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(926351232)))];
tensor<fp16, [1280]> layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(929628096)))];
tensor<fp16, [1, 1280, 1, 1500]> value_47_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = value_47_dilations_0, groups = value_47_groups_0, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = value_47_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor<string, []>("value_47_cast_fp16")];
tensor<int32, [4]> var_2946 = const()[name = tensor<string, []>("op_2946"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_47_cast_fp16 = reshape(shape = var_2946, x = query_47_cast_fp16)[name = tensor<string, []>("mh_q_47_cast_fp16")];
tensor<fp16, []> var_2948_to_fp16 = const()[name = tensor<string, []>("op_2948_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_2949_cast_fp16 = mul(x = mh_q_47_cast_fp16, y = var_2948_to_fp16)[name = tensor<string, []>("op_2949_cast_fp16")];
tensor<int32, [4]> var_2950 = const()[name = tensor<string, []>("op_2950"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2951_cast_fp16 = reshape(shape = var_2950, x = key_47_cast_fp16)[name = tensor<string, []>("op_2951_cast_fp16")];
tensor<bool, []> mh_w_47_transpose_x_0 = const()[name = tensor<string, []>("mh_w_47_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_47_transpose_y_0 = const()[name = tensor<string, []>("mh_w_47_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_2949_cast_fp16, y = var_2951_cast_fp16)[name = tensor<string, []>("mh_w_47_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_2954_cast_fp16 = softmax(axis = var_2892, x = mh_w_47_cast_fp16)[name = tensor<string, []>("op_2954_cast_fp16")];
tensor<int32, [4]> var_2955 = const()[name = tensor<string, []>("op_2955"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_2956_cast_fp16 = reshape(shape = var_2955, x = value_47_cast_fp16)[name = tensor<string, []>("op_2956_cast_fp16")];
tensor<bool, []> attn_47_transpose_x_0 = const()[name = tensor<string, []>("attn_47_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_47_transpose_y_0 = const()[name = tensor<string, []>("attn_47_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2956_cast_fp16, y = var_2954_cast_fp16)[name = tensor<string, []>("attn_47_cast_fp16")];
tensor<int32, [4]> var_2959 = const()[name = tensor<string, []>("op_2959"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_185_cast_fp16 = reshape(shape = var_2959, x = attn_47_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
tensor<string, []> obj_95_pad_type_0 = const()[name = tensor<string, []>("obj_95_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_95_strides_0 = const()[name = tensor<string, []>("obj_95_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_95_pad_0 = const()[name = tensor<string, []>("obj_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_95_dilations_0 = const()[name = tensor<string, []>("obj_95_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_95_groups_0 = const()[name = tensor<string, []>("obj_95_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(929630720)))];
tensor<fp16, [1280]> layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(932907584)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_95_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("obj_95_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_95_cast_fp16)[name = tensor<string, []>("inputs_95_cast_fp16")];
tensor<int32, [1]> out_95_axes_0 = const()[name = tensor<string, []>("out_95_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_2977_to_fp16 = const()[name = tensor<string, []>("op_2977_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_2977_to_fp16, x = inputs_95_cast_fp16)[name = tensor<string, []>("out_95_cast_fp16")];
tensor<fp16, [1280]> input_187_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_187_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(932910208)))];
tensor<fp16, [1280]> input_187_beta_0_to_fp16 = const()[name = tensor<string, []>("input_187_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(932912832)))];
tensor<fp16, []> input_187_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_187_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_187_cast_fp16 = batch_norm(beta = input_187_beta_0_to_fp16, epsilon = input_187_epsilon_0_to_fp16, gamma = input_187_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
tensor<string, []> input_189_pad_type_0 = const()[name = tensor<string, []>("input_189_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_189_strides_0 = const()[name = tensor<string, []>("input_189_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_189_pad_0 = const()[name = tensor<string, []>("input_189_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_189_dilations_0 = const()[name = tensor<string, []>("input_189_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_189_groups_0 = const()[name = tensor<string, []>("input_189_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_23_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(932915456)))];
tensor<fp16, [5120]> layers_23_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(946022720)))];
tensor<fp16, [1, 5120, 1, 1500]> input_189_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = layers_23_fc1_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
tensor<string, []> input_191_mode_0 = const()[name = tensor<string, []>("input_191_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
tensor<string, []> hidden_states_51_pad_type_0 = const()[name = tensor<string, []>("hidden_states_51_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_51_strides_0 = const()[name = tensor<string, []>("hidden_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_51_pad_0 = const()[name = tensor<string, []>("hidden_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_51_dilations_0 = const()[name = tensor<string, []>("hidden_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_51_groups_0 = const()[name = tensor<string, []>("hidden_states_51_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_23_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(946033024)))];
tensor<fp16, [1280]> layers_23_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(959140288)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_51_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = hidden_states_51_dilations_0, groups = hidden_states_51_groups_0, pad = hidden_states_51_pad_0, pad_type = hidden_states_51_pad_type_0, strides = hidden_states_51_strides_0, weight = layers_23_fc2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("hidden_states_51_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor<string, []>("inputs_97_cast_fp16")];
tensor<int32, []> var_3010 = const()[name = tensor<string, []>("op_3010"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_97_axes_0 = const()[name = tensor<string, []>("out_97_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3029_to_fp16 = const()[name = tensor<string, []>("op_3029_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_3029_to_fp16, x = inputs_97_cast_fp16)[name = tensor<string, []>("out_97_cast_fp16")];
tensor<fp16, [1280]> obj_97_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_97_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(959142912)))];
tensor<fp16, [1280]> obj_97_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_97_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(959145536)))];
tensor<fp16, []> obj_97_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_97_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor<string, []>("obj_97_cast_fp16")];
tensor<string, []> query_49_pad_type_0 = const()[name = tensor<string, []>("query_49_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_49_strides_0 = const()[name = tensor<string, []>("query_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_49_pad_0 = const()[name = tensor<string, []>("query_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_49_dilations_0 = const()[name = tensor<string, []>("query_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_49_groups_0 = const()[name = tensor<string, []>("query_49_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(959148160)))];
tensor<fp16, [1280]> layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(962425024)))];
tensor<fp16, [1, 1280, 1, 1500]> query_49_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_bias_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_24_self_attn_q_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor<string, []>("query_49_cast_fp16")];
tensor<string, []> key_49_pad_type_0 = const()[name = tensor<string, []>("key_49_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_49_strides_0 = const()[name = tensor<string, []>("key_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_49_pad_0 = const()[name = tensor<string, []>("key_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_49_dilations_0 = const()[name = tensor<string, []>("key_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_49_groups_0 = const()[name = tensor<string, []>("key_49_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(962427648)))];
tensor<fp16, [1, 1280, 1, 1500]> key_49_cast_fp16 = conv(dilations = key_49_dilations_0, groups = key_49_groups_0, pad = key_49_pad_0, pad_type = key_49_pad_type_0, strides = key_49_strides_0, weight = layers_24_self_attn_k_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor<string, []>("key_49_cast_fp16")];
tensor<string, []> value_49_pad_type_0 = const()[name = tensor<string, []>("value_49_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_49_strides_0 = const()[name = tensor<string, []>("value_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_49_pad_0 = const()[name = tensor<string, []>("value_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_49_dilations_0 = const()[name = tensor<string, []>("value_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_49_groups_0 = const()[name = tensor<string, []>("value_49_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(965704512)))];
tensor<fp16, [1280]> layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(968981376)))];
tensor<fp16, [1, 1280, 1, 1500]> value_49_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_bias_to_fp16, dilations = value_49_dilations_0, groups = value_49_groups_0, pad = value_49_pad_0, pad_type = value_49_pad_type_0, strides = value_49_strides_0, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor<string, []>("value_49_cast_fp16")];
tensor<int32, [4]> var_3064 = const()[name = tensor<string, []>("op_3064"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_49_cast_fp16 = reshape(shape = var_3064, x = query_49_cast_fp16)[name = tensor<string, []>("mh_q_49_cast_fp16")];
tensor<fp16, []> var_3066_to_fp16 = const()[name = tensor<string, []>("op_3066_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_3067_cast_fp16 = mul(x = mh_q_49_cast_fp16, y = var_3066_to_fp16)[name = tensor<string, []>("op_3067_cast_fp16")];
tensor<int32, [4]> var_3068 = const()[name = tensor<string, []>("op_3068"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3069_cast_fp16 = reshape(shape = var_3068, x = key_49_cast_fp16)[name = tensor<string, []>("op_3069_cast_fp16")];
tensor<bool, []> mh_w_49_transpose_x_0 = const()[name = tensor<string, []>("mh_w_49_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_49_transpose_y_0 = const()[name = tensor<string, []>("mh_w_49_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3067_cast_fp16, y = var_3069_cast_fp16)[name = tensor<string, []>("mh_w_49_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_3072_cast_fp16 = softmax(axis = var_3010, x = mh_w_49_cast_fp16)[name = tensor<string, []>("op_3072_cast_fp16")];
tensor<int32, [4]> var_3073 = const()[name = tensor<string, []>("op_3073"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3074_cast_fp16 = reshape(shape = var_3073, x = value_49_cast_fp16)[name = tensor<string, []>("op_3074_cast_fp16")];
tensor<bool, []> attn_49_transpose_x_0 = const()[name = tensor<string, []>("attn_49_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_49_transpose_y_0 = const()[name = tensor<string, []>("attn_49_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_3074_cast_fp16, y = var_3072_cast_fp16)[name = tensor<string, []>("attn_49_cast_fp16")];
tensor<int32, [4]> var_3077 = const()[name = tensor<string, []>("op_3077"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_193_cast_fp16 = reshape(shape = var_3077, x = attn_49_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
tensor<string, []> obj_99_pad_type_0 = const()[name = tensor<string, []>("obj_99_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_99_strides_0 = const()[name = tensor<string, []>("obj_99_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_99_pad_0 = const()[name = tensor<string, []>("obj_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_99_dilations_0 = const()[name = tensor<string, []>("obj_99_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_99_groups_0 = const()[name = tensor<string, []>("obj_99_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(968984000)))];
tensor<fp16, [1280]> layers_24_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(972260864)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_99_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_bias_to_fp16, dilations = obj_99_dilations_0, groups = obj_99_groups_0, pad = obj_99_pad_0, pad_type = obj_99_pad_type_0, strides = obj_99_strides_0, weight = layers_24_self_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("obj_99_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_99_cast_fp16)[name = tensor<string, []>("inputs_99_cast_fp16")];
tensor<int32, [1]> out_99_axes_0 = const()[name = tensor<string, []>("out_99_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3095_to_fp16 = const()[name = tensor<string, []>("op_3095_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_3095_to_fp16, x = inputs_99_cast_fp16)[name = tensor<string, []>("out_99_cast_fp16")];
tensor<fp16, [1280]> input_195_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_195_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(972263488)))];
tensor<fp16, [1280]> input_195_beta_0_to_fp16 = const()[name = tensor<string, []>("input_195_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(972266112)))];
tensor<fp16, []> input_195_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_195_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
tensor<string, []> input_197_pad_type_0 = const()[name = tensor<string, []>("input_197_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_197_strides_0 = const()[name = tensor<string, []>("input_197_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_197_pad_0 = const()[name = tensor<string, []>("input_197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_197_dilations_0 = const()[name = tensor<string, []>("input_197_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_197_groups_0 = const()[name = tensor<string, []>("input_197_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_24_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(972268736)))];
tensor<fp16, [5120]> layers_24_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(985376000)))];
tensor<fp16, [1, 5120, 1, 1500]> input_197_cast_fp16 = conv(bias = layers_24_fc1_bias_to_fp16, dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = layers_24_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
tensor<string, []> input_199_mode_0 = const()[name = tensor<string, []>("input_199_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
tensor<string, []> hidden_states_53_pad_type_0 = const()[name = tensor<string, []>("hidden_states_53_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = tensor<string, []>("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = tensor<string, []>("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = tensor<string, []>("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_53_groups_0 = const()[name = tensor<string, []>("hidden_states_53_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_24_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(985386304)))];
tensor<fp16, [1280]> layers_24_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(998493568)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_53_cast_fp16 = conv(bias = layers_24_fc2_bias_to_fp16, dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_24_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("hidden_states_53_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor<string, []>("inputs_101_cast_fp16")];
tensor<int32, []> var_3128 = const()[name = tensor<string, []>("op_3128"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_101_axes_0 = const()[name = tensor<string, []>("out_101_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3147_to_fp16 = const()[name = tensor<string, []>("op_3147_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_3147_to_fp16, x = inputs_101_cast_fp16)[name = tensor<string, []>("out_101_cast_fp16")];
tensor<fp16, [1280]> obj_101_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_101_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(998496192)))];
tensor<fp16, [1280]> obj_101_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_101_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(998498816)))];
tensor<fp16, []> obj_101_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_101_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_101_cast_fp16 = batch_norm(beta = obj_101_beta_0_to_fp16, epsilon = obj_101_epsilon_0_to_fp16, gamma = obj_101_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor<string, []>("obj_101_cast_fp16")];
tensor<string, []> query_51_pad_type_0 = const()[name = tensor<string, []>("query_51_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_51_strides_0 = const()[name = tensor<string, []>("query_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_51_pad_0 = const()[name = tensor<string, []>("query_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_51_dilations_0 = const()[name = tensor<string, []>("query_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_51_groups_0 = const()[name = tensor<string, []>("query_51_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(998501440)))];
tensor<fp16, [1280]> layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1001778304)))];
tensor<fp16, [1, 1280, 1, 1500]> query_51_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_bias_to_fp16, dilations = query_51_dilations_0, groups = query_51_groups_0, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = query_51_strides_0, weight = layers_25_self_attn_q_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor<string, []>("query_51_cast_fp16")];
tensor<string, []> key_51_pad_type_0 = const()[name = tensor<string, []>("key_51_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_51_strides_0 = const()[name = tensor<string, []>("key_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_51_pad_0 = const()[name = tensor<string, []>("key_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_51_dilations_0 = const()[name = tensor<string, []>("key_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_51_groups_0 = const()[name = tensor<string, []>("key_51_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1001780928)))];
tensor<fp16, [1, 1280, 1, 1500]> key_51_cast_fp16 = conv(dilations = key_51_dilations_0, groups = key_51_groups_0, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = key_51_strides_0, weight = layers_25_self_attn_k_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor<string, []>("key_51_cast_fp16")];
tensor<string, []> value_51_pad_type_0 = const()[name = tensor<string, []>("value_51_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_51_strides_0 = const()[name = tensor<string, []>("value_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_51_pad_0 = const()[name = tensor<string, []>("value_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_51_dilations_0 = const()[name = tensor<string, []>("value_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_51_groups_0 = const()[name = tensor<string, []>("value_51_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1005057792)))];
tensor<fp16, [1280]> layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1008334656)))];
tensor<fp16, [1, 1280, 1, 1500]> value_51_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_bias_to_fp16, dilations = value_51_dilations_0, groups = value_51_groups_0, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = value_51_strides_0, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor<string, []>("value_51_cast_fp16")];
tensor<int32, [4]> var_3182 = const()[name = tensor<string, []>("op_3182"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_51_cast_fp16 = reshape(shape = var_3182, x = query_51_cast_fp16)[name = tensor<string, []>("mh_q_51_cast_fp16")];
tensor<fp16, []> var_3184_to_fp16 = const()[name = tensor<string, []>("op_3184_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_3185_cast_fp16 = mul(x = mh_q_51_cast_fp16, y = var_3184_to_fp16)[name = tensor<string, []>("op_3185_cast_fp16")];
tensor<int32, [4]> var_3186 = const()[name = tensor<string, []>("op_3186"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3187_cast_fp16 = reshape(shape = var_3186, x = key_51_cast_fp16)[name = tensor<string, []>("op_3187_cast_fp16")];
tensor<bool, []> mh_w_51_transpose_x_0 = const()[name = tensor<string, []>("mh_w_51_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_51_transpose_y_0 = const()[name = tensor<string, []>("mh_w_51_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_51_cast_fp16 = matmul(transpose_x = mh_w_51_transpose_x_0, transpose_y = mh_w_51_transpose_y_0, x = var_3185_cast_fp16, y = var_3187_cast_fp16)[name = tensor<string, []>("mh_w_51_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_3190_cast_fp16 = softmax(axis = var_3128, x = mh_w_51_cast_fp16)[name = tensor<string, []>("op_3190_cast_fp16")];
tensor<int32, [4]> var_3191 = const()[name = tensor<string, []>("op_3191"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3192_cast_fp16 = reshape(shape = var_3191, x = value_51_cast_fp16)[name = tensor<string, []>("op_3192_cast_fp16")];
tensor<bool, []> attn_51_transpose_x_0 = const()[name = tensor<string, []>("attn_51_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_51_transpose_y_0 = const()[name = tensor<string, []>("attn_51_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_3192_cast_fp16, y = var_3190_cast_fp16)[name = tensor<string, []>("attn_51_cast_fp16")];
tensor<int32, [4]> var_3195 = const()[name = tensor<string, []>("op_3195"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_201_cast_fp16 = reshape(shape = var_3195, x = attn_51_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")];
tensor<string, []> obj_103_pad_type_0 = const()[name = tensor<string, []>("obj_103_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_103_strides_0 = const()[name = tensor<string, []>("obj_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_103_pad_0 = const()[name = tensor<string, []>("obj_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_103_dilations_0 = const()[name = tensor<string, []>("obj_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_103_groups_0 = const()[name = tensor<string, []>("obj_103_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1008337280)))];
tensor<fp16, [1280]> layers_25_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1011614144)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_103_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_bias_to_fp16, dilations = obj_103_dilations_0, groups = obj_103_groups_0, pad = obj_103_pad_0, pad_type = obj_103_pad_type_0, strides = obj_103_strides_0, weight = layers_25_self_attn_o_proj_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("obj_103_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = obj_103_cast_fp16)[name = tensor<string, []>("inputs_103_cast_fp16")];
tensor<int32, [1]> out_103_axes_0 = const()[name = tensor<string, []>("out_103_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3213_to_fp16 = const()[name = tensor<string, []>("op_3213_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3213_to_fp16, x = inputs_103_cast_fp16)[name = tensor<string, []>("out_103_cast_fp16")];
tensor<fp16, [1280]> input_203_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_203_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1011616768)))];
tensor<fp16, [1280]> input_203_beta_0_to_fp16 = const()[name = tensor<string, []>("input_203_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1011619392)))];
tensor<fp16, []> input_203_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_203_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_203_cast_fp16 = batch_norm(beta = input_203_beta_0_to_fp16, epsilon = input_203_epsilon_0_to_fp16, gamma = input_203_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
tensor<string, []> input_205_pad_type_0 = const()[name = tensor<string, []>("input_205_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_205_strides_0 = const()[name = tensor<string, []>("input_205_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_205_pad_0 = const()[name = tensor<string, []>("input_205_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_205_dilations_0 = const()[name = tensor<string, []>("input_205_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_205_groups_0 = const()[name = tensor<string, []>("input_205_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_25_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1011622016)))];
tensor<fp16, [5120]> layers_25_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1024729280)))];
tensor<fp16, [1, 5120, 1, 1500]> input_205_cast_fp16 = conv(bias = layers_25_fc1_bias_to_fp16, dilations = input_205_dilations_0, groups = input_205_groups_0, pad = input_205_pad_0, pad_type = input_205_pad_type_0, strides = input_205_strides_0, weight = layers_25_fc1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
tensor<string, []> input_207_mode_0 = const()[name = tensor<string, []>("input_207_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_207_cast_fp16 = gelu(mode = input_207_mode_0, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
tensor<string, []> hidden_states_55_pad_type_0 = const()[name = tensor<string, []>("hidden_states_55_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_55_strides_0 = const()[name = tensor<string, []>("hidden_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_55_pad_0 = const()[name = tensor<string, []>("hidden_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_55_dilations_0 = const()[name = tensor<string, []>("hidden_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_55_groups_0 = const()[name = tensor<string, []>("hidden_states_55_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_25_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1024739584)))];
tensor<fp16, [1280]> layers_25_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1037846848)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_55_cast_fp16 = conv(bias = layers_25_fc2_bias_to_fp16, dilations = hidden_states_55_dilations_0, groups = hidden_states_55_groups_0, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = hidden_states_55_strides_0, weight = layers_25_fc2_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("hidden_states_55_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor<string, []>("inputs_105_cast_fp16")];
tensor<int32, []> var_3246 = const()[name = tensor<string, []>("op_3246"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_105_axes_0 = const()[name = tensor<string, []>("out_105_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3265_to_fp16 = const()[name = tensor<string, []>("op_3265_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3265_to_fp16, x = inputs_105_cast_fp16)[name = tensor<string, []>("out_105_cast_fp16")];
tensor<fp16, [1280]> obj_105_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_105_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1037849472)))];
tensor<fp16, [1280]> obj_105_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_105_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1037852096)))];
tensor<fp16, []> obj_105_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_105_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor<string, []>("obj_105_cast_fp16")];
tensor<string, []> query_53_pad_type_0 = const()[name = tensor<string, []>("query_53_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_53_strides_0 = const()[name = tensor<string, []>("query_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_53_pad_0 = const()[name = tensor<string, []>("query_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_53_dilations_0 = const()[name = tensor<string, []>("query_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_53_groups_0 = const()[name = tensor<string, []>("query_53_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1037854720)))];
tensor<fp16, [1280]> layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1041131584)))];
tensor<fp16, [1, 1280, 1, 1500]> query_53_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_bias_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_26_self_attn_q_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor<string, []>("query_53_cast_fp16")];
tensor<string, []> key_53_pad_type_0 = const()[name = tensor<string, []>("key_53_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_53_strides_0 = const()[name = tensor<string, []>("key_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_53_pad_0 = const()[name = tensor<string, []>("key_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_53_dilations_0 = const()[name = tensor<string, []>("key_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_53_groups_0 = const()[name = tensor<string, []>("key_53_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1041134208)))];
tensor<fp16, [1, 1280, 1, 1500]> key_53_cast_fp16 = conv(dilations = key_53_dilations_0, groups = key_53_groups_0, pad = key_53_pad_0, pad_type = key_53_pad_type_0, strides = key_53_strides_0, weight = layers_26_self_attn_k_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor<string, []>("key_53_cast_fp16")];
tensor<string, []> value_53_pad_type_0 = const()[name = tensor<string, []>("value_53_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_53_strides_0 = const()[name = tensor<string, []>("value_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_53_pad_0 = const()[name = tensor<string, []>("value_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_53_dilations_0 = const()[name = tensor<string, []>("value_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_53_groups_0 = const()[name = tensor<string, []>("value_53_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1044411072)))];
tensor<fp16, [1280]> layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1047687936)))];
tensor<fp16, [1, 1280, 1, 1500]> value_53_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_bias_to_fp16, dilations = value_53_dilations_0, groups = value_53_groups_0, pad = value_53_pad_0, pad_type = value_53_pad_type_0, strides = value_53_strides_0, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor<string, []>("value_53_cast_fp16")];
tensor<int32, [4]> var_3300 = const()[name = tensor<string, []>("op_3300"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_53_cast_fp16 = reshape(shape = var_3300, x = query_53_cast_fp16)[name = tensor<string, []>("mh_q_53_cast_fp16")];
tensor<fp16, []> var_3302_to_fp16 = const()[name = tensor<string, []>("op_3302_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_3303_cast_fp16 = mul(x = mh_q_53_cast_fp16, y = var_3302_to_fp16)[name = tensor<string, []>("op_3303_cast_fp16")];
tensor<int32, [4]> var_3304 = const()[name = tensor<string, []>("op_3304"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3305_cast_fp16 = reshape(shape = var_3304, x = key_53_cast_fp16)[name = tensor<string, []>("op_3305_cast_fp16")];
tensor<bool, []> mh_w_53_transpose_x_0 = const()[name = tensor<string, []>("mh_w_53_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_53_transpose_y_0 = const()[name = tensor<string, []>("mh_w_53_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_3303_cast_fp16, y = var_3305_cast_fp16)[name = tensor<string, []>("mh_w_53_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_3308_cast_fp16 = softmax(axis = var_3246, x = mh_w_53_cast_fp16)[name = tensor<string, []>("op_3308_cast_fp16")];
tensor<int32, [4]> var_3309 = const()[name = tensor<string, []>("op_3309"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3310_cast_fp16 = reshape(shape = var_3309, x = value_53_cast_fp16)[name = tensor<string, []>("op_3310_cast_fp16")];
tensor<bool, []> attn_53_transpose_x_0 = const()[name = tensor<string, []>("attn_53_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_53_transpose_y_0 = const()[name = tensor<string, []>("attn_53_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3310_cast_fp16, y = var_3308_cast_fp16)[name = tensor<string, []>("attn_53_cast_fp16")];
tensor<int32, [4]> var_3313 = const()[name = tensor<string, []>("op_3313"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_209_cast_fp16 = reshape(shape = var_3313, x = attn_53_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
tensor<string, []> obj_107_pad_type_0 = const()[name = tensor<string, []>("obj_107_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_107_strides_0 = const()[name = tensor<string, []>("obj_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_107_pad_0 = const()[name = tensor<string, []>("obj_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_107_dilations_0 = const()[name = tensor<string, []>("obj_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_107_groups_0 = const()[name = tensor<string, []>("obj_107_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1047690560)))];
tensor<fp16, [1280]> layers_26_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050967424)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_107_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_bias_to_fp16, dilations = obj_107_dilations_0, groups = obj_107_groups_0, pad = obj_107_pad_0, pad_type = obj_107_pad_type_0, strides = obj_107_strides_0, weight = layers_26_self_attn_o_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("obj_107_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_107_cast_fp16)[name = tensor<string, []>("inputs_107_cast_fp16")];
tensor<int32, [1]> out_107_axes_0 = const()[name = tensor<string, []>("out_107_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3331_to_fp16 = const()[name = tensor<string, []>("op_3331_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_3331_to_fp16, x = inputs_107_cast_fp16)[name = tensor<string, []>("out_107_cast_fp16")];
tensor<fp16, [1280]> input_211_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_211_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050970048)))];
tensor<fp16, [1280]> input_211_beta_0_to_fp16 = const()[name = tensor<string, []>("input_211_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050972672)))];
tensor<fp16, []> input_211_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_211_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_211_cast_fp16 = batch_norm(beta = input_211_beta_0_to_fp16, epsilon = input_211_epsilon_0_to_fp16, gamma = input_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
tensor<string, []> input_213_pad_type_0 = const()[name = tensor<string, []>("input_213_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_213_strides_0 = const()[name = tensor<string, []>("input_213_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_213_pad_0 = const()[name = tensor<string, []>("input_213_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_213_dilations_0 = const()[name = tensor<string, []>("input_213_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_213_groups_0 = const()[name = tensor<string, []>("input_213_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_26_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050975296)))];
tensor<fp16, [5120]> layers_26_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1064082560)))];
tensor<fp16, [1, 5120, 1, 1500]> input_213_cast_fp16 = conv(bias = layers_26_fc1_bias_to_fp16, dilations = input_213_dilations_0, groups = input_213_groups_0, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = input_213_strides_0, weight = layers_26_fc1_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
tensor<string, []> input_215_mode_0 = const()[name = tensor<string, []>("input_215_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
tensor<string, []> hidden_states_57_pad_type_0 = const()[name = tensor<string, []>("hidden_states_57_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = tensor<string, []>("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = tensor<string, []>("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = tensor<string, []>("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_57_groups_0 = const()[name = tensor<string, []>("hidden_states_57_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_26_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1064092864)))];
tensor<fp16, [1280]> layers_26_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1077200128)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_57_cast_fp16 = conv(bias = layers_26_fc2_bias_to_fp16, dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_26_fc2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("hidden_states_57_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor<string, []>("inputs_109_cast_fp16")];
tensor<int32, []> var_3364 = const()[name = tensor<string, []>("op_3364"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_109_axes_0 = const()[name = tensor<string, []>("out_109_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3383_to_fp16 = const()[name = tensor<string, []>("op_3383_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_3383_to_fp16, x = inputs_109_cast_fp16)[name = tensor<string, []>("out_109_cast_fp16")];
tensor<fp16, [1280]> obj_109_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_109_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1077202752)))];
tensor<fp16, [1280]> obj_109_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_109_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1077205376)))];
tensor<fp16, []> obj_109_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_109_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor<string, []>("obj_109_cast_fp16")];
tensor<string, []> query_55_pad_type_0 = const()[name = tensor<string, []>("query_55_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_55_strides_0 = const()[name = tensor<string, []>("query_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_55_pad_0 = const()[name = tensor<string, []>("query_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_55_dilations_0 = const()[name = tensor<string, []>("query_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_55_groups_0 = const()[name = tensor<string, []>("query_55_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1077208000)))];
tensor<fp16, [1280]> layers_27_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1080484864)))];
tensor<fp16, [1, 1280, 1, 1500]> query_55_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_bias_to_fp16, dilations = query_55_dilations_0, groups = query_55_groups_0, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = query_55_strides_0, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor<string, []>("query_55_cast_fp16")];
tensor<string, []> key_55_pad_type_0 = const()[name = tensor<string, []>("key_55_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_55_strides_0 = const()[name = tensor<string, []>("key_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_55_pad_0 = const()[name = tensor<string, []>("key_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_55_dilations_0 = const()[name = tensor<string, []>("key_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_55_groups_0 = const()[name = tensor<string, []>("key_55_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1080487488)))];
tensor<fp16, [1, 1280, 1, 1500]> key_55_cast_fp16 = conv(dilations = key_55_dilations_0, groups = key_55_groups_0, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = key_55_strides_0, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor<string, []>("key_55_cast_fp16")];
tensor<string, []> value_55_pad_type_0 = const()[name = tensor<string, []>("value_55_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_55_strides_0 = const()[name = tensor<string, []>("value_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_55_pad_0 = const()[name = tensor<string, []>("value_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_55_dilations_0 = const()[name = tensor<string, []>("value_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_55_groups_0 = const()[name = tensor<string, []>("value_55_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1083764352)))];
tensor<fp16, [1280]> layers_27_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1087041216)))];
tensor<fp16, [1, 1280, 1, 1500]> value_55_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_bias_to_fp16, dilations = value_55_dilations_0, groups = value_55_groups_0, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = value_55_strides_0, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor<string, []>("value_55_cast_fp16")];
tensor<int32, [4]> var_3418 = const()[name = tensor<string, []>("op_3418"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_55_cast_fp16 = reshape(shape = var_3418, x = query_55_cast_fp16)[name = tensor<string, []>("mh_q_55_cast_fp16")];
tensor<fp16, []> var_3420_to_fp16 = const()[name = tensor<string, []>("op_3420_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_3421_cast_fp16 = mul(x = mh_q_55_cast_fp16, y = var_3420_to_fp16)[name = tensor<string, []>("op_3421_cast_fp16")];
tensor<int32, [4]> var_3422 = const()[name = tensor<string, []>("op_3422"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3423_cast_fp16 = reshape(shape = var_3422, x = key_55_cast_fp16)[name = tensor<string, []>("op_3423_cast_fp16")];
tensor<bool, []> mh_w_55_transpose_x_0 = const()[name = tensor<string, []>("mh_w_55_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_55_transpose_y_0 = const()[name = tensor<string, []>("mh_w_55_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_3421_cast_fp16, y = var_3423_cast_fp16)[name = tensor<string, []>("mh_w_55_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_3426_cast_fp16 = softmax(axis = var_3364, x = mh_w_55_cast_fp16)[name = tensor<string, []>("op_3426_cast_fp16")];
tensor<int32, [4]> var_3427 = const()[name = tensor<string, []>("op_3427"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3428_cast_fp16 = reshape(shape = var_3427, x = value_55_cast_fp16)[name = tensor<string, []>("op_3428_cast_fp16")];
tensor<bool, []> attn_55_transpose_x_0 = const()[name = tensor<string, []>("attn_55_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_55_transpose_y_0 = const()[name = tensor<string, []>("attn_55_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3428_cast_fp16, y = var_3426_cast_fp16)[name = tensor<string, []>("attn_55_cast_fp16")];
tensor<int32, [4]> var_3431 = const()[name = tensor<string, []>("op_3431"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_217_cast_fp16 = reshape(shape = var_3431, x = attn_55_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
tensor<string, []> obj_111_pad_type_0 = const()[name = tensor<string, []>("obj_111_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_111_strides_0 = const()[name = tensor<string, []>("obj_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_111_pad_0 = const()[name = tensor<string, []>("obj_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_111_dilations_0 = const()[name = tensor<string, []>("obj_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_111_groups_0 = const()[name = tensor<string, []>("obj_111_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1087043840)))];
tensor<fp16, [1280]> layers_27_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090320704)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_111_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_bias_to_fp16, dilations = obj_111_dilations_0, groups = obj_111_groups_0, pad = obj_111_pad_0, pad_type = obj_111_pad_type_0, strides = obj_111_strides_0, weight = layers_27_self_attn_o_proj_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("obj_111_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_111_cast_fp16)[name = tensor<string, []>("inputs_111_cast_fp16")];
tensor<int32, [1]> out_111_axes_0 = const()[name = tensor<string, []>("out_111_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3449_to_fp16 = const()[name = tensor<string, []>("op_3449_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_3449_to_fp16, x = inputs_111_cast_fp16)[name = tensor<string, []>("out_111_cast_fp16")];
tensor<fp16, [1280]> input_219_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_219_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090323328)))];
tensor<fp16, [1280]> input_219_beta_0_to_fp16 = const()[name = tensor<string, []>("input_219_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090325952)))];
tensor<fp16, []> input_219_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_219_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_219_cast_fp16 = batch_norm(beta = input_219_beta_0_to_fp16, epsilon = input_219_epsilon_0_to_fp16, gamma = input_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
tensor<string, []> input_221_pad_type_0 = const()[name = tensor<string, []>("input_221_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_221_strides_0 = const()[name = tensor<string, []>("input_221_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_221_pad_0 = const()[name = tensor<string, []>("input_221_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_221_dilations_0 = const()[name = tensor<string, []>("input_221_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_221_groups_0 = const()[name = tensor<string, []>("input_221_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_27_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090328576)))];
tensor<fp16, [5120]> layers_27_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1103435840)))];
tensor<fp16, [1, 5120, 1, 1500]> input_221_cast_fp16 = conv(bias = layers_27_fc1_bias_to_fp16, dilations = input_221_dilations_0, groups = input_221_groups_0, pad = input_221_pad_0, pad_type = input_221_pad_type_0, strides = input_221_strides_0, weight = layers_27_fc1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
tensor<string, []> input_223_mode_0 = const()[name = tensor<string, []>("input_223_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_223_cast_fp16 = gelu(mode = input_223_mode_0, x = input_221_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
tensor<string, []> hidden_states_59_pad_type_0 = const()[name = tensor<string, []>("hidden_states_59_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_59_strides_0 = const()[name = tensor<string, []>("hidden_states_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_59_pad_0 = const()[name = tensor<string, []>("hidden_states_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_59_dilations_0 = const()[name = tensor<string, []>("hidden_states_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_59_groups_0 = const()[name = tensor<string, []>("hidden_states_59_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_27_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1103446144)))];
tensor<fp16, [1280]> layers_27_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1116553408)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_59_cast_fp16 = conv(bias = layers_27_fc2_bias_to_fp16, dilations = hidden_states_59_dilations_0, groups = hidden_states_59_groups_0, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = hidden_states_59_strides_0, weight = layers_27_fc2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor<string, []>("hidden_states_59_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor<string, []>("inputs_113_cast_fp16")];
tensor<int32, []> var_3482 = const()[name = tensor<string, []>("op_3482"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_113_axes_0 = const()[name = tensor<string, []>("out_113_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3501_to_fp16 = const()[name = tensor<string, []>("op_3501_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_3501_to_fp16, x = inputs_113_cast_fp16)[name = tensor<string, []>("out_113_cast_fp16")];
tensor<fp16, [1280]> obj_113_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_113_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1116556032)))];
tensor<fp16, [1280]> obj_113_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_113_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1116558656)))];
tensor<fp16, []> obj_113_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_113_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor<string, []>("obj_113_cast_fp16")];
tensor<string, []> query_57_pad_type_0 = const()[name = tensor<string, []>("query_57_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_57_strides_0 = const()[name = tensor<string, []>("query_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_57_pad_0 = const()[name = tensor<string, []>("query_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_57_dilations_0 = const()[name = tensor<string, []>("query_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_57_groups_0 = const()[name = tensor<string, []>("query_57_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1116561280)))];
tensor<fp16, [1280]> layers_28_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1119838144)))];
tensor<fp16, [1, 1280, 1, 1500]> query_57_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_bias_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_28_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("query_57_cast_fp16")];
tensor<string, []> key_57_pad_type_0 = const()[name = tensor<string, []>("key_57_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_57_strides_0 = const()[name = tensor<string, []>("key_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_57_pad_0 = const()[name = tensor<string, []>("key_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_57_dilations_0 = const()[name = tensor<string, []>("key_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_57_groups_0 = const()[name = tensor<string, []>("key_57_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1119840768)))];
tensor<fp16, [1, 1280, 1, 1500]> key_57_cast_fp16 = conv(dilations = key_57_dilations_0, groups = key_57_groups_0, pad = key_57_pad_0, pad_type = key_57_pad_type_0, strides = key_57_strides_0, weight = layers_28_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("key_57_cast_fp16")];
tensor<string, []> value_57_pad_type_0 = const()[name = tensor<string, []>("value_57_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_57_strides_0 = const()[name = tensor<string, []>("value_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_57_pad_0 = const()[name = tensor<string, []>("value_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_57_dilations_0 = const()[name = tensor<string, []>("value_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_57_groups_0 = const()[name = tensor<string, []>("value_57_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1123117632)))];
tensor<fp16, [1280]> layers_28_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1126394496)))];
tensor<fp16, [1, 1280, 1, 1500]> value_57_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_bias_to_fp16, dilations = value_57_dilations_0, groups = value_57_groups_0, pad = value_57_pad_0, pad_type = value_57_pad_type_0, strides = value_57_strides_0, weight = layers_28_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("value_57_cast_fp16")];
tensor<int32, [4]> var_3536 = const()[name = tensor<string, []>("op_3536"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_57_cast_fp16 = reshape(shape = var_3536, x = query_57_cast_fp16)[name = tensor<string, []>("mh_q_57_cast_fp16")];
tensor<fp16, []> var_3538_to_fp16 = const()[name = tensor<string, []>("op_3538_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_3539_cast_fp16 = mul(x = mh_q_57_cast_fp16, y = var_3538_to_fp16)[name = tensor<string, []>("op_3539_cast_fp16")];
tensor<int32, [4]> var_3540 = const()[name = tensor<string, []>("op_3540"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3541_cast_fp16 = reshape(shape = var_3540, x = key_57_cast_fp16)[name = tensor<string, []>("op_3541_cast_fp16")];
tensor<bool, []> mh_w_57_transpose_x_0 = const()[name = tensor<string, []>("mh_w_57_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_57_transpose_y_0 = const()[name = tensor<string, []>("mh_w_57_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_3539_cast_fp16, y = var_3541_cast_fp16)[name = tensor<string, []>("mh_w_57_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_3544_cast_fp16 = softmax(axis = var_3482, x = mh_w_57_cast_fp16)[name = tensor<string, []>("op_3544_cast_fp16")];
tensor<int32, [4]> var_3545 = const()[name = tensor<string, []>("op_3545"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3546_cast_fp16 = reshape(shape = var_3545, x = value_57_cast_fp16)[name = tensor<string, []>("op_3546_cast_fp16")];
tensor<bool, []> attn_57_transpose_x_0 = const()[name = tensor<string, []>("attn_57_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_57_transpose_y_0 = const()[name = tensor<string, []>("attn_57_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3546_cast_fp16, y = var_3544_cast_fp16)[name = tensor<string, []>("attn_57_cast_fp16")];
tensor<int32, [4]> var_3549 = const()[name = tensor<string, []>("op_3549"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_225_cast_fp16 = reshape(shape = var_3549, x = attn_57_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
tensor<string, []> obj_115_pad_type_0 = const()[name = tensor<string, []>("obj_115_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_115_strides_0 = const()[name = tensor<string, []>("obj_115_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_115_pad_0 = const()[name = tensor<string, []>("obj_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_115_dilations_0 = const()[name = tensor<string, []>("obj_115_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_115_groups_0 = const()[name = tensor<string, []>("obj_115_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1126397120)))];
tensor<fp16, [1280]> layers_28_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1129673984)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_115_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_bias_to_fp16, dilations = obj_115_dilations_0, groups = obj_115_groups_0, pad = obj_115_pad_0, pad_type = obj_115_pad_type_0, strides = obj_115_strides_0, weight = layers_28_self_attn_o_proj_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("obj_115_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = obj_115_cast_fp16)[name = tensor<string, []>("inputs_115_cast_fp16")];
tensor<int32, [1]> out_115_axes_0 = const()[name = tensor<string, []>("out_115_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3567_to_fp16 = const()[name = tensor<string, []>("op_3567_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_3567_to_fp16, x = inputs_115_cast_fp16)[name = tensor<string, []>("out_115_cast_fp16")];
tensor<fp16, [1280]> input_227_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_227_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1129676608)))];
tensor<fp16, [1280]> input_227_beta_0_to_fp16 = const()[name = tensor<string, []>("input_227_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1129679232)))];
tensor<fp16, []> input_227_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_227_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_227_cast_fp16 = batch_norm(beta = input_227_beta_0_to_fp16, epsilon = input_227_epsilon_0_to_fp16, gamma = input_227_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
tensor<string, []> input_229_pad_type_0 = const()[name = tensor<string, []>("input_229_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_229_strides_0 = const()[name = tensor<string, []>("input_229_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_229_pad_0 = const()[name = tensor<string, []>("input_229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_229_dilations_0 = const()[name = tensor<string, []>("input_229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_229_groups_0 = const()[name = tensor<string, []>("input_229_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_28_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_28_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1129681856)))];
tensor<fp16, [5120]> layers_28_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1142789120)))];
tensor<fp16, [1, 5120, 1, 1500]> input_229_cast_fp16 = conv(bias = layers_28_fc1_bias_to_fp16, dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = layers_28_fc1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
tensor<string, []> input_231_mode_0 = const()[name = tensor<string, []>("input_231_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
tensor<string, []> hidden_states_61_pad_type_0 = const()[name = tensor<string, []>("hidden_states_61_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_61_strides_0 = const()[name = tensor<string, []>("hidden_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_61_pad_0 = const()[name = tensor<string, []>("hidden_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_61_dilations_0 = const()[name = tensor<string, []>("hidden_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_61_groups_0 = const()[name = tensor<string, []>("hidden_states_61_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_28_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_28_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1142799424)))];
tensor<fp16, [1280]> layers_28_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1155906688)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_61_cast_fp16 = conv(bias = layers_28_fc2_bias_to_fp16, dilations = hidden_states_61_dilations_0, groups = hidden_states_61_groups_0, pad = hidden_states_61_pad_0, pad_type = hidden_states_61_pad_type_0, strides = hidden_states_61_strides_0, weight = layers_28_fc2_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("hidden_states_61_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor<string, []>("inputs_117_cast_fp16")];
tensor<int32, []> var_3600 = const()[name = tensor<string, []>("op_3600"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_117_axes_0 = const()[name = tensor<string, []>("out_117_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3619_to_fp16 = const()[name = tensor<string, []>("op_3619_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_3619_to_fp16, x = inputs_117_cast_fp16)[name = tensor<string, []>("out_117_cast_fp16")];
tensor<fp16, [1280]> obj_117_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_117_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1155909312)))];
tensor<fp16, [1280]> obj_117_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_117_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1155911936)))];
tensor<fp16, []> obj_117_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_117_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor<string, []>("obj_117_cast_fp16")];
tensor<string, []> query_59_pad_type_0 = const()[name = tensor<string, []>("query_59_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_59_strides_0 = const()[name = tensor<string, []>("query_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_59_pad_0 = const()[name = tensor<string, []>("query_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_59_dilations_0 = const()[name = tensor<string, []>("query_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_59_groups_0 = const()[name = tensor<string, []>("query_59_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1155914560)))];
tensor<fp16, [1280]> layers_29_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1159191424)))];
tensor<fp16, [1, 1280, 1, 1500]> query_59_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_bias_to_fp16, dilations = query_59_dilations_0, groups = query_59_groups_0, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = query_59_strides_0, weight = layers_29_self_attn_q_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor<string, []>("query_59_cast_fp16")];
tensor<string, []> key_59_pad_type_0 = const()[name = tensor<string, []>("key_59_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_59_strides_0 = const()[name = tensor<string, []>("key_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_59_pad_0 = const()[name = tensor<string, []>("key_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_59_dilations_0 = const()[name = tensor<string, []>("key_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_59_groups_0 = const()[name = tensor<string, []>("key_59_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1159194048)))];
tensor<fp16, [1, 1280, 1, 1500]> key_59_cast_fp16 = conv(dilations = key_59_dilations_0, groups = key_59_groups_0, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = key_59_strides_0, weight = layers_29_self_attn_k_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor<string, []>("key_59_cast_fp16")];
tensor<string, []> value_59_pad_type_0 = const()[name = tensor<string, []>("value_59_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_59_strides_0 = const()[name = tensor<string, []>("value_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_59_pad_0 = const()[name = tensor<string, []>("value_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_59_dilations_0 = const()[name = tensor<string, []>("value_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_59_groups_0 = const()[name = tensor<string, []>("value_59_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1162470912)))];
tensor<fp16, [1280]> layers_29_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1165747776)))];
tensor<fp16, [1, 1280, 1, 1500]> value_59_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_bias_to_fp16, dilations = value_59_dilations_0, groups = value_59_groups_0, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = value_59_strides_0, weight = layers_29_self_attn_v_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor<string, []>("value_59_cast_fp16")];
tensor<int32, [4]> var_3654 = const()[name = tensor<string, []>("op_3654"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_59_cast_fp16 = reshape(shape = var_3654, x = query_59_cast_fp16)[name = tensor<string, []>("mh_q_59_cast_fp16")];
tensor<fp16, []> var_3656_to_fp16 = const()[name = tensor<string, []>("op_3656_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_3657_cast_fp16 = mul(x = mh_q_59_cast_fp16, y = var_3656_to_fp16)[name = tensor<string, []>("op_3657_cast_fp16")];
tensor<int32, [4]> var_3658 = const()[name = tensor<string, []>("op_3658"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3659_cast_fp16 = reshape(shape = var_3658, x = key_59_cast_fp16)[name = tensor<string, []>("op_3659_cast_fp16")];
tensor<bool, []> mh_w_59_transpose_x_0 = const()[name = tensor<string, []>("mh_w_59_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_59_transpose_y_0 = const()[name = tensor<string, []>("mh_w_59_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_3657_cast_fp16, y = var_3659_cast_fp16)[name = tensor<string, []>("mh_w_59_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_3662_cast_fp16 = softmax(axis = var_3600, x = mh_w_59_cast_fp16)[name = tensor<string, []>("op_3662_cast_fp16")];
tensor<int32, [4]> var_3663 = const()[name = tensor<string, []>("op_3663"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3664_cast_fp16 = reshape(shape = var_3663, x = value_59_cast_fp16)[name = tensor<string, []>("op_3664_cast_fp16")];
tensor<bool, []> attn_59_transpose_x_0 = const()[name = tensor<string, []>("attn_59_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_59_transpose_y_0 = const()[name = tensor<string, []>("attn_59_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3664_cast_fp16, y = var_3662_cast_fp16)[name = tensor<string, []>("attn_59_cast_fp16")];
tensor<int32, [4]> var_3667 = const()[name = tensor<string, []>("op_3667"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_233_cast_fp16 = reshape(shape = var_3667, x = attn_59_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
tensor<string, []> obj_119_pad_type_0 = const()[name = tensor<string, []>("obj_119_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_119_strides_0 = const()[name = tensor<string, []>("obj_119_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_119_pad_0 = const()[name = tensor<string, []>("obj_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_119_dilations_0 = const()[name = tensor<string, []>("obj_119_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_119_groups_0 = const()[name = tensor<string, []>("obj_119_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1165750400)))];
tensor<fp16, [1280]> layers_29_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1169027264)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_119_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_29_self_attn_o_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("obj_119_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_119_cast_fp16)[name = tensor<string, []>("inputs_119_cast_fp16")];
tensor<int32, [1]> out_119_axes_0 = const()[name = tensor<string, []>("out_119_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3685_to_fp16 = const()[name = tensor<string, []>("op_3685_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_3685_to_fp16, x = inputs_119_cast_fp16)[name = tensor<string, []>("out_119_cast_fp16")];
tensor<fp16, [1280]> input_235_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_235_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1169029888)))];
tensor<fp16, [1280]> input_235_beta_0_to_fp16 = const()[name = tensor<string, []>("input_235_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1169032512)))];
tensor<fp16, []> input_235_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_235_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
tensor<string, []> input_237_pad_type_0 = const()[name = tensor<string, []>("input_237_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_237_strides_0 = const()[name = tensor<string, []>("input_237_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_237_pad_0 = const()[name = tensor<string, []>("input_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_237_dilations_0 = const()[name = tensor<string, []>("input_237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_237_groups_0 = const()[name = tensor<string, []>("input_237_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_29_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_29_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1169035136)))];
tensor<fp16, [5120]> layers_29_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1182142400)))];
tensor<fp16, [1, 5120, 1, 1500]> input_237_cast_fp16 = conv(bias = layers_29_fc1_bias_to_fp16, dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = layers_29_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")];
tensor<string, []> input_239_mode_0 = const()[name = tensor<string, []>("input_239_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = input_237_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
tensor<string, []> hidden_states_63_pad_type_0 = const()[name = tensor<string, []>("hidden_states_63_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = tensor<string, []>("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = tensor<string, []>("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = tensor<string, []>("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_63_groups_0 = const()[name = tensor<string, []>("hidden_states_63_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_29_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_29_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1182152704)))];
tensor<fp16, [1280]> layers_29_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195259968)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_63_cast_fp16 = conv(bias = layers_29_fc2_bias_to_fp16, dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_29_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("hidden_states_63_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor<string, []>("inputs_121_cast_fp16")];
tensor<int32, []> var_3718 = const()[name = tensor<string, []>("op_3718"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_121_axes_0 = const()[name = tensor<string, []>("out_121_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3737_to_fp16 = const()[name = tensor<string, []>("op_3737_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_3737_to_fp16, x = inputs_121_cast_fp16)[name = tensor<string, []>("out_121_cast_fp16")];
tensor<fp16, [1280]> obj_121_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_121_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195262592)))];
tensor<fp16, [1280]> obj_121_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_121_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195265216)))];
tensor<fp16, []> obj_121_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_121_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor<string, []>("obj_121_cast_fp16")];
tensor<string, []> query_61_pad_type_0 = const()[name = tensor<string, []>("query_61_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_61_strides_0 = const()[name = tensor<string, []>("query_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_61_pad_0 = const()[name = tensor<string, []>("query_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_61_dilations_0 = const()[name = tensor<string, []>("query_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_61_groups_0 = const()[name = tensor<string, []>("query_61_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195267840)))];
tensor<fp16, [1280]> layers_30_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1198544704)))];
tensor<fp16, [1, 1280, 1, 1500]> query_61_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_bias_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_30_self_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor<string, []>("query_61_cast_fp16")];
tensor<string, []> key_61_pad_type_0 = const()[name = tensor<string, []>("key_61_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_61_strides_0 = const()[name = tensor<string, []>("key_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_61_pad_0 = const()[name = tensor<string, []>("key_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_61_dilations_0 = const()[name = tensor<string, []>("key_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_61_groups_0 = const()[name = tensor<string, []>("key_61_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1198547328)))];
tensor<fp16, [1, 1280, 1, 1500]> key_61_cast_fp16 = conv(dilations = key_61_dilations_0, groups = key_61_groups_0, pad = key_61_pad_0, pad_type = key_61_pad_type_0, strides = key_61_strides_0, weight = layers_30_self_attn_k_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor<string, []>("key_61_cast_fp16")];
tensor<string, []> value_61_pad_type_0 = const()[name = tensor<string, []>("value_61_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_61_strides_0 = const()[name = tensor<string, []>("value_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_61_pad_0 = const()[name = tensor<string, []>("value_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_61_dilations_0 = const()[name = tensor<string, []>("value_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_61_groups_0 = const()[name = tensor<string, []>("value_61_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1201824192)))];
tensor<fp16, [1280]> layers_30_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1205101056)))];
tensor<fp16, [1, 1280, 1, 1500]> value_61_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_bias_to_fp16, dilations = value_61_dilations_0, groups = value_61_groups_0, pad = value_61_pad_0, pad_type = value_61_pad_type_0, strides = value_61_strides_0, weight = layers_30_self_attn_v_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor<string, []>("value_61_cast_fp16")];
tensor<int32, [4]> var_3772 = const()[name = tensor<string, []>("op_3772"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_61_cast_fp16 = reshape(shape = var_3772, x = query_61_cast_fp16)[name = tensor<string, []>("mh_q_61_cast_fp16")];
tensor<fp16, []> var_3774_to_fp16 = const()[name = tensor<string, []>("op_3774_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_3775_cast_fp16 = mul(x = mh_q_61_cast_fp16, y = var_3774_to_fp16)[name = tensor<string, []>("op_3775_cast_fp16")];
tensor<int32, [4]> var_3776 = const()[name = tensor<string, []>("op_3776"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3777_cast_fp16 = reshape(shape = var_3776, x = key_61_cast_fp16)[name = tensor<string, []>("op_3777_cast_fp16")];
tensor<bool, []> mh_w_61_transpose_x_0 = const()[name = tensor<string, []>("mh_w_61_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_61_transpose_y_0 = const()[name = tensor<string, []>("mh_w_61_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_3775_cast_fp16, y = var_3777_cast_fp16)[name = tensor<string, []>("mh_w_61_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_3780_cast_fp16 = softmax(axis = var_3718, x = mh_w_61_cast_fp16)[name = tensor<string, []>("op_3780_cast_fp16")];
tensor<int32, [4]> var_3781 = const()[name = tensor<string, []>("op_3781"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3782_cast_fp16 = reshape(shape = var_3781, x = value_61_cast_fp16)[name = tensor<string, []>("op_3782_cast_fp16")];
tensor<bool, []> attn_61_transpose_x_0 = const()[name = tensor<string, []>("attn_61_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_61_transpose_y_0 = const()[name = tensor<string, []>("attn_61_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3782_cast_fp16, y = var_3780_cast_fp16)[name = tensor<string, []>("attn_61_cast_fp16")];
tensor<int32, [4]> var_3785 = const()[name = tensor<string, []>("op_3785"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_241_cast_fp16 = reshape(shape = var_3785, x = attn_61_cast_fp16)[name = tensor<string, []>("input_241_cast_fp16")];
tensor<string, []> obj_123_pad_type_0 = const()[name = tensor<string, []>("obj_123_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_123_strides_0 = const()[name = tensor<string, []>("obj_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_123_pad_0 = const()[name = tensor<string, []>("obj_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_123_dilations_0 = const()[name = tensor<string, []>("obj_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_123_groups_0 = const()[name = tensor<string, []>("obj_123_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1205103680)))];
tensor<fp16, [1280]> layers_30_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1208380544)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_123_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_30_self_attn_o_proj_weight_to_fp16, x = input_241_cast_fp16)[name = tensor<string, []>("obj_123_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_123_cast_fp16)[name = tensor<string, []>("inputs_123_cast_fp16")];
tensor<int32, [1]> out_123_axes_0 = const()[name = tensor<string, []>("out_123_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3803_to_fp16 = const()[name = tensor<string, []>("op_3803_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_3803_to_fp16, x = inputs_123_cast_fp16)[name = tensor<string, []>("out_123_cast_fp16")];
tensor<fp16, [1280]> input_243_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_243_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1208383168)))];
tensor<fp16, [1280]> input_243_beta_0_to_fp16 = const()[name = tensor<string, []>("input_243_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1208385792)))];
tensor<fp16, []> input_243_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_243_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_243_cast_fp16 = batch_norm(beta = input_243_beta_0_to_fp16, epsilon = input_243_epsilon_0_to_fp16, gamma = input_243_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor<string, []>("input_243_cast_fp16")];
tensor<string, []> input_245_pad_type_0 = const()[name = tensor<string, []>("input_245_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_245_strides_0 = const()[name = tensor<string, []>("input_245_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_245_pad_0 = const()[name = tensor<string, []>("input_245_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_245_dilations_0 = const()[name = tensor<string, []>("input_245_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_245_groups_0 = const()[name = tensor<string, []>("input_245_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_30_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1208388416)))];
tensor<fp16, [5120]> layers_30_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1221495680)))];
tensor<fp16, [1, 5120, 1, 1500]> input_245_cast_fp16 = conv(bias = layers_30_fc1_bias_to_fp16, dilations = input_245_dilations_0, groups = input_245_groups_0, pad = input_245_pad_0, pad_type = input_245_pad_type_0, strides = input_245_strides_0, weight = layers_30_fc1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
tensor<string, []> input_247_mode_0 = const()[name = tensor<string, []>("input_247_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_247_cast_fp16 = gelu(mode = input_247_mode_0, x = input_245_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
tensor<string, []> hidden_states_65_pad_type_0 = const()[name = tensor<string, []>("hidden_states_65_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_65_strides_0 = const()[name = tensor<string, []>("hidden_states_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_65_pad_0 = const()[name = tensor<string, []>("hidden_states_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_65_dilations_0 = const()[name = tensor<string, []>("hidden_states_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_65_groups_0 = const()[name = tensor<string, []>("hidden_states_65_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_30_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1221505984)))];
tensor<fp16, [1280]> layers_30_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1234613248)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_65_cast_fp16 = conv(bias = layers_30_fc2_bias_to_fp16, dilations = hidden_states_65_dilations_0, groups = hidden_states_65_groups_0, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = hidden_states_65_strides_0, weight = layers_30_fc2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("hidden_states_65_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor<string, []>("inputs_125_cast_fp16")];
tensor<int32, []> var_3836 = const()[name = tensor<string, []>("op_3836"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_125_axes_0 = const()[name = tensor<string, []>("out_125_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3855_to_fp16 = const()[name = tensor<string, []>("op_3855_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_3855_to_fp16, x = inputs_125_cast_fp16)[name = tensor<string, []>("out_125_cast_fp16")];
tensor<fp16, [1280]> obj_125_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_125_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1234615872)))];
tensor<fp16, [1280]> obj_125_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_125_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1234618496)))];
tensor<fp16, []> obj_125_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_125_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> obj_125_cast_fp16 = batch_norm(beta = obj_125_beta_0_to_fp16, epsilon = obj_125_epsilon_0_to_fp16, gamma = obj_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor<string, []>("obj_125_cast_fp16")];
tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_strides_0 = const()[name = tensor<string, []>("query_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_dilations_0 = const()[name = tensor<string, []>("query_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_groups_0 = const()[name = tensor<string, []>("query_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1234621120)))];
tensor<fp16, [1280]> layers_31_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1237897984)))];
tensor<fp16, [1, 1280, 1, 1500]> query_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_31_self_attn_q_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_strides_0 = const()[name = tensor<string, []>("key_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_dilations_0 = const()[name = tensor<string, []>("key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_groups_0 = const()[name = tensor<string, []>("key_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1237900608)))];
tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_31_self_attn_k_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_strides_0 = const()[name = tensor<string, []>("value_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_dilations_0 = const()[name = tensor<string, []>("value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_groups_0 = const()[name = tensor<string, []>("value_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1241177472)))];
tensor<fp16, [1280]> layers_31_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1244454336)))];
tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_31_self_attn_v_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor<string, []>("value_cast_fp16")];
tensor<int32, [4]> var_3890 = const()[name = tensor<string, []>("op_3890"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> mh_q_cast_fp16 = reshape(shape = var_3890, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")];
tensor<fp16, []> var_3892_to_fp16 = const()[name = tensor<string, []>("op_3892_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 20, 64, 1500]> var_3893_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_3892_to_fp16)[name = tensor<string, []>("op_3893_cast_fp16")];
tensor<int32, [4]> var_3894 = const()[name = tensor<string, []>("op_3894"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3895_cast_fp16 = reshape(shape = var_3894, x = key_cast_fp16)[name = tensor<string, []>("op_3895_cast_fp16")];
tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 20, 1500, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_3893_cast_fp16, y = var_3895_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
tensor<fp16, [1, 20, 1500, 1500]> var_3898_cast_fp16 = softmax(axis = var_3836, x = mh_w_cast_fp16)[name = tensor<string, []>("op_3898_cast_fp16")];
tensor<int32, [4]> var_3899 = const()[name = tensor<string, []>("op_3899"), val = tensor<int32, [4]>([1, 20, 64, -1])];
tensor<fp16, [1, 20, 64, 1500]> var_3900_cast_fp16 = reshape(shape = var_3899, x = value_cast_fp16)[name = tensor<string, []>("op_3900_cast_fp16")];
tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 20, 64, 1500]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_3900_cast_fp16, y = var_3898_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
tensor<int32, [4]> var_3903 = const()[name = tensor<string, []>("op_3903"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
tensor<fp16, [1, 1280, 1, 1500]> input_249_cast_fp16 = reshape(shape = var_3903, x = attn_cast_fp16)[name = tensor<string, []>("input_249_cast_fp16")];
tensor<string, []> obj_pad_type_0 = const()[name = tensor<string, []>("obj_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_strides_0 = const()[name = tensor<string, []>("obj_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_pad_0 = const()[name = tensor<string, []>("obj_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_dilations_0 = const()[name = tensor<string, []>("obj_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_groups_0 = const()[name = tensor<string, []>("obj_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1244456960)))];
tensor<fp16, [1280]> layers_31_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1247733824)))];
tensor<fp16, [1, 1280, 1, 1500]> obj_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_bias_to_fp16, dilations = obj_dilations_0, groups = obj_groups_0, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = obj_strides_0, weight = layers_31_self_attn_o_proj_weight_to_fp16, x = input_249_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = obj_cast_fp16)[name = tensor<string, []>("inputs_127_cast_fp16")];
tensor<int32, [1]> out_127_axes_0 = const()[name = tensor<string, []>("out_127_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3921_to_fp16 = const()[name = tensor<string, []>("op_3921_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_3921_to_fp16, x = inputs_127_cast_fp16)[name = tensor<string, []>("out_127_cast_fp16")];
tensor<fp16, [1280]> input_251_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_251_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1247736448)))];
tensor<fp16, [1280]> input_251_beta_0_to_fp16 = const()[name = tensor<string, []>("input_251_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1247739072)))];
tensor<fp16, []> input_251_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_251_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> input_251_cast_fp16 = batch_norm(beta = input_251_beta_0_to_fp16, epsilon = input_251_epsilon_0_to_fp16, gamma = input_251_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
tensor<string, []> input_253_pad_type_0 = const()[name = tensor<string, []>("input_253_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_253_strides_0 = const()[name = tensor<string, []>("input_253_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_253_pad_0 = const()[name = tensor<string, []>("input_253_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_253_dilations_0 = const()[name = tensor<string, []>("input_253_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_253_groups_0 = const()[name = tensor<string, []>("input_253_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [5120, 1280, 1, 1]> layers_31_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_31_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1247741696)))];
tensor<fp16, [5120]> layers_31_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1260848960)))];
tensor<fp16, [1, 5120, 1, 1500]> input_253_cast_fp16 = conv(bias = layers_31_fc1_bias_to_fp16, dilations = input_253_dilations_0, groups = input_253_groups_0, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = input_253_strides_0, weight = layers_31_fc1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor<string, []>("input_253_cast_fp16")];
tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 5120, 1, 1500]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_253_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
tensor<string, []> hidden_states_pad_type_0 = const()[name = tensor<string, []>("hidden_states_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_strides_0 = const()[name = tensor<string, []>("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_pad_0 = const()[name = tensor<string, []>("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_dilations_0 = const()[name = tensor<string, []>("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_groups_0 = const()[name = tensor<string, []>("hidden_states_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1280, 5120, 1, 1]> layers_31_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_31_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1260859264)))];
tensor<fp16, [1280]> layers_31_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1273966528)))];
tensor<fp16, [1, 1280, 1, 1500]> hidden_states_cast_fp16 = conv(bias = layers_31_fc2_bias_to_fp16, dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_31_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
tensor<fp16, [1, 1280, 1, 1500]> inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = hidden_states_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_3959_to_fp16 = const()[name = tensor<string, []>("op_3959_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3959_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
tensor<fp16, [1280]> encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1273969152)))];
tensor<fp16, [1280]> encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1273971776)))];
tensor<fp16, []> encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 1280, 1, 1500]> encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("encoder_output_embeds_type_fp32_cast_fp16")];
} -> (encoder_output_embeds);
}