program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3400.43.1"}, {"coremlc-version", "3400.58.2"}})]
{
    func main<ios18>(tensor<fp16, [1, 128, 1, 3000]> melspectrogram_features) {
            string var_124_pad_type_0 = const()[name = string("op_124_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_124_pad_0 = const()[name = string("op_124_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> var_124_strides_0 = const()[name = string("op_124_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_124_dilations_0 = const()[name = string("op_124_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_124_groups_0 = const()[name = string("op_124_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 128, 1, 3]> var_93_to_fp16 = const()[name = string("op_93_to_fp16"), val = tensor<fp16, [1280, 128, 1, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [1280]> var_105_to_fp16 = const()[name = string("op_105_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(983168)))];
            tensor<fp16, [1, 1280, 1, 3000]> var_124_cast_fp16 = conv(bias = var_105_to_fp16, dilations = var_124_dilations_0, groups = var_124_groups_0, pad = var_124_pad_0, pad_type = var_124_pad_type_0, strides = var_124_strides_0, weight = var_93_to_fp16, x = melspectrogram_features)[name = string("op_124_cast_fp16")];
            string var_162_pad_type_0 = const()[name = string("op_162_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_162_pad_0 = const()[name = string("op_162_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> var_162_strides_0 = const()[name = string("op_162_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_162_dilations_0 = const()[name = string("op_162_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_162_groups_0 = const()[name = string("op_162_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 128, 1, 3]> op_137_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 128, 1, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(985792))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1231616))))[name = string("op_137_to_fp16_palettized")];
            tensor<fp16, [1280]> var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1231744)))];
            tensor<fp16, [1, 1280, 1, 3000]> var_162_cast_fp16 = conv(bias = var_143_to_fp16, dilations = var_162_dilations_0, groups = var_162_groups_0, pad = var_162_pad_0, pad_type = var_162_pad_type_0, strides = var_162_strides_0, weight = op_137_to_fp16_palettized, x = melspectrogram_features)[name = string("op_162_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 3000]> var_164_cast_fp16 = add(x = var_124_cast_fp16, y = var_162_cast_fp16)[name = string("op_164_cast_fp16")];
            string hidden_states_1_mode_0 = const()[name = string("hidden_states_1_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1280, 1, 3000]> hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_164_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            string var_210_pad_type_0 = const()[name = string("op_210_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_210_pad_0 = const()[name = string("op_210_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> var_210_strides_0 = const()[name = string("op_210_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_210_dilations_0 = const()[name = string("op_210_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_210_groups_0 = const()[name = string("op_210_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 3]> var_179_to_fp16 = const()[name = string("op_179_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1234368)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_210_cast_fp16 = conv(bias = var_105_to_fp16, dilations = var_210_dilations_0, groups = var_210_groups_0, pad = var_210_pad_0, pad_type = var_210_pad_type_0, strides = var_210_strides_0, weight = var_179_to_fp16, x = hidden_states_1_cast_fp16)[name = string("op_210_cast_fp16")];
            string var_248_pad_type_0 = const()[name = string("op_248_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_248_pad_0 = const()[name = string("op_248_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> var_248_strides_0 = const()[name = string("op_248_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_248_dilations_0 = const()[name = string("op_248_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_248_groups_0 = const()[name = string("op_248_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 3]> op_223_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11064832))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13522496))))[name = string("op_223_to_fp16_palettized")];
            tensor<fp16, [1280]> var_229_to_fp16 = const()[name = string("op_229_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13522624)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_248_cast_fp16 = conv(bias = var_229_to_fp16, dilations = var_248_dilations_0, groups = var_248_groups_0, pad = var_248_pad_0, pad_type = var_248_pad_type_0, strides = var_248_strides_0, weight = op_223_to_fp16_palettized, x = hidden_states_1_cast_fp16)[name = string("op_248_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> var_250_cast_fp16 = add(x = var_210_cast_fp16, y = var_248_cast_fp16)[name = string("op_250_cast_fp16")];
            string hidden_states_3_mode_0 = const()[name = string("hidden_states_3_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_250_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> var_270_to_fp16 = const()[name = string("op_270_to_fp16"), val = tensor<fp16, [1, 1280, 1, 1500]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13525248)))];
            tensor<fp16, [1, 1280, 1, 1500]> inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_270_to_fp16)[name = string("inputs_1_cast_fp16")];
            int32 var_284 = const()[name = string("op_284"), val = int32(3)];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_303_to_fp16 = const()[name = string("op_303_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_303_to_fp16, x = inputs_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [1280]> obj_1_variance_0_to_fp16 = const()[name = string("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17365312)))];
            tensor<fp16, [1280]> obj_1_gamma_0_to_fp16 = const()[name = string("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17367936)))];
            tensor<fp16, [1280]> obj_1_beta_0_to_fp16 = const()[name = string("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17370560)))];
            fp16 obj_1_epsilon_0_to_fp16 = const()[name = string("obj_1_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = string("obj_1_cast_fp16")];
            string var_325_pad_type_0 = const()[name = string("op_325_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_325_strides_0 = const()[name = string("op_325_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_325_pad_0 = const()[name = string("op_325_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_325_dilations_0 = const()[name = string("op_325_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_325_groups_0 = const()[name = string("op_325_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17373184))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18192448))))[name = string("layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_0_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18192576)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_325_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_325_dilations_0, groups = var_325_groups_0, pad = var_325_pad_0, pad_type = var_325_pad_type_0, strides = var_325_strides_0, weight = layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = string("op_325_cast_fp16")];
            string var_331_pad_type_0 = const()[name = string("op_331_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_331_strides_0 = const()[name = string("op_331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_331_pad_0 = const()[name = string("op_331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_331_dilations_0 = const()[name = string("op_331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_331_groups_0 = const()[name = string("op_331_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18275392))), nonzero_data = tensor<fp16, [40058]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18195200))))[name = string("layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_331_cast_fp16 = conv(dilations = var_331_dilations_0, groups = var_331_groups_0, pad = var_331_pad_0, pad_type = var_331_pad_type_0, strides = var_331_strides_0, weight = layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = string("op_331_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_1_cast_fp16 = add(x = var_325_cast_fp16, y = var_331_cast_fp16)[name = string("query_1_cast_fp16")];
            string var_340_pad_type_0 = const()[name = string("op_340_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_340_strides_0 = const()[name = string("op_340_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_340_pad_0 = const()[name = string("op_340_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_340_dilations_0 = const()[name = string("op_340_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_340_groups_0 = const()[name = string("op_340_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18480256))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19299520))))[name = string("layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_340_cast_fp16 = conv(dilations = var_340_dilations_0, groups = var_340_groups_0, pad = var_340_pad_0, pad_type = var_340_pad_type_0, strides = var_340_strides_0, weight = layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = string("op_340_cast_fp16")];
            string var_346_pad_type_0 = const()[name = string("op_346_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_346_strides_0 = const()[name = string("op_346_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_346_pad_0 = const()[name = string("op_346_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_346_dilations_0 = const()[name = string("op_346_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_346_groups_0 = const()[name = string("op_346_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19363520))), nonzero_data = tensor<fp16, [31892]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19299648))))[name = string("layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_346_cast_fp16 = conv(dilations = var_346_dilations_0, groups = var_346_groups_0, pad = var_346_pad_0, pad_type = var_346_pad_type_0, strides = var_346_strides_0, weight = layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = string("op_346_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_1_cast_fp16 = add(x = var_340_cast_fp16, y = var_346_cast_fp16)[name = string("key_1_cast_fp16")];
            string var_356_pad_type_0 = const()[name = string("op_356_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_356_strides_0 = const()[name = string("op_356_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_356_pad_0 = const()[name = string("op_356_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_356_dilations_0 = const()[name = string("op_356_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_356_groups_0 = const()[name = string("op_356_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19568384))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20387648))))[name = string("layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_0_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20387776)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_356_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_356_dilations_0, groups = var_356_groups_0, pad = var_356_pad_0, pad_type = var_356_pad_type_0, strides = var_356_strides_0, weight = layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = string("op_356_cast_fp16")];
            string var_362_pad_type_0 = const()[name = string("op_362_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_362_strides_0 = const()[name = string("op_362_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_362_pad_0 = const()[name = string("op_362_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_362_dilations_0 = const()[name = string("op_362_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_362_groups_0 = const()[name = string("op_362_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20474048))), nonzero_data = tensor<fp16, [41792]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20390400))))[name = string("layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_362_cast_fp16 = conv(dilations = var_362_dilations_0, groups = var_362_groups_0, pad = var_362_pad_0, pad_type = var_362_pad_type_0, strides = var_362_strides_0, weight = layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = string("op_362_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_1_cast_fp16 = add(x = var_356_cast_fp16, y = var_362_cast_fp16)[name = string("value_1_cast_fp16")];
            tensor<int32, [4]> var_365 = const()[name = string("op_365"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_1_cast_fp16 = reshape(shape = var_365, x = query_1_cast_fp16)[name = string("mh_q_1_cast_fp16")];
            fp16 var_367_to_fp16 = const()[name = string("op_367_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_368_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_367_to_fp16)[name = string("op_368_cast_fp16")];
            tensor<int32, [4]> var_369 = const()[name = string("op_369"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_370_cast_fp16 = reshape(shape = var_369, x = key_1_cast_fp16)[name = string("op_370_cast_fp16")];
            bool mh_w_1_transpose_x_0 = const()[name = string("mh_w_1_transpose_x_0"), val = bool(true)];
            bool mh_w_1_transpose_y_0 = const()[name = string("mh_w_1_transpose_y_0"), val = 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_368_cast_fp16, y = var_370_cast_fp16)[name = string("mh_w_1_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_373_cast_fp16 = softmax(axis = var_284, x = mh_w_1_cast_fp16)[name = string("op_373_cast_fp16")];
            tensor<int32, [4]> var_374 = const()[name = string("op_374"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_375_cast_fp16 = reshape(shape = var_374, x = value_1_cast_fp16)[name = string("op_375_cast_fp16")];
            bool attn_1_transpose_x_0 = const()[name = string("attn_1_transpose_x_0"), val = bool(false)];
            bool attn_1_transpose_y_0 = const()[name = string("attn_1_transpose_y_0"), val = 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_375_cast_fp16, y = var_373_cast_fp16)[name = string("attn_1_cast_fp16")];
            tensor<int32, [4]> var_378 = const()[name = string("op_378"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_1_cast_fp16 = reshape(shape = var_378, x = attn_1_cast_fp16)[name = string("input_1_cast_fp16")];
            string var_388_pad_type_0 = const()[name = string("op_388_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_388_strides_0 = const()[name = string("op_388_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_388_pad_0 = const()[name = string("op_388_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_388_dilations_0 = const()[name = string("op_388_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_388_groups_0 = const()[name = string("op_388_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20678912))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21498176))))[name = string("layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_0_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21498304)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_388_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_388_dilations_0, groups = var_388_groups_0, pad = var_388_pad_0, pad_type = var_388_pad_type_0, strides = var_388_strides_0, weight = layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = string("op_388_cast_fp16")];
            string var_394_pad_type_0 = const()[name = string("op_394_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_394_strides_0 = const()[name = string("op_394_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_394_pad_0 = const()[name = string("op_394_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_394_dilations_0 = const()[name = string("op_394_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_394_groups_0 = const()[name = string("op_394_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21553024))), nonzero_data = tensor<fp16, [25985]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21500928))))[name = string("layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_394_cast_fp16 = conv(dilations = var_394_dilations_0, groups = var_394_groups_0, pad = var_394_pad_0, pad_type = var_394_pad_type_0, strides = var_394_strides_0, weight = layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_1_cast_fp16)[name = string("op_394_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_3_cast_fp16 = add(x = var_388_cast_fp16, y = var_394_cast_fp16)[name = 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 = string("inputs_3_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_405_to_fp16 = const()[name = string("op_405_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_405_to_fp16, x = inputs_3_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [1280]> input_3_gamma_0_to_fp16 = const()[name = string("input_3_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21757888)))];
            tensor<fp16, [1280]> input_3_beta_0_to_fp16 = const()[name = string("input_3_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21760512)))];
            fp16 input_3_epsilon_0_to_fp16 = const()[name = string("input_3_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            string var_423_pad_type_0 = const()[name = string("op_423_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_423_strides_0 = const()[name = string("op_423_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_423_pad_0 = const()[name = string("op_423_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_423_dilations_0 = const()[name = string("op_423_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_423_groups_0 = const()[name = string("op_423_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21763136))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25040000))))[name = string("layers_0_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_0_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_0_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25040128)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_423_cast_fp16 = conv(bias = layers_0_fc1_inlier_module_bias_to_fp16, dilations = var_423_dilations_0, groups = var_423_groups_0, pad = var_423_pad_0, pad_type = var_423_pad_type_0, strides = var_423_strides_0, weight = layers_0_fc1_inlier_module_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = string("op_423_cast_fp16")];
            string var_429_pad_type_0 = const()[name = string("op_429_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_429_strides_0 = const()[name = string("op_429_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_429_pad_0 = const()[name = string("op_429_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_429_dilations_0 = const()[name = string("op_429_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_429_groups_0 = const()[name = string("op_429_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25353152))), nonzero_data = tensor<fp16, [151326]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25050432))))[name = string("layers_0_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_429_cast_fp16 = conv(dilations = var_429_dilations_0, groups = var_429_groups_0, pad = var_429_pad_0, pad_type = var_429_pad_type_0, strides = var_429_strides_0, weight = layers_0_fc1_outlier_module_weight_to_fp16_sparsified, x = input_3_cast_fp16)[name = string("op_429_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_5_cast_fp16 = add(x = var_423_cast_fp16, y = var_429_cast_fp16)[name = string("input_5_cast_fp16")];
            string input_7_mode_0 = const()[name = string("input_7_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
            string var_440_pad_type_0 = const()[name = string("op_440_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_440_strides_0 = const()[name = string("op_440_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_440_pad_0 = const()[name = string("op_440_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_440_dilations_0 = const()[name = string("op_440_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_440_groups_0 = const()[name = string("op_440_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26172416))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29449280))))[name = string("layers_0_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_0_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_0_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29449408)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_440_cast_fp16 = conv(bias = layers_0_fc2_inlier_module_bias_to_fp16, dilations = var_440_dilations_0, groups = var_440_groups_0, pad = var_440_pad_0, pad_type = var_440_pad_type_0, strides = var_440_strides_0, weight = layers_0_fc2_inlier_module_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = string("op_440_cast_fp16")];
            string var_446_pad_type_0 = const()[name = string("op_446_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_446_strides_0 = const()[name = string("op_446_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_446_pad_0 = const()[name = string("op_446_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_446_dilations_0 = const()[name = string("op_446_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_446_groups_0 = const()[name = string("op_446_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29689088))), nonzero_data = tensor<fp16, [118467]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29452032))))[name = string("layers_0_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_446_cast_fp16 = conv(dilations = var_446_dilations_0, groups = var_446_groups_0, pad = var_446_pad_0, pad_type = var_446_pad_type_0, strides = var_446_strides_0, weight = layers_0_fc2_outlier_module_weight_to_fp16_sparsified, x = input_7_cast_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_5_cast_fp16 = add(x = var_440_cast_fp16, y = var_446_cast_fp16)[name = 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 = string("inputs_5_cast_fp16")];
            int32 var_456 = const()[name = string("op_456"), val = int32(3)];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_475_to_fp16 = const()[name = string("op_475_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_475_to_fp16, x = inputs_5_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [1280]> obj_5_gamma_0_to_fp16 = const()[name = string("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30508352)))];
            tensor<fp16, [1280]> obj_5_beta_0_to_fp16 = const()[name = string("obj_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30510976)))];
            fp16 obj_5_epsilon_0_to_fp16 = const()[name = string("obj_5_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = string("obj_5_cast_fp16")];
            string var_497_pad_type_0 = const()[name = string("op_497_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_497_strides_0 = const()[name = string("op_497_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_497_pad_0 = const()[name = string("op_497_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_497_dilations_0 = const()[name = string("op_497_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_497_groups_0 = const()[name = string("op_497_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30513600))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31332864))))[name = string("layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_1_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31332992)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_497_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_497_dilations_0, groups = var_497_groups_0, pad = var_497_pad_0, pad_type = var_497_pad_type_0, strides = var_497_strides_0, weight = layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = string("op_497_cast_fp16")];
            string var_503_pad_type_0 = const()[name = string("op_503_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_503_strides_0 = const()[name = string("op_503_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_503_pad_0 = const()[name = string("op_503_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_503_dilations_0 = const()[name = string("op_503_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_503_groups_0 = const()[name = string("op_503_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31399552))), nonzero_data = tensor<fp16, [31911]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31335616))))[name = string("layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_503_cast_fp16 = conv(dilations = var_503_dilations_0, groups = var_503_groups_0, pad = var_503_pad_0, pad_type = var_503_pad_type_0, strides = var_503_strides_0, weight = layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_5_cast_fp16)[name = string("op_503_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_3_cast_fp16 = add(x = var_497_cast_fp16, y = var_503_cast_fp16)[name = string("query_3_cast_fp16")];
            string var_512_pad_type_0 = const()[name = string("op_512_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_512_strides_0 = const()[name = string("op_512_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_512_pad_0 = const()[name = string("op_512_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_512_dilations_0 = const()[name = string("op_512_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_512_groups_0 = const()[name = string("op_512_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31604416))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32423680))))[name = string("layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_512_cast_fp16 = conv(dilations = var_512_dilations_0, groups = var_512_groups_0, pad = var_512_pad_0, pad_type = var_512_pad_type_0, strides = var_512_strides_0, weight = layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = string("op_512_cast_fp16")];
            string var_518_pad_type_0 = const()[name = string("op_518_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_518_strides_0 = const()[name = string("op_518_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_518_pad_0 = const()[name = string("op_518_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_518_dilations_0 = const()[name = string("op_518_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_518_groups_0 = const()[name = string("op_518_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32475264))), nonzero_data = tensor<fp16, [25674]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32423808))))[name = string("layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_518_cast_fp16 = conv(dilations = var_518_dilations_0, groups = var_518_groups_0, pad = var_518_pad_0, pad_type = var_518_pad_type_0, strides = var_518_strides_0, weight = layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_5_cast_fp16)[name = string("op_518_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = add(x = var_512_cast_fp16, y = var_518_cast_fp16)[name = string("key_3_cast_fp16")];
            string var_528_pad_type_0 = const()[name = string("op_528_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_528_strides_0 = const()[name = string("op_528_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_528_pad_0 = const()[name = string("op_528_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_528_dilations_0 = const()[name = string("op_528_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_528_groups_0 = const()[name = string("op_528_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32680128))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33499392))))[name = string("layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_1_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33499520)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_528_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_528_dilations_0, groups = var_528_groups_0, pad = var_528_pad_0, pad_type = var_528_pad_type_0, strides = var_528_strides_0, weight = layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = string("op_528_cast_fp16")];
            string var_534_pad_type_0 = const()[name = string("op_534_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_534_strides_0 = const()[name = string("op_534_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_534_pad_0 = const()[name = string("op_534_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_534_dilations_0 = const()[name = string("op_534_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_534_groups_0 = const()[name = string("op_534_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33549632))), nonzero_data = tensor<fp16, [23687]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33502144))))[name = string("layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_534_cast_fp16 = conv(dilations = var_534_dilations_0, groups = var_534_groups_0, pad = var_534_pad_0, pad_type = var_534_pad_type_0, strides = var_534_strides_0, weight = layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_5_cast_fp16)[name = string("op_534_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = add(x = var_528_cast_fp16, y = var_534_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<int32, [4]> var_537 = const()[name = string("op_537"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_3_cast_fp16 = reshape(shape = var_537, x = query_3_cast_fp16)[name = string("mh_q_3_cast_fp16")];
            fp16 var_539_to_fp16 = const()[name = string("op_539_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_540_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_539_to_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_542_cast_fp16 = reshape(shape = var_541, x = key_3_cast_fp16)[name = string("op_542_cast_fp16")];
            bool mh_w_3_transpose_x_0 = const()[name = string("mh_w_3_transpose_x_0"), val = bool(true)];
            bool mh_w_3_transpose_y_0 = const()[name = string("mh_w_3_transpose_y_0"), val = 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_540_cast_fp16, y = var_542_cast_fp16)[name = string("mh_w_3_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_545_cast_fp16 = softmax(axis = var_456, x = mh_w_3_cast_fp16)[name = string("op_545_cast_fp16")];
            tensor<int32, [4]> var_546 = const()[name = string("op_546"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_547_cast_fp16 = reshape(shape = var_546, x = value_3_cast_fp16)[name = string("op_547_cast_fp16")];
            bool attn_3_transpose_x_0 = const()[name = string("attn_3_transpose_x_0"), val = bool(false)];
            bool attn_3_transpose_y_0 = const()[name = string("attn_3_transpose_y_0"), val = 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_547_cast_fp16, y = var_545_cast_fp16)[name = string("attn_3_cast_fp16")];
            tensor<int32, [4]> var_550 = const()[name = string("op_550"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_9_cast_fp16 = reshape(shape = var_550, x = attn_3_cast_fp16)[name = string("input_9_cast_fp16")];
            string var_560_pad_type_0 = const()[name = string("op_560_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_560_strides_0 = const()[name = string("op_560_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_560_pad_0 = const()[name = string("op_560_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_560_dilations_0 = const()[name = string("op_560_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_560_groups_0 = const()[name = string("op_560_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33754496))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34573760))))[name = string("layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_1_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34573888)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_560_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_560_dilations_0, groups = var_560_groups_0, pad = var_560_pad_0, pad_type = var_560_pad_type_0, strides = var_560_strides_0, weight = layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("op_560_cast_fp16")];
            string var_566_pad_type_0 = const()[name = string("op_566_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_566_strides_0 = const()[name = string("op_566_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_566_pad_0 = const()[name = string("op_566_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_566_dilations_0 = const()[name = string("op_566_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_566_groups_0 = const()[name = string("op_566_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34612736))), nonzero_data = tensor<fp16, [18054]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34576512))))[name = string("layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_566_cast_fp16 = conv(dilations = var_566_dilations_0, groups = var_566_groups_0, pad = var_566_pad_0, pad_type = var_566_pad_type_0, strides = var_566_strides_0, weight = layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_9_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_7_cast_fp16 = add(x = var_560_cast_fp16, y = var_566_cast_fp16)[name = 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 = string("inputs_7_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_577_to_fp16 = const()[name = string("op_577_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_577_to_fp16, x = inputs_7_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [1280]> input_11_gamma_0_to_fp16 = const()[name = string("input_11_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34817600)))];
            tensor<fp16, [1280]> input_11_beta_0_to_fp16 = const()[name = string("input_11_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34820224)))];
            fp16 input_11_epsilon_0_to_fp16 = const()[name = string("input_11_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            string var_595_pad_type_0 = const()[name = string("op_595_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_595_strides_0 = const()[name = string("op_595_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_595_pad_0 = const()[name = string("op_595_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_595_dilations_0 = const()[name = string("op_595_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_595_groups_0 = const()[name = string("op_595_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34822848))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38099712))))[name = string("layers_1_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_1_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_1_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38099840)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_595_cast_fp16 = conv(bias = layers_1_fc1_inlier_module_bias_to_fp16, dilations = var_595_dilations_0, groups = var_595_groups_0, pad = var_595_pad_0, pad_type = var_595_pad_type_0, strides = var_595_strides_0, weight = layers_1_fc1_inlier_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = string("op_595_cast_fp16")];
            string var_601_pad_type_0 = const()[name = string("op_601_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_601_strides_0 = const()[name = string("op_601_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_601_pad_0 = const()[name = string("op_601_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_601_dilations_0 = const()[name = string("op_601_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_601_groups_0 = const()[name = string("op_601_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38256320))), nonzero_data = tensor<fp16, [73044]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38110144))))[name = string("layers_1_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_601_cast_fp16 = conv(dilations = var_601_dilations_0, groups = var_601_groups_0, pad = var_601_pad_0, pad_type = var_601_pad_type_0, strides = var_601_strides_0, weight = layers_1_fc1_outlier_module_weight_to_fp16_sparsified, x = input_11_cast_fp16)[name = string("op_601_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_13_cast_fp16 = add(x = var_595_cast_fp16, y = var_601_cast_fp16)[name = string("input_13_cast_fp16")];
            string input_15_mode_0 = const()[name = string("input_15_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")];
            string var_612_pad_type_0 = const()[name = string("op_612_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_612_strides_0 = const()[name = string("op_612_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_612_pad_0 = const()[name = string("op_612_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_612_dilations_0 = const()[name = string("op_612_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_612_groups_0 = const()[name = string("op_612_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39075584))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42352448))))[name = string("layers_1_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_1_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_1_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42352576)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_612_cast_fp16 = conv(bias = layers_1_fc2_inlier_module_bias_to_fp16, dilations = var_612_dilations_0, groups = var_612_groups_0, pad = var_612_pad_0, pad_type = var_612_pad_type_0, strides = var_612_strides_0, weight = layers_1_fc2_inlier_module_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = string("op_612_cast_fp16")];
            string var_618_pad_type_0 = const()[name = string("op_618_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_618_strides_0 = const()[name = string("op_618_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_618_pad_0 = const()[name = string("op_618_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_618_dilations_0 = const()[name = string("op_618_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_618_groups_0 = const()[name = string("op_618_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42567552))), nonzero_data = tensor<fp16, [106134]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42355200))))[name = string("layers_1_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_618_cast_fp16 = conv(dilations = var_618_dilations_0, groups = var_618_groups_0, pad = var_618_pad_0, pad_type = var_618_pad_type_0, strides = var_618_strides_0, weight = layers_1_fc2_outlier_module_weight_to_fp16_sparsified, x = input_15_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_7_cast_fp16 = add(x = var_612_cast_fp16, y = var_618_cast_fp16)[name = 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 = string("inputs_9_cast_fp16")];
            int32 var_628 = const()[name = string("op_628"), val = int32(3)];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_647_to_fp16 = const()[name = string("op_647_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_647_to_fp16, x = inputs_9_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [1280]> obj_9_gamma_0_to_fp16 = const()[name = string("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43386816)))];
            tensor<fp16, [1280]> obj_9_beta_0_to_fp16 = const()[name = string("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43389440)))];
            fp16 obj_9_epsilon_0_to_fp16 = const()[name = string("obj_9_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = string("obj_9_cast_fp16")];
            string var_669_pad_type_0 = const()[name = string("op_669_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_669_strides_0 = const()[name = string("op_669_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_669_pad_0 = const()[name = string("op_669_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_669_dilations_0 = const()[name = string("op_669_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_669_groups_0 = const()[name = string("op_669_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43392064))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44211328))))[name = string("layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_2_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44211456)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_669_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_669_dilations_0, groups = var_669_groups_0, pad = var_669_pad_0, pad_type = var_669_pad_type_0, strides = var_669_strides_0, weight = layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = string("op_669_cast_fp16")];
            string var_675_pad_type_0 = const()[name = string("op_675_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_675_strides_0 = const()[name = string("op_675_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_675_pad_0 = const()[name = string("op_675_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_675_dilations_0 = const()[name = string("op_675_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_675_groups_0 = const()[name = string("op_675_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44271552))), nonzero_data = tensor<fp16, [28690]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44214080))))[name = string("layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_675_cast_fp16 = conv(dilations = var_675_dilations_0, groups = var_675_groups_0, pad = var_675_pad_0, pad_type = var_675_pad_type_0, strides = var_675_strides_0, weight = layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_9_cast_fp16)[name = string("op_675_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_5_cast_fp16 = add(x = var_669_cast_fp16, y = var_675_cast_fp16)[name = string("query_5_cast_fp16")];
            string var_684_pad_type_0 = const()[name = string("op_684_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_684_strides_0 = const()[name = string("op_684_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_684_pad_0 = const()[name = string("op_684_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_684_dilations_0 = const()[name = string("op_684_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_684_groups_0 = const()[name = string("op_684_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44476416))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45295680))))[name = string("layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_684_cast_fp16 = conv(dilations = var_684_dilations_0, groups = var_684_groups_0, pad = var_684_pad_0, pad_type = var_684_pad_type_0, strides = var_684_strides_0, weight = layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = string("op_684_cast_fp16")];
            string var_690_pad_type_0 = const()[name = string("op_690_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_690_strides_0 = const()[name = string("op_690_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_690_pad_0 = const()[name = string("op_690_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_690_dilations_0 = const()[name = string("op_690_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_690_groups_0 = const()[name = string("op_690_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45344832))), nonzero_data = tensor<fp16, [24459]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45295808))))[name = string("layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_690_cast_fp16 = conv(dilations = var_690_dilations_0, groups = var_690_groups_0, pad = var_690_pad_0, pad_type = var_690_pad_type_0, strides = var_690_strides_0, weight = layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_9_cast_fp16)[name = string("op_690_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_5_cast_fp16 = add(x = var_684_cast_fp16, y = var_690_cast_fp16)[name = string("key_5_cast_fp16")];
            string var_700_pad_type_0 = const()[name = string("op_700_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_700_strides_0 = const()[name = string("op_700_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_700_pad_0 = const()[name = string("op_700_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_700_dilations_0 = const()[name = string("op_700_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_700_groups_0 = const()[name = string("op_700_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45549696))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46368960))))[name = string("layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_2_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46369088)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_700_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_700_dilations_0, groups = var_700_groups_0, pad = var_700_pad_0, pad_type = var_700_pad_type_0, strides = var_700_strides_0, weight = layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = string("op_700_cast_fp16")];
            string var_706_pad_type_0 = const()[name = string("op_706_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_706_strides_0 = const()[name = string("op_706_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_706_pad_0 = const()[name = string("op_706_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_706_dilations_0 = const()[name = string("op_706_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_706_groups_0 = const()[name = string("op_706_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46410240))), nonzero_data = tensor<fp16, [19205]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46371712))))[name = string("layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_706_cast_fp16 = conv(dilations = var_706_dilations_0, groups = var_706_groups_0, pad = var_706_pad_0, pad_type = var_706_pad_type_0, strides = var_706_strides_0, weight = layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_9_cast_fp16)[name = string("op_706_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_5_cast_fp16 = add(x = var_700_cast_fp16, y = var_706_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<int32, [4]> var_709 = const()[name = string("op_709"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_5_cast_fp16 = reshape(shape = var_709, x = query_5_cast_fp16)[name = string("mh_q_5_cast_fp16")];
            fp16 var_711_to_fp16 = const()[name = string("op_711_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_712_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_711_to_fp16)[name = string("op_712_cast_fp16")];
            tensor<int32, [4]> var_713 = const()[name = 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 = key_5_cast_fp16)[name = string("op_714_cast_fp16")];
            bool mh_w_5_transpose_x_0 = const()[name = string("mh_w_5_transpose_x_0"), val = bool(true)];
            bool mh_w_5_transpose_y_0 = const()[name = string("mh_w_5_transpose_y_0"), val = 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_712_cast_fp16, y = var_714_cast_fp16)[name = string("mh_w_5_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_717_cast_fp16 = softmax(axis = var_628, x = mh_w_5_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<int32, [4]> var_718 = const()[name = string("op_718"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_719_cast_fp16 = reshape(shape = var_718, x = value_5_cast_fp16)[name = string("op_719_cast_fp16")];
            bool attn_5_transpose_x_0 = const()[name = string("attn_5_transpose_x_0"), val = bool(false)];
            bool attn_5_transpose_y_0 = const()[name = string("attn_5_transpose_y_0"), val = 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_719_cast_fp16, y = var_717_cast_fp16)[name = string("attn_5_cast_fp16")];
            tensor<int32, [4]> var_722 = const()[name = string("op_722"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_17_cast_fp16 = reshape(shape = var_722, x = attn_5_cast_fp16)[name = string("input_17_cast_fp16")];
            string var_732_pad_type_0 = const()[name = string("op_732_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_732_strides_0 = const()[name = string("op_732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_732_pad_0 = const()[name = string("op_732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_732_dilations_0 = const()[name = string("op_732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_732_groups_0 = const()[name = string("op_732_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46615104))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47434368))))[name = string("layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_2_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47434496)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_732_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_732_dilations_0, groups = var_732_groups_0, pad = var_732_pad_0, pad_type = var_732_pad_type_0, strides = var_732_strides_0, weight = layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = string("op_732_cast_fp16")];
            string var_738_pad_type_0 = const()[name = string("op_738_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_738_strides_0 = const()[name = string("op_738_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_738_pad_0 = const()[name = string("op_738_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_738_dilations_0 = const()[name = string("op_738_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_738_groups_0 = const()[name = string("op_738_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47467328))), nonzero_data = tensor<fp16, [15072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47437120))))[name = string("layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_738_cast_fp16 = conv(dilations = var_738_dilations_0, groups = var_738_groups_0, pad = var_738_pad_0, pad_type = var_738_pad_type_0, strides = var_738_strides_0, weight = layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_17_cast_fp16)[name = string("op_738_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_11_cast_fp16 = add(x = var_732_cast_fp16, y = var_738_cast_fp16)[name = 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 = string("inputs_11_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_749_to_fp16 = const()[name = string("op_749_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_749_to_fp16, x = inputs_11_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [1280]> input_19_gamma_0_to_fp16 = const()[name = string("input_19_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47672192)))];
            tensor<fp16, [1280]> input_19_beta_0_to_fp16 = const()[name = string("input_19_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47674816)))];
            fp16 input_19_epsilon_0_to_fp16 = const()[name = string("input_19_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            string var_767_pad_type_0 = const()[name = string("op_767_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_767_strides_0 = const()[name = string("op_767_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_767_pad_0 = const()[name = string("op_767_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_767_dilations_0 = const()[name = string("op_767_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_767_groups_0 = const()[name = string("op_767_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47677440))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50954304))))[name = string("layers_2_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_2_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_2_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50954432)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_767_cast_fp16 = conv(bias = layers_2_fc1_inlier_module_bias_to_fp16, dilations = var_767_dilations_0, groups = var_767_groups_0, pad = var_767_pad_0, pad_type = var_767_pad_type_0, strides = var_767_strides_0, weight = layers_2_fc1_inlier_module_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = string("op_767_cast_fp16")];
            string var_773_pad_type_0 = const()[name = string("op_773_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_773_strides_0 = const()[name = string("op_773_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_773_pad_0 = const()[name = string("op_773_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_773_dilations_0 = const()[name = string("op_773_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_773_groups_0 = const()[name = string("op_773_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51003008))), nonzero_data = tensor<fp16, [19084]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50964736))))[name = string("layers_2_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_773_cast_fp16 = conv(dilations = var_773_dilations_0, groups = var_773_groups_0, pad = var_773_pad_0, pad_type = var_773_pad_type_0, strides = var_773_strides_0, weight = layers_2_fc1_outlier_module_weight_to_fp16_sparsified, x = input_19_cast_fp16)[name = string("op_773_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_21_cast_fp16 = add(x = var_767_cast_fp16, y = var_773_cast_fp16)[name = string("input_21_cast_fp16")];
            string input_23_mode_0 = const()[name = string("input_23_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = input_21_cast_fp16)[name = string("input_23_cast_fp16")];
            string var_784_pad_type_0 = const()[name = string("op_784_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_784_strides_0 = const()[name = string("op_784_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_784_pad_0 = const()[name = string("op_784_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_784_dilations_0 = const()[name = string("op_784_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_784_groups_0 = const()[name = string("op_784_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51822272))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55099136))))[name = string("layers_2_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_2_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_2_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55099264)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_784_cast_fp16 = conv(bias = layers_2_fc2_inlier_module_bias_to_fp16, dilations = var_784_dilations_0, groups = var_784_groups_0, pad = var_784_pad_0, pad_type = var_784_pad_type_0, strides = var_784_strides_0, weight = layers_2_fc2_inlier_module_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = string("op_784_cast_fp16")];
            string var_790_pad_type_0 = const()[name = string("op_790_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_790_strides_0 = const()[name = string("op_790_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_790_pad_0 = const()[name = string("op_790_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_790_dilations_0 = const()[name = string("op_790_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_790_groups_0 = const()[name = string("op_790_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55315392))), nonzero_data = tensor<fp16, [106720]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55101888))))[name = string("layers_2_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_790_cast_fp16 = conv(dilations = var_790_dilations_0, groups = var_790_groups_0, pad = var_790_pad_0, pad_type = var_790_pad_type_0, strides = var_790_strides_0, weight = layers_2_fc2_outlier_module_weight_to_fp16_sparsified, x = input_23_cast_fp16)[name = string("op_790_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_9_cast_fp16 = add(x = var_784_cast_fp16, y = var_790_cast_fp16)[name = 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 = string("inputs_13_cast_fp16")];
            int32 var_800 = const()[name = string("op_800"), val = int32(3)];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_819_to_fp16 = const()[name = string("op_819_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_819_to_fp16, x = inputs_13_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [1280]> obj_13_gamma_0_to_fp16 = const()[name = string("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56134656)))];
            tensor<fp16, [1280]> obj_13_beta_0_to_fp16 = const()[name = string("obj_13_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56137280)))];
            fp16 obj_13_epsilon_0_to_fp16 = const()[name = string("obj_13_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = string("obj_13_cast_fp16")];
            string var_841_pad_type_0 = const()[name = string("op_841_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_841_strides_0 = const()[name = string("op_841_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_841_pad_0 = const()[name = string("op_841_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_841_dilations_0 = const()[name = string("op_841_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_841_groups_0 = const()[name = string("op_841_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56139904))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56959168))))[name = string("layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_3_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56959296)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_841_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_841_dilations_0, groups = var_841_groups_0, pad = var_841_pad_0, pad_type = var_841_pad_type_0, strides = var_841_strides_0, weight = layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = string("op_841_cast_fp16")];
            string var_847_pad_type_0 = const()[name = string("op_847_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_847_strides_0 = const()[name = string("op_847_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_847_pad_0 = const()[name = string("op_847_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_847_dilations_0 = const()[name = string("op_847_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_847_groups_0 = const()[name = string("op_847_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57013184))), nonzero_data = tensor<fp16, [25575]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56961920))))[name = string("layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_847_cast_fp16 = conv(dilations = var_847_dilations_0, groups = var_847_groups_0, pad = var_847_pad_0, pad_type = var_847_pad_type_0, strides = var_847_strides_0, weight = layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_13_cast_fp16)[name = string("op_847_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_7_cast_fp16 = add(x = var_841_cast_fp16, y = var_847_cast_fp16)[name = string("query_7_cast_fp16")];
            string var_856_pad_type_0 = const()[name = string("op_856_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_856_strides_0 = const()[name = string("op_856_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_856_pad_0 = const()[name = string("op_856_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_856_dilations_0 = const()[name = string("op_856_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_856_groups_0 = const()[name = string("op_856_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57218048))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58037312))))[name = string("layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_856_cast_fp16 = conv(dilations = var_856_dilations_0, groups = var_856_groups_0, pad = var_856_pad_0, pad_type = var_856_pad_type_0, strides = var_856_strides_0, weight = layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = string("op_856_cast_fp16")];
            string var_862_pad_type_0 = const()[name = string("op_862_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_862_strides_0 = const()[name = string("op_862_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_862_pad_0 = const()[name = string("op_862_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_862_dilations_0 = const()[name = string("op_862_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_862_groups_0 = const()[name = string("op_862_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58077888))), nonzero_data = tensor<fp16, [20191]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58037440))))[name = string("layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_862_cast_fp16 = conv(dilations = var_862_dilations_0, groups = var_862_groups_0, pad = var_862_pad_0, pad_type = var_862_pad_type_0, strides = var_862_strides_0, weight = layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_13_cast_fp16)[name = string("op_862_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_7_cast_fp16 = add(x = var_856_cast_fp16, y = var_862_cast_fp16)[name = string("key_7_cast_fp16")];
            string var_872_pad_type_0 = const()[name = string("op_872_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_872_strides_0 = const()[name = string("op_872_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_872_pad_0 = const()[name = string("op_872_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_872_dilations_0 = const()[name = string("op_872_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_872_groups_0 = const()[name = string("op_872_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58282752))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59102016))))[name = string("layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_3_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59102144)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_872_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_872_dilations_0, groups = var_872_groups_0, pad = var_872_pad_0, pad_type = var_872_pad_type_0, strides = var_872_strides_0, weight = layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = string("op_872_cast_fp16")];
            string var_878_pad_type_0 = const()[name = string("op_878_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_878_strides_0 = const()[name = string("op_878_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_878_pad_0 = const()[name = string("op_878_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_878_dilations_0 = const()[name = string("op_878_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_878_groups_0 = const()[name = string("op_878_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59138432))), nonzero_data = tensor<fp16, [16773]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59104768))))[name = string("layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_878_cast_fp16 = conv(dilations = var_878_dilations_0, groups = var_878_groups_0, pad = var_878_pad_0, pad_type = var_878_pad_type_0, strides = var_878_strides_0, weight = layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_13_cast_fp16)[name = string("op_878_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_7_cast_fp16 = add(x = var_872_cast_fp16, y = var_878_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<int32, [4]> var_881 = const()[name = string("op_881"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_7_cast_fp16 = reshape(shape = var_881, x = query_7_cast_fp16)[name = string("mh_q_7_cast_fp16")];
            fp16 var_883_to_fp16 = const()[name = string("op_883_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_884_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_883_to_fp16)[name = string("op_884_cast_fp16")];
            tensor<int32, [4]> var_885 = const()[name = string("op_885"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_886_cast_fp16 = reshape(shape = var_885, x = key_7_cast_fp16)[name = string("op_886_cast_fp16")];
            bool mh_w_7_transpose_x_0 = const()[name = string("mh_w_7_transpose_x_0"), val = bool(true)];
            bool mh_w_7_transpose_y_0 = const()[name = string("mh_w_7_transpose_y_0"), val = 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_884_cast_fp16, y = var_886_cast_fp16)[name = string("mh_w_7_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_889_cast_fp16 = softmax(axis = var_800, x = mh_w_7_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<int32, [4]> var_890 = const()[name = string("op_890"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_891_cast_fp16 = reshape(shape = var_890, x = value_7_cast_fp16)[name = string("op_891_cast_fp16")];
            bool attn_7_transpose_x_0 = const()[name = string("attn_7_transpose_x_0"), val = bool(false)];
            bool attn_7_transpose_y_0 = const()[name = string("attn_7_transpose_y_0"), val = 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_891_cast_fp16, y = var_889_cast_fp16)[name = string("attn_7_cast_fp16")];
            tensor<int32, [4]> var_894 = const()[name = string("op_894"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_25_cast_fp16 = reshape(shape = var_894, x = attn_7_cast_fp16)[name = string("input_25_cast_fp16")];
            string var_904_pad_type_0 = const()[name = string("op_904_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_904_strides_0 = const()[name = string("op_904_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_904_pad_0 = const()[name = string("op_904_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_904_dilations_0 = const()[name = string("op_904_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_904_groups_0 = const()[name = string("op_904_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59343296))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60162560))))[name = string("layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_3_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60162688)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_904_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_904_dilations_0, groups = var_904_groups_0, pad = var_904_pad_0, pad_type = var_904_pad_type_0, strides = var_904_strides_0, weight = layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = string("op_904_cast_fp16")];
            string var_910_pad_type_0 = const()[name = string("op_910_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_910_strides_0 = const()[name = string("op_910_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_910_pad_0 = const()[name = string("op_910_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_910_dilations_0 = const()[name = string("op_910_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_910_groups_0 = const()[name = string("op_910_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60188096))), nonzero_data = tensor<fp16, [11350]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60165312))))[name = string("layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_910_cast_fp16 = conv(dilations = var_910_dilations_0, groups = var_910_groups_0, pad = var_910_pad_0, pad_type = var_910_pad_type_0, strides = var_910_strides_0, weight = layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_25_cast_fp16)[name = string("op_910_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_15_cast_fp16 = add(x = var_904_cast_fp16, y = var_910_cast_fp16)[name = 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 = string("inputs_15_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_921_to_fp16 = const()[name = string("op_921_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_921_to_fp16, x = inputs_15_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [1280]> input_27_gamma_0_to_fp16 = const()[name = string("input_27_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60392960)))];
            tensor<fp16, [1280]> input_27_beta_0_to_fp16 = const()[name = string("input_27_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60395584)))];
            fp16 input_27_epsilon_0_to_fp16 = const()[name = string("input_27_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            string var_939_pad_type_0 = const()[name = string("op_939_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_939_strides_0 = const()[name = string("op_939_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_939_pad_0 = const()[name = string("op_939_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_939_dilations_0 = const()[name = string("op_939_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_939_groups_0 = const()[name = string("op_939_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60398208))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63675072))))[name = string("layers_3_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_3_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_3_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63675200)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_939_cast_fp16 = conv(bias = layers_3_fc1_inlier_module_bias_to_fp16, dilations = var_939_dilations_0, groups = var_939_groups_0, pad = var_939_pad_0, pad_type = var_939_pad_type_0, strides = var_939_strides_0, weight = layers_3_fc1_inlier_module_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = string("op_939_cast_fp16")];
            string var_945_pad_type_0 = const()[name = string("op_945_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_945_strides_0 = const()[name = string("op_945_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_945_pad_0 = const()[name = string("op_945_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_945_dilations_0 = const()[name = string("op_945_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_945_groups_0 = const()[name = string("op_945_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63740864))), nonzero_data = tensor<fp16, [27622]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63685504))))[name = string("layers_3_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_945_cast_fp16 = conv(dilations = var_945_dilations_0, groups = var_945_groups_0, pad = var_945_pad_0, pad_type = var_945_pad_type_0, strides = var_945_strides_0, weight = layers_3_fc1_outlier_module_weight_to_fp16_sparsified, x = input_27_cast_fp16)[name = string("op_945_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_29_cast_fp16 = add(x = var_939_cast_fp16, y = var_945_cast_fp16)[name = string("input_29_cast_fp16")];
            string input_31_mode_0 = const()[name = string("input_31_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")];
            string var_956_pad_type_0 = const()[name = string("op_956_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_956_strides_0 = const()[name = string("op_956_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_956_pad_0 = const()[name = string("op_956_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_956_dilations_0 = const()[name = string("op_956_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_956_groups_0 = const()[name = string("op_956_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64560128))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67836992))))[name = string("layers_3_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_3_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_3_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67837120)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_956_cast_fp16 = conv(bias = layers_3_fc2_inlier_module_bias_to_fp16, dilations = var_956_dilations_0, groups = var_956_groups_0, pad = var_956_pad_0, pad_type = var_956_pad_type_0, strides = var_956_strides_0, weight = layers_3_fc2_inlier_module_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = string("op_956_cast_fp16")];
            string var_962_pad_type_0 = const()[name = string("op_962_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_962_strides_0 = const()[name = string("op_962_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_962_pad_0 = const()[name = string("op_962_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_962_dilations_0 = const()[name = string("op_962_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_962_groups_0 = const()[name = string("op_962_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68064960))), nonzero_data = tensor<fp16, [112545]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67839744))))[name = string("layers_3_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_962_cast_fp16 = conv(dilations = var_962_dilations_0, groups = var_962_groups_0, pad = var_962_pad_0, pad_type = var_962_pad_type_0, strides = var_962_strides_0, weight = layers_3_fc2_outlier_module_weight_to_fp16_sparsified, x = input_31_cast_fp16)[name = string("op_962_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_11_cast_fp16 = add(x = var_956_cast_fp16, y = var_962_cast_fp16)[name = 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 = string("inputs_17_cast_fp16")];
            int32 var_972 = const()[name = string("op_972"), val = int32(3)];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_991_to_fp16 = const()[name = string("op_991_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_991_to_fp16, x = inputs_17_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [1280]> obj_17_gamma_0_to_fp16 = const()[name = string("obj_17_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68884224)))];
            tensor<fp16, [1280]> obj_17_beta_0_to_fp16 = const()[name = string("obj_17_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68886848)))];
            fp16 obj_17_epsilon_0_to_fp16 = const()[name = string("obj_17_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = string("obj_17_cast_fp16")];
            string var_1013_pad_type_0 = const()[name = string("op_1013_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1013_strides_0 = const()[name = string("op_1013_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1013_pad_0 = const()[name = string("op_1013_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1013_dilations_0 = const()[name = string("op_1013_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1013_groups_0 = const()[name = string("op_1013_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68889472))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69708736))))[name = string("layers_4_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_4_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69708864)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1013_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1013_dilations_0, groups = var_1013_groups_0, pad = var_1013_pad_0, pad_type = var_1013_pad_type_0, strides = var_1013_strides_0, weight = layers_4_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = string("op_1013_cast_fp16")];
            string var_1019_pad_type_0 = const()[name = string("op_1019_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1019_strides_0 = const()[name = string("op_1019_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1019_pad_0 = const()[name = string("op_1019_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1019_dilations_0 = const()[name = string("op_1019_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1019_groups_0 = const()[name = string("op_1019_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69759808))), nonzero_data = tensor<fp16, [24120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69711488))))[name = string("layers_4_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1019_cast_fp16 = conv(dilations = var_1019_dilations_0, groups = var_1019_groups_0, pad = var_1019_pad_0, pad_type = var_1019_pad_type_0, strides = var_1019_strides_0, weight = layers_4_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_17_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_9_cast_fp16 = add(x = var_1013_cast_fp16, y = var_1019_cast_fp16)[name = string("query_9_cast_fp16")];
            string var_1028_pad_type_0 = const()[name = string("op_1028_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1028_strides_0 = const()[name = string("op_1028_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1028_pad_0 = const()[name = string("op_1028_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1028_dilations_0 = const()[name = string("op_1028_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1028_groups_0 = const()[name = string("op_1028_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69964672))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70783936))))[name = string("layers_4_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1028_cast_fp16 = conv(dilations = var_1028_dilations_0, groups = var_1028_groups_0, pad = var_1028_pad_0, pad_type = var_1028_pad_type_0, strides = var_1028_strides_0, weight = layers_4_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = string("op_1028_cast_fp16")];
            string var_1034_pad_type_0 = const()[name = string("op_1034_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1034_strides_0 = const()[name = string("op_1034_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1034_pad_0 = const()[name = string("op_1034_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1034_dilations_0 = const()[name = string("op_1034_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1034_groups_0 = const()[name = string("op_1034_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70823680))), nonzero_data = tensor<fp16, [19761]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70784064))))[name = string("layers_4_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1034_cast_fp16 = conv(dilations = var_1034_dilations_0, groups = var_1034_groups_0, pad = var_1034_pad_0, pad_type = var_1034_pad_type_0, strides = var_1034_strides_0, weight = layers_4_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_17_cast_fp16)[name = string("op_1034_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_9_cast_fp16 = add(x = var_1028_cast_fp16, y = var_1034_cast_fp16)[name = string("key_9_cast_fp16")];
            string var_1044_pad_type_0 = const()[name = string("op_1044_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1044_strides_0 = const()[name = string("op_1044_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1044_pad_0 = const()[name = string("op_1044_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1044_dilations_0 = const()[name = string("op_1044_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1044_groups_0 = const()[name = string("op_1044_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71028544))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71847808))))[name = string("layers_4_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_4_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71847936)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1044_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1044_dilations_0, groups = var_1044_groups_0, pad = var_1044_pad_0, pad_type = var_1044_pad_type_0, strides = var_1044_strides_0, weight = layers_4_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = string("op_1044_cast_fp16")];
            string var_1050_pad_type_0 = const()[name = string("op_1050_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1050_strides_0 = const()[name = string("op_1050_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1050_pad_0 = const()[name = string("op_1050_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1050_dilations_0 = const()[name = string("op_1050_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1050_groups_0 = const()[name = string("op_1050_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71883776))), nonzero_data = tensor<fp16, [16547]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71850560))))[name = string("layers_4_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1050_cast_fp16 = conv(dilations = var_1050_dilations_0, groups = var_1050_groups_0, pad = var_1050_pad_0, pad_type = var_1050_pad_type_0, strides = var_1050_strides_0, weight = layers_4_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_17_cast_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_9_cast_fp16 = add(x = var_1044_cast_fp16, y = var_1050_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<int32, [4]> var_1053 = const()[name = string("op_1053"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_9_cast_fp16 = reshape(shape = var_1053, x = query_9_cast_fp16)[name = string("mh_q_9_cast_fp16")];
            fp16 var_1055_to_fp16 = const()[name = string("op_1055_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_1056_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_1055_to_fp16)[name = string("op_1056_cast_fp16")];
            tensor<int32, [4]> var_1057 = const()[name = string("op_1057"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1058_cast_fp16 = reshape(shape = var_1057, x = key_9_cast_fp16)[name = string("op_1058_cast_fp16")];
            bool mh_w_9_transpose_x_0 = const()[name = string("mh_w_9_transpose_x_0"), val = bool(true)];
            bool mh_w_9_transpose_y_0 = const()[name = string("mh_w_9_transpose_y_0"), val = 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_1056_cast_fp16, y = var_1058_cast_fp16)[name = string("mh_w_9_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_1061_cast_fp16 = softmax(axis = var_972, x = mh_w_9_cast_fp16)[name = string("op_1061_cast_fp16")];
            tensor<int32, [4]> var_1062 = const()[name = 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 = value_9_cast_fp16)[name = string("op_1063_cast_fp16")];
            bool attn_9_transpose_x_0 = const()[name = string("attn_9_transpose_x_0"), val = bool(false)];
            bool attn_9_transpose_y_0 = const()[name = string("attn_9_transpose_y_0"), val = 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_1063_cast_fp16, y = var_1061_cast_fp16)[name = string("attn_9_cast_fp16")];
            tensor<int32, [4]> var_1066 = const()[name = string("op_1066"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_33_cast_fp16 = reshape(shape = var_1066, x = attn_9_cast_fp16)[name = string("input_33_cast_fp16")];
            string var_1076_pad_type_0 = const()[name = string("op_1076_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1076_strides_0 = const()[name = string("op_1076_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1076_pad_0 = const()[name = string("op_1076_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1076_dilations_0 = const()[name = string("op_1076_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1076_groups_0 = const()[name = string("op_1076_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72088640))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72907904))))[name = string("layers_4_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_4_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_4_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72908032)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1076_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1076_dilations_0, groups = var_1076_groups_0, pad = var_1076_pad_0, pad_type = var_1076_pad_type_0, strides = var_1076_strides_0, weight = layers_4_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = string("op_1076_cast_fp16")];
            string var_1082_pad_type_0 = const()[name = string("op_1082_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1082_strides_0 = const()[name = string("op_1082_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1082_pad_0 = const()[name = string("op_1082_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1082_dilations_0 = const()[name = string("op_1082_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1082_groups_0 = const()[name = string("op_1082_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72933952))), nonzero_data = tensor<fp16, [11593]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72910656))))[name = string("layers_4_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1082_cast_fp16 = conv(dilations = var_1082_dilations_0, groups = var_1082_groups_0, pad = var_1082_pad_0, pad_type = var_1082_pad_type_0, strides = var_1082_strides_0, weight = layers_4_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_33_cast_fp16)[name = string("op_1082_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_19_cast_fp16 = add(x = var_1076_cast_fp16, y = var_1082_cast_fp16)[name = 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 = string("inputs_19_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1093_to_fp16 = const()[name = string("op_1093_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1093_to_fp16, x = inputs_19_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [1280]> input_35_gamma_0_to_fp16 = const()[name = string("input_35_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73138816)))];
            tensor<fp16, [1280]> input_35_beta_0_to_fp16 = const()[name = string("input_35_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73141440)))];
            fp16 input_35_epsilon_0_to_fp16 = const()[name = string("input_35_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            string var_1111_pad_type_0 = const()[name = string("op_1111_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1111_strides_0 = const()[name = string("op_1111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1111_pad_0 = const()[name = string("op_1111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1111_dilations_0 = const()[name = string("op_1111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1111_groups_0 = const()[name = string("op_1111_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_4_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73144064))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76420928))))[name = string("layers_4_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_4_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_4_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76421056)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_1111_cast_fp16 = conv(bias = layers_4_fc1_inlier_module_bias_to_fp16, dilations = var_1111_dilations_0, groups = var_1111_groups_0, pad = var_1111_pad_0, pad_type = var_1111_pad_type_0, strides = var_1111_strides_0, weight = layers_4_fc1_inlier_module_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("op_1111_cast_fp16")];
            string var_1117_pad_type_0 = const()[name = string("op_1117_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1117_strides_0 = const()[name = string("op_1117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1117_pad_0 = const()[name = string("op_1117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1117_dilations_0 = const()[name = string("op_1117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1117_groups_0 = const()[name = string("op_1117_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_4_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76440192))), nonzero_data = tensor<fp16, [4361]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76431360))))[name = string("layers_4_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_1117_cast_fp16 = conv(dilations = var_1117_dilations_0, groups = var_1117_groups_0, pad = var_1117_pad_0, pad_type = var_1117_pad_type_0, strides = var_1117_strides_0, weight = layers_4_fc1_outlier_module_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_37_cast_fp16 = add(x = var_1111_cast_fp16, y = var_1117_cast_fp16)[name = string("input_37_cast_fp16")];
            string input_39_mode_0 = const()[name = string("input_39_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = string("input_39_cast_fp16")];
            string var_1128_pad_type_0 = const()[name = string("op_1128_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1128_strides_0 = const()[name = string("op_1128_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1128_pad_0 = const()[name = string("op_1128_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1128_dilations_0 = const()[name = string("op_1128_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1128_groups_0 = const()[name = string("op_1128_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_4_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77259456))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80536320))))[name = string("layers_4_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_4_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_4_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80536448)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1128_cast_fp16 = conv(bias = layers_4_fc2_inlier_module_bias_to_fp16, dilations = var_1128_dilations_0, groups = var_1128_groups_0, pad = var_1128_pad_0, pad_type = var_1128_pad_type_0, strides = var_1128_strides_0, weight = layers_4_fc2_inlier_module_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = string("op_1128_cast_fp16")];
            string var_1134_pad_type_0 = const()[name = string("op_1134_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1134_strides_0 = const()[name = string("op_1134_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1134_pad_0 = const()[name = string("op_1134_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1134_dilations_0 = const()[name = string("op_1134_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1134_groups_0 = const()[name = string("op_1134_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_4_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80778304))), nonzero_data = tensor<fp16, [119569]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80539072))))[name = string("layers_4_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1134_cast_fp16 = conv(dilations = var_1134_dilations_0, groups = var_1134_groups_0, pad = var_1134_pad_0, pad_type = var_1134_pad_type_0, strides = var_1134_strides_0, weight = layers_4_fc2_outlier_module_weight_to_fp16_sparsified, x = input_39_cast_fp16)[name = string("op_1134_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_13_cast_fp16 = add(x = var_1128_cast_fp16, y = var_1134_cast_fp16)[name = 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 = string("inputs_21_cast_fp16")];
            int32 var_1144 = const()[name = string("op_1144"), val = int32(3)];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1163_to_fp16 = const()[name = string("op_1163_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1163_to_fp16, x = inputs_21_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [1280]> obj_21_gamma_0_to_fp16 = const()[name = string("obj_21_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81597568)))];
            tensor<fp16, [1280]> obj_21_beta_0_to_fp16 = const()[name = string("obj_21_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81600192)))];
            fp16 obj_21_epsilon_0_to_fp16 = const()[name = string("obj_21_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = string("obj_21_cast_fp16")];
            string var_1185_pad_type_0 = const()[name = string("op_1185_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1185_strides_0 = const()[name = string("op_1185_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1185_pad_0 = const()[name = string("op_1185_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1185_dilations_0 = const()[name = string("op_1185_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1185_groups_0 = const()[name = string("op_1185_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81602816))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82422080))))[name = string("layers_5_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_5_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82422208)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1185_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1185_dilations_0, groups = var_1185_groups_0, pad = var_1185_pad_0, pad_type = var_1185_pad_type_0, strides = var_1185_strides_0, weight = layers_5_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = string("op_1185_cast_fp16")];
            string var_1191_pad_type_0 = const()[name = string("op_1191_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1191_strides_0 = const()[name = string("op_1191_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1191_pad_0 = const()[name = string("op_1191_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1191_dilations_0 = const()[name = string("op_1191_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1191_groups_0 = const()[name = string("op_1191_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82473856))), nonzero_data = tensor<fp16, [24480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82424832))))[name = string("layers_5_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1191_cast_fp16 = conv(dilations = var_1191_dilations_0, groups = var_1191_groups_0, pad = var_1191_pad_0, pad_type = var_1191_pad_type_0, strides = var_1191_strides_0, weight = layers_5_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_21_cast_fp16)[name = string("op_1191_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_11_cast_fp16 = add(x = var_1185_cast_fp16, y = var_1191_cast_fp16)[name = string("query_11_cast_fp16")];
            string var_1200_pad_type_0 = const()[name = string("op_1200_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1200_strides_0 = const()[name = string("op_1200_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1200_pad_0 = const()[name = string("op_1200_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1200_dilations_0 = const()[name = string("op_1200_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1200_groups_0 = const()[name = string("op_1200_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82678720))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83497984))))[name = string("layers_5_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1200_cast_fp16 = conv(dilations = var_1200_dilations_0, groups = var_1200_groups_0, pad = var_1200_pad_0, pad_type = var_1200_pad_type_0, strides = var_1200_strides_0, weight = layers_5_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = string("op_1200_cast_fp16")];
            string var_1206_pad_type_0 = const()[name = string("op_1206_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1206_strides_0 = const()[name = string("op_1206_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1206_pad_0 = const()[name = string("op_1206_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1206_dilations_0 = const()[name = string("op_1206_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1206_groups_0 = const()[name = string("op_1206_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83533248))), nonzero_data = tensor<fp16, [17508]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83498112))))[name = string("layers_5_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1206_cast_fp16 = conv(dilations = var_1206_dilations_0, groups = var_1206_groups_0, pad = var_1206_pad_0, pad_type = var_1206_pad_type_0, strides = var_1206_strides_0, weight = layers_5_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_21_cast_fp16)[name = string("op_1206_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_11_cast_fp16 = add(x = var_1200_cast_fp16, y = var_1206_cast_fp16)[name = string("key_11_cast_fp16")];
            string var_1216_pad_type_0 = const()[name = string("op_1216_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1216_strides_0 = const()[name = string("op_1216_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1216_pad_0 = const()[name = string("op_1216_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1216_dilations_0 = const()[name = string("op_1216_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1216_groups_0 = const()[name = string("op_1216_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83738112))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84557376))))[name = string("layers_5_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_5_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84557504)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1216_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1216_dilations_0, groups = var_1216_groups_0, pad = var_1216_pad_0, pad_type = var_1216_pad_type_0, strides = var_1216_strides_0, weight = layers_5_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = string("op_1216_cast_fp16")];
            string var_1222_pad_type_0 = const()[name = string("op_1222_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1222_strides_0 = const()[name = string("op_1222_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1222_pad_0 = const()[name = string("op_1222_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1222_dilations_0 = const()[name = string("op_1222_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1222_groups_0 = const()[name = string("op_1222_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84586624))), nonzero_data = tensor<fp16, [13195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84560128))))[name = string("layers_5_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1222_cast_fp16 = conv(dilations = var_1222_dilations_0, groups = var_1222_groups_0, pad = var_1222_pad_0, pad_type = var_1222_pad_type_0, strides = var_1222_strides_0, weight = layers_5_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_21_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_11_cast_fp16 = add(x = var_1216_cast_fp16, y = var_1222_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<int32, [4]> var_1225 = const()[name = string("op_1225"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_11_cast_fp16 = reshape(shape = var_1225, x = query_11_cast_fp16)[name = string("mh_q_11_cast_fp16")];
            fp16 var_1227_to_fp16 = const()[name = string("op_1227_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_1228_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1227_to_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> var_1229 = const()[name = string("op_1229"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1230_cast_fp16 = reshape(shape = var_1229, x = key_11_cast_fp16)[name = string("op_1230_cast_fp16")];
            bool mh_w_11_transpose_x_0 = const()[name = string("mh_w_11_transpose_x_0"), val = bool(true)];
            bool mh_w_11_transpose_y_0 = const()[name = string("mh_w_11_transpose_y_0"), val = 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_1228_cast_fp16, y = var_1230_cast_fp16)[name = string("mh_w_11_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_1233_cast_fp16 = softmax(axis = var_1144, x = mh_w_11_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [4]> var_1234 = const()[name = string("op_1234"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1235_cast_fp16 = reshape(shape = var_1234, x = value_11_cast_fp16)[name = string("op_1235_cast_fp16")];
            bool attn_11_transpose_x_0 = const()[name = string("attn_11_transpose_x_0"), val = bool(false)];
            bool attn_11_transpose_y_0 = const()[name = string("attn_11_transpose_y_0"), val = 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_1235_cast_fp16, y = var_1233_cast_fp16)[name = string("attn_11_cast_fp16")];
            tensor<int32, [4]> var_1238 = const()[name = string("op_1238"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_41_cast_fp16 = reshape(shape = var_1238, x = attn_11_cast_fp16)[name = string("input_41_cast_fp16")];
            string var_1248_pad_type_0 = const()[name = string("op_1248_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1248_strides_0 = const()[name = string("op_1248_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1248_pad_0 = const()[name = string("op_1248_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1248_dilations_0 = const()[name = string("op_1248_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1248_groups_0 = const()[name = string("op_1248_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84791488))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85610752))))[name = string("layers_5_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_5_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85610880)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1248_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1248_dilations_0, groups = var_1248_groups_0, pad = var_1248_pad_0, pad_type = var_1248_pad_type_0, strides = var_1248_strides_0, weight = layers_5_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = string("op_1248_cast_fp16")];
            string var_1254_pad_type_0 = const()[name = string("op_1254_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1254_strides_0 = const()[name = string("op_1254_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1254_pad_0 = const()[name = string("op_1254_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1254_dilations_0 = const()[name = string("op_1254_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1254_groups_0 = const()[name = string("op_1254_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85634944))), nonzero_data = tensor<fp16, [10679]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85613504))))[name = string("layers_5_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1254_cast_fp16 = conv(dilations = var_1254_dilations_0, groups = var_1254_groups_0, pad = var_1254_pad_0, pad_type = var_1254_pad_type_0, strides = var_1254_strides_0, weight = layers_5_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_41_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_23_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1254_cast_fp16)[name = 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 = string("inputs_23_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1265_to_fp16 = const()[name = string("op_1265_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_1265_to_fp16, x = inputs_23_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [1280]> input_43_gamma_0_to_fp16 = const()[name = string("input_43_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85839808)))];
            tensor<fp16, [1280]> input_43_beta_0_to_fp16 = const()[name = string("input_43_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85842432)))];
            fp16 input_43_epsilon_0_to_fp16 = const()[name = string("input_43_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            string var_1283_pad_type_0 = const()[name = string("op_1283_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1283_strides_0 = const()[name = string("op_1283_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1283_pad_0 = const()[name = string("op_1283_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1283_dilations_0 = const()[name = string("op_1283_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1283_groups_0 = const()[name = string("op_1283_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_5_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85845056))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89121920))))[name = string("layers_5_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_5_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_5_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89122048)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_1283_cast_fp16 = conv(bias = layers_5_fc1_inlier_module_bias_to_fp16, dilations = var_1283_dilations_0, groups = var_1283_groups_0, pad = var_1283_pad_0, pad_type = var_1283_pad_type_0, strides = var_1283_strides_0, weight = layers_5_fc1_inlier_module_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = string("op_1283_cast_fp16")];
            string var_1289_pad_type_0 = const()[name = string("op_1289_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1289_strides_0 = const()[name = string("op_1289_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1289_pad_0 = const()[name = string("op_1289_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1289_dilations_0 = const()[name = string("op_1289_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1289_groups_0 = const()[name = string("op_1289_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_5_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89166464))), nonzero_data = tensor<fp16, [17005]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89132352))))[name = string("layers_5_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_1289_cast_fp16 = conv(dilations = var_1289_dilations_0, groups = var_1289_groups_0, pad = var_1289_pad_0, pad_type = var_1289_pad_type_0, strides = var_1289_strides_0, weight = layers_5_fc1_outlier_module_weight_to_fp16_sparsified, x = input_43_cast_fp16)[name = string("op_1289_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_45_cast_fp16 = add(x = var_1283_cast_fp16, y = var_1289_cast_fp16)[name = string("input_45_cast_fp16")];
            string input_47_mode_0 = const()[name = string("input_47_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            string var_1300_pad_type_0 = const()[name = string("op_1300_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1300_strides_0 = const()[name = string("op_1300_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1300_pad_0 = const()[name = string("op_1300_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1300_dilations_0 = const()[name = string("op_1300_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1300_groups_0 = const()[name = string("op_1300_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_5_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89985728))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93262592))))[name = string("layers_5_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_5_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_5_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93262720)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1300_cast_fp16 = conv(bias = layers_5_fc2_inlier_module_bias_to_fp16, dilations = var_1300_dilations_0, groups = var_1300_groups_0, pad = var_1300_pad_0, pad_type = var_1300_pad_type_0, strides = var_1300_strides_0, weight = layers_5_fc2_inlier_module_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = string("op_1300_cast_fp16")];
            string var_1306_pad_type_0 = const()[name = string("op_1306_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1306_strides_0 = const()[name = string("op_1306_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1306_pad_0 = const()[name = string("op_1306_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1306_dilations_0 = const()[name = string("op_1306_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1306_groups_0 = const()[name = string("op_1306_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_5_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93508032))), nonzero_data = tensor<fp16, [121288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93265344))))[name = string("layers_5_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1306_cast_fp16 = conv(dilations = var_1306_dilations_0, groups = var_1306_groups_0, pad = var_1306_pad_0, pad_type = var_1306_pad_type_0, strides = var_1306_strides_0, weight = layers_5_fc2_outlier_module_weight_to_fp16_sparsified, x = input_47_cast_fp16)[name = string("op_1306_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_15_cast_fp16 = add(x = var_1300_cast_fp16, y = var_1306_cast_fp16)[name = 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 = string("inputs_25_cast_fp16")];
            int32 var_1316 = const()[name = string("op_1316"), val = int32(3)];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1335_to_fp16 = const()[name = string("op_1335_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_1335_to_fp16, x = inputs_25_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [1280]> obj_25_gamma_0_to_fp16 = const()[name = string("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94327296)))];
            tensor<fp16, [1280]> obj_25_beta_0_to_fp16 = const()[name = string("obj_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94329920)))];
            fp16 obj_25_epsilon_0_to_fp16 = const()[name = string("obj_25_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = string("obj_25_cast_fp16")];
            string var_1357_pad_type_0 = const()[name = string("op_1357_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1357_strides_0 = const()[name = string("op_1357_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1357_pad_0 = const()[name = string("op_1357_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1357_dilations_0 = const()[name = string("op_1357_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1357_groups_0 = const()[name = string("op_1357_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94332544))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95151808))))[name = string("layers_6_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_6_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95151936)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1357_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1357_dilations_0, groups = var_1357_groups_0, pad = var_1357_pad_0, pad_type = var_1357_pad_type_0, strides = var_1357_strides_0, weight = layers_6_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = string("op_1357_cast_fp16")];
            string var_1363_pad_type_0 = const()[name = string("op_1363_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1363_strides_0 = const()[name = string("op_1363_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1363_pad_0 = const()[name = string("op_1363_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1363_dilations_0 = const()[name = string("op_1363_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1363_groups_0 = const()[name = string("op_1363_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95204224))), nonzero_data = tensor<fp16, [24786]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95154560))))[name = string("layers_6_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1363_cast_fp16 = conv(dilations = var_1363_dilations_0, groups = var_1363_groups_0, pad = var_1363_pad_0, pad_type = var_1363_pad_type_0, strides = var_1363_strides_0, weight = layers_6_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_25_cast_fp16)[name = string("op_1363_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_13_cast_fp16 = add(x = var_1357_cast_fp16, y = var_1363_cast_fp16)[name = string("query_13_cast_fp16")];
            string var_1372_pad_type_0 = const()[name = string("op_1372_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1372_strides_0 = const()[name = string("op_1372_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1372_pad_0 = const()[name = string("op_1372_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1372_dilations_0 = const()[name = string("op_1372_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1372_groups_0 = const()[name = string("op_1372_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95409088))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96228352))))[name = string("layers_6_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1372_cast_fp16 = conv(dilations = var_1372_dilations_0, groups = var_1372_groups_0, pad = var_1372_pad_0, pad_type = var_1372_pad_type_0, strides = var_1372_strides_0, weight = layers_6_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = string("op_1372_cast_fp16")];
            string var_1378_pad_type_0 = const()[name = string("op_1378_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1378_strides_0 = const()[name = string("op_1378_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1378_pad_0 = const()[name = string("op_1378_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1378_dilations_0 = const()[name = string("op_1378_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1378_groups_0 = const()[name = string("op_1378_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96263168))), nonzero_data = tensor<fp16, [17281]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96228480))))[name = string("layers_6_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1378_cast_fp16 = conv(dilations = var_1378_dilations_0, groups = var_1378_groups_0, pad = var_1378_pad_0, pad_type = var_1378_pad_type_0, strides = var_1378_strides_0, weight = layers_6_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_25_cast_fp16)[name = string("op_1378_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_13_cast_fp16 = add(x = var_1372_cast_fp16, y = var_1378_cast_fp16)[name = string("key_13_cast_fp16")];
            string var_1388_pad_type_0 = const()[name = string("op_1388_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1388_strides_0 = const()[name = string("op_1388_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1388_pad_0 = const()[name = string("op_1388_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1388_dilations_0 = const()[name = string("op_1388_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1388_groups_0 = const()[name = string("op_1388_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96468032))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97287296))))[name = string("layers_6_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_6_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97287424)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1388_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1388_dilations_0, groups = var_1388_groups_0, pad = var_1388_pad_0, pad_type = var_1388_pad_type_0, strides = var_1388_strides_0, weight = layers_6_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = string("op_1388_cast_fp16")];
            string var_1394_pad_type_0 = const()[name = string("op_1394_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1394_strides_0 = const()[name = string("op_1394_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1394_pad_0 = const()[name = string("op_1394_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1394_dilations_0 = const()[name = string("op_1394_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1394_groups_0 = const()[name = string("op_1394_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97313920))), nonzero_data = tensor<fp16, [11874]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97290048))))[name = string("layers_6_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1394_cast_fp16 = conv(dilations = var_1394_dilations_0, groups = var_1394_groups_0, pad = var_1394_pad_0, pad_type = var_1394_pad_type_0, strides = var_1394_strides_0, weight = layers_6_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_25_cast_fp16)[name = string("op_1394_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_13_cast_fp16 = add(x = var_1388_cast_fp16, y = var_1394_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<int32, [4]> var_1397 = const()[name = string("op_1397"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_13_cast_fp16 = reshape(shape = var_1397, x = query_13_cast_fp16)[name = string("mh_q_13_cast_fp16")];
            fp16 var_1399_to_fp16 = const()[name = string("op_1399_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_1400_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1399_to_fp16)[name = string("op_1400_cast_fp16")];
            tensor<int32, [4]> var_1401 = const()[name = string("op_1401"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1402_cast_fp16 = reshape(shape = var_1401, x = key_13_cast_fp16)[name = string("op_1402_cast_fp16")];
            bool mh_w_13_transpose_x_0 = const()[name = string("mh_w_13_transpose_x_0"), val = bool(true)];
            bool mh_w_13_transpose_y_0 = const()[name = string("mh_w_13_transpose_y_0"), val = 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_1400_cast_fp16, y = var_1402_cast_fp16)[name = string("mh_w_13_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_1405_cast_fp16 = softmax(axis = var_1316, x = mh_w_13_cast_fp16)[name = string("op_1405_cast_fp16")];
            tensor<int32, [4]> var_1406 = const()[name = string("op_1406"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1407_cast_fp16 = reshape(shape = var_1406, x = value_13_cast_fp16)[name = string("op_1407_cast_fp16")];
            bool attn_13_transpose_x_0 = const()[name = string("attn_13_transpose_x_0"), val = bool(false)];
            bool attn_13_transpose_y_0 = const()[name = string("attn_13_transpose_y_0"), val = 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_1407_cast_fp16, y = var_1405_cast_fp16)[name = string("attn_13_cast_fp16")];
            tensor<int32, [4]> var_1410 = const()[name = string("op_1410"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_49_cast_fp16 = reshape(shape = var_1410, x = attn_13_cast_fp16)[name = string("input_49_cast_fp16")];
            string var_1420_pad_type_0 = const()[name = string("op_1420_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1420_strides_0 = const()[name = string("op_1420_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1420_pad_0 = const()[name = string("op_1420_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1420_dilations_0 = const()[name = string("op_1420_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1420_groups_0 = const()[name = string("op_1420_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97518784))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98338048))))[name = string("layers_6_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_6_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_6_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98338176)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1420_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1420_dilations_0, groups = var_1420_groups_0, pad = var_1420_pad_0, pad_type = var_1420_pad_type_0, strides = var_1420_strides_0, weight = layers_6_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = string("op_1420_cast_fp16")];
            string var_1426_pad_type_0 = const()[name = string("op_1426_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1426_strides_0 = const()[name = string("op_1426_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1426_pad_0 = const()[name = string("op_1426_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1426_dilations_0 = const()[name = string("op_1426_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1426_groups_0 = const()[name = string("op_1426_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98362688))), nonzero_data = tensor<fp16, [10911]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98340800))))[name = string("layers_6_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1426_cast_fp16 = conv(dilations = var_1426_dilations_0, groups = var_1426_groups_0, pad = var_1426_pad_0, pad_type = var_1426_pad_type_0, strides = var_1426_strides_0, weight = layers_6_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_49_cast_fp16)[name = string("op_1426_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_27_cast_fp16 = add(x = var_1420_cast_fp16, y = var_1426_cast_fp16)[name = 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 = string("inputs_27_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1437_to_fp16 = const()[name = string("op_1437_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1437_to_fp16, x = inputs_27_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [1280]> input_51_gamma_0_to_fp16 = const()[name = string("input_51_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98567552)))];
            tensor<fp16, [1280]> input_51_beta_0_to_fp16 = const()[name = string("input_51_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98570176)))];
            fp16 input_51_epsilon_0_to_fp16 = const()[name = string("input_51_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            string var_1455_pad_type_0 = const()[name = string("op_1455_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1455_strides_0 = const()[name = string("op_1455_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1455_pad_0 = const()[name = string("op_1455_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1455_dilations_0 = const()[name = string("op_1455_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1455_groups_0 = const()[name = string("op_1455_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_6_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98572800))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101849664))))[name = string("layers_6_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_6_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_6_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101849792)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_1455_cast_fp16 = conv(bias = layers_6_fc1_inlier_module_bias_to_fp16, dilations = var_1455_dilations_0, groups = var_1455_groups_0, pad = var_1455_pad_0, pad_type = var_1455_pad_type_0, strides = var_1455_strides_0, weight = layers_6_fc1_inlier_module_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("op_1455_cast_fp16")];
            string var_1461_pad_type_0 = const()[name = string("op_1461_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1461_strides_0 = const()[name = string("op_1461_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1461_pad_0 = const()[name = string("op_1461_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1461_dilations_0 = const()[name = string("op_1461_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1461_groups_0 = const()[name = string("op_1461_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_6_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101881792))), nonzero_data = tensor<fp16, [10794]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101860096))))[name = string("layers_6_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_1461_cast_fp16 = conv(dilations = var_1461_dilations_0, groups = var_1461_groups_0, pad = var_1461_pad_0, pad_type = var_1461_pad_type_0, strides = var_1461_strides_0, weight = layers_6_fc1_outlier_module_weight_to_fp16_sparsified, x = input_51_cast_fp16)[name = string("op_1461_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_53_cast_fp16 = add(x = var_1455_cast_fp16, y = var_1461_cast_fp16)[name = string("input_53_cast_fp16")];
            string input_55_mode_0 = const()[name = string("input_55_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = input_53_cast_fp16)[name = string("input_55_cast_fp16")];
            string var_1472_pad_type_0 = const()[name = string("op_1472_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1472_strides_0 = const()[name = string("op_1472_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1472_pad_0 = const()[name = string("op_1472_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1472_dilations_0 = const()[name = string("op_1472_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1472_groups_0 = const()[name = string("op_1472_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_6_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102701056))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105977920))))[name = string("layers_6_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_6_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_6_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105978048)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1472_cast_fp16 = conv(bias = layers_6_fc2_inlier_module_bias_to_fp16, dilations = var_1472_dilations_0, groups = var_1472_groups_0, pad = var_1472_pad_0, pad_type = var_1472_pad_type_0, strides = var_1472_strides_0, weight = layers_6_fc2_inlier_module_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = string("op_1472_cast_fp16")];
            string var_1478_pad_type_0 = const()[name = string("op_1478_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1478_strides_0 = const()[name = string("op_1478_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1478_pad_0 = const()[name = string("op_1478_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1478_dilations_0 = const()[name = string("op_1478_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1478_groups_0 = const()[name = string("op_1478_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_6_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106227584))), nonzero_data = tensor<fp16, [123417]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105980672))))[name = string("layers_6_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1478_cast_fp16 = conv(dilations = var_1478_dilations_0, groups = var_1478_groups_0, pad = var_1478_pad_0, pad_type = var_1478_pad_type_0, strides = var_1478_strides_0, weight = layers_6_fc2_outlier_module_weight_to_fp16_sparsified, x = input_55_cast_fp16)[name = string("op_1478_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_17_cast_fp16 = add(x = var_1472_cast_fp16, y = var_1478_cast_fp16)[name = 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 = string("inputs_29_cast_fp16")];
            int32 var_1488 = const()[name = string("op_1488"), val = int32(3)];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1507_to_fp16 = const()[name = string("op_1507_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1507_to_fp16, x = inputs_29_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [1280]> obj_29_gamma_0_to_fp16 = const()[name = string("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107046848)))];
            tensor<fp16, [1280]> obj_29_beta_0_to_fp16 = const()[name = string("obj_29_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107049472)))];
            fp16 obj_29_epsilon_0_to_fp16 = const()[name = string("obj_29_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = string("obj_29_cast_fp16")];
            string var_1529_pad_type_0 = const()[name = string("op_1529_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1529_strides_0 = const()[name = string("op_1529_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1529_pad_0 = const()[name = string("op_1529_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1529_dilations_0 = const()[name = string("op_1529_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1529_groups_0 = const()[name = string("op_1529_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107052096))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107871360))))[name = string("layers_7_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_7_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107871488)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1529_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1529_dilations_0, groups = var_1529_groups_0, pad = var_1529_pad_0, pad_type = var_1529_pad_type_0, strides = var_1529_strides_0, weight = layers_7_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = string("op_1529_cast_fp16")];
            string var_1535_pad_type_0 = const()[name = string("op_1535_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1535_strides_0 = const()[name = string("op_1535_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1535_pad_0 = const()[name = string("op_1535_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1535_dilations_0 = const()[name = string("op_1535_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1535_groups_0 = const()[name = string("op_1535_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107923264))), nonzero_data = tensor<fp16, [24530]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107874112))))[name = string("layers_7_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1535_cast_fp16 = conv(dilations = var_1535_dilations_0, groups = var_1535_groups_0, pad = var_1535_pad_0, pad_type = var_1535_pad_type_0, strides = var_1535_strides_0, weight = layers_7_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = string("op_1535_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_15_cast_fp16 = add(x = var_1529_cast_fp16, y = var_1535_cast_fp16)[name = string("query_15_cast_fp16")];
            string var_1544_pad_type_0 = const()[name = string("op_1544_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1544_strides_0 = const()[name = string("op_1544_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1544_pad_0 = const()[name = string("op_1544_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1544_dilations_0 = const()[name = string("op_1544_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1544_groups_0 = const()[name = string("op_1544_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108128128))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108947392))))[name = string("layers_7_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1544_cast_fp16 = conv(dilations = var_1544_dilations_0, groups = var_1544_groups_0, pad = var_1544_pad_0, pad_type = var_1544_pad_type_0, strides = var_1544_strides_0, weight = layers_7_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = string("op_1544_cast_fp16")];
            string var_1550_pad_type_0 = const()[name = string("op_1550_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1550_strides_0 = const()[name = string("op_1550_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1550_pad_0 = const()[name = string("op_1550_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1550_dilations_0 = const()[name = string("op_1550_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1550_groups_0 = const()[name = string("op_1550_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108983872))), nonzero_data = tensor<fp16, [18142]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108947520))))[name = string("layers_7_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1550_cast_fp16 = conv(dilations = var_1550_dilations_0, groups = var_1550_groups_0, pad = var_1550_pad_0, pad_type = var_1550_pad_type_0, strides = var_1550_strides_0, weight = layers_7_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_15_cast_fp16 = add(x = var_1544_cast_fp16, y = var_1550_cast_fp16)[name = string("key_15_cast_fp16")];
            string var_1560_pad_type_0 = const()[name = string("op_1560_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1560_strides_0 = const()[name = string("op_1560_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1560_pad_0 = const()[name = string("op_1560_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1560_dilations_0 = const()[name = string("op_1560_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1560_groups_0 = const()[name = string("op_1560_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109188736))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110008000))))[name = string("layers_7_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_7_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110008128)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1560_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1560_dilations_0, groups = var_1560_groups_0, pad = var_1560_pad_0, pad_type = var_1560_pad_type_0, strides = var_1560_strides_0, weight = layers_7_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = string("op_1560_cast_fp16")];
            string var_1566_pad_type_0 = const()[name = string("op_1566_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1566_strides_0 = const()[name = string("op_1566_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1566_pad_0 = const()[name = string("op_1566_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1566_dilations_0 = const()[name = string("op_1566_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1566_groups_0 = const()[name = string("op_1566_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110034304))), nonzero_data = tensor<fp16, [11740]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110010752))))[name = string("layers_7_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1566_cast_fp16 = conv(dilations = var_1566_dilations_0, groups = var_1566_groups_0, pad = var_1566_pad_0, pad_type = var_1566_pad_type_0, strides = var_1566_strides_0, weight = layers_7_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = string("op_1566_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_15_cast_fp16 = add(x = var_1560_cast_fp16, y = var_1566_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<int32, [4]> var_1569 = const()[name = string("op_1569"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_15_cast_fp16 = reshape(shape = var_1569, x = query_15_cast_fp16)[name = string("mh_q_15_cast_fp16")];
            fp16 var_1571_to_fp16 = const()[name = string("op_1571_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_1572_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_1571_to_fp16)[name = string("op_1572_cast_fp16")];
            tensor<int32, [4]> var_1573 = const()[name = string("op_1573"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1574_cast_fp16 = reshape(shape = var_1573, x = key_15_cast_fp16)[name = string("op_1574_cast_fp16")];
            bool mh_w_15_transpose_x_0 = const()[name = string("mh_w_15_transpose_x_0"), val = bool(true)];
            bool mh_w_15_transpose_y_0 = const()[name = string("mh_w_15_transpose_y_0"), val = 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_1572_cast_fp16, y = var_1574_cast_fp16)[name = string("mh_w_15_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_1577_cast_fp16 = softmax(axis = var_1488, x = mh_w_15_cast_fp16)[name = string("op_1577_cast_fp16")];
            tensor<int32, [4]> var_1578 = const()[name = string("op_1578"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1579_cast_fp16 = reshape(shape = var_1578, x = value_15_cast_fp16)[name = string("op_1579_cast_fp16")];
            bool attn_15_transpose_x_0 = const()[name = string("attn_15_transpose_x_0"), val = bool(false)];
            bool attn_15_transpose_y_0 = const()[name = string("attn_15_transpose_y_0"), val = 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_1579_cast_fp16, y = var_1577_cast_fp16)[name = string("attn_15_cast_fp16")];
            tensor<int32, [4]> var_1582 = const()[name = string("op_1582"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_57_cast_fp16 = reshape(shape = var_1582, x = attn_15_cast_fp16)[name = string("input_57_cast_fp16")];
            string var_1592_pad_type_0 = const()[name = string("op_1592_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1592_strides_0 = const()[name = string("op_1592_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1592_pad_0 = const()[name = string("op_1592_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1592_dilations_0 = const()[name = string("op_1592_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1592_groups_0 = const()[name = string("op_1592_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110239168))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111058432))))[name = string("layers_7_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_7_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_7_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111058560)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1592_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1592_dilations_0, groups = var_1592_groups_0, pad = var_1592_pad_0, pad_type = var_1592_pad_type_0, strides = var_1592_strides_0, weight = layers_7_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = string("op_1592_cast_fp16")];
            string var_1598_pad_type_0 = const()[name = string("op_1598_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1598_strides_0 = const()[name = string("op_1598_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1598_pad_0 = const()[name = string("op_1598_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1598_dilations_0 = const()[name = string("op_1598_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1598_groups_0 = const()[name = string("op_1598_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111080064))), nonzero_data = tensor<fp16, [9395]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111061184))))[name = string("layers_7_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1598_cast_fp16 = conv(dilations = var_1598_dilations_0, groups = var_1598_groups_0, pad = var_1598_pad_0, pad_type = var_1598_pad_type_0, strides = var_1598_strides_0, weight = layers_7_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_57_cast_fp16)[name = string("op_1598_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_31_cast_fp16 = add(x = var_1592_cast_fp16, y = var_1598_cast_fp16)[name = 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 = string("inputs_31_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1609_to_fp16 = const()[name = string("op_1609_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1609_to_fp16, x = inputs_31_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [1280]> input_59_gamma_0_to_fp16 = const()[name = string("input_59_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111284928)))];
            tensor<fp16, [1280]> input_59_beta_0_to_fp16 = const()[name = string("input_59_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111287552)))];
            fp16 input_59_epsilon_0_to_fp16 = const()[name = string("input_59_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            string var_1627_pad_type_0 = const()[name = string("op_1627_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1627_strides_0 = const()[name = string("op_1627_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1627_pad_0 = const()[name = string("op_1627_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1627_dilations_0 = const()[name = string("op_1627_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1627_groups_0 = const()[name = string("op_1627_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_7_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111290176))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114567040))))[name = string("layers_7_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_7_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_7_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114567168)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_1627_cast_fp16 = conv(bias = layers_7_fc1_inlier_module_bias_to_fp16, dilations = var_1627_dilations_0, groups = var_1627_groups_0, pad = var_1627_pad_0, pad_type = var_1627_pad_type_0, strides = var_1627_strides_0, weight = layers_7_fc1_inlier_module_weight_to_fp16_palettized, x = input_59_cast_fp16)[name = string("op_1627_cast_fp16")];
            string var_1633_pad_type_0 = const()[name = string("op_1633_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1633_strides_0 = const()[name = string("op_1633_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1633_pad_0 = const()[name = string("op_1633_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1633_dilations_0 = const()[name = string("op_1633_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1633_groups_0 = const()[name = string("op_1633_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_7_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114597440))), nonzero_data = tensor<fp16, [9922]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114577472))))[name = string("layers_7_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_1633_cast_fp16 = conv(dilations = var_1633_dilations_0, groups = var_1633_groups_0, pad = var_1633_pad_0, pad_type = var_1633_pad_type_0, strides = var_1633_strides_0, weight = layers_7_fc1_outlier_module_weight_to_fp16_sparsified, x = input_59_cast_fp16)[name = string("op_1633_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_61_cast_fp16 = add(x = var_1627_cast_fp16, y = var_1633_cast_fp16)[name = string("input_61_cast_fp16")];
            string input_63_mode_0 = const()[name = string("input_63_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_63_cast_fp16 = gelu(mode = input_63_mode_0, x = input_61_cast_fp16)[name = string("input_63_cast_fp16")];
            string var_1644_pad_type_0 = const()[name = string("op_1644_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1644_strides_0 = const()[name = string("op_1644_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1644_pad_0 = const()[name = string("op_1644_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1644_dilations_0 = const()[name = string("op_1644_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1644_groups_0 = const()[name = string("op_1644_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_7_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115416704))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118693568))))[name = string("layers_7_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_7_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_7_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118693696)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1644_cast_fp16 = conv(bias = layers_7_fc2_inlier_module_bias_to_fp16, dilations = var_1644_dilations_0, groups = var_1644_groups_0, pad = var_1644_pad_0, pad_type = var_1644_pad_type_0, strides = var_1644_strides_0, weight = layers_7_fc2_inlier_module_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = string("op_1644_cast_fp16")];
            string var_1650_pad_type_0 = const()[name = string("op_1650_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1650_strides_0 = const()[name = string("op_1650_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1650_pad_0 = const()[name = string("op_1650_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1650_dilations_0 = const()[name = string("op_1650_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1650_groups_0 = const()[name = string("op_1650_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_7_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118968000))), nonzero_data = tensor<fp16, [135797]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118696320))))[name = string("layers_7_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1650_cast_fp16 = conv(dilations = var_1650_dilations_0, groups = var_1650_groups_0, pad = var_1650_pad_0, pad_type = var_1650_pad_type_0, strides = var_1650_strides_0, weight = layers_7_fc2_outlier_module_weight_to_fp16_sparsified, x = input_63_cast_fp16)[name = string("op_1650_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_19_cast_fp16 = add(x = var_1644_cast_fp16, y = var_1650_cast_fp16)[name = 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 = string("inputs_33_cast_fp16")];
            int32 var_1660 = const()[name = string("op_1660"), val = int32(3)];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1679_to_fp16 = const()[name = string("op_1679_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1679_to_fp16, x = inputs_33_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [1280]> obj_33_gamma_0_to_fp16 = const()[name = string("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119787264)))];
            tensor<fp16, [1280]> obj_33_beta_0_to_fp16 = const()[name = string("obj_33_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119789888)))];
            fp16 obj_33_epsilon_0_to_fp16 = const()[name = string("obj_33_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = string("obj_33_cast_fp16")];
            string var_1701_pad_type_0 = const()[name = string("op_1701_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1701_strides_0 = const()[name = string("op_1701_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1701_pad_0 = const()[name = string("op_1701_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1701_dilations_0 = const()[name = string("op_1701_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1701_groups_0 = const()[name = string("op_1701_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119792512))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120611776))))[name = string("layers_8_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_8_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120611904)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1701_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1701_dilations_0, groups = var_1701_groups_0, pad = var_1701_pad_0, pad_type = var_1701_pad_type_0, strides = var_1701_strides_0, weight = layers_8_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = string("op_1701_cast_fp16")];
            string var_1707_pad_type_0 = const()[name = string("op_1707_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1707_strides_0 = const()[name = string("op_1707_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1707_pad_0 = const()[name = string("op_1707_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1707_dilations_0 = const()[name = string("op_1707_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1707_groups_0 = const()[name = string("op_1707_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120662848))), nonzero_data = tensor<fp16, [24111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120614528))))[name = string("layers_8_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1707_cast_fp16 = conv(dilations = var_1707_dilations_0, groups = var_1707_groups_0, pad = var_1707_pad_0, pad_type = var_1707_pad_type_0, strides = var_1707_strides_0, weight = layers_8_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_33_cast_fp16)[name = string("op_1707_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_17_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1707_cast_fp16)[name = string("query_17_cast_fp16")];
            string var_1716_pad_type_0 = const()[name = string("op_1716_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1716_strides_0 = const()[name = string("op_1716_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1716_pad_0 = const()[name = string("op_1716_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1716_dilations_0 = const()[name = string("op_1716_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1716_groups_0 = const()[name = string("op_1716_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120867712))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121686976))))[name = string("layers_8_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1716_cast_fp16 = conv(dilations = var_1716_dilations_0, groups = var_1716_groups_0, pad = var_1716_pad_0, pad_type = var_1716_pad_type_0, strides = var_1716_strides_0, weight = layers_8_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = string("op_1716_cast_fp16")];
            string var_1722_pad_type_0 = const()[name = string("op_1722_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1722_strides_0 = const()[name = string("op_1722_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1722_pad_0 = const()[name = string("op_1722_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1722_dilations_0 = const()[name = string("op_1722_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1722_groups_0 = const()[name = string("op_1722_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121721856))), nonzero_data = tensor<fp16, [17344]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121687104))))[name = string("layers_8_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1722_cast_fp16 = conv(dilations = var_1722_dilations_0, groups = var_1722_groups_0, pad = var_1722_pad_0, pad_type = var_1722_pad_type_0, strides = var_1722_strides_0, weight = layers_8_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_33_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_17_cast_fp16 = add(x = var_1716_cast_fp16, y = var_1722_cast_fp16)[name = string("key_17_cast_fp16")];
            string var_1732_pad_type_0 = const()[name = string("op_1732_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1732_strides_0 = const()[name = string("op_1732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1732_pad_0 = const()[name = string("op_1732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1732_dilations_0 = const()[name = string("op_1732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1732_groups_0 = const()[name = string("op_1732_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121926720))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122745984))))[name = string("layers_8_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_8_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122746112)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1732_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1732_dilations_0, groups = var_1732_groups_0, pad = var_1732_pad_0, pad_type = var_1732_pad_type_0, strides = var_1732_strides_0, weight = layers_8_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = string("op_1732_cast_fp16")];
            string var_1738_pad_type_0 = const()[name = string("op_1738_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1738_strides_0 = const()[name = string("op_1738_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1738_pad_0 = const()[name = string("op_1738_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1738_dilations_0 = const()[name = string("op_1738_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1738_groups_0 = const()[name = string("op_1738_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122771968))), nonzero_data = tensor<fp16, [11555]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122748736))))[name = string("layers_8_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1738_cast_fp16 = conv(dilations = var_1738_dilations_0, groups = var_1738_groups_0, pad = var_1738_pad_0, pad_type = var_1738_pad_type_0, strides = var_1738_strides_0, weight = layers_8_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_33_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_17_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1738_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<int32, [4]> var_1741 = const()[name = string("op_1741"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_17_cast_fp16 = reshape(shape = var_1741, x = query_17_cast_fp16)[name = string("mh_q_17_cast_fp16")];
            fp16 var_1743_to_fp16 = const()[name = string("op_1743_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_1744_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1743_to_fp16)[name = string("op_1744_cast_fp16")];
            tensor<int32, [4]> var_1745 = const()[name = string("op_1745"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1746_cast_fp16 = reshape(shape = var_1745, x = key_17_cast_fp16)[name = string("op_1746_cast_fp16")];
            bool mh_w_17_transpose_x_0 = const()[name = string("mh_w_17_transpose_x_0"), val = bool(true)];
            bool mh_w_17_transpose_y_0 = const()[name = string("mh_w_17_transpose_y_0"), val = 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_1744_cast_fp16, y = var_1746_cast_fp16)[name = string("mh_w_17_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_1749_cast_fp16 = softmax(axis = var_1660, x = mh_w_17_cast_fp16)[name = string("op_1749_cast_fp16")];
            tensor<int32, [4]> var_1750 = const()[name = string("op_1750"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1751_cast_fp16 = reshape(shape = var_1750, x = value_17_cast_fp16)[name = string("op_1751_cast_fp16")];
            bool attn_17_transpose_x_0 = const()[name = string("attn_17_transpose_x_0"), val = bool(false)];
            bool attn_17_transpose_y_0 = const()[name = string("attn_17_transpose_y_0"), val = 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_1751_cast_fp16, y = var_1749_cast_fp16)[name = string("attn_17_cast_fp16")];
            tensor<int32, [4]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_65_cast_fp16 = reshape(shape = var_1754, x = attn_17_cast_fp16)[name = string("input_65_cast_fp16")];
            string var_1764_pad_type_0 = const()[name = string("op_1764_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1764_strides_0 = const()[name = string("op_1764_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1764_pad_0 = const()[name = string("op_1764_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1764_dilations_0 = const()[name = string("op_1764_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1764_groups_0 = const()[name = string("op_1764_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122976832))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123796096))))[name = string("layers_8_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_8_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123796224)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1764_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1764_dilations_0, groups = var_1764_groups_0, pad = var_1764_pad_0, pad_type = var_1764_pad_type_0, strides = var_1764_strides_0, weight = layers_8_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = string("op_1764_cast_fp16")];
            string var_1770_pad_type_0 = const()[name = string("op_1770_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1770_strides_0 = const()[name = string("op_1770_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1770_pad_0 = const()[name = string("op_1770_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1770_dilations_0 = const()[name = string("op_1770_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1770_groups_0 = const()[name = string("op_1770_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123819136))), nonzero_data = tensor<fp16, [10093]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123798848))))[name = string("layers_8_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1770_cast_fp16 = conv(dilations = var_1770_dilations_0, groups = var_1770_groups_0, pad = var_1770_pad_0, pad_type = var_1770_pad_type_0, strides = var_1770_strides_0, weight = layers_8_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_65_cast_fp16)[name = string("op_1770_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_35_cast_fp16 = add(x = var_1764_cast_fp16, y = var_1770_cast_fp16)[name = 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 = string("inputs_35_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1781_to_fp16 = const()[name = string("op_1781_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1781_to_fp16, x = inputs_35_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [1280]> input_67_gamma_0_to_fp16 = const()[name = string("input_67_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124024000)))];
            tensor<fp16, [1280]> input_67_beta_0_to_fp16 = const()[name = string("input_67_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124026624)))];
            fp16 input_67_epsilon_0_to_fp16 = const()[name = string("input_67_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            string var_1799_pad_type_0 = const()[name = string("op_1799_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1799_strides_0 = const()[name = string("op_1799_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1799_pad_0 = const()[name = string("op_1799_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1799_dilations_0 = const()[name = string("op_1799_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1799_groups_0 = const()[name = string("op_1799_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_8_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124029248))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127306112))))[name = string("layers_8_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_8_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_8_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127306240)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_1799_cast_fp16 = conv(bias = layers_8_fc1_inlier_module_bias_to_fp16, dilations = var_1799_dilations_0, groups = var_1799_groups_0, pad = var_1799_pad_0, pad_type = var_1799_pad_type_0, strides = var_1799_strides_0, weight = layers_8_fc1_inlier_module_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = string("op_1799_cast_fp16")];
            string var_1805_pad_type_0 = const()[name = string("op_1805_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1805_strides_0 = const()[name = string("op_1805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1805_pad_0 = const()[name = string("op_1805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1805_dilations_0 = const()[name = string("op_1805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1805_groups_0 = const()[name = string("op_1805_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_8_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127346368))), nonzero_data = tensor<fp16, [14856]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127316544))))[name = string("layers_8_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_1805_cast_fp16 = conv(dilations = var_1805_dilations_0, groups = var_1805_groups_0, pad = var_1805_pad_0, pad_type = var_1805_pad_type_0, strides = var_1805_strides_0, weight = layers_8_fc1_outlier_module_weight_to_fp16_sparsified, x = input_67_cast_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_69_cast_fp16 = add(x = var_1799_cast_fp16, y = var_1805_cast_fp16)[name = string("input_69_cast_fp16")];
            string input_71_mode_0 = const()[name = string("input_71_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = string("input_71_cast_fp16")];
            string var_1816_pad_type_0 = const()[name = string("op_1816_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1816_strides_0 = const()[name = string("op_1816_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1816_pad_0 = const()[name = string("op_1816_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1816_dilations_0 = const()[name = string("op_1816_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1816_groups_0 = const()[name = string("op_1816_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_8_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128165632))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131442496))))[name = string("layers_8_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_8_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_8_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131442624)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1816_cast_fp16 = conv(bias = layers_8_fc2_inlier_module_bias_to_fp16, dilations = var_1816_dilations_0, groups = var_1816_groups_0, pad = var_1816_pad_0, pad_type = var_1816_pad_type_0, strides = var_1816_strides_0, weight = layers_8_fc2_inlier_module_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = string("op_1816_cast_fp16")];
            string var_1822_pad_type_0 = const()[name = string("op_1822_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1822_strides_0 = const()[name = string("op_1822_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1822_pad_0 = const()[name = string("op_1822_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1822_dilations_0 = const()[name = string("op_1822_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1822_groups_0 = const()[name = string("op_1822_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_8_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131699584))), nonzero_data = tensor<fp16, [127131]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131445248))))[name = string("layers_8_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1822_cast_fp16 = conv(dilations = var_1822_dilations_0, groups = var_1822_groups_0, pad = var_1822_pad_0, pad_type = var_1822_pad_type_0, strides = var_1822_strides_0, weight = layers_8_fc2_outlier_module_weight_to_fp16_sparsified, x = input_71_cast_fp16)[name = string("op_1822_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_21_cast_fp16 = add(x = var_1816_cast_fp16, y = var_1822_cast_fp16)[name = 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 = string("inputs_37_cast_fp16")];
            int32 var_1832 = const()[name = string("op_1832"), val = int32(3)];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1851_to_fp16 = const()[name = string("op_1851_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1851_to_fp16, x = inputs_37_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [1280]> obj_37_gamma_0_to_fp16 = const()[name = string("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132518848)))];
            tensor<fp16, [1280]> obj_37_beta_0_to_fp16 = const()[name = string("obj_37_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132521472)))];
            fp16 obj_37_epsilon_0_to_fp16 = const()[name = string("obj_37_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = string("obj_37_cast_fp16")];
            string var_1873_pad_type_0 = const()[name = string("op_1873_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1873_strides_0 = const()[name = string("op_1873_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1873_pad_0 = const()[name = string("op_1873_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1873_dilations_0 = const()[name = string("op_1873_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1873_groups_0 = const()[name = string("op_1873_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132524096))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133343360))))[name = string("layers_9_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_9_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133343488)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1873_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1873_dilations_0, groups = var_1873_groups_0, pad = var_1873_pad_0, pad_type = var_1873_pad_type_0, strides = var_1873_strides_0, weight = layers_9_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = string("op_1873_cast_fp16")];
            string var_1879_pad_type_0 = const()[name = string("op_1879_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1879_strides_0 = const()[name = string("op_1879_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1879_pad_0 = const()[name = string("op_1879_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1879_dilations_0 = const()[name = string("op_1879_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1879_groups_0 = const()[name = string("op_1879_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133392896))), nonzero_data = tensor<fp16, [23337]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133346112))))[name = string("layers_9_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1879_cast_fp16 = conv(dilations = var_1879_dilations_0, groups = var_1879_groups_0, pad = var_1879_pad_0, pad_type = var_1879_pad_type_0, strides = var_1879_strides_0, weight = layers_9_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_37_cast_fp16)[name = string("op_1879_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_19_cast_fp16 = add(x = var_1873_cast_fp16, y = var_1879_cast_fp16)[name = string("query_19_cast_fp16")];
            string var_1888_pad_type_0 = const()[name = string("op_1888_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1888_strides_0 = const()[name = string("op_1888_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1888_pad_0 = const()[name = string("op_1888_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1888_dilations_0 = const()[name = string("op_1888_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1888_groups_0 = const()[name = string("op_1888_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133597760))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134417024))))[name = string("layers_9_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1888_cast_fp16 = conv(dilations = var_1888_dilations_0, groups = var_1888_groups_0, pad = var_1888_pad_0, pad_type = var_1888_pad_type_0, strides = var_1888_strides_0, weight = layers_9_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = string("op_1888_cast_fp16")];
            string var_1894_pad_type_0 = const()[name = string("op_1894_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1894_strides_0 = const()[name = string("op_1894_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1894_pad_0 = const()[name = string("op_1894_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1894_dilations_0 = const()[name = string("op_1894_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1894_groups_0 = const()[name = string("op_1894_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134451584))), nonzero_data = tensor<fp16, [17153]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134417152))))[name = string("layers_9_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1894_cast_fp16 = conv(dilations = var_1894_dilations_0, groups = var_1894_groups_0, pad = var_1894_pad_0, pad_type = var_1894_pad_type_0, strides = var_1894_strides_0, weight = layers_9_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_37_cast_fp16)[name = string("op_1894_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_19_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1894_cast_fp16)[name = string("key_19_cast_fp16")];
            string var_1904_pad_type_0 = const()[name = string("op_1904_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1904_strides_0 = const()[name = string("op_1904_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1904_pad_0 = const()[name = string("op_1904_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1904_dilations_0 = const()[name = string("op_1904_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1904_groups_0 = const()[name = string("op_1904_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134656448))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135475712))))[name = string("layers_9_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_9_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135475840)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1904_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1904_dilations_0, groups = var_1904_groups_0, pad = var_1904_pad_0, pad_type = var_1904_pad_type_0, strides = var_1904_strides_0, weight = layers_9_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = string("op_1904_cast_fp16")];
            string var_1910_pad_type_0 = const()[name = string("op_1910_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1910_strides_0 = const()[name = string("op_1910_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1910_pad_0 = const()[name = string("op_1910_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1910_dilations_0 = const()[name = string("op_1910_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1910_groups_0 = const()[name = string("op_1910_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135500160))), nonzero_data = tensor<fp16, [10813]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135478464))))[name = string("layers_9_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1910_cast_fp16 = conv(dilations = var_1910_dilations_0, groups = var_1910_groups_0, pad = var_1910_pad_0, pad_type = var_1910_pad_type_0, strides = var_1910_strides_0, weight = layers_9_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_37_cast_fp16)[name = string("op_1910_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_19_cast_fp16 = add(x = var_1904_cast_fp16, y = var_1910_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<int32, [4]> var_1913 = const()[name = string("op_1913"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_19_cast_fp16 = reshape(shape = var_1913, x = query_19_cast_fp16)[name = string("mh_q_19_cast_fp16")];
            fp16 var_1915_to_fp16 = const()[name = string("op_1915_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_1916_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1915_to_fp16)[name = string("op_1916_cast_fp16")];
            tensor<int32, [4]> var_1917 = const()[name = string("op_1917"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1918_cast_fp16 = reshape(shape = var_1917, x = key_19_cast_fp16)[name = string("op_1918_cast_fp16")];
            bool mh_w_19_transpose_x_0 = const()[name = string("mh_w_19_transpose_x_0"), val = bool(true)];
            bool mh_w_19_transpose_y_0 = const()[name = string("mh_w_19_transpose_y_0"), val = 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_1916_cast_fp16, y = var_1918_cast_fp16)[name = string("mh_w_19_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_1921_cast_fp16 = softmax(axis = var_1832, x = mh_w_19_cast_fp16)[name = string("op_1921_cast_fp16")];
            tensor<int32, [4]> var_1922 = const()[name = string("op_1922"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1923_cast_fp16 = reshape(shape = var_1922, x = value_19_cast_fp16)[name = string("op_1923_cast_fp16")];
            bool attn_19_transpose_x_0 = const()[name = string("attn_19_transpose_x_0"), val = bool(false)];
            bool attn_19_transpose_y_0 = const()[name = string("attn_19_transpose_y_0"), val = 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_1923_cast_fp16, y = var_1921_cast_fp16)[name = string("attn_19_cast_fp16")];
            tensor<int32, [4]> var_1926 = const()[name = string("op_1926"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_73_cast_fp16 = reshape(shape = var_1926, x = attn_19_cast_fp16)[name = string("input_73_cast_fp16")];
            string var_1936_pad_type_0 = const()[name = string("op_1936_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1936_strides_0 = const()[name = string("op_1936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1936_pad_0 = const()[name = string("op_1936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1936_dilations_0 = const()[name = string("op_1936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1936_groups_0 = const()[name = string("op_1936_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135705024))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136524288))))[name = string("layers_9_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_9_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_9_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136524416)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1936_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1936_dilations_0, groups = var_1936_groups_0, pad = var_1936_pad_0, pad_type = var_1936_pad_type_0, strides = var_1936_strides_0, weight = layers_9_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_73_cast_fp16)[name = string("op_1936_cast_fp16")];
            string var_1942_pad_type_0 = const()[name = string("op_1942_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1942_strides_0 = const()[name = string("op_1942_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1942_pad_0 = const()[name = string("op_1942_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1942_dilations_0 = const()[name = string("op_1942_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1942_groups_0 = const()[name = string("op_1942_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136545344))), nonzero_data = tensor<fp16, [9107]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136527040))))[name = string("layers_9_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1942_cast_fp16 = conv(dilations = var_1942_dilations_0, groups = var_1942_groups_0, pad = var_1942_pad_0, pad_type = var_1942_pad_type_0, strides = var_1942_strides_0, weight = layers_9_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_73_cast_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_39_cast_fp16 = add(x = var_1936_cast_fp16, y = var_1942_cast_fp16)[name = 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 = string("inputs_39_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1953_to_fp16 = const()[name = string("op_1953_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1953_to_fp16, x = inputs_39_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [1280]> input_75_gamma_0_to_fp16 = const()[name = string("input_75_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136750208)))];
            tensor<fp16, [1280]> input_75_beta_0_to_fp16 = const()[name = string("input_75_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136752832)))];
            fp16 input_75_epsilon_0_to_fp16 = const()[name = string("input_75_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            string var_1971_pad_type_0 = const()[name = string("op_1971_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1971_strides_0 = const()[name = string("op_1971_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1971_pad_0 = const()[name = string("op_1971_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1971_dilations_0 = const()[name = string("op_1971_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1971_groups_0 = const()[name = string("op_1971_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_9_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136755456))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140032320))))[name = string("layers_9_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_9_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_9_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140032448)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_1971_cast_fp16 = conv(bias = layers_9_fc1_inlier_module_bias_to_fp16, dilations = var_1971_dilations_0, groups = var_1971_groups_0, pad = var_1971_pad_0, pad_type = var_1971_pad_type_0, strides = var_1971_strides_0, weight = layers_9_fc1_inlier_module_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("op_1971_cast_fp16")];
            string var_1977_pad_type_0 = const()[name = string("op_1977_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1977_strides_0 = const()[name = string("op_1977_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1977_pad_0 = const()[name = string("op_1977_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1977_dilations_0 = const()[name = string("op_1977_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1977_groups_0 = const()[name = string("op_1977_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_9_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140073152))), nonzero_data = tensor<fp16, [15166]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140042752))))[name = string("layers_9_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_1977_cast_fp16 = conv(dilations = var_1977_dilations_0, groups = var_1977_groups_0, pad = var_1977_pad_0, pad_type = var_1977_pad_type_0, strides = var_1977_strides_0, weight = layers_9_fc1_outlier_module_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_77_cast_fp16 = add(x = var_1971_cast_fp16, y = var_1977_cast_fp16)[name = string("input_77_cast_fp16")];
            string input_79_mode_0 = const()[name = string("input_79_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")];
            string var_1988_pad_type_0 = const()[name = string("op_1988_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1988_strides_0 = const()[name = string("op_1988_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1988_pad_0 = const()[name = string("op_1988_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1988_dilations_0 = const()[name = string("op_1988_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1988_groups_0 = const()[name = string("op_1988_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_9_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140892416))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144169280))))[name = string("layers_9_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_9_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_9_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144169408)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_1988_cast_fp16 = conv(bias = layers_9_fc2_inlier_module_bias_to_fp16, dilations = var_1988_dilations_0, groups = var_1988_groups_0, pad = var_1988_pad_0, pad_type = var_1988_pad_type_0, strides = var_1988_strides_0, weight = layers_9_fc2_inlier_module_weight_to_fp16_palettized, x = input_79_cast_fp16)[name = string("op_1988_cast_fp16")];
            string var_1994_pad_type_0 = const()[name = string("op_1994_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1994_strides_0 = const()[name = string("op_1994_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1994_pad_0 = const()[name = string("op_1994_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1994_dilations_0 = const()[name = string("op_1994_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1994_groups_0 = const()[name = string("op_1994_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_9_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144438144))), nonzero_data = tensor<fp16, [133022]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144172032))))[name = string("layers_9_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_1994_cast_fp16 = conv(dilations = var_1994_dilations_0, groups = var_1994_groups_0, pad = var_1994_pad_0, pad_type = var_1994_pad_type_0, strides = var_1994_strides_0, weight = layers_9_fc2_outlier_module_weight_to_fp16_sparsified, x = input_79_cast_fp16)[name = string("op_1994_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_23_cast_fp16 = add(x = var_1988_cast_fp16, y = var_1994_cast_fp16)[name = 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 = string("inputs_41_cast_fp16")];
            int32 var_2004 = const()[name = string("op_2004"), val = int32(3)];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2023_to_fp16 = const()[name = string("op_2023_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_2023_to_fp16, x = inputs_41_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [1280]> obj_41_gamma_0_to_fp16 = const()[name = string("obj_41_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145257408)))];
            tensor<fp16, [1280]> obj_41_beta_0_to_fp16 = const()[name = string("obj_41_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145260032)))];
            fp16 obj_41_epsilon_0_to_fp16 = const()[name = string("obj_41_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = string("obj_41_cast_fp16")];
            string var_2045_pad_type_0 = const()[name = string("op_2045_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2045_strides_0 = const()[name = string("op_2045_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2045_pad_0 = const()[name = string("op_2045_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2045_dilations_0 = const()[name = string("op_2045_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2045_groups_0 = const()[name = string("op_2045_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145262656))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146081920))))[name = string("layers_10_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_10_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146082048)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2045_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_2045_dilations_0, groups = var_2045_groups_0, pad = var_2045_pad_0, pad_type = var_2045_pad_type_0, strides = var_2045_strides_0, weight = layers_10_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = string("op_2045_cast_fp16")];
            string var_2051_pad_type_0 = const()[name = string("op_2051_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2051_strides_0 = const()[name = string("op_2051_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2051_pad_0 = const()[name = string("op_2051_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2051_dilations_0 = const()[name = string("op_2051_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2051_groups_0 = const()[name = string("op_2051_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146130624))), nonzero_data = tensor<fp16, [22915]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146084672))))[name = string("layers_10_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2051_cast_fp16 = conv(dilations = var_2051_dilations_0, groups = var_2051_groups_0, pad = var_2051_pad_0, pad_type = var_2051_pad_type_0, strides = var_2051_strides_0, weight = layers_10_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_41_cast_fp16)[name = string("op_2051_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_21_cast_fp16 = add(x = var_2045_cast_fp16, y = var_2051_cast_fp16)[name = string("query_21_cast_fp16")];
            string var_2060_pad_type_0 = const()[name = string("op_2060_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2060_strides_0 = const()[name = string("op_2060_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2060_pad_0 = const()[name = string("op_2060_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2060_dilations_0 = const()[name = string("op_2060_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2060_groups_0 = const()[name = string("op_2060_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146335488))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147154752))))[name = string("layers_10_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2060_cast_fp16 = conv(dilations = var_2060_dilations_0, groups = var_2060_groups_0, pad = var_2060_pad_0, pad_type = var_2060_pad_type_0, strides = var_2060_strides_0, weight = layers_10_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = string("op_2060_cast_fp16")];
            string var_2066_pad_type_0 = const()[name = string("op_2066_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2066_strides_0 = const()[name = string("op_2066_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2066_pad_0 = const()[name = string("op_2066_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2066_dilations_0 = const()[name = string("op_2066_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2066_groups_0 = const()[name = string("op_2066_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147187392))), nonzero_data = tensor<fp16, [16203]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147154880))))[name = string("layers_10_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2066_cast_fp16 = conv(dilations = var_2066_dilations_0, groups = var_2066_groups_0, pad = var_2066_pad_0, pad_type = var_2066_pad_type_0, strides = var_2066_strides_0, weight = layers_10_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_41_cast_fp16)[name = string("op_2066_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_21_cast_fp16 = add(x = var_2060_cast_fp16, y = var_2066_cast_fp16)[name = string("key_21_cast_fp16")];
            string var_2076_pad_type_0 = const()[name = string("op_2076_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2076_strides_0 = const()[name = string("op_2076_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2076_pad_0 = const()[name = string("op_2076_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2076_dilations_0 = const()[name = string("op_2076_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2076_groups_0 = const()[name = string("op_2076_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147392256))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148211520))))[name = string("layers_10_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_10_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148211648)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2076_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_2076_dilations_0, groups = var_2076_groups_0, pad = var_2076_pad_0, pad_type = var_2076_pad_type_0, strides = var_2076_strides_0, weight = layers_10_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = string("op_2076_cast_fp16")];
            string var_2082_pad_type_0 = const()[name = string("op_2082_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2082_strides_0 = const()[name = string("op_2082_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2082_pad_0 = const()[name = string("op_2082_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2082_dilations_0 = const()[name = string("op_2082_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2082_groups_0 = const()[name = string("op_2082_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148234688))), nonzero_data = tensor<fp16, [10173]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148214272))))[name = string("layers_10_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2082_cast_fp16 = conv(dilations = var_2082_dilations_0, groups = var_2082_groups_0, pad = var_2082_pad_0, pad_type = var_2082_pad_type_0, strides = var_2082_strides_0, weight = layers_10_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_41_cast_fp16)[name = string("op_2082_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_21_cast_fp16 = add(x = var_2076_cast_fp16, y = var_2082_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<int32, [4]> var_2085 = const()[name = string("op_2085"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_21_cast_fp16 = reshape(shape = var_2085, x = query_21_cast_fp16)[name = string("mh_q_21_cast_fp16")];
            fp16 var_2087_to_fp16 = const()[name = string("op_2087_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_2088_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_2087_to_fp16)[name = string("op_2088_cast_fp16")];
            tensor<int32, [4]> var_2089 = const()[name = string("op_2089"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2090_cast_fp16 = reshape(shape = var_2089, x = key_21_cast_fp16)[name = string("op_2090_cast_fp16")];
            bool mh_w_21_transpose_x_0 = const()[name = string("mh_w_21_transpose_x_0"), val = bool(true)];
            bool mh_w_21_transpose_y_0 = const()[name = string("mh_w_21_transpose_y_0"), val = 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_2088_cast_fp16, y = var_2090_cast_fp16)[name = string("mh_w_21_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_2093_cast_fp16 = softmax(axis = var_2004, x = mh_w_21_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [4]> var_2094 = const()[name = string("op_2094"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2095_cast_fp16 = reshape(shape = var_2094, x = value_21_cast_fp16)[name = string("op_2095_cast_fp16")];
            bool attn_21_transpose_x_0 = const()[name = string("attn_21_transpose_x_0"), val = bool(false)];
            bool attn_21_transpose_y_0 = const()[name = string("attn_21_transpose_y_0"), val = 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_2095_cast_fp16, y = var_2093_cast_fp16)[name = string("attn_21_cast_fp16")];
            tensor<int32, [4]> var_2098 = const()[name = string("op_2098"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_81_cast_fp16 = reshape(shape = var_2098, x = attn_21_cast_fp16)[name = string("input_81_cast_fp16")];
            string var_2108_pad_type_0 = const()[name = string("op_2108_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2108_strides_0 = const()[name = string("op_2108_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2108_pad_0 = const()[name = string("op_2108_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2108_dilations_0 = const()[name = string("op_2108_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2108_groups_0 = const()[name = string("op_2108_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148439552))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149258816))))[name = string("layers_10_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_10_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_10_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149258944)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2108_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_2108_dilations_0, groups = var_2108_groups_0, pad = var_2108_pad_0, pad_type = var_2108_pad_type_0, strides = var_2108_strides_0, weight = layers_10_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = string("op_2108_cast_fp16")];
            string var_2114_pad_type_0 = const()[name = string("op_2114_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2114_strides_0 = const()[name = string("op_2114_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2114_pad_0 = const()[name = string("op_2114_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2114_dilations_0 = const()[name = string("op_2114_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2114_groups_0 = const()[name = string("op_2114_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149280384))), nonzero_data = tensor<fp16, [9361]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149261568))))[name = string("layers_10_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2114_cast_fp16 = conv(dilations = var_2114_dilations_0, groups = var_2114_groups_0, pad = var_2114_pad_0, pad_type = var_2114_pad_type_0, strides = var_2114_strides_0, weight = layers_10_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_81_cast_fp16)[name = string("op_2114_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_43_cast_fp16 = add(x = var_2108_cast_fp16, y = var_2114_cast_fp16)[name = 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 = string("inputs_43_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2125_to_fp16 = const()[name = string("op_2125_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_2125_to_fp16, x = inputs_43_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [1280]> input_83_gamma_0_to_fp16 = const()[name = string("input_83_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149485248)))];
            tensor<fp16, [1280]> input_83_beta_0_to_fp16 = const()[name = string("input_83_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149487872)))];
            fp16 input_83_epsilon_0_to_fp16 = const()[name = string("input_83_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            string var_2143_pad_type_0 = const()[name = string("op_2143_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2143_strides_0 = const()[name = string("op_2143_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2143_pad_0 = const()[name = string("op_2143_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2143_dilations_0 = const()[name = string("op_2143_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2143_groups_0 = const()[name = string("op_2143_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_10_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149490496))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152767360))))[name = string("layers_10_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_10_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_10_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152767488)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_2143_cast_fp16 = conv(bias = layers_10_fc1_inlier_module_bias_to_fp16, dilations = var_2143_dilations_0, groups = var_2143_groups_0, pad = var_2143_pad_0, pad_type = var_2143_pad_type_0, strides = var_2143_strides_0, weight = layers_10_fc1_inlier_module_weight_to_fp16_palettized, x = input_83_cast_fp16)[name = string("op_2143_cast_fp16")];
            string var_2149_pad_type_0 = const()[name = string("op_2149_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2149_strides_0 = const()[name = string("op_2149_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2149_pad_0 = const()[name = string("op_2149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2149_dilations_0 = const()[name = string("op_2149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2149_groups_0 = const()[name = string("op_2149_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_10_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152808704))), nonzero_data = tensor<fp16, [15418]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152777792))))[name = string("layers_10_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_2149_cast_fp16 = conv(dilations = var_2149_dilations_0, groups = var_2149_groups_0, pad = var_2149_pad_0, pad_type = var_2149_pad_type_0, strides = var_2149_strides_0, weight = layers_10_fc1_outlier_module_weight_to_fp16_sparsified, x = input_83_cast_fp16)[name = string("op_2149_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_85_cast_fp16 = add(x = var_2143_cast_fp16, y = var_2149_cast_fp16)[name = string("input_85_cast_fp16")];
            string input_87_mode_0 = const()[name = string("input_87_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = string("input_87_cast_fp16")];
            string var_2160_pad_type_0 = const()[name = string("op_2160_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2160_strides_0 = const()[name = string("op_2160_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2160_pad_0 = const()[name = string("op_2160_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2160_dilations_0 = const()[name = string("op_2160_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2160_groups_0 = const()[name = string("op_2160_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_10_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153627968))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156904832))))[name = string("layers_10_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_10_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_10_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156904960)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2160_cast_fp16 = conv(bias = layers_10_fc2_inlier_module_bias_to_fp16, dilations = var_2160_dilations_0, groups = var_2160_groups_0, pad = var_2160_pad_0, pad_type = var_2160_pad_type_0, strides = var_2160_strides_0, weight = layers_10_fc2_inlier_module_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = string("op_2160_cast_fp16")];
            string var_2166_pad_type_0 = const()[name = string("op_2166_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2166_strides_0 = const()[name = string("op_2166_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2166_pad_0 = const()[name = string("op_2166_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2166_dilations_0 = const()[name = string("op_2166_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2166_groups_0 = const()[name = string("op_2166_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_10_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157157056))), nonzero_data = tensor<fp16, [124674]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156907584))))[name = string("layers_10_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2166_cast_fp16 = conv(dilations = var_2166_dilations_0, groups = var_2166_groups_0, pad = var_2166_pad_0, pad_type = var_2166_pad_type_0, strides = var_2166_strides_0, weight = layers_10_fc2_outlier_module_weight_to_fp16_sparsified, x = input_87_cast_fp16)[name = string("op_2166_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_25_cast_fp16 = add(x = var_2160_cast_fp16, y = var_2166_cast_fp16)[name = 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 = string("inputs_45_cast_fp16")];
            int32 var_2176 = const()[name = string("op_2176"), val = int32(3)];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2195_to_fp16 = const()[name = string("op_2195_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_2195_to_fp16, x = inputs_45_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [1280]> obj_45_gamma_0_to_fp16 = const()[name = string("obj_45_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157976320)))];
            tensor<fp16, [1280]> obj_45_beta_0_to_fp16 = const()[name = string("obj_45_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157978944)))];
            fp16 obj_45_epsilon_0_to_fp16 = const()[name = string("obj_45_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = string("obj_45_cast_fp16")];
            string var_2217_pad_type_0 = const()[name = string("op_2217_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2217_strides_0 = const()[name = string("op_2217_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2217_pad_0 = const()[name = string("op_2217_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2217_dilations_0 = const()[name = string("op_2217_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2217_groups_0 = const()[name = string("op_2217_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157981568))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158800832))))[name = string("layers_11_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_11_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_11_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158800960)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2217_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_2217_dilations_0, groups = var_2217_groups_0, pad = var_2217_pad_0, pad_type = var_2217_pad_type_0, strides = var_2217_strides_0, weight = layers_11_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = string("op_2217_cast_fp16")];
            string var_2223_pad_type_0 = const()[name = string("op_2223_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2223_strides_0 = const()[name = string("op_2223_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2223_pad_0 = const()[name = string("op_2223_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2223_dilations_0 = const()[name = string("op_2223_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2223_groups_0 = const()[name = string("op_2223_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158852864))), nonzero_data = tensor<fp16, [24598]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158803584))))[name = string("layers_11_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2223_cast_fp16 = conv(dilations = var_2223_dilations_0, groups = var_2223_groups_0, pad = var_2223_pad_0, pad_type = var_2223_pad_type_0, strides = var_2223_strides_0, weight = layers_11_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_45_cast_fp16)[name = string("op_2223_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_23_cast_fp16 = add(x = var_2217_cast_fp16, y = var_2223_cast_fp16)[name = string("query_23_cast_fp16")];
            string var_2232_pad_type_0 = const()[name = string("op_2232_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2232_strides_0 = const()[name = string("op_2232_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2232_pad_0 = const()[name = string("op_2232_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2232_dilations_0 = const()[name = string("op_2232_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2232_groups_0 = const()[name = string("op_2232_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159057728))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159876992))))[name = string("layers_11_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2232_cast_fp16 = conv(dilations = var_2232_dilations_0, groups = var_2232_groups_0, pad = var_2232_pad_0, pad_type = var_2232_pad_type_0, strides = var_2232_strides_0, weight = layers_11_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = string("op_2232_cast_fp16")];
            string var_2238_pad_type_0 = const()[name = string("op_2238_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2238_strides_0 = const()[name = string("op_2238_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2238_pad_0 = const()[name = string("op_2238_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2238_dilations_0 = const()[name = string("op_2238_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2238_groups_0 = const()[name = string("op_2238_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159906048))), nonzero_data = tensor<fp16, [14404]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159877120))))[name = string("layers_11_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2238_cast_fp16 = conv(dilations = var_2238_dilations_0, groups = var_2238_groups_0, pad = var_2238_pad_0, pad_type = var_2238_pad_type_0, strides = var_2238_strides_0, weight = layers_11_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_45_cast_fp16)[name = string("op_2238_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_23_cast_fp16 = add(x = var_2232_cast_fp16, y = var_2238_cast_fp16)[name = string("key_23_cast_fp16")];
            string var_2248_pad_type_0 = const()[name = string("op_2248_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2248_strides_0 = const()[name = string("op_2248_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2248_pad_0 = const()[name = string("op_2248_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2248_dilations_0 = const()[name = string("op_2248_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2248_groups_0 = const()[name = string("op_2248_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160110912))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160930176))))[name = string("layers_11_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_11_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160930304)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2248_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_2248_dilations_0, groups = var_2248_groups_0, pad = var_2248_pad_0, pad_type = var_2248_pad_type_0, strides = var_2248_strides_0, weight = layers_11_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = string("op_2248_cast_fp16")];
            string var_2254_pad_type_0 = const()[name = string("op_2254_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2254_strides_0 = const()[name = string("op_2254_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2254_pad_0 = const()[name = string("op_2254_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2254_dilations_0 = const()[name = string("op_2254_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2254_groups_0 = const()[name = string("op_2254_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160953216))), nonzero_data = tensor<fp16, [10097]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160932928))))[name = string("layers_11_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2254_cast_fp16 = conv(dilations = var_2254_dilations_0, groups = var_2254_groups_0, pad = var_2254_pad_0, pad_type = var_2254_pad_type_0, strides = var_2254_strides_0, weight = layers_11_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_45_cast_fp16)[name = string("op_2254_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_23_cast_fp16 = add(x = var_2248_cast_fp16, y = var_2254_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<int32, [4]> var_2257 = const()[name = string("op_2257"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_23_cast_fp16 = reshape(shape = var_2257, x = query_23_cast_fp16)[name = string("mh_q_23_cast_fp16")];
            fp16 var_2259_to_fp16 = const()[name = string("op_2259_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_2260_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_2259_to_fp16)[name = string("op_2260_cast_fp16")];
            tensor<int32, [4]> var_2261 = const()[name = string("op_2261"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2262_cast_fp16 = reshape(shape = var_2261, x = key_23_cast_fp16)[name = string("op_2262_cast_fp16")];
            bool mh_w_23_transpose_x_0 = const()[name = string("mh_w_23_transpose_x_0"), val = bool(true)];
            bool mh_w_23_transpose_y_0 = const()[name = string("mh_w_23_transpose_y_0"), val = 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_2260_cast_fp16, y = var_2262_cast_fp16)[name = string("mh_w_23_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_2265_cast_fp16 = softmax(axis = var_2176, x = mh_w_23_cast_fp16)[name = string("op_2265_cast_fp16")];
            tensor<int32, [4]> var_2266 = const()[name = string("op_2266"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2267_cast_fp16 = reshape(shape = var_2266, x = value_23_cast_fp16)[name = string("op_2267_cast_fp16")];
            bool attn_23_transpose_x_0 = const()[name = string("attn_23_transpose_x_0"), val = bool(false)];
            bool attn_23_transpose_y_0 = const()[name = string("attn_23_transpose_y_0"), val = 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_2267_cast_fp16, y = var_2265_cast_fp16)[name = string("attn_23_cast_fp16")];
            tensor<int32, [4]> var_2270 = const()[name = string("op_2270"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_89_cast_fp16 = reshape(shape = var_2270, x = attn_23_cast_fp16)[name = string("input_89_cast_fp16")];
            string var_2280_pad_type_0 = const()[name = string("op_2280_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2280_strides_0 = const()[name = string("op_2280_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2280_pad_0 = const()[name = string("op_2280_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2280_dilations_0 = const()[name = string("op_2280_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2280_groups_0 = const()[name = string("op_2280_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161158080))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161977344))))[name = string("layers_11_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_11_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_11_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161977472)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2280_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_2280_dilations_0, groups = var_2280_groups_0, pad = var_2280_pad_0, pad_type = var_2280_pad_type_0, strides = var_2280_strides_0, weight = layers_11_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = string("op_2280_cast_fp16")];
            string var_2286_pad_type_0 = const()[name = string("op_2286_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2286_strides_0 = const()[name = string("op_2286_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2286_pad_0 = const()[name = string("op_2286_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2286_dilations_0 = const()[name = string("op_2286_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2286_groups_0 = const()[name = string("op_2286_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161997632))), nonzero_data = tensor<fp16, [8713]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161980096))))[name = string("layers_11_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2286_cast_fp16 = conv(dilations = var_2286_dilations_0, groups = var_2286_groups_0, pad = var_2286_pad_0, pad_type = var_2286_pad_type_0, strides = var_2286_strides_0, weight = layers_11_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_89_cast_fp16)[name = string("op_2286_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_47_cast_fp16 = add(x = var_2280_cast_fp16, y = var_2286_cast_fp16)[name = 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 = string("inputs_47_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2297_to_fp16 = const()[name = string("op_2297_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_2297_to_fp16, x = inputs_47_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [1280]> input_91_gamma_0_to_fp16 = const()[name = string("input_91_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162202496)))];
            tensor<fp16, [1280]> input_91_beta_0_to_fp16 = const()[name = string("input_91_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162205120)))];
            fp16 input_91_epsilon_0_to_fp16 = const()[name = string("input_91_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            string var_2315_pad_type_0 = const()[name = string("op_2315_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2315_strides_0 = const()[name = string("op_2315_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2315_pad_0 = const()[name = string("op_2315_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2315_dilations_0 = const()[name = string("op_2315_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2315_groups_0 = const()[name = string("op_2315_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_11_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162207744))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165484608))))[name = string("layers_11_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_11_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_11_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165484736)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_2315_cast_fp16 = conv(bias = layers_11_fc1_inlier_module_bias_to_fp16, dilations = var_2315_dilations_0, groups = var_2315_groups_0, pad = var_2315_pad_0, pad_type = var_2315_pad_type_0, strides = var_2315_strides_0, weight = layers_11_fc1_inlier_module_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("op_2315_cast_fp16")];
            string var_2321_pad_type_0 = const()[name = string("op_2321_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2321_strides_0 = const()[name = string("op_2321_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2321_pad_0 = const()[name = string("op_2321_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2321_dilations_0 = const()[name = string("op_2321_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2321_groups_0 = const()[name = string("op_2321_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_11_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165526016))), nonzero_data = tensor<fp16, [15430]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165495040))))[name = string("layers_11_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_2321_cast_fp16 = conv(dilations = var_2321_dilations_0, groups = var_2321_groups_0, pad = var_2321_pad_0, pad_type = var_2321_pad_type_0, strides = var_2321_strides_0, weight = layers_11_fc1_outlier_module_weight_to_fp16_sparsified, x = input_91_cast_fp16)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_93_cast_fp16 = add(x = var_2315_cast_fp16, y = var_2321_cast_fp16)[name = string("input_93_cast_fp16")];
            string input_95_mode_0 = const()[name = string("input_95_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = input_93_cast_fp16)[name = string("input_95_cast_fp16")];
            string var_2332_pad_type_0 = const()[name = string("op_2332_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2332_strides_0 = const()[name = string("op_2332_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2332_pad_0 = const()[name = string("op_2332_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2332_dilations_0 = const()[name = string("op_2332_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2332_groups_0 = const()[name = string("op_2332_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_11_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166345280))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169622144))))[name = string("layers_11_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_11_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_11_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169622272)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2332_cast_fp16 = conv(bias = layers_11_fc2_inlier_module_bias_to_fp16, dilations = var_2332_dilations_0, groups = var_2332_groups_0, pad = var_2332_pad_0, pad_type = var_2332_pad_type_0, strides = var_2332_strides_0, weight = layers_11_fc2_inlier_module_weight_to_fp16_palettized, x = input_95_cast_fp16)[name = string("op_2332_cast_fp16")];
            string var_2338_pad_type_0 = const()[name = string("op_2338_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2338_strides_0 = const()[name = string("op_2338_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2338_pad_0 = const()[name = string("op_2338_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2338_dilations_0 = const()[name = string("op_2338_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2338_groups_0 = const()[name = string("op_2338_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_11_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169852352))), nonzero_data = tensor<fp16, [113683]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169624896))))[name = string("layers_11_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2338_cast_fp16 = conv(dilations = var_2338_dilations_0, groups = var_2338_groups_0, pad = var_2338_pad_0, pad_type = var_2338_pad_type_0, strides = var_2338_strides_0, weight = layers_11_fc2_outlier_module_weight_to_fp16_sparsified, x = input_95_cast_fp16)[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_27_cast_fp16 = add(x = var_2332_cast_fp16, y = var_2338_cast_fp16)[name = 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 = string("inputs_49_cast_fp16")];
            int32 var_2348 = const()[name = string("op_2348"), val = int32(3)];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2367_to_fp16 = const()[name = string("op_2367_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_2367_to_fp16, x = inputs_49_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [1280]> obj_49_gamma_0_to_fp16 = const()[name = string("obj_49_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170671616)))];
            tensor<fp16, [1280]> obj_49_beta_0_to_fp16 = const()[name = string("obj_49_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170674240)))];
            fp16 obj_49_epsilon_0_to_fp16 = const()[name = string("obj_49_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = string("obj_49_cast_fp16")];
            string var_2389_pad_type_0 = const()[name = string("op_2389_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2389_strides_0 = const()[name = string("op_2389_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2389_pad_0 = const()[name = string("op_2389_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2389_dilations_0 = const()[name = string("op_2389_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2389_groups_0 = const()[name = string("op_2389_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170676864))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171496128))))[name = string("layers_12_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_12_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_12_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171496256)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2389_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_2389_dilations_0, groups = var_2389_groups_0, pad = var_2389_pad_0, pad_type = var_2389_pad_type_0, strides = var_2389_strides_0, weight = layers_12_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = string("op_2389_cast_fp16")];
            string var_2395_pad_type_0 = const()[name = string("op_2395_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2395_strides_0 = const()[name = string("op_2395_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2395_pad_0 = const()[name = string("op_2395_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2395_dilations_0 = const()[name = string("op_2395_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2395_groups_0 = const()[name = string("op_2395_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171547328))), nonzero_data = tensor<fp16, [24167]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171498880))))[name = string("layers_12_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2395_cast_fp16 = conv(dilations = var_2395_dilations_0, groups = var_2395_groups_0, pad = var_2395_pad_0, pad_type = var_2395_pad_type_0, strides = var_2395_strides_0, weight = layers_12_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_49_cast_fp16)[name = string("op_2395_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_25_cast_fp16 = add(x = var_2389_cast_fp16, y = var_2395_cast_fp16)[name = string("query_25_cast_fp16")];
            string var_2404_pad_type_0 = const()[name = string("op_2404_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2404_strides_0 = const()[name = string("op_2404_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2404_pad_0 = const()[name = string("op_2404_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2404_dilations_0 = const()[name = string("op_2404_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2404_groups_0 = const()[name = string("op_2404_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171752192))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172571456))))[name = string("layers_12_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2404_cast_fp16 = conv(dilations = var_2404_dilations_0, groups = var_2404_groups_0, pad = var_2404_pad_0, pad_type = var_2404_pad_type_0, strides = var_2404_strides_0, weight = layers_12_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = string("op_2404_cast_fp16")];
            string var_2410_pad_type_0 = const()[name = string("op_2410_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2410_strides_0 = const()[name = string("op_2410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2410_pad_0 = const()[name = string("op_2410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2410_dilations_0 = const()[name = string("op_2410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2410_groups_0 = const()[name = string("op_2410_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172602112))), nonzero_data = tensor<fp16, [15212]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172571584))))[name = string("layers_12_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2410_cast_fp16 = conv(dilations = var_2410_dilations_0, groups = var_2410_groups_0, pad = var_2410_pad_0, pad_type = var_2410_pad_type_0, strides = var_2410_strides_0, weight = layers_12_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_49_cast_fp16)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_25_cast_fp16 = add(x = var_2404_cast_fp16, y = var_2410_cast_fp16)[name = string("key_25_cast_fp16")];
            string var_2420_pad_type_0 = const()[name = string("op_2420_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2420_strides_0 = const()[name = string("op_2420_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2420_pad_0 = const()[name = string("op_2420_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2420_dilations_0 = const()[name = string("op_2420_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2420_groups_0 = const()[name = string("op_2420_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172806976))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173626240))))[name = string("layers_12_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_12_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173626368)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2420_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_2420_dilations_0, groups = var_2420_groups_0, pad = var_2420_pad_0, pad_type = var_2420_pad_type_0, strides = var_2420_strides_0, weight = layers_12_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = string("op_2420_cast_fp16")];
            string var_2426_pad_type_0 = const()[name = string("op_2426_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2426_strides_0 = const()[name = string("op_2426_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2426_pad_0 = const()[name = string("op_2426_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2426_dilations_0 = const()[name = string("op_2426_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2426_groups_0 = const()[name = string("op_2426_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173648128))), nonzero_data = tensor<fp16, [9512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173628992))))[name = string("layers_12_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2426_cast_fp16 = conv(dilations = var_2426_dilations_0, groups = var_2426_groups_0, pad = var_2426_pad_0, pad_type = var_2426_pad_type_0, strides = var_2426_strides_0, weight = layers_12_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_49_cast_fp16)[name = string("op_2426_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_25_cast_fp16 = add(x = var_2420_cast_fp16, y = var_2426_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_25_cast_fp16 = reshape(shape = var_2429, x = query_25_cast_fp16)[name = string("mh_q_25_cast_fp16")];
            fp16 var_2431_to_fp16 = const()[name = string("op_2431_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_2432_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_2431_to_fp16)[name = string("op_2432_cast_fp16")];
            tensor<int32, [4]> var_2433 = const()[name = string("op_2433"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2434_cast_fp16 = reshape(shape = var_2433, x = key_25_cast_fp16)[name = string("op_2434_cast_fp16")];
            bool mh_w_25_transpose_x_0 = const()[name = string("mh_w_25_transpose_x_0"), val = bool(true)];
            bool mh_w_25_transpose_y_0 = const()[name = string("mh_w_25_transpose_y_0"), val = 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_2432_cast_fp16, y = var_2434_cast_fp16)[name = string("mh_w_25_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_2437_cast_fp16 = softmax(axis = var_2348, x = mh_w_25_cast_fp16)[name = string("op_2437_cast_fp16")];
            tensor<int32, [4]> var_2438 = const()[name = string("op_2438"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2439_cast_fp16 = reshape(shape = var_2438, x = value_25_cast_fp16)[name = string("op_2439_cast_fp16")];
            bool attn_25_transpose_x_0 = const()[name = string("attn_25_transpose_x_0"), val = bool(false)];
            bool attn_25_transpose_y_0 = const()[name = string("attn_25_transpose_y_0"), val = 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_2439_cast_fp16, y = var_2437_cast_fp16)[name = string("attn_25_cast_fp16")];
            tensor<int32, [4]> var_2442 = const()[name = string("op_2442"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_97_cast_fp16 = reshape(shape = var_2442, x = attn_25_cast_fp16)[name = string("input_97_cast_fp16")];
            string var_2452_pad_type_0 = const()[name = string("op_2452_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2452_strides_0 = const()[name = string("op_2452_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2452_pad_0 = const()[name = string("op_2452_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2452_dilations_0 = const()[name = string("op_2452_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2452_groups_0 = const()[name = string("op_2452_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173852992))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174672256))))[name = string("layers_12_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_12_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_12_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174672384)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2452_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_2452_dilations_0, groups = var_2452_groups_0, pad = var_2452_pad_0, pad_type = var_2452_pad_type_0, strides = var_2452_strides_0, weight = layers_12_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = string("op_2452_cast_fp16")];
            string var_2458_pad_type_0 = const()[name = string("op_2458_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2458_strides_0 = const()[name = string("op_2458_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2458_pad_0 = const()[name = string("op_2458_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2458_dilations_0 = const()[name = string("op_2458_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2458_groups_0 = const()[name = string("op_2458_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174690880))), nonzero_data = tensor<fp16, [7893]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174675008))))[name = string("layers_12_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2458_cast_fp16 = conv(dilations = var_2458_dilations_0, groups = var_2458_groups_0, pad = var_2458_pad_0, pad_type = var_2458_pad_type_0, strides = var_2458_strides_0, weight = layers_12_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_97_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_51_cast_fp16 = add(x = var_2452_cast_fp16, y = var_2458_cast_fp16)[name = 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 = string("inputs_51_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2469_to_fp16 = const()[name = string("op_2469_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_2469_to_fp16, x = inputs_51_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [1280]> input_99_gamma_0_to_fp16 = const()[name = string("input_99_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174895744)))];
            tensor<fp16, [1280]> input_99_beta_0_to_fp16 = const()[name = string("input_99_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174898368)))];
            fp16 input_99_epsilon_0_to_fp16 = const()[name = string("input_99_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            string var_2487_pad_type_0 = const()[name = string("op_2487_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2487_strides_0 = const()[name = string("op_2487_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2487_pad_0 = const()[name = string("op_2487_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2487_dilations_0 = const()[name = string("op_2487_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2487_groups_0 = const()[name = string("op_2487_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_12_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174900992))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178177856))))[name = string("layers_12_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_12_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_12_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178177984)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_2487_cast_fp16 = conv(bias = layers_12_fc1_inlier_module_bias_to_fp16, dilations = var_2487_dilations_0, groups = var_2487_groups_0, pad = var_2487_pad_0, pad_type = var_2487_pad_type_0, strides = var_2487_strides_0, weight = layers_12_fc1_inlier_module_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = string("op_2487_cast_fp16")];
            string var_2493_pad_type_0 = const()[name = string("op_2493_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2493_strides_0 = const()[name = string("op_2493_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2493_pad_0 = const()[name = string("op_2493_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2493_dilations_0 = const()[name = string("op_2493_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2493_groups_0 = const()[name = string("op_2493_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_12_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178232512))), nonzero_data = tensor<fp16, [22050]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178188288))))[name = string("layers_12_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_2493_cast_fp16 = conv(dilations = var_2493_dilations_0, groups = var_2493_groups_0, pad = var_2493_pad_0, pad_type = var_2493_pad_type_0, strides = var_2493_strides_0, weight = layers_12_fc1_outlier_module_weight_to_fp16_sparsified, x = input_99_cast_fp16)[name = string("op_2493_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_101_cast_fp16 = add(x = var_2487_cast_fp16, y = var_2493_cast_fp16)[name = string("input_101_cast_fp16")];
            string input_103_mode_0 = const()[name = string("input_103_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
            string var_2504_pad_type_0 = const()[name = string("op_2504_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2504_strides_0 = const()[name = string("op_2504_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2504_pad_0 = const()[name = string("op_2504_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2504_dilations_0 = const()[name = string("op_2504_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2504_groups_0 = const()[name = string("op_2504_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_12_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179051776))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182328640))))[name = string("layers_12_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_12_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_12_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182328768)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2504_cast_fp16 = conv(bias = layers_12_fc2_inlier_module_bias_to_fp16, dilations = var_2504_dilations_0, groups = var_2504_groups_0, pad = var_2504_pad_0, pad_type = var_2504_pad_type_0, strides = var_2504_strides_0, weight = layers_12_fc2_inlier_module_weight_to_fp16_palettized, x = input_103_cast_fp16)[name = string("op_2504_cast_fp16")];
            string var_2510_pad_type_0 = const()[name = string("op_2510_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2510_strides_0 = const()[name = string("op_2510_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2510_pad_0 = const()[name = string("op_2510_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2510_dilations_0 = const()[name = string("op_2510_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2510_groups_0 = const()[name = string("op_2510_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_12_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182509056))), nonzero_data = tensor<fp16, [88781]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182331392))))[name = string("layers_12_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2510_cast_fp16 = conv(dilations = var_2510_dilations_0, groups = var_2510_groups_0, pad = var_2510_pad_0, pad_type = var_2510_pad_type_0, strides = var_2510_strides_0, weight = layers_12_fc2_outlier_module_weight_to_fp16_sparsified, x = input_103_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_29_cast_fp16 = add(x = var_2504_cast_fp16, y = var_2510_cast_fp16)[name = 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 = string("inputs_53_cast_fp16")];
            int32 var_2520 = const()[name = string("op_2520"), val = int32(3)];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2539_to_fp16 = const()[name = string("op_2539_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_2539_to_fp16, x = inputs_53_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [1280]> obj_53_gamma_0_to_fp16 = const()[name = string("obj_53_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183328320)))];
            tensor<fp16, [1280]> obj_53_beta_0_to_fp16 = const()[name = string("obj_53_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183330944)))];
            fp16 obj_53_epsilon_0_to_fp16 = const()[name = string("obj_53_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = string("obj_53_cast_fp16")];
            string var_2561_pad_type_0 = const()[name = string("op_2561_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2561_strides_0 = const()[name = string("op_2561_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2561_pad_0 = const()[name = string("op_2561_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2561_dilations_0 = const()[name = string("op_2561_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2561_groups_0 = const()[name = string("op_2561_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183333568))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184152832))))[name = string("layers_13_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_13_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184152960)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2561_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_2561_dilations_0, groups = var_2561_groups_0, pad = var_2561_pad_0, pad_type = var_2561_pad_type_0, strides = var_2561_strides_0, weight = layers_13_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = string("op_2561_cast_fp16")];
            string var_2567_pad_type_0 = const()[name = string("op_2567_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2567_strides_0 = const()[name = string("op_2567_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2567_pad_0 = const()[name = string("op_2567_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2567_dilations_0 = const()[name = string("op_2567_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2567_groups_0 = const()[name = string("op_2567_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184209280))), nonzero_data = tensor<fp16, [26789]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184155584))))[name = string("layers_13_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2567_cast_fp16 = conv(dilations = var_2567_dilations_0, groups = var_2567_groups_0, pad = var_2567_pad_0, pad_type = var_2567_pad_type_0, strides = var_2567_strides_0, weight = layers_13_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_53_cast_fp16)[name = string("op_2567_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_27_cast_fp16 = add(x = var_2561_cast_fp16, y = var_2567_cast_fp16)[name = string("query_27_cast_fp16")];
            string var_2576_pad_type_0 = const()[name = string("op_2576_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2576_strides_0 = const()[name = string("op_2576_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2576_pad_0 = const()[name = string("op_2576_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2576_dilations_0 = const()[name = string("op_2576_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2576_groups_0 = const()[name = string("op_2576_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184414144))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185233408))))[name = string("layers_13_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2576_cast_fp16 = conv(dilations = var_2576_dilations_0, groups = var_2576_groups_0, pad = var_2576_pad_0, pad_type = var_2576_pad_type_0, strides = var_2576_strides_0, weight = layers_13_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = string("op_2576_cast_fp16")];
            string var_2582_pad_type_0 = const()[name = string("op_2582_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2582_strides_0 = const()[name = string("op_2582_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2582_pad_0 = const()[name = string("op_2582_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2582_dilations_0 = const()[name = string("op_2582_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2582_groups_0 = const()[name = string("op_2582_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185265728))), nonzero_data = tensor<fp16, [16047]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185233536))))[name = string("layers_13_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2582_cast_fp16 = conv(dilations = var_2582_dilations_0, groups = var_2582_groups_0, pad = var_2582_pad_0, pad_type = var_2582_pad_type_0, strides = var_2582_strides_0, weight = layers_13_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_53_cast_fp16)[name = string("op_2582_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_27_cast_fp16 = add(x = var_2576_cast_fp16, y = var_2582_cast_fp16)[name = string("key_27_cast_fp16")];
            string var_2592_pad_type_0 = const()[name = string("op_2592_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2592_strides_0 = const()[name = string("op_2592_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2592_pad_0 = const()[name = string("op_2592_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2592_dilations_0 = const()[name = string("op_2592_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2592_groups_0 = const()[name = string("op_2592_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185470592))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186289856))))[name = string("layers_13_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_13_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186289984)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2592_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_2592_dilations_0, groups = var_2592_groups_0, pad = var_2592_pad_0, pad_type = var_2592_pad_type_0, strides = var_2592_strides_0, weight = layers_13_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = string("op_2592_cast_fp16")];
            string var_2598_pad_type_0 = const()[name = string("op_2598_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2598_strides_0 = const()[name = string("op_2598_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2598_pad_0 = const()[name = string("op_2598_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2598_dilations_0 = const()[name = string("op_2598_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2598_groups_0 = const()[name = string("op_2598_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186311936))), nonzero_data = tensor<fp16, [9625]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186292608))))[name = string("layers_13_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2598_cast_fp16 = conv(dilations = var_2598_dilations_0, groups = var_2598_groups_0, pad = var_2598_pad_0, pad_type = var_2598_pad_type_0, strides = var_2598_strides_0, weight = layers_13_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_53_cast_fp16)[name = string("op_2598_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_27_cast_fp16 = add(x = var_2592_cast_fp16, y = var_2598_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<int32, [4]> var_2601 = const()[name = string("op_2601"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_27_cast_fp16 = reshape(shape = var_2601, x = query_27_cast_fp16)[name = string("mh_q_27_cast_fp16")];
            fp16 var_2603_to_fp16 = const()[name = string("op_2603_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_2604_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_2603_to_fp16)[name = string("op_2604_cast_fp16")];
            tensor<int32, [4]> var_2605 = const()[name = string("op_2605"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2606_cast_fp16 = reshape(shape = var_2605, x = key_27_cast_fp16)[name = string("op_2606_cast_fp16")];
            bool mh_w_27_transpose_x_0 = const()[name = string("mh_w_27_transpose_x_0"), val = bool(true)];
            bool mh_w_27_transpose_y_0 = const()[name = string("mh_w_27_transpose_y_0"), val = 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_2604_cast_fp16, y = var_2606_cast_fp16)[name = string("mh_w_27_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_2609_cast_fp16 = softmax(axis = var_2520, x = mh_w_27_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [4]> var_2610 = const()[name = string("op_2610"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2611_cast_fp16 = reshape(shape = var_2610, x = value_27_cast_fp16)[name = string("op_2611_cast_fp16")];
            bool attn_27_transpose_x_0 = const()[name = string("attn_27_transpose_x_0"), val = bool(false)];
            bool attn_27_transpose_y_0 = const()[name = string("attn_27_transpose_y_0"), val = 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_2611_cast_fp16, y = var_2609_cast_fp16)[name = string("attn_27_cast_fp16")];
            tensor<int32, [4]> var_2614 = const()[name = string("op_2614"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_105_cast_fp16 = reshape(shape = var_2614, x = attn_27_cast_fp16)[name = string("input_105_cast_fp16")];
            string var_2624_pad_type_0 = const()[name = string("op_2624_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2624_strides_0 = const()[name = string("op_2624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2624_pad_0 = const()[name = string("op_2624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2624_dilations_0 = const()[name = string("op_2624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2624_groups_0 = const()[name = string("op_2624_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186516800))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187336064))))[name = string("layers_13_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_13_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187336192)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2624_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_2624_dilations_0, groups = var_2624_groups_0, pad = var_2624_pad_0, pad_type = var_2624_pad_type_0, strides = var_2624_strides_0, weight = layers_13_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_105_cast_fp16)[name = string("op_2624_cast_fp16")];
            string var_2630_pad_type_0 = const()[name = string("op_2630_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2630_strides_0 = const()[name = string("op_2630_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2630_pad_0 = const()[name = string("op_2630_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2630_dilations_0 = const()[name = string("op_2630_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2630_groups_0 = const()[name = string("op_2630_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187358784))), nonzero_data = tensor<fp16, [9926]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187338816))))[name = string("layers_13_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2630_cast_fp16 = conv(dilations = var_2630_dilations_0, groups = var_2630_groups_0, pad = var_2630_pad_0, pad_type = var_2630_pad_type_0, strides = var_2630_strides_0, weight = layers_13_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_105_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_55_cast_fp16 = add(x = var_2624_cast_fp16, y = var_2630_cast_fp16)[name = 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 = string("inputs_55_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2641_to_fp16 = const()[name = string("op_2641_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_2641_to_fp16, x = inputs_55_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [1280]> input_107_gamma_0_to_fp16 = const()[name = string("input_107_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187563648)))];
            tensor<fp16, [1280]> input_107_beta_0_to_fp16 = const()[name = string("input_107_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187566272)))];
            fp16 input_107_epsilon_0_to_fp16 = const()[name = string("input_107_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            string var_2659_pad_type_0 = const()[name = string("op_2659_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2659_strides_0 = const()[name = string("op_2659_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2659_pad_0 = const()[name = string("op_2659_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2659_dilations_0 = const()[name = string("op_2659_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2659_groups_0 = const()[name = string("op_2659_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_13_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187568896))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190845760))))[name = string("layers_13_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_13_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_13_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190845888)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_2659_cast_fp16 = conv(bias = layers_13_fc1_inlier_module_bias_to_fp16, dilations = var_2659_dilations_0, groups = var_2659_groups_0, pad = var_2659_pad_0, pad_type = var_2659_pad_type_0, strides = var_2659_strides_0, weight = layers_13_fc1_inlier_module_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = string("op_2659_cast_fp16")];
            string var_2665_pad_type_0 = const()[name = string("op_2665_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2665_strides_0 = const()[name = string("op_2665_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2665_pad_0 = const()[name = string("op_2665_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2665_dilations_0 = const()[name = string("op_2665_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2665_groups_0 = const()[name = string("op_2665_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_13_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190903232))), nonzero_data = tensor<fp16, [23480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190856192))))[name = string("layers_13_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_2665_cast_fp16 = conv(dilations = var_2665_dilations_0, groups = var_2665_groups_0, pad = var_2665_pad_0, pad_type = var_2665_pad_type_0, strides = var_2665_strides_0, weight = layers_13_fc1_outlier_module_weight_to_fp16_sparsified, x = input_107_cast_fp16)[name = string("op_2665_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_109_cast_fp16 = add(x = var_2659_cast_fp16, y = var_2665_cast_fp16)[name = string("input_109_cast_fp16")];
            string input_111_mode_0 = const()[name = string("input_111_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
            string var_2676_pad_type_0 = const()[name = string("op_2676_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2676_strides_0 = const()[name = string("op_2676_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2676_pad_0 = const()[name = string("op_2676_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2676_dilations_0 = const()[name = string("op_2676_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2676_groups_0 = const()[name = string("op_2676_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_13_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191722496))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194999360))))[name = string("layers_13_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_13_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_13_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194999488)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2676_cast_fp16 = conv(bias = layers_13_fc2_inlier_module_bias_to_fp16, dilations = var_2676_dilations_0, groups = var_2676_groups_0, pad = var_2676_pad_0, pad_type = var_2676_pad_type_0, strides = var_2676_strides_0, weight = layers_13_fc2_inlier_module_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("op_2676_cast_fp16")];
            string var_2682_pad_type_0 = const()[name = string("op_2682_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2682_strides_0 = const()[name = string("op_2682_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2682_pad_0 = const()[name = string("op_2682_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2682_dilations_0 = const()[name = string("op_2682_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2682_groups_0 = const()[name = string("op_2682_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_13_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195176960))), nonzero_data = tensor<fp16, [87377]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195002112))))[name = string("layers_13_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2682_cast_fp16 = conv(dilations = var_2682_dilations_0, groups = var_2682_groups_0, pad = var_2682_pad_0, pad_type = var_2682_pad_type_0, strides = var_2682_strides_0, weight = layers_13_fc2_outlier_module_weight_to_fp16_sparsified, x = input_111_cast_fp16)[name = string("op_2682_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_31_cast_fp16 = add(x = var_2676_cast_fp16, y = var_2682_cast_fp16)[name = 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 = string("inputs_57_cast_fp16")];
            int32 var_2692 = const()[name = string("op_2692"), val = int32(3)];
            tensor<int32, [1]> out_57_axes_0 = const()[name = string("out_57_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2711_to_fp16 = const()[name = string("op_2711_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_2711_to_fp16, x = inputs_57_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [1280]> obj_57_gamma_0_to_fp16 = const()[name = string("obj_57_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195996224)))];
            tensor<fp16, [1280]> obj_57_beta_0_to_fp16 = const()[name = string("obj_57_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195998848)))];
            fp16 obj_57_epsilon_0_to_fp16 = const()[name = string("obj_57_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = string("obj_57_cast_fp16")];
            string var_2733_pad_type_0 = const()[name = string("op_2733_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2733_strides_0 = const()[name = string("op_2733_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2733_pad_0 = const()[name = string("op_2733_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2733_dilations_0 = const()[name = string("op_2733_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2733_groups_0 = const()[name = string("op_2733_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196001472))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196820736))))[name = string("layers_14_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_14_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_14_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196820864)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2733_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_2733_dilations_0, groups = var_2733_groups_0, pad = var_2733_pad_0, pad_type = var_2733_pad_type_0, strides = var_2733_strides_0, weight = layers_14_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = string("op_2733_cast_fp16")];
            string var_2739_pad_type_0 = const()[name = string("op_2739_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2739_strides_0 = const()[name = string("op_2739_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2739_pad_0 = const()[name = string("op_2739_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2739_dilations_0 = const()[name = string("op_2739_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2739_groups_0 = const()[name = string("op_2739_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196868352))), nonzero_data = tensor<fp16, [22369]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196823488))))[name = string("layers_14_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2739_cast_fp16 = conv(dilations = var_2739_dilations_0, groups = var_2739_groups_0, pad = var_2739_pad_0, pad_type = var_2739_pad_type_0, strides = var_2739_strides_0, weight = layers_14_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_57_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_29_cast_fp16 = add(x = var_2733_cast_fp16, y = var_2739_cast_fp16)[name = string("query_29_cast_fp16")];
            string var_2748_pad_type_0 = const()[name = string("op_2748_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2748_strides_0 = const()[name = string("op_2748_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2748_pad_0 = const()[name = string("op_2748_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2748_dilations_0 = const()[name = string("op_2748_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2748_groups_0 = const()[name = string("op_2748_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197073216))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197892480))))[name = string("layers_14_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2748_cast_fp16 = conv(dilations = var_2748_dilations_0, groups = var_2748_groups_0, pad = var_2748_pad_0, pad_type = var_2748_pad_type_0, strides = var_2748_strides_0, weight = layers_14_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = string("op_2748_cast_fp16")];
            string var_2754_pad_type_0 = const()[name = string("op_2754_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2754_strides_0 = const()[name = string("op_2754_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2754_pad_0 = const()[name = string("op_2754_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2754_dilations_0 = const()[name = string("op_2754_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2754_groups_0 = const()[name = string("op_2754_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197921344))), nonzero_data = tensor<fp16, [14324]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197892608))))[name = string("layers_14_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2754_cast_fp16 = conv(dilations = var_2754_dilations_0, groups = var_2754_groups_0, pad = var_2754_pad_0, pad_type = var_2754_pad_type_0, strides = var_2754_strides_0, weight = layers_14_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_57_cast_fp16)[name = string("op_2754_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_29_cast_fp16 = add(x = var_2748_cast_fp16, y = var_2754_cast_fp16)[name = string("key_29_cast_fp16")];
            string var_2764_pad_type_0 = const()[name = string("op_2764_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2764_strides_0 = const()[name = string("op_2764_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2764_pad_0 = const()[name = string("op_2764_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2764_dilations_0 = const()[name = string("op_2764_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2764_groups_0 = const()[name = string("op_2764_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198126208))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198945472))))[name = string("layers_14_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_14_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_14_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198945600)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2764_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_2764_dilations_0, groups = var_2764_groups_0, pad = var_2764_pad_0, pad_type = var_2764_pad_type_0, strides = var_2764_strides_0, weight = layers_14_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = string("op_2764_cast_fp16")];
            string var_2770_pad_type_0 = const()[name = string("op_2770_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2770_strides_0 = const()[name = string("op_2770_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2770_pad_0 = const()[name = string("op_2770_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2770_dilations_0 = const()[name = string("op_2770_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2770_groups_0 = const()[name = string("op_2770_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198963968))), nonzero_data = tensor<fp16, [7837]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198948224))))[name = string("layers_14_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2770_cast_fp16 = conv(dilations = var_2770_dilations_0, groups = var_2770_groups_0, pad = var_2770_pad_0, pad_type = var_2770_pad_type_0, strides = var_2770_strides_0, weight = layers_14_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_57_cast_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_29_cast_fp16 = add(x = var_2764_cast_fp16, y = var_2770_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<int32, [4]> var_2773 = const()[name = string("op_2773"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_29_cast_fp16 = reshape(shape = var_2773, x = query_29_cast_fp16)[name = string("mh_q_29_cast_fp16")];
            fp16 var_2775_to_fp16 = const()[name = string("op_2775_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_2776_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_2775_to_fp16)[name = string("op_2776_cast_fp16")];
            tensor<int32, [4]> var_2777 = const()[name = string("op_2777"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2778_cast_fp16 = reshape(shape = var_2777, x = key_29_cast_fp16)[name = string("op_2778_cast_fp16")];
            bool mh_w_29_transpose_x_0 = const()[name = string("mh_w_29_transpose_x_0"), val = bool(true)];
            bool mh_w_29_transpose_y_0 = const()[name = string("mh_w_29_transpose_y_0"), val = 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_2776_cast_fp16, y = var_2778_cast_fp16)[name = string("mh_w_29_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_2781_cast_fp16 = softmax(axis = var_2692, x = mh_w_29_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [4]> var_2782 = const()[name = string("op_2782"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2783_cast_fp16 = reshape(shape = var_2782, x = value_29_cast_fp16)[name = string("op_2783_cast_fp16")];
            bool attn_29_transpose_x_0 = const()[name = string("attn_29_transpose_x_0"), val = bool(false)];
            bool attn_29_transpose_y_0 = const()[name = string("attn_29_transpose_y_0"), val = 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_2783_cast_fp16, y = var_2781_cast_fp16)[name = string("attn_29_cast_fp16")];
            tensor<int32, [4]> var_2786 = const()[name = string("op_2786"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_113_cast_fp16 = reshape(shape = var_2786, x = attn_29_cast_fp16)[name = string("input_113_cast_fp16")];
            string var_2796_pad_type_0 = const()[name = string("op_2796_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2796_strides_0 = const()[name = string("op_2796_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2796_pad_0 = const()[name = string("op_2796_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2796_dilations_0 = const()[name = string("op_2796_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2796_groups_0 = const()[name = string("op_2796_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199168832))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199988096))))[name = string("layers_14_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_14_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_14_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199988224)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2796_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_2796_dilations_0, groups = var_2796_groups_0, pad = var_2796_pad_0, pad_type = var_2796_pad_type_0, strides = var_2796_strides_0, weight = layers_14_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = string("op_2796_cast_fp16")];
            string var_2802_pad_type_0 = const()[name = string("op_2802_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2802_strides_0 = const()[name = string("op_2802_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2802_pad_0 = const()[name = string("op_2802_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2802_dilations_0 = const()[name = string("op_2802_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2802_groups_0 = const()[name = string("op_2802_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200005632))), nonzero_data = tensor<fp16, [7359]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199990848))))[name = string("layers_14_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2802_cast_fp16 = conv(dilations = var_2802_dilations_0, groups = var_2802_groups_0, pad = var_2802_pad_0, pad_type = var_2802_pad_type_0, strides = var_2802_strides_0, weight = layers_14_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_113_cast_fp16)[name = string("op_2802_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_59_cast_fp16 = add(x = var_2796_cast_fp16, y = var_2802_cast_fp16)[name = 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 = string("inputs_59_cast_fp16")];
            tensor<int32, [1]> out_59_axes_0 = const()[name = string("out_59_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2813_to_fp16 = const()[name = string("op_2813_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_2813_to_fp16, x = inputs_59_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [1280]> input_115_gamma_0_to_fp16 = const()[name = string("input_115_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200210496)))];
            tensor<fp16, [1280]> input_115_beta_0_to_fp16 = const()[name = string("input_115_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200213120)))];
            fp16 input_115_epsilon_0_to_fp16 = const()[name = string("input_115_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            string var_2831_pad_type_0 = const()[name = string("op_2831_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2831_strides_0 = const()[name = string("op_2831_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2831_pad_0 = const()[name = string("op_2831_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2831_dilations_0 = const()[name = string("op_2831_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2831_groups_0 = const()[name = string("op_2831_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_14_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200215744))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203492608))))[name = string("layers_14_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_14_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_14_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203492736)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_2831_cast_fp16 = conv(bias = layers_14_fc1_inlier_module_bias_to_fp16, dilations = var_2831_dilations_0, groups = var_2831_groups_0, pad = var_2831_pad_0, pad_type = var_2831_pad_type_0, strides = var_2831_strides_0, weight = layers_14_fc1_inlier_module_weight_to_fp16_palettized, x = input_115_cast_fp16)[name = string("op_2831_cast_fp16")];
            string var_2837_pad_type_0 = const()[name = string("op_2837_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2837_strides_0 = const()[name = string("op_2837_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2837_pad_0 = const()[name = string("op_2837_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2837_dilations_0 = const()[name = string("op_2837_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2837_groups_0 = const()[name = string("op_2837_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_14_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203552448))), nonzero_data = tensor<fp16, [24642]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203503040))))[name = string("layers_14_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_2837_cast_fp16 = conv(dilations = var_2837_dilations_0, groups = var_2837_groups_0, pad = var_2837_pad_0, pad_type = var_2837_pad_type_0, strides = var_2837_strides_0, weight = layers_14_fc1_outlier_module_weight_to_fp16_sparsified, x = input_115_cast_fp16)[name = string("op_2837_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_117_cast_fp16 = add(x = var_2831_cast_fp16, y = var_2837_cast_fp16)[name = string("input_117_cast_fp16")];
            string input_119_mode_0 = const()[name = string("input_119_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = string("input_119_cast_fp16")];
            string var_2848_pad_type_0 = const()[name = string("op_2848_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2848_strides_0 = const()[name = string("op_2848_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2848_pad_0 = const()[name = string("op_2848_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2848_dilations_0 = const()[name = string("op_2848_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2848_groups_0 = const()[name = string("op_2848_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_14_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204371712))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207648576))))[name = string("layers_14_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_14_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_14_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207648704)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2848_cast_fp16 = conv(bias = layers_14_fc2_inlier_module_bias_to_fp16, dilations = var_2848_dilations_0, groups = var_2848_groups_0, pad = var_2848_pad_0, pad_type = var_2848_pad_type_0, strides = var_2848_strides_0, weight = layers_14_fc2_inlier_module_weight_to_fp16_palettized, x = input_119_cast_fp16)[name = string("op_2848_cast_fp16")];
            string var_2854_pad_type_0 = const()[name = string("op_2854_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2854_strides_0 = const()[name = string("op_2854_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2854_pad_0 = const()[name = string("op_2854_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2854_dilations_0 = const()[name = string("op_2854_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2854_groups_0 = const()[name = string("op_2854_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_14_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207800128))), nonzero_data = tensor<fp16, [74342]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207651328))))[name = string("layers_14_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2854_cast_fp16 = conv(dilations = var_2854_dilations_0, groups = var_2854_groups_0, pad = var_2854_pad_0, pad_type = var_2854_pad_type_0, strides = var_2854_strides_0, weight = layers_14_fc2_outlier_module_weight_to_fp16_sparsified, x = input_119_cast_fp16)[name = string("op_2854_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_33_cast_fp16 = add(x = var_2848_cast_fp16, y = var_2854_cast_fp16)[name = 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 = string("inputs_61_cast_fp16")];
            int32 var_2864 = const()[name = string("op_2864"), val = int32(3)];
            tensor<int32, [1]> out_61_axes_0 = const()[name = string("out_61_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2883_to_fp16 = const()[name = string("op_2883_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_2883_to_fp16, x = inputs_61_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [1280]> obj_61_gamma_0_to_fp16 = const()[name = string("obj_61_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208619392)))];
            tensor<fp16, [1280]> obj_61_beta_0_to_fp16 = const()[name = string("obj_61_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208622016)))];
            fp16 obj_61_epsilon_0_to_fp16 = const()[name = string("obj_61_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = string("obj_61_cast_fp16")];
            string var_2905_pad_type_0 = const()[name = string("op_2905_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2905_strides_0 = const()[name = string("op_2905_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2905_pad_0 = const()[name = string("op_2905_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2905_dilations_0 = const()[name = string("op_2905_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2905_groups_0 = const()[name = string("op_2905_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208624640))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209443904))))[name = string("layers_15_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_15_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_15_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209444032)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2905_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_2905_dilations_0, groups = var_2905_groups_0, pad = var_2905_pad_0, pad_type = var_2905_pad_type_0, strides = var_2905_strides_0, weight = layers_15_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = string("op_2905_cast_fp16")];
            string var_2911_pad_type_0 = const()[name = string("op_2911_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2911_strides_0 = const()[name = string("op_2911_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2911_pad_0 = const()[name = string("op_2911_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2911_dilations_0 = const()[name = string("op_2911_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2911_groups_0 = const()[name = string("op_2911_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209492416))), nonzero_data = tensor<fp16, [22840]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209446656))))[name = string("layers_15_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2911_cast_fp16 = conv(dilations = var_2911_dilations_0, groups = var_2911_groups_0, pad = var_2911_pad_0, pad_type = var_2911_pad_type_0, strides = var_2911_strides_0, weight = layers_15_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_61_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_31_cast_fp16 = add(x = var_2905_cast_fp16, y = var_2911_cast_fp16)[name = string("query_31_cast_fp16")];
            string var_2920_pad_type_0 = const()[name = string("op_2920_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2920_strides_0 = const()[name = string("op_2920_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2920_pad_0 = const()[name = string("op_2920_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2920_dilations_0 = const()[name = string("op_2920_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2920_groups_0 = const()[name = string("op_2920_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209697280))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210516544))))[name = string("layers_15_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2920_cast_fp16 = conv(dilations = var_2920_dilations_0, groups = var_2920_groups_0, pad = var_2920_pad_0, pad_type = var_2920_pad_type_0, strides = var_2920_strides_0, weight = layers_15_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = string("op_2920_cast_fp16")];
            string var_2926_pad_type_0 = const()[name = string("op_2926_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2926_strides_0 = const()[name = string("op_2926_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2926_pad_0 = const()[name = string("op_2926_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2926_dilations_0 = const()[name = string("op_2926_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2926_groups_0 = const()[name = string("op_2926_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210542144))), nonzero_data = tensor<fp16, [12688]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210516672))))[name = string("layers_15_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2926_cast_fp16 = conv(dilations = var_2926_dilations_0, groups = var_2926_groups_0, pad = var_2926_pad_0, pad_type = var_2926_pad_type_0, strides = var_2926_strides_0, weight = layers_15_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_61_cast_fp16)[name = string("op_2926_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_31_cast_fp16 = add(x = var_2920_cast_fp16, y = var_2926_cast_fp16)[name = string("key_31_cast_fp16")];
            string var_2936_pad_type_0 = const()[name = string("op_2936_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2936_strides_0 = const()[name = string("op_2936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2936_pad_0 = const()[name = string("op_2936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2936_dilations_0 = const()[name = string("op_2936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2936_groups_0 = const()[name = string("op_2936_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210747008))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211566272))))[name = string("layers_15_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_15_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_15_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211566400)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2936_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_2936_dilations_0, groups = var_2936_groups_0, pad = var_2936_pad_0, pad_type = var_2936_pad_type_0, strides = var_2936_strides_0, weight = layers_15_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = string("op_2936_cast_fp16")];
            string var_2942_pad_type_0 = const()[name = string("op_2942_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2942_strides_0 = const()[name = string("op_2942_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2942_pad_0 = const()[name = string("op_2942_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2942_dilations_0 = const()[name = string("op_2942_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2942_groups_0 = const()[name = string("op_2942_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211586240))), nonzero_data = tensor<fp16, [8563]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211569024))))[name = string("layers_15_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2942_cast_fp16 = conv(dilations = var_2942_dilations_0, groups = var_2942_groups_0, pad = var_2942_pad_0, pad_type = var_2942_pad_type_0, strides = var_2942_strides_0, weight = layers_15_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_61_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_31_cast_fp16 = add(x = var_2936_cast_fp16, y = var_2942_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<int32, [4]> var_2945 = const()[name = string("op_2945"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_31_cast_fp16 = reshape(shape = var_2945, x = query_31_cast_fp16)[name = string("mh_q_31_cast_fp16")];
            fp16 var_2947_to_fp16 = const()[name = string("op_2947_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_2948_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_2947_to_fp16)[name = string("op_2948_cast_fp16")];
            tensor<int32, [4]> var_2949 = const()[name = string("op_2949"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2950_cast_fp16 = reshape(shape = var_2949, x = key_31_cast_fp16)[name = string("op_2950_cast_fp16")];
            bool mh_w_31_transpose_x_0 = const()[name = string("mh_w_31_transpose_x_0"), val = bool(true)];
            bool mh_w_31_transpose_y_0 = const()[name = string("mh_w_31_transpose_y_0"), val = 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_2948_cast_fp16, y = var_2950_cast_fp16)[name = string("mh_w_31_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_2953_cast_fp16 = softmax(axis = var_2864, x = mh_w_31_cast_fp16)[name = string("op_2953_cast_fp16")];
            tensor<int32, [4]> var_2954 = const()[name = string("op_2954"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2955_cast_fp16 = reshape(shape = var_2954, x = value_31_cast_fp16)[name = string("op_2955_cast_fp16")];
            bool attn_31_transpose_x_0 = const()[name = string("attn_31_transpose_x_0"), val = bool(false)];
            bool attn_31_transpose_y_0 = const()[name = string("attn_31_transpose_y_0"), val = 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_2955_cast_fp16, y = var_2953_cast_fp16)[name = string("attn_31_cast_fp16")];
            tensor<int32, [4]> var_2958 = const()[name = string("op_2958"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_121_cast_fp16 = reshape(shape = var_2958, x = attn_31_cast_fp16)[name = string("input_121_cast_fp16")];
            string var_2968_pad_type_0 = const()[name = string("op_2968_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2968_strides_0 = const()[name = string("op_2968_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2968_pad_0 = const()[name = string("op_2968_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2968_dilations_0 = const()[name = string("op_2968_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2968_groups_0 = const()[name = string("op_2968_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211791104))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212610368))))[name = string("layers_15_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_15_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_15_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212610496)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_2968_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_2968_dilations_0, groups = var_2968_groups_0, pad = var_2968_pad_0, pad_type = var_2968_pad_type_0, strides = var_2968_strides_0, weight = layers_15_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("op_2968_cast_fp16")];
            string var_2974_pad_type_0 = const()[name = string("op_2974_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2974_strides_0 = const()[name = string("op_2974_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2974_pad_0 = const()[name = string("op_2974_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2974_dilations_0 = const()[name = string("op_2974_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2974_groups_0 = const()[name = string("op_2974_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212633664))), nonzero_data = tensor<fp16, [10240]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212613120))))[name = string("layers_15_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_2974_cast_fp16 = conv(dilations = var_2974_dilations_0, groups = var_2974_groups_0, pad = var_2974_pad_0, pad_type = var_2974_pad_type_0, strides = var_2974_strides_0, weight = layers_15_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("op_2974_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_63_cast_fp16 = add(x = var_2968_cast_fp16, y = var_2974_cast_fp16)[name = 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 = string("inputs_63_cast_fp16")];
            tensor<int32, [1]> out_63_axes_0 = const()[name = string("out_63_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2985_to_fp16 = const()[name = string("op_2985_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2985_to_fp16, x = inputs_63_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [1280]> input_123_gamma_0_to_fp16 = const()[name = string("input_123_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212838528)))];
            tensor<fp16, [1280]> input_123_beta_0_to_fp16 = const()[name = string("input_123_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212841152)))];
            fp16 input_123_epsilon_0_to_fp16 = const()[name = string("input_123_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            string var_3003_pad_type_0 = const()[name = string("op_3003_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3003_strides_0 = const()[name = string("op_3003_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3003_pad_0 = const()[name = string("op_3003_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3003_dilations_0 = const()[name = string("op_3003_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3003_groups_0 = const()[name = string("op_3003_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_15_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212843776))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216120640))))[name = string("layers_15_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_15_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_15_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216120768)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_3003_cast_fp16 = conv(bias = layers_15_fc1_inlier_module_bias_to_fp16, dilations = var_3003_dilations_0, groups = var_3003_groups_0, pad = var_3003_pad_0, pad_type = var_3003_pad_type_0, strides = var_3003_strides_0, weight = layers_15_fc1_inlier_module_weight_to_fp16_palettized, x = input_123_cast_fp16)[name = string("op_3003_cast_fp16")];
            string var_3009_pad_type_0 = const()[name = string("op_3009_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3009_strides_0 = const()[name = string("op_3009_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3009_pad_0 = const()[name = string("op_3009_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3009_dilations_0 = const()[name = string("op_3009_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3009_groups_0 = const()[name = string("op_3009_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_15_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216176896))), nonzero_data = tensor<fp16, [22855]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216131072))))[name = string("layers_15_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_3009_cast_fp16 = conv(dilations = var_3009_dilations_0, groups = var_3009_groups_0, pad = var_3009_pad_0, pad_type = var_3009_pad_type_0, strides = var_3009_strides_0, weight = layers_15_fc1_outlier_module_weight_to_fp16_sparsified, x = input_123_cast_fp16)[name = string("op_3009_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_125_cast_fp16 = add(x = var_3003_cast_fp16, y = var_3009_cast_fp16)[name = string("input_125_cast_fp16")];
            string input_127_mode_0 = const()[name = string("input_127_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = input_125_cast_fp16)[name = string("input_127_cast_fp16")];
            string var_3020_pad_type_0 = const()[name = string("op_3020_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3020_strides_0 = const()[name = string("op_3020_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3020_pad_0 = const()[name = string("op_3020_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3020_dilations_0 = const()[name = string("op_3020_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3020_groups_0 = const()[name = string("op_3020_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_15_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216996160))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220273024))))[name = string("layers_15_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_15_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_15_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220273152)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3020_cast_fp16 = conv(bias = layers_15_fc2_inlier_module_bias_to_fp16, dilations = var_3020_dilations_0, groups = var_3020_groups_0, pad = var_3020_pad_0, pad_type = var_3020_pad_type_0, strides = var_3020_strides_0, weight = layers_15_fc2_inlier_module_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = string("op_3020_cast_fp16")];
            string var_3026_pad_type_0 = const()[name = string("op_3026_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3026_strides_0 = const()[name = string("op_3026_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3026_pad_0 = const()[name = string("op_3026_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3026_dilations_0 = const()[name = string("op_3026_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3026_groups_0 = const()[name = string("op_3026_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_15_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220449024))), nonzero_data = tensor<fp16, [86580]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220275776))))[name = string("layers_15_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3026_cast_fp16 = conv(dilations = var_3026_dilations_0, groups = var_3026_groups_0, pad = var_3026_pad_0, pad_type = var_3026_pad_type_0, strides = var_3026_strides_0, weight = layers_15_fc2_outlier_module_weight_to_fp16_sparsified, x = input_127_cast_fp16)[name = string("op_3026_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_35_cast_fp16 = add(x = var_3020_cast_fp16, y = var_3026_cast_fp16)[name = 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 = string("inputs_65_cast_fp16")];
            int32 var_3036 = const()[name = string("op_3036"), val = int32(3)];
            tensor<int32, [1]> out_65_axes_0 = const()[name = string("out_65_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3055_to_fp16 = const()[name = string("op_3055_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_3055_to_fp16, x = inputs_65_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [1280]> obj_65_gamma_0_to_fp16 = const()[name = string("obj_65_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221268288)))];
            tensor<fp16, [1280]> obj_65_beta_0_to_fp16 = const()[name = string("obj_65_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221270912)))];
            fp16 obj_65_epsilon_0_to_fp16 = const()[name = string("obj_65_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = string("obj_65_cast_fp16")];
            string var_3077_pad_type_0 = const()[name = string("op_3077_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3077_strides_0 = const()[name = string("op_3077_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3077_pad_0 = const()[name = string("op_3077_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3077_dilations_0 = const()[name = string("op_3077_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3077_groups_0 = const()[name = string("op_3077_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221273536))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222092800))))[name = string("layers_16_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_16_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_16_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222092928)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3077_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_3077_dilations_0, groups = var_3077_groups_0, pad = var_3077_pad_0, pad_type = var_3077_pad_type_0, strides = var_3077_strides_0, weight = layers_16_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = string("op_3077_cast_fp16")];
            string var_3083_pad_type_0 = const()[name = string("op_3083_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3083_strides_0 = const()[name = string("op_3083_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3083_pad_0 = const()[name = string("op_3083_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3083_dilations_0 = const()[name = string("op_3083_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3083_groups_0 = const()[name = string("op_3083_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222164416))), nonzero_data = tensor<fp16, [34395]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222095552))))[name = string("layers_16_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3083_cast_fp16 = conv(dilations = var_3083_dilations_0, groups = var_3083_groups_0, pad = var_3083_pad_0, pad_type = var_3083_pad_type_0, strides = var_3083_strides_0, weight = layers_16_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_65_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_33_cast_fp16 = add(x = var_3077_cast_fp16, y = var_3083_cast_fp16)[name = string("query_33_cast_fp16")];
            string var_3092_pad_type_0 = const()[name = string("op_3092_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3092_strides_0 = const()[name = string("op_3092_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3092_pad_0 = const()[name = string("op_3092_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3092_dilations_0 = const()[name = string("op_3092_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3092_groups_0 = const()[name = string("op_3092_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222369280))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223188544))))[name = string("layers_16_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3092_cast_fp16 = conv(dilations = var_3092_dilations_0, groups = var_3092_groups_0, pad = var_3092_pad_0, pad_type = var_3092_pad_type_0, strides = var_3092_strides_0, weight = layers_16_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = string("op_3092_cast_fp16")];
            string var_3098_pad_type_0 = const()[name = string("op_3098_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3098_strides_0 = const()[name = string("op_3098_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3098_pad_0 = const()[name = string("op_3098_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3098_dilations_0 = const()[name = string("op_3098_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3098_groups_0 = const()[name = string("op_3098_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223217792))), nonzero_data = tensor<fp16, [14518]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223188672))))[name = string("layers_16_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3098_cast_fp16 = conv(dilations = var_3098_dilations_0, groups = var_3098_groups_0, pad = var_3098_pad_0, pad_type = var_3098_pad_type_0, strides = var_3098_strides_0, weight = layers_16_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_65_cast_fp16)[name = string("op_3098_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_33_cast_fp16 = add(x = var_3092_cast_fp16, y = var_3098_cast_fp16)[name = string("key_33_cast_fp16")];
            string var_3108_pad_type_0 = const()[name = string("op_3108_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3108_strides_0 = const()[name = string("op_3108_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3108_pad_0 = const()[name = string("op_3108_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3108_dilations_0 = const()[name = string("op_3108_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3108_groups_0 = const()[name = string("op_3108_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223422656))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224241920))))[name = string("layers_16_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_16_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_16_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224242048)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3108_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_3108_dilations_0, groups = var_3108_groups_0, pad = var_3108_pad_0, pad_type = var_3108_pad_type_0, strides = var_3108_strides_0, weight = layers_16_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = string("op_3108_cast_fp16")];
            string var_3114_pad_type_0 = const()[name = string("op_3114_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3114_strides_0 = const()[name = string("op_3114_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3114_pad_0 = const()[name = string("op_3114_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3114_dilations_0 = const()[name = string("op_3114_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3114_groups_0 = const()[name = string("op_3114_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224261888))), nonzero_data = tensor<fp16, [8562]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224244672))))[name = string("layers_16_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3114_cast_fp16 = conv(dilations = var_3114_dilations_0, groups = var_3114_groups_0, pad = var_3114_pad_0, pad_type = var_3114_pad_type_0, strides = var_3114_strides_0, weight = layers_16_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_65_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_33_cast_fp16 = add(x = var_3108_cast_fp16, y = var_3114_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<int32, [4]> var_3117 = const()[name = string("op_3117"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_33_cast_fp16 = reshape(shape = var_3117, x = query_33_cast_fp16)[name = string("mh_q_33_cast_fp16")];
            fp16 var_3119_to_fp16 = const()[name = string("op_3119_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_3120_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_3119_to_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [4]> var_3121 = const()[name = string("op_3121"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3122_cast_fp16 = reshape(shape = var_3121, x = key_33_cast_fp16)[name = string("op_3122_cast_fp16")];
            bool mh_w_33_transpose_x_0 = const()[name = string("mh_w_33_transpose_x_0"), val = bool(true)];
            bool mh_w_33_transpose_y_0 = const()[name = string("mh_w_33_transpose_y_0"), val = 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_3120_cast_fp16, y = var_3122_cast_fp16)[name = string("mh_w_33_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_3125_cast_fp16 = softmax(axis = var_3036, x = mh_w_33_cast_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [4]> var_3126 = const()[name = string("op_3126"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3127_cast_fp16 = reshape(shape = var_3126, x = value_33_cast_fp16)[name = string("op_3127_cast_fp16")];
            bool attn_33_transpose_x_0 = const()[name = string("attn_33_transpose_x_0"), val = bool(false)];
            bool attn_33_transpose_y_0 = const()[name = string("attn_33_transpose_y_0"), val = 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_3127_cast_fp16, y = var_3125_cast_fp16)[name = string("attn_33_cast_fp16")];
            tensor<int32, [4]> var_3130 = const()[name = string("op_3130"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_129_cast_fp16 = reshape(shape = var_3130, x = attn_33_cast_fp16)[name = string("input_129_cast_fp16")];
            string var_3140_pad_type_0 = const()[name = string("op_3140_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3140_strides_0 = const()[name = string("op_3140_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3140_pad_0 = const()[name = string("op_3140_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3140_dilations_0 = const()[name = string("op_3140_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3140_groups_0 = const()[name = string("op_3140_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224466752))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225286016))))[name = string("layers_16_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_16_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_16_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225286144)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3140_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_3140_dilations_0, groups = var_3140_groups_0, pad = var_3140_pad_0, pad_type = var_3140_pad_type_0, strides = var_3140_strides_0, weight = layers_16_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_129_cast_fp16)[name = string("op_3140_cast_fp16")];
            string var_3146_pad_type_0 = const()[name = string("op_3146_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3146_strides_0 = const()[name = string("op_3146_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3146_pad_0 = const()[name = string("op_3146_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3146_dilations_0 = const()[name = string("op_3146_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3146_groups_0 = const()[name = string("op_3146_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225307136))), nonzero_data = tensor<fp16, [9123]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225288768))))[name = string("layers_16_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3146_cast_fp16 = conv(dilations = var_3146_dilations_0, groups = var_3146_groups_0, pad = var_3146_pad_0, pad_type = var_3146_pad_type_0, strides = var_3146_strides_0, weight = layers_16_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_129_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_67_cast_fp16 = add(x = var_3140_cast_fp16, y = var_3146_cast_fp16)[name = 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 = string("inputs_67_cast_fp16")];
            tensor<int32, [1]> out_67_axes_0 = const()[name = string("out_67_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3157_to_fp16 = const()[name = string("op_3157_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_3157_to_fp16, x = inputs_67_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [1280]> input_131_gamma_0_to_fp16 = const()[name = string("input_131_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225512000)))];
            tensor<fp16, [1280]> input_131_beta_0_to_fp16 = const()[name = string("input_131_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225514624)))];
            fp16 input_131_epsilon_0_to_fp16 = const()[name = string("input_131_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            string var_3175_pad_type_0 = const()[name = string("op_3175_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3175_strides_0 = const()[name = string("op_3175_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3175_pad_0 = const()[name = string("op_3175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3175_dilations_0 = const()[name = string("op_3175_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3175_groups_0 = const()[name = string("op_3175_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_16_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225517248))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228794112))))[name = string("layers_16_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_16_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_16_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228794240)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_3175_cast_fp16 = conv(bias = layers_16_fc1_inlier_module_bias_to_fp16, dilations = var_3175_dilations_0, groups = var_3175_groups_0, pad = var_3175_pad_0, pad_type = var_3175_pad_type_0, strides = var_3175_strides_0, weight = layers_16_fc1_inlier_module_weight_to_fp16_palettized, x = input_131_cast_fp16)[name = string("op_3175_cast_fp16")];
            string var_3181_pad_type_0 = const()[name = string("op_3181_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3181_strides_0 = const()[name = string("op_3181_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3181_pad_0 = const()[name = string("op_3181_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3181_dilations_0 = const()[name = string("op_3181_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3181_groups_0 = const()[name = string("op_3181_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_16_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228879488))), nonzero_data = tensor<fp16, [37435]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228804544))))[name = string("layers_16_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_3181_cast_fp16 = conv(dilations = var_3181_dilations_0, groups = var_3181_groups_0, pad = var_3181_pad_0, pad_type = var_3181_pad_type_0, strides = var_3181_strides_0, weight = layers_16_fc1_outlier_module_weight_to_fp16_sparsified, x = input_131_cast_fp16)[name = string("op_3181_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_133_cast_fp16 = add(x = var_3175_cast_fp16, y = var_3181_cast_fp16)[name = string("input_133_cast_fp16")];
            string input_135_mode_0 = const()[name = string("input_135_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_135_cast_fp16 = gelu(mode = input_135_mode_0, x = input_133_cast_fp16)[name = string("input_135_cast_fp16")];
            string var_3192_pad_type_0 = const()[name = string("op_3192_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3192_strides_0 = const()[name = string("op_3192_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3192_pad_0 = const()[name = string("op_3192_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3192_dilations_0 = const()[name = string("op_3192_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3192_groups_0 = const()[name = string("op_3192_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_16_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229698752))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232975616))))[name = string("layers_16_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_16_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_16_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232975744)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3192_cast_fp16 = conv(bias = layers_16_fc2_inlier_module_bias_to_fp16, dilations = var_3192_dilations_0, groups = var_3192_groups_0, pad = var_3192_pad_0, pad_type = var_3192_pad_type_0, strides = var_3192_strides_0, weight = layers_16_fc2_inlier_module_weight_to_fp16_palettized, x = input_135_cast_fp16)[name = string("op_3192_cast_fp16")];
            string var_3198_pad_type_0 = const()[name = string("op_3198_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3198_strides_0 = const()[name = string("op_3198_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3198_pad_0 = const()[name = string("op_3198_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3198_dilations_0 = const()[name = string("op_3198_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3198_groups_0 = const()[name = string("op_3198_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_16_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233104448))), nonzero_data = tensor<fp16, [62994]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232978368))))[name = string("layers_16_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3198_cast_fp16 = conv(dilations = var_3198_dilations_0, groups = var_3198_groups_0, pad = var_3198_pad_0, pad_type = var_3198_pad_type_0, strides = var_3198_strides_0, weight = layers_16_fc2_outlier_module_weight_to_fp16_sparsified, x = input_135_cast_fp16)[name = string("op_3198_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_37_cast_fp16 = add(x = var_3192_cast_fp16, y = var_3198_cast_fp16)[name = 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 = string("inputs_69_cast_fp16")];
            int32 var_3208 = const()[name = string("op_3208"), val = int32(3)];
            tensor<int32, [1]> out_69_axes_0 = const()[name = string("out_69_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3227_to_fp16 = const()[name = string("op_3227_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_3227_to_fp16, x = inputs_69_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [1280]> obj_69_gamma_0_to_fp16 = const()[name = string("obj_69_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233923712)))];
            tensor<fp16, [1280]> obj_69_beta_0_to_fp16 = const()[name = string("obj_69_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233926336)))];
            fp16 obj_69_epsilon_0_to_fp16 = const()[name = string("obj_69_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = string("obj_69_cast_fp16")];
            string var_3249_pad_type_0 = const()[name = string("op_3249_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3249_strides_0 = const()[name = string("op_3249_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3249_pad_0 = const()[name = string("op_3249_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3249_dilations_0 = const()[name = string("op_3249_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3249_groups_0 = const()[name = string("op_3249_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233928960))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234748224))))[name = string("layers_17_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_17_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_17_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234748352)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3249_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_3249_dilations_0, groups = var_3249_groups_0, pad = var_3249_pad_0, pad_type = var_3249_pad_type_0, strides = var_3249_strides_0, weight = layers_17_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = string("op_3249_cast_fp16")];
            string var_3255_pad_type_0 = const()[name = string("op_3255_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3255_strides_0 = const()[name = string("op_3255_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3255_pad_0 = const()[name = string("op_3255_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3255_dilations_0 = const()[name = string("op_3255_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3255_groups_0 = const()[name = string("op_3255_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234801152))), nonzero_data = tensor<fp16, [25054]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234750976))))[name = string("layers_17_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3255_cast_fp16 = conv(dilations = var_3255_dilations_0, groups = var_3255_groups_0, pad = var_3255_pad_0, pad_type = var_3255_pad_type_0, strides = var_3255_strides_0, weight = layers_17_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_69_cast_fp16)[name = string("op_3255_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_35_cast_fp16 = add(x = var_3249_cast_fp16, y = var_3255_cast_fp16)[name = string("query_35_cast_fp16")];
            string var_3264_pad_type_0 = const()[name = string("op_3264_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3264_strides_0 = const()[name = string("op_3264_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3264_pad_0 = const()[name = string("op_3264_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3264_dilations_0 = const()[name = string("op_3264_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3264_groups_0 = const()[name = string("op_3264_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235006016))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235825280))))[name = string("layers_17_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3264_cast_fp16 = conv(dilations = var_3264_dilations_0, groups = var_3264_groups_0, pad = var_3264_pad_0, pad_type = var_3264_pad_type_0, strides = var_3264_strides_0, weight = layers_17_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = string("op_3264_cast_fp16")];
            string var_3270_pad_type_0 = const()[name = string("op_3270_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3270_strides_0 = const()[name = string("op_3270_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3270_pad_0 = const()[name = string("op_3270_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3270_dilations_0 = const()[name = string("op_3270_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3270_groups_0 = const()[name = string("op_3270_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235854656))), nonzero_data = tensor<fp16, [14566]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235825408))))[name = string("layers_17_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3270_cast_fp16 = conv(dilations = var_3270_dilations_0, groups = var_3270_groups_0, pad = var_3270_pad_0, pad_type = var_3270_pad_type_0, strides = var_3270_strides_0, weight = layers_17_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_69_cast_fp16)[name = string("op_3270_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_35_cast_fp16 = add(x = var_3264_cast_fp16, y = var_3270_cast_fp16)[name = string("key_35_cast_fp16")];
            string var_3280_pad_type_0 = const()[name = string("op_3280_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3280_strides_0 = const()[name = string("op_3280_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3280_pad_0 = const()[name = string("op_3280_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3280_dilations_0 = const()[name = string("op_3280_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3280_groups_0 = const()[name = string("op_3280_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236059520))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236878784))))[name = string("layers_17_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_17_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_17_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236878912)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3280_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_3280_dilations_0, groups = var_3280_groups_0, pad = var_3280_pad_0, pad_type = var_3280_pad_type_0, strides = var_3280_strides_0, weight = layers_17_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = string("op_3280_cast_fp16")];
            string var_3286_pad_type_0 = const()[name = string("op_3286_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3286_strides_0 = const()[name = string("op_3286_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3286_pad_0 = const()[name = string("op_3286_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3286_dilations_0 = const()[name = string("op_3286_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3286_groups_0 = const()[name = string("op_3286_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236897920))), nonzero_data = tensor<fp16, [8130]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236881536))))[name = string("layers_17_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3286_cast_fp16 = conv(dilations = var_3286_dilations_0, groups = var_3286_groups_0, pad = var_3286_pad_0, pad_type = var_3286_pad_type_0, strides = var_3286_strides_0, weight = layers_17_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_69_cast_fp16)[name = string("op_3286_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_35_cast_fp16 = add(x = var_3280_cast_fp16, y = var_3286_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<int32, [4]> var_3289 = const()[name = string("op_3289"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_35_cast_fp16 = reshape(shape = var_3289, x = query_35_cast_fp16)[name = string("mh_q_35_cast_fp16")];
            fp16 var_3291_to_fp16 = const()[name = string("op_3291_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_3292_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_3291_to_fp16)[name = string("op_3292_cast_fp16")];
            tensor<int32, [4]> var_3293 = const()[name = string("op_3293"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3294_cast_fp16 = reshape(shape = var_3293, x = key_35_cast_fp16)[name = string("op_3294_cast_fp16")];
            bool mh_w_35_transpose_x_0 = const()[name = string("mh_w_35_transpose_x_0"), val = bool(true)];
            bool mh_w_35_transpose_y_0 = const()[name = string("mh_w_35_transpose_y_0"), val = 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_3292_cast_fp16, y = var_3294_cast_fp16)[name = string("mh_w_35_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_3297_cast_fp16 = softmax(axis = var_3208, x = mh_w_35_cast_fp16)[name = string("op_3297_cast_fp16")];
            tensor<int32, [4]> var_3298 = const()[name = string("op_3298"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3299_cast_fp16 = reshape(shape = var_3298, x = value_35_cast_fp16)[name = string("op_3299_cast_fp16")];
            bool attn_35_transpose_x_0 = const()[name = string("attn_35_transpose_x_0"), val = bool(false)];
            bool attn_35_transpose_y_0 = const()[name = string("attn_35_transpose_y_0"), val = 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_3299_cast_fp16, y = var_3297_cast_fp16)[name = string("attn_35_cast_fp16")];
            tensor<int32, [4]> var_3302 = const()[name = string("op_3302"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_137_cast_fp16 = reshape(shape = var_3302, x = attn_35_cast_fp16)[name = string("input_137_cast_fp16")];
            string var_3312_pad_type_0 = const()[name = string("op_3312_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3312_strides_0 = const()[name = string("op_3312_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3312_pad_0 = const()[name = string("op_3312_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3312_dilations_0 = const()[name = string("op_3312_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3312_groups_0 = const()[name = string("op_3312_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237102784))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237922048))))[name = string("layers_17_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_17_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_17_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237922176)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3312_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_3312_dilations_0, groups = var_3312_groups_0, pad = var_3312_pad_0, pad_type = var_3312_pad_type_0, strides = var_3312_strides_0, weight = layers_17_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_137_cast_fp16)[name = string("op_3312_cast_fp16")];
            string var_3318_pad_type_0 = const()[name = string("op_3318_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3318_strides_0 = const()[name = string("op_3318_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3318_pad_0 = const()[name = string("op_3318_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3318_dilations_0 = const()[name = string("op_3318_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3318_groups_0 = const()[name = string("op_3318_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237940736))), nonzero_data = tensor<fp16, [7928]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237924800))))[name = string("layers_17_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3318_cast_fp16 = conv(dilations = var_3318_dilations_0, groups = var_3318_groups_0, pad = var_3318_pad_0, pad_type = var_3318_pad_type_0, strides = var_3318_strides_0, weight = layers_17_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_137_cast_fp16)[name = string("op_3318_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_71_cast_fp16 = add(x = var_3312_cast_fp16, y = var_3318_cast_fp16)[name = 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 = string("inputs_71_cast_fp16")];
            tensor<int32, [1]> out_71_axes_0 = const()[name = string("out_71_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3329_to_fp16 = const()[name = string("op_3329_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_3329_to_fp16, x = inputs_71_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [1280]> input_139_gamma_0_to_fp16 = const()[name = string("input_139_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238145600)))];
            tensor<fp16, [1280]> input_139_beta_0_to_fp16 = const()[name = string("input_139_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238148224)))];
            fp16 input_139_epsilon_0_to_fp16 = const()[name = string("input_139_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            string var_3347_pad_type_0 = const()[name = string("op_3347_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3347_strides_0 = const()[name = string("op_3347_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3347_pad_0 = const()[name = string("op_3347_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3347_dilations_0 = const()[name = string("op_3347_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3347_groups_0 = const()[name = string("op_3347_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_17_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238150848))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241427712))))[name = string("layers_17_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_17_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_17_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241427840)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_3347_cast_fp16 = conv(bias = layers_17_fc1_inlier_module_bias_to_fp16, dilations = var_3347_dilations_0, groups = var_3347_groups_0, pad = var_3347_pad_0, pad_type = var_3347_pad_type_0, strides = var_3347_strides_0, weight = layers_17_fc1_inlier_module_weight_to_fp16_palettized, x = input_139_cast_fp16)[name = string("op_3347_cast_fp16")];
            string var_3353_pad_type_0 = const()[name = string("op_3353_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3353_strides_0 = const()[name = string("op_3353_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3353_pad_0 = const()[name = string("op_3353_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3353_dilations_0 = const()[name = string("op_3353_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3353_groups_0 = const()[name = string("op_3353_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_17_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241516992))), nonzero_data = tensor<fp16, [39374]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241438144))))[name = string("layers_17_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_3353_cast_fp16 = conv(dilations = var_3353_dilations_0, groups = var_3353_groups_0, pad = var_3353_pad_0, pad_type = var_3353_pad_type_0, strides = var_3353_strides_0, weight = layers_17_fc1_outlier_module_weight_to_fp16_sparsified, x = input_139_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_141_cast_fp16 = add(x = var_3347_cast_fp16, y = var_3353_cast_fp16)[name = string("input_141_cast_fp16")];
            string input_143_mode_0 = const()[name = string("input_143_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = input_141_cast_fp16)[name = string("input_143_cast_fp16")];
            string var_3364_pad_type_0 = const()[name = string("op_3364_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3364_strides_0 = const()[name = string("op_3364_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3364_pad_0 = const()[name = string("op_3364_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3364_dilations_0 = const()[name = string("op_3364_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3364_groups_0 = const()[name = string("op_3364_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_17_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242336256))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245613120))))[name = string("layers_17_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_17_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_17_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245613248)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3364_cast_fp16 = conv(bias = layers_17_fc2_inlier_module_bias_to_fp16, dilations = var_3364_dilations_0, groups = var_3364_groups_0, pad = var_3364_pad_0, pad_type = var_3364_pad_type_0, strides = var_3364_strides_0, weight = layers_17_fc2_inlier_module_weight_to_fp16_palettized, x = input_143_cast_fp16)[name = string("op_3364_cast_fp16")];
            string var_3370_pad_type_0 = const()[name = string("op_3370_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3370_strides_0 = const()[name = string("op_3370_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3370_pad_0 = const()[name = string("op_3370_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3370_dilations_0 = const()[name = string("op_3370_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3370_groups_0 = const()[name = string("op_3370_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_17_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245717952))), nonzero_data = tensor<fp16, [51001]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245615872))))[name = string("layers_17_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3370_cast_fp16 = conv(dilations = var_3370_dilations_0, groups = var_3370_groups_0, pad = var_3370_pad_0, pad_type = var_3370_pad_type_0, strides = var_3370_strides_0, weight = layers_17_fc2_outlier_module_weight_to_fp16_sparsified, x = input_143_cast_fp16)[name = string("op_3370_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_39_cast_fp16 = add(x = var_3364_cast_fp16, y = var_3370_cast_fp16)[name = 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 = string("inputs_73_cast_fp16")];
            int32 var_3380 = const()[name = string("op_3380"), val = int32(3)];
            tensor<int32, [1]> out_73_axes_0 = const()[name = string("out_73_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3399_to_fp16 = const()[name = string("op_3399_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_3399_to_fp16, x = inputs_73_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [1280]> obj_73_gamma_0_to_fp16 = const()[name = string("obj_73_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246537216)))];
            tensor<fp16, [1280]> obj_73_beta_0_to_fp16 = const()[name = string("obj_73_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246539840)))];
            fp16 obj_73_epsilon_0_to_fp16 = const()[name = string("obj_73_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = string("obj_73_cast_fp16")];
            string var_3421_pad_type_0 = const()[name = string("op_3421_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3421_strides_0 = const()[name = string("op_3421_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3421_pad_0 = const()[name = string("op_3421_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3421_dilations_0 = const()[name = string("op_3421_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3421_groups_0 = const()[name = string("op_3421_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246542464))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247361728))))[name = string("layers_18_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_18_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_18_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247361856)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3421_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_3421_dilations_0, groups = var_3421_groups_0, pad = var_3421_pad_0, pad_type = var_3421_pad_type_0, strides = var_3421_strides_0, weight = layers_18_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = string("op_3421_cast_fp16")];
            string var_3427_pad_type_0 = const()[name = string("op_3427_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3427_strides_0 = const()[name = string("op_3427_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3427_pad_0 = const()[name = string("op_3427_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3427_dilations_0 = const()[name = string("op_3427_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3427_groups_0 = const()[name = string("op_3427_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247417600))), nonzero_data = tensor<fp16, [26500]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247364480))))[name = string("layers_18_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3427_cast_fp16 = conv(dilations = var_3427_dilations_0, groups = var_3427_groups_0, pad = var_3427_pad_0, pad_type = var_3427_pad_type_0, strides = var_3427_strides_0, weight = layers_18_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_73_cast_fp16)[name = string("op_3427_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_37_cast_fp16 = add(x = var_3421_cast_fp16, y = var_3427_cast_fp16)[name = string("query_37_cast_fp16")];
            string var_3436_pad_type_0 = const()[name = string("op_3436_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3436_strides_0 = const()[name = string("op_3436_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3436_pad_0 = const()[name = string("op_3436_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3436_dilations_0 = const()[name = string("op_3436_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3436_groups_0 = const()[name = string("op_3436_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247622464))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248441728))))[name = string("layers_18_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3436_cast_fp16 = conv(dilations = var_3436_dilations_0, groups = var_3436_groups_0, pad = var_3436_pad_0, pad_type = var_3436_pad_type_0, strides = var_3436_strides_0, weight = layers_18_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = string("op_3436_cast_fp16")];
            string var_3442_pad_type_0 = const()[name = string("op_3442_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3442_strides_0 = const()[name = string("op_3442_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3442_pad_0 = const()[name = string("op_3442_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3442_dilations_0 = const()[name = string("op_3442_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3442_groups_0 = const()[name = string("op_3442_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248472512))), nonzero_data = tensor<fp16, [15282]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248441856))))[name = string("layers_18_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3442_cast_fp16 = conv(dilations = var_3442_dilations_0, groups = var_3442_groups_0, pad = var_3442_pad_0, pad_type = var_3442_pad_type_0, strides = var_3442_strides_0, weight = layers_18_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_73_cast_fp16)[name = string("op_3442_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_37_cast_fp16 = add(x = var_3436_cast_fp16, y = var_3442_cast_fp16)[name = string("key_37_cast_fp16")];
            string var_3452_pad_type_0 = const()[name = string("op_3452_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3452_strides_0 = const()[name = string("op_3452_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3452_pad_0 = const()[name = string("op_3452_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3452_dilations_0 = const()[name = string("op_3452_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3452_groups_0 = const()[name = string("op_3452_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248677376))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249496640))))[name = string("layers_18_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_18_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_18_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249496768)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3452_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_3452_dilations_0, groups = var_3452_groups_0, pad = var_3452_pad_0, pad_type = var_3452_pad_type_0, strides = var_3452_strides_0, weight = layers_18_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = string("op_3452_cast_fp16")];
            string var_3458_pad_type_0 = const()[name = string("op_3458_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3458_strides_0 = const()[name = string("op_3458_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3458_pad_0 = const()[name = string("op_3458_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3458_dilations_0 = const()[name = string("op_3458_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3458_groups_0 = const()[name = string("op_3458_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249515520))), nonzero_data = tensor<fp16, [8030]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249499392))))[name = string("layers_18_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3458_cast_fp16 = conv(dilations = var_3458_dilations_0, groups = var_3458_groups_0, pad = var_3458_pad_0, pad_type = var_3458_pad_type_0, strides = var_3458_strides_0, weight = layers_18_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_73_cast_fp16)[name = string("op_3458_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_37_cast_fp16 = add(x = var_3452_cast_fp16, y = var_3458_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<int32, [4]> var_3461 = const()[name = string("op_3461"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_37_cast_fp16 = reshape(shape = var_3461, x = query_37_cast_fp16)[name = string("mh_q_37_cast_fp16")];
            fp16 var_3463_to_fp16 = const()[name = string("op_3463_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_3464_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_3463_to_fp16)[name = string("op_3464_cast_fp16")];
            tensor<int32, [4]> var_3465 = const()[name = string("op_3465"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3466_cast_fp16 = reshape(shape = var_3465, x = key_37_cast_fp16)[name = string("op_3466_cast_fp16")];
            bool mh_w_37_transpose_x_0 = const()[name = string("mh_w_37_transpose_x_0"), val = bool(true)];
            bool mh_w_37_transpose_y_0 = const()[name = string("mh_w_37_transpose_y_0"), val = 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_3464_cast_fp16, y = var_3466_cast_fp16)[name = string("mh_w_37_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_3469_cast_fp16 = softmax(axis = var_3380, x = mh_w_37_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<int32, [4]> var_3470 = const()[name = string("op_3470"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3471_cast_fp16 = reshape(shape = var_3470, x = value_37_cast_fp16)[name = string("op_3471_cast_fp16")];
            bool attn_37_transpose_x_0 = const()[name = string("attn_37_transpose_x_0"), val = bool(false)];
            bool attn_37_transpose_y_0 = const()[name = string("attn_37_transpose_y_0"), val = 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_3471_cast_fp16, y = var_3469_cast_fp16)[name = string("attn_37_cast_fp16")];
            tensor<int32, [4]> var_3474 = const()[name = string("op_3474"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_145_cast_fp16 = reshape(shape = var_3474, x = attn_37_cast_fp16)[name = string("input_145_cast_fp16")];
            string var_3484_pad_type_0 = const()[name = string("op_3484_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3484_strides_0 = const()[name = string("op_3484_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3484_pad_0 = const()[name = string("op_3484_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3484_dilations_0 = const()[name = string("op_3484_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3484_groups_0 = const()[name = string("op_3484_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249720384))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250539648))))[name = string("layers_18_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_18_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_18_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250539776)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3484_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_3484_dilations_0, groups = var_3484_groups_0, pad = var_3484_pad_0, pad_type = var_3484_pad_type_0, strides = var_3484_strides_0, weight = layers_18_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_145_cast_fp16)[name = string("op_3484_cast_fp16")];
            string var_3490_pad_type_0 = const()[name = string("op_3490_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3490_strides_0 = const()[name = string("op_3490_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3490_pad_0 = const()[name = string("op_3490_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3490_dilations_0 = const()[name = string("op_3490_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3490_groups_0 = const()[name = string("op_3490_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250558144))), nonzero_data = tensor<fp16, [7810]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250542400))))[name = string("layers_18_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3490_cast_fp16 = conv(dilations = var_3490_dilations_0, groups = var_3490_groups_0, pad = var_3490_pad_0, pad_type = var_3490_pad_type_0, strides = var_3490_strides_0, weight = layers_18_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_145_cast_fp16)[name = string("op_3490_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_75_cast_fp16 = add(x = var_3484_cast_fp16, y = var_3490_cast_fp16)[name = 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 = string("inputs_75_cast_fp16")];
            tensor<int32, [1]> out_75_axes_0 = const()[name = string("out_75_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3501_to_fp16 = const()[name = string("op_3501_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_3501_to_fp16, x = inputs_75_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [1280]> input_147_gamma_0_to_fp16 = const()[name = string("input_147_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250763008)))];
            tensor<fp16, [1280]> input_147_beta_0_to_fp16 = const()[name = string("input_147_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250765632)))];
            fp16 input_147_epsilon_0_to_fp16 = const()[name = string("input_147_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            string var_3519_pad_type_0 = const()[name = string("op_3519_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3519_strides_0 = const()[name = string("op_3519_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3519_pad_0 = const()[name = string("op_3519_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3519_dilations_0 = const()[name = string("op_3519_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3519_groups_0 = const()[name = string("op_3519_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_18_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250768256))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254045120))))[name = string("layers_18_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_18_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_18_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254045248)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_3519_cast_fp16 = conv(bias = layers_18_fc1_inlier_module_bias_to_fp16, dilations = var_3519_dilations_0, groups = var_3519_groups_0, pad = var_3519_pad_0, pad_type = var_3519_pad_type_0, strides = var_3519_strides_0, weight = layers_18_fc1_inlier_module_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("op_3519_cast_fp16")];
            string var_3525_pad_type_0 = const()[name = string("op_3525_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3525_strides_0 = const()[name = string("op_3525_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3525_pad_0 = const()[name = string("op_3525_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3525_dilations_0 = const()[name = string("op_3525_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3525_groups_0 = const()[name = string("op_3525_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_18_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254130816))), nonzero_data = tensor<fp16, [37590]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254055552))))[name = string("layers_18_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_3525_cast_fp16 = conv(dilations = var_3525_dilations_0, groups = var_3525_groups_0, pad = var_3525_pad_0, pad_type = var_3525_pad_type_0, strides = var_3525_strides_0, weight = layers_18_fc1_outlier_module_weight_to_fp16_sparsified, x = input_147_cast_fp16)[name = string("op_3525_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_149_cast_fp16 = add(x = var_3519_cast_fp16, y = var_3525_cast_fp16)[name = string("input_149_cast_fp16")];
            string input_151_mode_0 = const()[name = string("input_151_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = string("input_151_cast_fp16")];
            string var_3536_pad_type_0 = const()[name = string("op_3536_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3536_strides_0 = const()[name = string("op_3536_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3536_pad_0 = const()[name = string("op_3536_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3536_dilations_0 = const()[name = string("op_3536_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3536_groups_0 = const()[name = string("op_3536_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_18_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254950080))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258226944))))[name = string("layers_18_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_18_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_18_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258227072)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3536_cast_fp16 = conv(bias = layers_18_fc2_inlier_module_bias_to_fp16, dilations = var_3536_dilations_0, groups = var_3536_groups_0, pad = var_3536_pad_0, pad_type = var_3536_pad_type_0, strides = var_3536_strides_0, weight = layers_18_fc2_inlier_module_weight_to_fp16_palettized, x = input_151_cast_fp16)[name = string("op_3536_cast_fp16")];
            string var_3542_pad_type_0 = const()[name = string("op_3542_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3542_strides_0 = const()[name = string("op_3542_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3542_pad_0 = const()[name = string("op_3542_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3542_dilations_0 = const()[name = string("op_3542_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3542_groups_0 = const()[name = string("op_3542_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_18_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258318976))), nonzero_data = tensor<fp16, [44591]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258229696))))[name = string("layers_18_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3542_cast_fp16 = conv(dilations = var_3542_dilations_0, groups = var_3542_groups_0, pad = var_3542_pad_0, pad_type = var_3542_pad_type_0, strides = var_3542_strides_0, weight = layers_18_fc2_outlier_module_weight_to_fp16_sparsified, x = input_151_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_41_cast_fp16 = add(x = var_3536_cast_fp16, y = var_3542_cast_fp16)[name = 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 = string("inputs_77_cast_fp16")];
            int32 var_3552 = const()[name = string("op_3552"), val = int32(3)];
            tensor<int32, [1]> out_77_axes_0 = const()[name = string("out_77_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3571_to_fp16 = const()[name = string("op_3571_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_3571_to_fp16, x = inputs_77_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [1280]> obj_77_gamma_0_to_fp16 = const()[name = string("obj_77_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259138240)))];
            tensor<fp16, [1280]> obj_77_beta_0_to_fp16 = const()[name = string("obj_77_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259140864)))];
            fp16 obj_77_epsilon_0_to_fp16 = const()[name = string("obj_77_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = string("obj_77_cast_fp16")];
            string var_3593_pad_type_0 = const()[name = string("op_3593_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3593_strides_0 = const()[name = string("op_3593_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3593_pad_0 = const()[name = string("op_3593_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3593_dilations_0 = const()[name = string("op_3593_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3593_groups_0 = const()[name = string("op_3593_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259143488))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259962752))))[name = string("layers_19_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_19_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_19_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259962880)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3593_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_3593_dilations_0, groups = var_3593_groups_0, pad = var_3593_pad_0, pad_type = var_3593_pad_type_0, strides = var_3593_strides_0, weight = layers_19_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = string("op_3593_cast_fp16")];
            string var_3599_pad_type_0 = const()[name = string("op_3599_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3599_strides_0 = const()[name = string("op_3599_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3599_pad_0 = const()[name = string("op_3599_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3599_dilations_0 = const()[name = string("op_3599_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3599_groups_0 = const()[name = string("op_3599_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260020736))), nonzero_data = tensor<fp16, [27561]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259965504))))[name = string("layers_19_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3599_cast_fp16 = conv(dilations = var_3599_dilations_0, groups = var_3599_groups_0, pad = var_3599_pad_0, pad_type = var_3599_pad_type_0, strides = var_3599_strides_0, weight = layers_19_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_77_cast_fp16)[name = string("op_3599_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_39_cast_fp16 = add(x = var_3593_cast_fp16, y = var_3599_cast_fp16)[name = string("query_39_cast_fp16")];
            string var_3608_pad_type_0 = const()[name = string("op_3608_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3608_strides_0 = const()[name = string("op_3608_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3608_pad_0 = const()[name = string("op_3608_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3608_dilations_0 = const()[name = string("op_3608_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3608_groups_0 = const()[name = string("op_3608_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260225600))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261044864))))[name = string("layers_19_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3608_cast_fp16 = conv(dilations = var_3608_dilations_0, groups = var_3608_groups_0, pad = var_3608_pad_0, pad_type = var_3608_pad_type_0, strides = var_3608_strides_0, weight = layers_19_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = string("op_3608_cast_fp16")];
            string var_3614_pad_type_0 = const()[name = string("op_3614_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3614_strides_0 = const()[name = string("op_3614_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3614_pad_0 = const()[name = string("op_3614_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3614_dilations_0 = const()[name = string("op_3614_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3614_groups_0 = const()[name = string("op_3614_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261079936))), nonzero_data = tensor<fp16, [17431]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261044992))))[name = string("layers_19_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3614_cast_fp16 = conv(dilations = var_3614_dilations_0, groups = var_3614_groups_0, pad = var_3614_pad_0, pad_type = var_3614_pad_type_0, strides = var_3614_strides_0, weight = layers_19_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_77_cast_fp16)[name = string("op_3614_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_39_cast_fp16 = add(x = var_3608_cast_fp16, y = var_3614_cast_fp16)[name = string("key_39_cast_fp16")];
            string var_3624_pad_type_0 = const()[name = string("op_3624_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3624_strides_0 = const()[name = string("op_3624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3624_pad_0 = const()[name = string("op_3624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3624_dilations_0 = const()[name = string("op_3624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3624_groups_0 = const()[name = string("op_3624_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261284800))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262104064))))[name = string("layers_19_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_19_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_19_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262104192)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3624_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_3624_dilations_0, groups = var_3624_groups_0, pad = var_3624_pad_0, pad_type = var_3624_pad_type_0, strides = var_3624_strides_0, weight = layers_19_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = string("op_3624_cast_fp16")];
            string var_3630_pad_type_0 = const()[name = string("op_3630_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3630_strides_0 = const()[name = string("op_3630_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3630_pad_0 = const()[name = string("op_3630_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3630_dilations_0 = const()[name = string("op_3630_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3630_groups_0 = const()[name = string("op_3630_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262121984))), nonzero_data = tensor<fp16, [7538]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262106816))))[name = string("layers_19_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3630_cast_fp16 = conv(dilations = var_3630_dilations_0, groups = var_3630_groups_0, pad = var_3630_pad_0, pad_type = var_3630_pad_type_0, strides = var_3630_strides_0, weight = layers_19_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_77_cast_fp16)[name = string("op_3630_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_39_cast_fp16 = add(x = var_3624_cast_fp16, y = var_3630_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<int32, [4]> var_3633 = const()[name = string("op_3633"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_39_cast_fp16 = reshape(shape = var_3633, x = query_39_cast_fp16)[name = string("mh_q_39_cast_fp16")];
            fp16 var_3635_to_fp16 = const()[name = string("op_3635_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_3636_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_3635_to_fp16)[name = string("op_3636_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3638_cast_fp16 = reshape(shape = var_3637, x = key_39_cast_fp16)[name = string("op_3638_cast_fp16")];
            bool mh_w_39_transpose_x_0 = const()[name = string("mh_w_39_transpose_x_0"), val = bool(true)];
            bool mh_w_39_transpose_y_0 = const()[name = string("mh_w_39_transpose_y_0"), val = 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_3636_cast_fp16, y = var_3638_cast_fp16)[name = string("mh_w_39_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_3641_cast_fp16 = softmax(axis = var_3552, x = mh_w_39_cast_fp16)[name = string("op_3641_cast_fp16")];
            tensor<int32, [4]> var_3642 = const()[name = string("op_3642"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3643_cast_fp16 = reshape(shape = var_3642, x = value_39_cast_fp16)[name = string("op_3643_cast_fp16")];
            bool attn_39_transpose_x_0 = const()[name = string("attn_39_transpose_x_0"), val = bool(false)];
            bool attn_39_transpose_y_0 = const()[name = string("attn_39_transpose_y_0"), val = 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_3643_cast_fp16, y = var_3641_cast_fp16)[name = string("attn_39_cast_fp16")];
            tensor<int32, [4]> var_3646 = const()[name = string("op_3646"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_153_cast_fp16 = reshape(shape = var_3646, x = attn_39_cast_fp16)[name = string("input_153_cast_fp16")];
            string var_3656_pad_type_0 = const()[name = string("op_3656_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3656_strides_0 = const()[name = string("op_3656_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3656_pad_0 = const()[name = string("op_3656_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3656_dilations_0 = const()[name = string("op_3656_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3656_groups_0 = const()[name = string("op_3656_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262326848))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263146112))))[name = string("layers_19_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_19_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_19_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263146240)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3656_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_3656_dilations_0, groups = var_3656_groups_0, pad = var_3656_pad_0, pad_type = var_3656_pad_type_0, strides = var_3656_strides_0, weight = layers_19_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_153_cast_fp16)[name = string("op_3656_cast_fp16")];
            string var_3662_pad_type_0 = const()[name = string("op_3662_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3662_strides_0 = const()[name = string("op_3662_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3662_pad_0 = const()[name = string("op_3662_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3662_dilations_0 = const()[name = string("op_3662_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3662_groups_0 = const()[name = string("op_3662_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263162496))), nonzero_data = tensor<fp16, [6783]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263148864))))[name = string("layers_19_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3662_cast_fp16 = conv(dilations = var_3662_dilations_0, groups = var_3662_groups_0, pad = var_3662_pad_0, pad_type = var_3662_pad_type_0, strides = var_3662_strides_0, weight = layers_19_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_153_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_79_cast_fp16 = add(x = var_3656_cast_fp16, y = var_3662_cast_fp16)[name = 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 = string("inputs_79_cast_fp16")];
            tensor<int32, [1]> out_79_axes_0 = const()[name = string("out_79_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3673_to_fp16 = const()[name = string("op_3673_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_3673_to_fp16, x = inputs_79_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [1280]> input_155_gamma_0_to_fp16 = const()[name = string("input_155_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263367360)))];
            tensor<fp16, [1280]> input_155_beta_0_to_fp16 = const()[name = string("input_155_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263369984)))];
            fp16 input_155_epsilon_0_to_fp16 = const()[name = string("input_155_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            string var_3691_pad_type_0 = const()[name = string("op_3691_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3691_strides_0 = const()[name = string("op_3691_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3691_pad_0 = const()[name = string("op_3691_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3691_dilations_0 = const()[name = string("op_3691_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3691_groups_0 = const()[name = string("op_3691_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_19_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263372608))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266649472))))[name = string("layers_19_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_19_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_19_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266649600)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_3691_cast_fp16 = conv(bias = layers_19_fc1_inlier_module_bias_to_fp16, dilations = var_3691_dilations_0, groups = var_3691_groups_0, pad = var_3691_pad_0, pad_type = var_3691_pad_type_0, strides = var_3691_strides_0, weight = layers_19_fc1_inlier_module_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = string("op_3691_cast_fp16")];
            string var_3697_pad_type_0 = const()[name = string("op_3697_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3697_strides_0 = const()[name = string("op_3697_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3697_pad_0 = const()[name = string("op_3697_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3697_dilations_0 = const()[name = string("op_3697_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3697_groups_0 = const()[name = string("op_3697_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_19_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266736192))), nonzero_data = tensor<fp16, [38107]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266659904))))[name = string("layers_19_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_3697_cast_fp16 = conv(dilations = var_3697_dilations_0, groups = var_3697_groups_0, pad = var_3697_pad_0, pad_type = var_3697_pad_type_0, strides = var_3697_strides_0, weight = layers_19_fc1_outlier_module_weight_to_fp16_sparsified, x = input_155_cast_fp16)[name = string("op_3697_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_157_cast_fp16 = add(x = var_3691_cast_fp16, y = var_3697_cast_fp16)[name = string("input_157_cast_fp16")];
            string input_159_mode_0 = const()[name = string("input_159_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = string("input_159_cast_fp16")];
            string var_3708_pad_type_0 = const()[name = string("op_3708_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3708_strides_0 = const()[name = string("op_3708_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3708_pad_0 = const()[name = string("op_3708_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3708_dilations_0 = const()[name = string("op_3708_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3708_groups_0 = const()[name = string("op_3708_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_19_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267555456))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270832320))))[name = string("layers_19_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_19_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_19_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270832448)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3708_cast_fp16 = conv(bias = layers_19_fc2_inlier_module_bias_to_fp16, dilations = var_3708_dilations_0, groups = var_3708_groups_0, pad = var_3708_pad_0, pad_type = var_3708_pad_type_0, strides = var_3708_strides_0, weight = layers_19_fc2_inlier_module_weight_to_fp16_palettized, x = input_159_cast_fp16)[name = string("op_3708_cast_fp16")];
            string var_3714_pad_type_0 = const()[name = string("op_3714_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3714_strides_0 = const()[name = string("op_3714_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3714_pad_0 = const()[name = string("op_3714_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3714_dilations_0 = const()[name = string("op_3714_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3714_groups_0 = const()[name = string("op_3714_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_19_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270911040))), nonzero_data = tensor<fp16, [37935]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270835072))))[name = string("layers_19_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3714_cast_fp16 = conv(dilations = var_3714_dilations_0, groups = var_3714_groups_0, pad = var_3714_pad_0, pad_type = var_3714_pad_type_0, strides = var_3714_strides_0, weight = layers_19_fc2_outlier_module_weight_to_fp16_sparsified, x = input_159_cast_fp16)[name = string("op_3714_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_43_cast_fp16 = add(x = var_3708_cast_fp16, y = var_3714_cast_fp16)[name = 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 = string("inputs_81_cast_fp16")];
            int32 var_3724 = const()[name = string("op_3724"), val = int32(3)];
            tensor<int32, [1]> out_81_axes_0 = const()[name = string("out_81_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3743_to_fp16 = const()[name = string("op_3743_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_3743_to_fp16, x = inputs_81_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [1280]> obj_81_gamma_0_to_fp16 = const()[name = string("obj_81_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271730304)))];
            tensor<fp16, [1280]> obj_81_beta_0_to_fp16 = const()[name = string("obj_81_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271732928)))];
            fp16 obj_81_epsilon_0_to_fp16 = const()[name = string("obj_81_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = string("obj_81_cast_fp16")];
            string var_3765_pad_type_0 = const()[name = string("op_3765_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3765_strides_0 = const()[name = string("op_3765_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3765_pad_0 = const()[name = string("op_3765_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3765_dilations_0 = const()[name = string("op_3765_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3765_groups_0 = const()[name = string("op_3765_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271735552))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272554816))))[name = string("layers_20_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_20_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_20_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272554944)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3765_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_3765_dilations_0, groups = var_3765_groups_0, pad = var_3765_pad_0, pad_type = var_3765_pad_type_0, strides = var_3765_strides_0, weight = layers_20_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = string("op_3765_cast_fp16")];
            string var_3771_pad_type_0 = const()[name = string("op_3771_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3771_strides_0 = const()[name = string("op_3771_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3771_pad_0 = const()[name = string("op_3771_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3771_dilations_0 = const()[name = string("op_3771_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3771_groups_0 = const()[name = string("op_3771_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272594048))), nonzero_data = tensor<fp16, [18178]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272557568))))[name = string("layers_20_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3771_cast_fp16 = conv(dilations = var_3771_dilations_0, groups = var_3771_groups_0, pad = var_3771_pad_0, pad_type = var_3771_pad_type_0, strides = var_3771_strides_0, weight = layers_20_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_81_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_41_cast_fp16 = add(x = var_3765_cast_fp16, y = var_3771_cast_fp16)[name = string("query_41_cast_fp16")];
            string var_3780_pad_type_0 = const()[name = string("op_3780_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3780_strides_0 = const()[name = string("op_3780_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3780_pad_0 = const()[name = string("op_3780_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3780_dilations_0 = const()[name = string("op_3780_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3780_groups_0 = const()[name = string("op_3780_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272798912))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273618176))))[name = string("layers_20_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3780_cast_fp16 = conv(dilations = var_3780_dilations_0, groups = var_3780_groups_0, pad = var_3780_pad_0, pad_type = var_3780_pad_type_0, strides = var_3780_strides_0, weight = layers_20_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = string("op_3780_cast_fp16")];
            string var_3786_pad_type_0 = const()[name = string("op_3786_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3786_strides_0 = const()[name = string("op_3786_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3786_pad_0 = const()[name = string("op_3786_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3786_dilations_0 = const()[name = string("op_3786_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3786_groups_0 = const()[name = string("op_3786_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273644416))), nonzero_data = tensor<fp16, [13001]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273618304))))[name = string("layers_20_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3786_cast_fp16 = conv(dilations = var_3786_dilations_0, groups = var_3786_groups_0, pad = var_3786_pad_0, pad_type = var_3786_pad_type_0, strides = var_3786_strides_0, weight = layers_20_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_81_cast_fp16)[name = string("op_3786_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_41_cast_fp16 = add(x = var_3780_cast_fp16, y = var_3786_cast_fp16)[name = string("key_41_cast_fp16")];
            string var_3796_pad_type_0 = const()[name = string("op_3796_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3796_strides_0 = const()[name = string("op_3796_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3796_pad_0 = const()[name = string("op_3796_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3796_dilations_0 = const()[name = string("op_3796_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3796_groups_0 = const()[name = string("op_3796_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273849280))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274668544))))[name = string("layers_20_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_20_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_20_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274668672)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3796_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_3796_dilations_0, groups = var_3796_groups_0, pad = var_3796_pad_0, pad_type = var_3796_pad_type_0, strides = var_3796_strides_0, weight = layers_20_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = string("op_3796_cast_fp16")];
            string var_3802_pad_type_0 = const()[name = string("op_3802_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3802_strides_0 = const()[name = string("op_3802_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3802_pad_0 = const()[name = string("op_3802_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3802_dilations_0 = const()[name = string("op_3802_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3802_groups_0 = const()[name = string("op_3802_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274685312))), nonzero_data = tensor<fp16, [6955]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274671296))))[name = string("layers_20_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3802_cast_fp16 = conv(dilations = var_3802_dilations_0, groups = var_3802_groups_0, pad = var_3802_pad_0, pad_type = var_3802_pad_type_0, strides = var_3802_strides_0, weight = layers_20_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_81_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_41_cast_fp16 = add(x = var_3796_cast_fp16, y = var_3802_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<int32, [4]> var_3805 = const()[name = string("op_3805"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_41_cast_fp16 = reshape(shape = var_3805, x = query_41_cast_fp16)[name = string("mh_q_41_cast_fp16")];
            fp16 var_3807_to_fp16 = const()[name = string("op_3807_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_3808_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_3807_to_fp16)[name = string("op_3808_cast_fp16")];
            tensor<int32, [4]> var_3809 = const()[name = string("op_3809"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3810_cast_fp16 = reshape(shape = var_3809, x = key_41_cast_fp16)[name = string("op_3810_cast_fp16")];
            bool mh_w_41_transpose_x_0 = const()[name = string("mh_w_41_transpose_x_0"), val = bool(true)];
            bool mh_w_41_transpose_y_0 = const()[name = string("mh_w_41_transpose_y_0"), val = 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_3808_cast_fp16, y = var_3810_cast_fp16)[name = string("mh_w_41_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_3813_cast_fp16 = softmax(axis = var_3724, x = mh_w_41_cast_fp16)[name = string("op_3813_cast_fp16")];
            tensor<int32, [4]> var_3814 = const()[name = string("op_3814"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3815_cast_fp16 = reshape(shape = var_3814, x = value_41_cast_fp16)[name = string("op_3815_cast_fp16")];
            bool attn_41_transpose_x_0 = const()[name = string("attn_41_transpose_x_0"), val = bool(false)];
            bool attn_41_transpose_y_0 = const()[name = string("attn_41_transpose_y_0"), val = 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_3815_cast_fp16, y = var_3813_cast_fp16)[name = string("attn_41_cast_fp16")];
            tensor<int32, [4]> var_3818 = const()[name = string("op_3818"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_161_cast_fp16 = reshape(shape = var_3818, x = attn_41_cast_fp16)[name = string("input_161_cast_fp16")];
            string var_3828_pad_type_0 = const()[name = string("op_3828_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3828_strides_0 = const()[name = string("op_3828_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3828_pad_0 = const()[name = string("op_3828_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3828_dilations_0 = const()[name = string("op_3828_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3828_groups_0 = const()[name = string("op_3828_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274890176))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275709440))))[name = string("layers_20_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_20_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_20_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275709568)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3828_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_3828_dilations_0, groups = var_3828_groups_0, pad = var_3828_pad_0, pad_type = var_3828_pad_type_0, strides = var_3828_strides_0, weight = layers_20_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_161_cast_fp16)[name = string("op_3828_cast_fp16")];
            string var_3834_pad_type_0 = const()[name = string("op_3834_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3834_strides_0 = const()[name = string("op_3834_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3834_pad_0 = const()[name = string("op_3834_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3834_dilations_0 = const()[name = string("op_3834_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3834_groups_0 = const()[name = string("op_3834_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275726912))), nonzero_data = tensor<fp16, [7315]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275712192))))[name = string("layers_20_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3834_cast_fp16 = conv(dilations = var_3834_dilations_0, groups = var_3834_groups_0, pad = var_3834_pad_0, pad_type = var_3834_pad_type_0, strides = var_3834_strides_0, weight = layers_20_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_161_cast_fp16)[name = string("op_3834_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_83_cast_fp16 = add(x = var_3828_cast_fp16, y = var_3834_cast_fp16)[name = 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 = string("inputs_83_cast_fp16")];
            tensor<int32, [1]> out_83_axes_0 = const()[name = string("out_83_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3845_to_fp16 = const()[name = string("op_3845_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_3845_to_fp16, x = inputs_83_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [1280]> input_163_gamma_0_to_fp16 = const()[name = string("input_163_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275931776)))];
            tensor<fp16, [1280]> input_163_beta_0_to_fp16 = const()[name = string("input_163_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275934400)))];
            fp16 input_163_epsilon_0_to_fp16 = const()[name = string("input_163_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            string var_3863_pad_type_0 = const()[name = string("op_3863_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3863_strides_0 = const()[name = string("op_3863_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3863_pad_0 = const()[name = string("op_3863_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3863_dilations_0 = const()[name = string("op_3863_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3863_groups_0 = const()[name = string("op_3863_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_20_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275937024))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279213888))))[name = string("layers_20_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_20_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_20_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279214016)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_3863_cast_fp16 = conv(bias = layers_20_fc1_inlier_module_bias_to_fp16, dilations = var_3863_dilations_0, groups = var_3863_groups_0, pad = var_3863_pad_0, pad_type = var_3863_pad_type_0, strides = var_3863_strides_0, weight = layers_20_fc1_inlier_module_weight_to_fp16_palettized, x = input_163_cast_fp16)[name = string("op_3863_cast_fp16")];
            string var_3869_pad_type_0 = const()[name = string("op_3869_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3869_strides_0 = const()[name = string("op_3869_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3869_pad_0 = const()[name = string("op_3869_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3869_dilations_0 = const()[name = string("op_3869_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3869_groups_0 = const()[name = string("op_3869_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_20_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279311872))), nonzero_data = tensor<fp16, [43740]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279224320))))[name = string("layers_20_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_3869_cast_fp16 = conv(dilations = var_3869_dilations_0, groups = var_3869_groups_0, pad = var_3869_pad_0, pad_type = var_3869_pad_type_0, strides = var_3869_strides_0, weight = layers_20_fc1_outlier_module_weight_to_fp16_sparsified, x = input_163_cast_fp16)[name = string("op_3869_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_165_cast_fp16 = add(x = var_3863_cast_fp16, y = var_3869_cast_fp16)[name = string("input_165_cast_fp16")];
            string input_167_mode_0 = const()[name = string("input_167_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
            string var_3880_pad_type_0 = const()[name = string("op_3880_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3880_strides_0 = const()[name = string("op_3880_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3880_pad_0 = const()[name = string("op_3880_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3880_dilations_0 = const()[name = string("op_3880_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3880_groups_0 = const()[name = string("op_3880_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_20_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280131136))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283408000))))[name = string("layers_20_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_20_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_20_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283408128)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3880_cast_fp16 = conv(bias = layers_20_fc2_inlier_module_bias_to_fp16, dilations = var_3880_dilations_0, groups = var_3880_groups_0, pad = var_3880_pad_0, pad_type = var_3880_pad_type_0, strides = var_3880_strides_0, weight = layers_20_fc2_inlier_module_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("op_3880_cast_fp16")];
            string var_3886_pad_type_0 = const()[name = string("op_3886_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3886_strides_0 = const()[name = string("op_3886_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3886_pad_0 = const()[name = string("op_3886_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3886_dilations_0 = const()[name = string("op_3886_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3886_groups_0 = const()[name = string("op_3886_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_20_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283466304))), nonzero_data = tensor<fp16, [27714]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283410752))))[name = string("layers_20_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3886_cast_fp16 = conv(dilations = var_3886_dilations_0, groups = var_3886_groups_0, pad = var_3886_pad_0, pad_type = var_3886_pad_type_0, strides = var_3886_strides_0, weight = layers_20_fc2_outlier_module_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("op_3886_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_45_cast_fp16 = add(x = var_3880_cast_fp16, y = var_3886_cast_fp16)[name = 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 = string("inputs_85_cast_fp16")];
            int32 var_3896 = const()[name = string("op_3896"), val = int32(3)];
            tensor<int32, [1]> out_85_axes_0 = const()[name = string("out_85_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_3915_to_fp16 = const()[name = string("op_3915_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_3915_to_fp16, x = inputs_85_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [1280]> obj_85_gamma_0_to_fp16 = const()[name = string("obj_85_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284285568)))];
            tensor<fp16, [1280]> obj_85_beta_0_to_fp16 = const()[name = string("obj_85_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284288192)))];
            fp16 obj_85_epsilon_0_to_fp16 = const()[name = string("obj_85_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = string("obj_85_cast_fp16")];
            string var_3937_pad_type_0 = const()[name = string("op_3937_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3937_strides_0 = const()[name = string("op_3937_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3937_pad_0 = const()[name = string("op_3937_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3937_dilations_0 = const()[name = string("op_3937_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3937_groups_0 = const()[name = string("op_3937_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284290816))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285110080))))[name = string("layers_21_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_21_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_21_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285110208)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3937_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_3937_dilations_0, groups = var_3937_groups_0, pad = var_3937_pad_0, pad_type = var_3937_pad_type_0, strides = var_3937_strides_0, weight = layers_21_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = string("op_3937_cast_fp16")];
            string var_3943_pad_type_0 = const()[name = string("op_3943_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3943_strides_0 = const()[name = string("op_3943_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3943_pad_0 = const()[name = string("op_3943_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3943_dilations_0 = const()[name = string("op_3943_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3943_groups_0 = const()[name = string("op_3943_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285140544))), nonzero_data = tensor<fp16, [13818]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285112832))))[name = string("layers_21_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3943_cast_fp16 = conv(dilations = var_3943_dilations_0, groups = var_3943_groups_0, pad = var_3943_pad_0, pad_type = var_3943_pad_type_0, strides = var_3943_strides_0, weight = layers_21_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_85_cast_fp16)[name = string("op_3943_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_43_cast_fp16 = add(x = var_3937_cast_fp16, y = var_3943_cast_fp16)[name = string("query_43_cast_fp16")];
            string var_3952_pad_type_0 = const()[name = string("op_3952_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3952_strides_0 = const()[name = string("op_3952_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3952_pad_0 = const()[name = string("op_3952_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3952_dilations_0 = const()[name = string("op_3952_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3952_groups_0 = const()[name = string("op_3952_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285345408))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286164672))))[name = string("layers_21_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3952_cast_fp16 = conv(dilations = var_3952_dilations_0, groups = var_3952_groups_0, pad = var_3952_pad_0, pad_type = var_3952_pad_type_0, strides = var_3952_strides_0, weight = layers_21_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = string("op_3952_cast_fp16")];
            string var_3958_pad_type_0 = const()[name = string("op_3958_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3958_strides_0 = const()[name = string("op_3958_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3958_pad_0 = const()[name = string("op_3958_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3958_dilations_0 = const()[name = string("op_3958_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3958_groups_0 = const()[name = string("op_3958_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286193792))), nonzero_data = tensor<fp16, [14444]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286164800))))[name = string("layers_21_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3958_cast_fp16 = conv(dilations = var_3958_dilations_0, groups = var_3958_groups_0, pad = var_3958_pad_0, pad_type = var_3958_pad_type_0, strides = var_3958_strides_0, weight = layers_21_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_85_cast_fp16)[name = string("op_3958_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_43_cast_fp16 = add(x = var_3952_cast_fp16, y = var_3958_cast_fp16)[name = string("key_43_cast_fp16")];
            string var_3968_pad_type_0 = const()[name = string("op_3968_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3968_strides_0 = const()[name = string("op_3968_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3968_pad_0 = const()[name = string("op_3968_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3968_dilations_0 = const()[name = string("op_3968_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3968_groups_0 = const()[name = string("op_3968_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286398656))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287217920))))[name = string("layers_21_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_21_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_21_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287218048)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_3968_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_3968_dilations_0, groups = var_3968_groups_0, pad = var_3968_pad_0, pad_type = var_3968_pad_type_0, strides = var_3968_strides_0, weight = layers_21_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = string("op_3968_cast_fp16")];
            string var_3974_pad_type_0 = const()[name = string("op_3974_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3974_strides_0 = const()[name = string("op_3974_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3974_pad_0 = const()[name = string("op_3974_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3974_dilations_0 = const()[name = string("op_3974_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3974_groups_0 = const()[name = string("op_3974_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287233472))), nonzero_data = tensor<fp16, [6347]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287220672))))[name = string("layers_21_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_3974_cast_fp16 = conv(dilations = var_3974_dilations_0, groups = var_3974_groups_0, pad = var_3974_pad_0, pad_type = var_3974_pad_type_0, strides = var_3974_strides_0, weight = layers_21_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_85_cast_fp16)[name = string("op_3974_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_43_cast_fp16 = add(x = var_3968_cast_fp16, y = var_3974_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<int32, [4]> var_3977 = const()[name = string("op_3977"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_43_cast_fp16 = reshape(shape = var_3977, x = query_43_cast_fp16)[name = string("mh_q_43_cast_fp16")];
            fp16 var_3979_to_fp16 = const()[name = string("op_3979_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_3980_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_3979_to_fp16)[name = string("op_3980_cast_fp16")];
            tensor<int32, [4]> var_3981 = const()[name = string("op_3981"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3982_cast_fp16 = reshape(shape = var_3981, x = key_43_cast_fp16)[name = string("op_3982_cast_fp16")];
            bool mh_w_43_transpose_x_0 = const()[name = string("mh_w_43_transpose_x_0"), val = bool(true)];
            bool mh_w_43_transpose_y_0 = const()[name = string("mh_w_43_transpose_y_0"), val = 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_3980_cast_fp16, y = var_3982_cast_fp16)[name = string("mh_w_43_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_3985_cast_fp16 = softmax(axis = var_3896, x = mh_w_43_cast_fp16)[name = string("op_3985_cast_fp16")];
            tensor<int32, [4]> var_3986 = const()[name = string("op_3986"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3987_cast_fp16 = reshape(shape = var_3986, x = value_43_cast_fp16)[name = string("op_3987_cast_fp16")];
            bool attn_43_transpose_x_0 = const()[name = string("attn_43_transpose_x_0"), val = bool(false)];
            bool attn_43_transpose_y_0 = const()[name = string("attn_43_transpose_y_0"), val = 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_3987_cast_fp16, y = var_3985_cast_fp16)[name = string("attn_43_cast_fp16")];
            tensor<int32, [4]> var_3990 = const()[name = string("op_3990"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_169_cast_fp16 = reshape(shape = var_3990, x = attn_43_cast_fp16)[name = string("input_169_cast_fp16")];
            string var_4000_pad_type_0 = const()[name = string("op_4000_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4000_strides_0 = const()[name = string("op_4000_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4000_pad_0 = const()[name = string("op_4000_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4000_dilations_0 = const()[name = string("op_4000_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4000_groups_0 = const()[name = string("op_4000_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287438336))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288257600))))[name = string("layers_21_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_21_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_21_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288257728)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4000_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_4000_dilations_0, groups = var_4000_groups_0, pad = var_4000_pad_0, pad_type = var_4000_pad_type_0, strides = var_4000_strides_0, weight = layers_21_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("op_4000_cast_fp16")];
            string var_4006_pad_type_0 = const()[name = string("op_4006_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4006_strides_0 = const()[name = string("op_4006_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4006_pad_0 = const()[name = string("op_4006_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4006_dilations_0 = const()[name = string("op_4006_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4006_groups_0 = const()[name = string("op_4006_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288272320))), nonzero_data = tensor<fp16, [5922]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288260352))))[name = string("layers_21_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4006_cast_fp16 = conv(dilations = var_4006_dilations_0, groups = var_4006_groups_0, pad = var_4006_pad_0, pad_type = var_4006_pad_type_0, strides = var_4006_strides_0, weight = layers_21_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_169_cast_fp16)[name = string("op_4006_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_87_cast_fp16 = add(x = var_4000_cast_fp16, y = var_4006_cast_fp16)[name = 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 = string("inputs_87_cast_fp16")];
            tensor<int32, [1]> out_87_axes_0 = const()[name = string("out_87_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4017_to_fp16 = const()[name = string("op_4017_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_4017_to_fp16, x = inputs_87_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [1280]> input_171_gamma_0_to_fp16 = const()[name = string("input_171_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288477184)))];
            tensor<fp16, [1280]> input_171_beta_0_to_fp16 = const()[name = string("input_171_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288479808)))];
            fp16 input_171_epsilon_0_to_fp16 = const()[name = string("input_171_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            string var_4035_pad_type_0 = const()[name = string("op_4035_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4035_strides_0 = const()[name = string("op_4035_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4035_pad_0 = const()[name = string("op_4035_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4035_dilations_0 = const()[name = string("op_4035_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4035_groups_0 = const()[name = string("op_4035_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_21_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288482432))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291759296))))[name = string("layers_21_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_21_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_21_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291759424)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_4035_cast_fp16 = conv(bias = layers_21_fc1_inlier_module_bias_to_fp16, dilations = var_4035_dilations_0, groups = var_4035_groups_0, pad = var_4035_pad_0, pad_type = var_4035_pad_type_0, strides = var_4035_strides_0, weight = layers_21_fc1_inlier_module_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = string("op_4035_cast_fp16")];
            string var_4041_pad_type_0 = const()[name = string("op_4041_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4041_strides_0 = const()[name = string("op_4041_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4041_pad_0 = const()[name = string("op_4041_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4041_dilations_0 = const()[name = string("op_4041_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4041_groups_0 = const()[name = string("op_4041_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_21_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291830912))), nonzero_data = tensor<fp16, [30547]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291769728))))[name = string("layers_21_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_4041_cast_fp16 = conv(dilations = var_4041_dilations_0, groups = var_4041_groups_0, pad = var_4041_pad_0, pad_type = var_4041_pad_type_0, strides = var_4041_strides_0, weight = layers_21_fc1_outlier_module_weight_to_fp16_sparsified, x = input_171_cast_fp16)[name = string("op_4041_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_173_cast_fp16 = add(x = var_4035_cast_fp16, y = var_4041_cast_fp16)[name = string("input_173_cast_fp16")];
            string input_175_mode_0 = const()[name = string("input_175_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = input_173_cast_fp16)[name = string("input_175_cast_fp16")];
            string var_4052_pad_type_0 = const()[name = string("op_4052_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4052_strides_0 = const()[name = string("op_4052_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4052_pad_0 = const()[name = string("op_4052_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4052_dilations_0 = const()[name = string("op_4052_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4052_groups_0 = const()[name = string("op_4052_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_21_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292650176))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295927040))))[name = string("layers_21_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_21_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_21_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295927168)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4052_cast_fp16 = conv(bias = layers_21_fc2_inlier_module_bias_to_fp16, dilations = var_4052_dilations_0, groups = var_4052_groups_0, pad = var_4052_pad_0, pad_type = var_4052_pad_type_0, strides = var_4052_strides_0, weight = layers_21_fc2_inlier_module_weight_to_fp16_palettized, x = input_175_cast_fp16)[name = string("op_4052_cast_fp16")];
            string var_4058_pad_type_0 = const()[name = string("op_4058_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4058_strides_0 = const()[name = string("op_4058_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4058_pad_0 = const()[name = string("op_4058_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4058_dilations_0 = const()[name = string("op_4058_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4058_groups_0 = const()[name = string("op_4058_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_21_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295992512))), nonzero_data = tensor<fp16, [31326]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295929792))))[name = string("layers_21_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4058_cast_fp16 = conv(dilations = var_4058_dilations_0, groups = var_4058_groups_0, pad = var_4058_pad_0, pad_type = var_4058_pad_type_0, strides = var_4058_strides_0, weight = layers_21_fc2_outlier_module_weight_to_fp16_sparsified, x = input_175_cast_fp16)[name = string("op_4058_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_47_cast_fp16 = add(x = var_4052_cast_fp16, y = var_4058_cast_fp16)[name = 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 = string("inputs_89_cast_fp16")];
            int32 var_4068 = const()[name = string("op_4068"), val = int32(3)];
            tensor<int32, [1]> out_89_axes_0 = const()[name = string("out_89_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4087_to_fp16 = const()[name = string("op_4087_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_4087_to_fp16, x = inputs_89_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [1280]> obj_89_gamma_0_to_fp16 = const()[name = string("obj_89_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296811776)))];
            tensor<fp16, [1280]> obj_89_beta_0_to_fp16 = const()[name = string("obj_89_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296814400)))];
            fp16 obj_89_epsilon_0_to_fp16 = const()[name = string("obj_89_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = string("obj_89_cast_fp16")];
            string var_4109_pad_type_0 = const()[name = string("op_4109_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4109_strides_0 = const()[name = string("op_4109_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4109_pad_0 = const()[name = string("op_4109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4109_dilations_0 = const()[name = string("op_4109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4109_groups_0 = const()[name = string("op_4109_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296817024))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297636288))))[name = string("layers_22_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_22_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_22_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297636416)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4109_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_4109_dilations_0, groups = var_4109_groups_0, pad = var_4109_pad_0, pad_type = var_4109_pad_type_0, strides = var_4109_strides_0, weight = layers_22_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = string("op_4109_cast_fp16")];
            string var_4115_pad_type_0 = const()[name = string("op_4115_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4115_strides_0 = const()[name = string("op_4115_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4115_pad_0 = const()[name = string("op_4115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4115_dilations_0 = const()[name = string("op_4115_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4115_groups_0 = const()[name = string("op_4115_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297674560))), nonzero_data = tensor<fp16, [17724]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297639040))))[name = string("layers_22_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4115_cast_fp16 = conv(dilations = var_4115_dilations_0, groups = var_4115_groups_0, pad = var_4115_pad_0, pad_type = var_4115_pad_type_0, strides = var_4115_strides_0, weight = layers_22_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_89_cast_fp16)[name = string("op_4115_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_45_cast_fp16 = add(x = var_4109_cast_fp16, y = var_4115_cast_fp16)[name = string("query_45_cast_fp16")];
            string var_4124_pad_type_0 = const()[name = string("op_4124_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4124_strides_0 = const()[name = string("op_4124_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4124_pad_0 = const()[name = string("op_4124_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4124_dilations_0 = const()[name = string("op_4124_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4124_groups_0 = const()[name = string("op_4124_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297879424))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298698688))))[name = string("layers_22_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4124_cast_fp16 = conv(dilations = var_4124_dilations_0, groups = var_4124_groups_0, pad = var_4124_pad_0, pad_type = var_4124_pad_type_0, strides = var_4124_strides_0, weight = layers_22_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = string("op_4124_cast_fp16")];
            string var_4130_pad_type_0 = const()[name = string("op_4130_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4130_strides_0 = const()[name = string("op_4130_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4130_pad_0 = const()[name = string("op_4130_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4130_dilations_0 = const()[name = string("op_4130_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4130_groups_0 = const()[name = string("op_4130_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298728768))), nonzero_data = tensor<fp16, [14937]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298698816))))[name = string("layers_22_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4130_cast_fp16 = conv(dilations = var_4130_dilations_0, groups = var_4130_groups_0, pad = var_4130_pad_0, pad_type = var_4130_pad_type_0, strides = var_4130_strides_0, weight = layers_22_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_89_cast_fp16)[name = string("op_4130_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_45_cast_fp16 = add(x = var_4124_cast_fp16, y = var_4130_cast_fp16)[name = string("key_45_cast_fp16")];
            string var_4140_pad_type_0 = const()[name = string("op_4140_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4140_strides_0 = const()[name = string("op_4140_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4140_pad_0 = const()[name = string("op_4140_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4140_dilations_0 = const()[name = string("op_4140_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4140_groups_0 = const()[name = string("op_4140_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298933632))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299752896))))[name = string("layers_22_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_22_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_22_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299753024)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4140_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_4140_dilations_0, groups = var_4140_groups_0, pad = var_4140_pad_0, pad_type = var_4140_pad_type_0, strides = var_4140_strides_0, weight = layers_22_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = string("op_4140_cast_fp16")];
            string var_4146_pad_type_0 = const()[name = string("op_4146_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4146_strides_0 = const()[name = string("op_4146_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4146_pad_0 = const()[name = string("op_4146_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4146_dilations_0 = const()[name = string("op_4146_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4146_groups_0 = const()[name = string("op_4146_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299768064))), nonzero_data = tensor<fp16, [6148]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299755648))))[name = string("layers_22_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4146_cast_fp16 = conv(dilations = var_4146_dilations_0, groups = var_4146_groups_0, pad = var_4146_pad_0, pad_type = var_4146_pad_type_0, strides = var_4146_strides_0, weight = layers_22_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_89_cast_fp16)[name = string("op_4146_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_45_cast_fp16 = add(x = var_4140_cast_fp16, y = var_4146_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<int32, [4]> var_4149 = const()[name = string("op_4149"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_45_cast_fp16 = reshape(shape = var_4149, x = query_45_cast_fp16)[name = string("mh_q_45_cast_fp16")];
            fp16 var_4151_to_fp16 = const()[name = string("op_4151_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_4152_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_4151_to_fp16)[name = string("op_4152_cast_fp16")];
            tensor<int32, [4]> var_4153 = const()[name = string("op_4153"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4154_cast_fp16 = reshape(shape = var_4153, x = key_45_cast_fp16)[name = string("op_4154_cast_fp16")];
            bool mh_w_45_transpose_x_0 = const()[name = string("mh_w_45_transpose_x_0"), val = bool(true)];
            bool mh_w_45_transpose_y_0 = const()[name = string("mh_w_45_transpose_y_0"), val = 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_4152_cast_fp16, y = var_4154_cast_fp16)[name = string("mh_w_45_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_4157_cast_fp16 = softmax(axis = var_4068, x = mh_w_45_cast_fp16)[name = string("op_4157_cast_fp16")];
            tensor<int32, [4]> var_4158 = const()[name = string("op_4158"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4159_cast_fp16 = reshape(shape = var_4158, x = value_45_cast_fp16)[name = string("op_4159_cast_fp16")];
            bool attn_45_transpose_x_0 = const()[name = string("attn_45_transpose_x_0"), val = bool(false)];
            bool attn_45_transpose_y_0 = const()[name = string("attn_45_transpose_y_0"), val = 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_4159_cast_fp16, y = var_4157_cast_fp16)[name = string("attn_45_cast_fp16")];
            tensor<int32, [4]> var_4162 = const()[name = string("op_4162"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_177_cast_fp16 = reshape(shape = var_4162, x = attn_45_cast_fp16)[name = string("input_177_cast_fp16")];
            string var_4172_pad_type_0 = const()[name = string("op_4172_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4172_strides_0 = const()[name = string("op_4172_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4172_pad_0 = const()[name = string("op_4172_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4172_dilations_0 = const()[name = string("op_4172_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4172_groups_0 = const()[name = string("op_4172_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299972928))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300792192))))[name = string("layers_22_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_22_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_22_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300792320)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4172_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_4172_dilations_0, groups = var_4172_groups_0, pad = var_4172_pad_0, pad_type = var_4172_pad_type_0, strides = var_4172_strides_0, weight = layers_22_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = string("op_4172_cast_fp16")];
            string var_4178_pad_type_0 = const()[name = string("op_4178_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4178_strides_0 = const()[name = string("op_4178_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4178_pad_0 = const()[name = string("op_4178_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4178_dilations_0 = const()[name = string("op_4178_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4178_groups_0 = const()[name = string("op_4178_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300807232))), nonzero_data = tensor<fp16, [6088]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300794944))))[name = string("layers_22_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4178_cast_fp16 = conv(dilations = var_4178_dilations_0, groups = var_4178_groups_0, pad = var_4178_pad_0, pad_type = var_4178_pad_type_0, strides = var_4178_strides_0, weight = layers_22_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_177_cast_fp16)[name = string("op_4178_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_91_cast_fp16 = add(x = var_4172_cast_fp16, y = var_4178_cast_fp16)[name = 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 = string("inputs_91_cast_fp16")];
            tensor<int32, [1]> out_91_axes_0 = const()[name = string("out_91_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4189_to_fp16 = const()[name = string("op_4189_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_4189_to_fp16, x = inputs_91_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [1280]> input_179_gamma_0_to_fp16 = const()[name = string("input_179_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301012096)))];
            tensor<fp16, [1280]> input_179_beta_0_to_fp16 = const()[name = string("input_179_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301014720)))];
            fp16 input_179_epsilon_0_to_fp16 = const()[name = string("input_179_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            string var_4207_pad_type_0 = const()[name = string("op_4207_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4207_strides_0 = const()[name = string("op_4207_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4207_pad_0 = const()[name = string("op_4207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4207_dilations_0 = const()[name = string("op_4207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4207_groups_0 = const()[name = string("op_4207_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_22_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301017344))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304294208))))[name = string("layers_22_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_22_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_22_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304294336)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_4207_cast_fp16 = conv(bias = layers_22_fc1_inlier_module_bias_to_fp16, dilations = var_4207_dilations_0, groups = var_4207_groups_0, pad = var_4207_pad_0, pad_type = var_4207_pad_type_0, strides = var_4207_strides_0, weight = layers_22_fc1_inlier_module_weight_to_fp16_palettized, x = input_179_cast_fp16)[name = string("op_4207_cast_fp16")];
            string var_4213_pad_type_0 = const()[name = string("op_4213_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4213_strides_0 = const()[name = string("op_4213_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4213_pad_0 = const()[name = string("op_4213_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4213_dilations_0 = const()[name = string("op_4213_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4213_groups_0 = const()[name = string("op_4213_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_22_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304372736))), nonzero_data = tensor<fp16, [34004]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304304640))))[name = string("layers_22_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_4213_cast_fp16 = conv(dilations = var_4213_dilations_0, groups = var_4213_groups_0, pad = var_4213_pad_0, pad_type = var_4213_pad_type_0, strides = var_4213_strides_0, weight = layers_22_fc1_outlier_module_weight_to_fp16_sparsified, x = input_179_cast_fp16)[name = string("op_4213_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_181_cast_fp16 = add(x = var_4207_cast_fp16, y = var_4213_cast_fp16)[name = string("input_181_cast_fp16")];
            string input_183_mode_0 = const()[name = string("input_183_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_183_cast_fp16 = gelu(mode = input_183_mode_0, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")];
            string var_4224_pad_type_0 = const()[name = string("op_4224_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4224_strides_0 = const()[name = string("op_4224_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4224_pad_0 = const()[name = string("op_4224_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4224_dilations_0 = const()[name = string("op_4224_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4224_groups_0 = const()[name = string("op_4224_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_22_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305192000))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308468864))))[name = string("layers_22_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_22_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_22_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308468992)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4224_cast_fp16 = conv(bias = layers_22_fc2_inlier_module_bias_to_fp16, dilations = var_4224_dilations_0, groups = var_4224_groups_0, pad = var_4224_pad_0, pad_type = var_4224_pad_type_0, strides = var_4224_strides_0, weight = layers_22_fc2_inlier_module_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("op_4224_cast_fp16")];
            string var_4230_pad_type_0 = const()[name = string("op_4230_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4230_strides_0 = const()[name = string("op_4230_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4230_pad_0 = const()[name = string("op_4230_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4230_dilations_0 = const()[name = string("op_4230_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4230_groups_0 = const()[name = string("op_4230_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_22_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308533376))), nonzero_data = tensor<fp16, [30830]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308471616))))[name = string("layers_22_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4230_cast_fp16 = conv(dilations = var_4230_dilations_0, groups = var_4230_groups_0, pad = var_4230_pad_0, pad_type = var_4230_pad_type_0, strides = var_4230_strides_0, weight = layers_22_fc2_outlier_module_weight_to_fp16_sparsified, x = input_183_cast_fp16)[name = string("op_4230_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_49_cast_fp16 = add(x = var_4224_cast_fp16, y = var_4230_cast_fp16)[name = 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 = string("inputs_93_cast_fp16")];
            int32 var_4240 = const()[name = string("op_4240"), val = int32(3)];
            tensor<int32, [1]> out_93_axes_0 = const()[name = string("out_93_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4259_to_fp16 = const()[name = string("op_4259_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_4259_to_fp16, x = inputs_93_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [1280]> obj_93_gamma_0_to_fp16 = const()[name = string("obj_93_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309352640)))];
            tensor<fp16, [1280]> obj_93_beta_0_to_fp16 = const()[name = string("obj_93_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309355264)))];
            fp16 obj_93_epsilon_0_to_fp16 = const()[name = string("obj_93_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = string("obj_93_cast_fp16")];
            string var_4281_pad_type_0 = const()[name = string("op_4281_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4281_strides_0 = const()[name = string("op_4281_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4281_pad_0 = const()[name = string("op_4281_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4281_dilations_0 = const()[name = string("op_4281_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4281_groups_0 = const()[name = string("op_4281_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309357888))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310177152))))[name = string("layers_23_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_23_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_23_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310177280)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4281_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_4281_dilations_0, groups = var_4281_groups_0, pad = var_4281_pad_0, pad_type = var_4281_pad_type_0, strides = var_4281_strides_0, weight = layers_23_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = string("op_4281_cast_fp16")];
            string var_4287_pad_type_0 = const()[name = string("op_4287_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4287_strides_0 = const()[name = string("op_4287_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4287_pad_0 = const()[name = string("op_4287_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4287_dilations_0 = const()[name = string("op_4287_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4287_groups_0 = const()[name = string("op_4287_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310208448))), nonzero_data = tensor<fp16, [14215]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310179904))))[name = string("layers_23_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4287_cast_fp16 = conv(dilations = var_4287_dilations_0, groups = var_4287_groups_0, pad = var_4287_pad_0, pad_type = var_4287_pad_type_0, strides = var_4287_strides_0, weight = layers_23_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_93_cast_fp16)[name = string("op_4287_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_47_cast_fp16 = add(x = var_4281_cast_fp16, y = var_4287_cast_fp16)[name = string("query_47_cast_fp16")];
            string var_4296_pad_type_0 = const()[name = string("op_4296_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4296_strides_0 = const()[name = string("op_4296_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4296_pad_0 = const()[name = string("op_4296_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4296_dilations_0 = const()[name = string("op_4296_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4296_groups_0 = const()[name = string("op_4296_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310413312))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311232576))))[name = string("layers_23_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4296_cast_fp16 = conv(dilations = var_4296_dilations_0, groups = var_4296_groups_0, pad = var_4296_pad_0, pad_type = var_4296_pad_type_0, strides = var_4296_strides_0, weight = layers_23_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = string("op_4296_cast_fp16")];
            string var_4302_pad_type_0 = const()[name = string("op_4302_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4302_strides_0 = const()[name = string("op_4302_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4302_pad_0 = const()[name = string("op_4302_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4302_dilations_0 = const()[name = string("op_4302_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4302_groups_0 = const()[name = string("op_4302_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311257088))), nonzero_data = tensor<fp16, [12160]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311232704))))[name = string("layers_23_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4302_cast_fp16 = conv(dilations = var_4302_dilations_0, groups = var_4302_groups_0, pad = var_4302_pad_0, pad_type = var_4302_pad_type_0, strides = var_4302_strides_0, weight = layers_23_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_93_cast_fp16)[name = string("op_4302_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_47_cast_fp16 = add(x = var_4296_cast_fp16, y = var_4302_cast_fp16)[name = string("key_47_cast_fp16")];
            string var_4312_pad_type_0 = const()[name = string("op_4312_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4312_strides_0 = const()[name = string("op_4312_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4312_pad_0 = const()[name = string("op_4312_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4312_dilations_0 = const()[name = string("op_4312_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4312_groups_0 = const()[name = string("op_4312_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311461952))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312281216))))[name = string("layers_23_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_23_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_23_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312281344)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4312_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_4312_dilations_0, groups = var_4312_groups_0, pad = var_4312_pad_0, pad_type = var_4312_pad_type_0, strides = var_4312_strides_0, weight = layers_23_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = string("op_4312_cast_fp16")];
            string var_4318_pad_type_0 = const()[name = string("op_4318_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4318_strides_0 = const()[name = string("op_4318_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4318_pad_0 = const()[name = string("op_4318_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4318_dilations_0 = const()[name = string("op_4318_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4318_groups_0 = const()[name = string("op_4318_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312297152))), nonzero_data = tensor<fp16, [6559]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312283968))))[name = string("layers_23_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4318_cast_fp16 = conv(dilations = var_4318_dilations_0, groups = var_4318_groups_0, pad = var_4318_pad_0, pad_type = var_4318_pad_type_0, strides = var_4318_strides_0, weight = layers_23_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_93_cast_fp16)[name = string("op_4318_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_47_cast_fp16 = add(x = var_4312_cast_fp16, y = var_4318_cast_fp16)[name = string("value_47_cast_fp16")];
            tensor<int32, [4]> var_4321 = const()[name = string("op_4321"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_47_cast_fp16 = reshape(shape = var_4321, x = query_47_cast_fp16)[name = string("mh_q_47_cast_fp16")];
            fp16 var_4323_to_fp16 = const()[name = string("op_4323_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_4324_cast_fp16 = mul(x = mh_q_47_cast_fp16, y = var_4323_to_fp16)[name = string("op_4324_cast_fp16")];
            tensor<int32, [4]> var_4325 = const()[name = string("op_4325"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4326_cast_fp16 = reshape(shape = var_4325, x = key_47_cast_fp16)[name = string("op_4326_cast_fp16")];
            bool mh_w_47_transpose_x_0 = const()[name = string("mh_w_47_transpose_x_0"), val = bool(true)];
            bool mh_w_47_transpose_y_0 = const()[name = string("mh_w_47_transpose_y_0"), val = 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_4324_cast_fp16, y = var_4326_cast_fp16)[name = string("mh_w_47_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_4329_cast_fp16 = softmax(axis = var_4240, x = mh_w_47_cast_fp16)[name = string("op_4329_cast_fp16")];
            tensor<int32, [4]> var_4330 = const()[name = string("op_4330"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4331_cast_fp16 = reshape(shape = var_4330, x = value_47_cast_fp16)[name = string("op_4331_cast_fp16")];
            bool attn_47_transpose_x_0 = const()[name = string("attn_47_transpose_x_0"), val = bool(false)];
            bool attn_47_transpose_y_0 = const()[name = string("attn_47_transpose_y_0"), val = 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_4331_cast_fp16, y = var_4329_cast_fp16)[name = string("attn_47_cast_fp16")];
            tensor<int32, [4]> var_4334 = const()[name = string("op_4334"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_185_cast_fp16 = reshape(shape = var_4334, x = attn_47_cast_fp16)[name = string("input_185_cast_fp16")];
            string var_4344_pad_type_0 = const()[name = string("op_4344_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4344_strides_0 = const()[name = string("op_4344_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4344_pad_0 = const()[name = string("op_4344_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4344_dilations_0 = const()[name = string("op_4344_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4344_groups_0 = const()[name = string("op_4344_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312502016))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313321280))))[name = string("layers_23_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_23_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_23_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313321408)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4344_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_4344_dilations_0, groups = var_4344_groups_0, pad = var_4344_pad_0, pad_type = var_4344_pad_type_0, strides = var_4344_strides_0, weight = layers_23_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_185_cast_fp16)[name = string("op_4344_cast_fp16")];
            string var_4350_pad_type_0 = const()[name = string("op_4350_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4350_strides_0 = const()[name = string("op_4350_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4350_pad_0 = const()[name = string("op_4350_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4350_dilations_0 = const()[name = string("op_4350_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4350_groups_0 = const()[name = string("op_4350_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313337280))), nonzero_data = tensor<fp16, [6574]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313324032))))[name = string("layers_23_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4350_cast_fp16 = conv(dilations = var_4350_dilations_0, groups = var_4350_groups_0, pad = var_4350_pad_0, pad_type = var_4350_pad_type_0, strides = var_4350_strides_0, weight = layers_23_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_185_cast_fp16)[name = string("op_4350_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_95_cast_fp16 = add(x = var_4344_cast_fp16, y = var_4350_cast_fp16)[name = 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 = string("inputs_95_cast_fp16")];
            tensor<int32, [1]> out_95_axes_0 = const()[name = string("out_95_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4361_to_fp16 = const()[name = string("op_4361_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_4361_to_fp16, x = inputs_95_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [1280]> input_187_gamma_0_to_fp16 = const()[name = string("input_187_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313542144)))];
            tensor<fp16, [1280]> input_187_beta_0_to_fp16 = const()[name = string("input_187_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313544768)))];
            fp16 input_187_epsilon_0_to_fp16 = const()[name = string("input_187_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            string var_4379_pad_type_0 = const()[name = string("op_4379_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4379_strides_0 = const()[name = string("op_4379_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4379_pad_0 = const()[name = string("op_4379_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4379_dilations_0 = const()[name = string("op_4379_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4379_groups_0 = const()[name = string("op_4379_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_23_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313547392))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316824256))))[name = string("layers_23_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_23_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_23_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316824384)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_4379_cast_fp16 = conv(bias = layers_23_fc1_inlier_module_bias_to_fp16, dilations = var_4379_dilations_0, groups = var_4379_groups_0, pad = var_4379_pad_0, pad_type = var_4379_pad_type_0, strides = var_4379_strides_0, weight = layers_23_fc1_inlier_module_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("op_4379_cast_fp16")];
            string var_4385_pad_type_0 = const()[name = string("op_4385_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4385_strides_0 = const()[name = string("op_4385_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4385_pad_0 = const()[name = string("op_4385_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4385_dilations_0 = const()[name = string("op_4385_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4385_groups_0 = const()[name = string("op_4385_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_23_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316896960))), nonzero_data = tensor<fp16, [31092]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316834688))))[name = string("layers_23_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_4385_cast_fp16 = conv(dilations = var_4385_dilations_0, groups = var_4385_groups_0, pad = var_4385_pad_0, pad_type = var_4385_pad_type_0, strides = var_4385_strides_0, weight = layers_23_fc1_outlier_module_weight_to_fp16_sparsified, x = input_187_cast_fp16)[name = string("op_4385_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_189_cast_fp16 = add(x = var_4379_cast_fp16, y = var_4385_cast_fp16)[name = string("input_189_cast_fp16")];
            string input_191_mode_0 = const()[name = string("input_191_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = string("input_191_cast_fp16")];
            string var_4396_pad_type_0 = const()[name = string("op_4396_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4396_strides_0 = const()[name = string("op_4396_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4396_pad_0 = const()[name = string("op_4396_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4396_dilations_0 = const()[name = string("op_4396_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4396_groups_0 = const()[name = string("op_4396_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_23_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317716224))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320993088))))[name = string("layers_23_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_23_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_23_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320993216)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4396_cast_fp16 = conv(bias = layers_23_fc2_inlier_module_bias_to_fp16, dilations = var_4396_dilations_0, groups = var_4396_groups_0, pad = var_4396_pad_0, pad_type = var_4396_pad_type_0, strides = var_4396_strides_0, weight = layers_23_fc2_inlier_module_weight_to_fp16_palettized, x = input_191_cast_fp16)[name = string("op_4396_cast_fp16")];
            string var_4402_pad_type_0 = const()[name = string("op_4402_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4402_strides_0 = const()[name = string("op_4402_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4402_pad_0 = const()[name = string("op_4402_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4402_dilations_0 = const()[name = string("op_4402_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4402_groups_0 = const()[name = string("op_4402_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_23_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321052864))), nonzero_data = tensor<fp16, [28479]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320995840))))[name = string("layers_23_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4402_cast_fp16 = conv(dilations = var_4402_dilations_0, groups = var_4402_groups_0, pad = var_4402_pad_0, pad_type = var_4402_pad_type_0, strides = var_4402_strides_0, weight = layers_23_fc2_outlier_module_weight_to_fp16_sparsified, x = input_191_cast_fp16)[name = string("op_4402_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_51_cast_fp16 = add(x = var_4396_cast_fp16, y = var_4402_cast_fp16)[name = 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 = string("inputs_97_cast_fp16")];
            int32 var_4412 = const()[name = string("op_4412"), val = int32(3)];
            tensor<int32, [1]> out_97_axes_0 = const()[name = string("out_97_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4431_to_fp16 = const()[name = string("op_4431_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_4431_to_fp16, x = inputs_97_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [1280]> obj_97_gamma_0_to_fp16 = const()[name = string("obj_97_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321872128)))];
            tensor<fp16, [1280]> obj_97_beta_0_to_fp16 = const()[name = string("obj_97_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321874752)))];
            fp16 obj_97_epsilon_0_to_fp16 = const()[name = string("obj_97_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = string("obj_97_cast_fp16")];
            string var_4453_pad_type_0 = const()[name = string("op_4453_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4453_strides_0 = const()[name = string("op_4453_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4453_pad_0 = const()[name = string("op_4453_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4453_dilations_0 = const()[name = string("op_4453_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4453_groups_0 = const()[name = string("op_4453_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321877376))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322696640))))[name = string("layers_24_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_24_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_24_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322696768)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4453_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_4453_dilations_0, groups = var_4453_groups_0, pad = var_4453_pad_0, pad_type = var_4453_pad_type_0, strides = var_4453_strides_0, weight = layers_24_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = string("op_4453_cast_fp16")];
            string var_4459_pad_type_0 = const()[name = string("op_4459_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4459_strides_0 = const()[name = string("op_4459_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4459_pad_0 = const()[name = string("op_4459_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4459_dilations_0 = const()[name = string("op_4459_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4459_groups_0 = const()[name = string("op_4459_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322733056))), nonzero_data = tensor<fp16, [16797]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322699392))))[name = string("layers_24_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4459_cast_fp16 = conv(dilations = var_4459_dilations_0, groups = var_4459_groups_0, pad = var_4459_pad_0, pad_type = var_4459_pad_type_0, strides = var_4459_strides_0, weight = layers_24_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_97_cast_fp16)[name = string("op_4459_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_49_cast_fp16 = add(x = var_4453_cast_fp16, y = var_4459_cast_fp16)[name = string("query_49_cast_fp16")];
            string var_4468_pad_type_0 = const()[name = string("op_4468_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4468_strides_0 = const()[name = string("op_4468_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4468_pad_0 = const()[name = string("op_4468_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4468_dilations_0 = const()[name = string("op_4468_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4468_groups_0 = const()[name = string("op_4468_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322937920))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323757184))))[name = string("layers_24_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4468_cast_fp16 = conv(dilations = var_4468_dilations_0, groups = var_4468_groups_0, pad = var_4468_pad_0, pad_type = var_4468_pad_type_0, strides = var_4468_strides_0, weight = layers_24_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = string("op_4468_cast_fp16")];
            string var_4474_pad_type_0 = const()[name = string("op_4474_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4474_strides_0 = const()[name = string("op_4474_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4474_pad_0 = const()[name = string("op_4474_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4474_dilations_0 = const()[name = string("op_4474_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4474_groups_0 = const()[name = string("op_4474_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323783744))), nonzero_data = tensor<fp16, [13170]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323757312))))[name = string("layers_24_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4474_cast_fp16 = conv(dilations = var_4474_dilations_0, groups = var_4474_groups_0, pad = var_4474_pad_0, pad_type = var_4474_pad_type_0, strides = var_4474_strides_0, weight = layers_24_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_97_cast_fp16)[name = string("op_4474_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_49_cast_fp16 = add(x = var_4468_cast_fp16, y = var_4474_cast_fp16)[name = string("key_49_cast_fp16")];
            string var_4484_pad_type_0 = const()[name = string("op_4484_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4484_strides_0 = const()[name = string("op_4484_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4484_pad_0 = const()[name = string("op_4484_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4484_dilations_0 = const()[name = string("op_4484_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4484_groups_0 = const()[name = string("op_4484_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323988608))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324807872))))[name = string("layers_24_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_24_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_24_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324808000)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4484_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_4484_dilations_0, groups = var_4484_groups_0, pad = var_4484_pad_0, pad_type = var_4484_pad_type_0, strides = var_4484_strides_0, weight = layers_24_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = string("op_4484_cast_fp16")];
            string var_4490_pad_type_0 = const()[name = string("op_4490_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4490_strides_0 = const()[name = string("op_4490_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4490_pad_0 = const()[name = string("op_4490_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4490_dilations_0 = const()[name = string("op_4490_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4490_groups_0 = const()[name = string("op_4490_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324822080))), nonzero_data = tensor<fp16, [5691]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324810624))))[name = string("layers_24_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4490_cast_fp16 = conv(dilations = var_4490_dilations_0, groups = var_4490_groups_0, pad = var_4490_pad_0, pad_type = var_4490_pad_type_0, strides = var_4490_strides_0, weight = layers_24_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_97_cast_fp16)[name = string("op_4490_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_49_cast_fp16 = add(x = var_4484_cast_fp16, y = var_4490_cast_fp16)[name = string("value_49_cast_fp16")];
            tensor<int32, [4]> var_4493 = const()[name = string("op_4493"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_49_cast_fp16 = reshape(shape = var_4493, x = query_49_cast_fp16)[name = string("mh_q_49_cast_fp16")];
            fp16 var_4495_to_fp16 = const()[name = string("op_4495_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_4496_cast_fp16 = mul(x = mh_q_49_cast_fp16, y = var_4495_to_fp16)[name = string("op_4496_cast_fp16")];
            tensor<int32, [4]> var_4497 = const()[name = string("op_4497"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4498_cast_fp16 = reshape(shape = var_4497, x = key_49_cast_fp16)[name = string("op_4498_cast_fp16")];
            bool mh_w_49_transpose_x_0 = const()[name = string("mh_w_49_transpose_x_0"), val = bool(true)];
            bool mh_w_49_transpose_y_0 = const()[name = string("mh_w_49_transpose_y_0"), val = 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_4496_cast_fp16, y = var_4498_cast_fp16)[name = string("mh_w_49_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_4501_cast_fp16 = softmax(axis = var_4412, x = mh_w_49_cast_fp16)[name = string("op_4501_cast_fp16")];
            tensor<int32, [4]> var_4502 = const()[name = string("op_4502"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4503_cast_fp16 = reshape(shape = var_4502, x = value_49_cast_fp16)[name = string("op_4503_cast_fp16")];
            bool attn_49_transpose_x_0 = const()[name = string("attn_49_transpose_x_0"), val = bool(false)];
            bool attn_49_transpose_y_0 = const()[name = string("attn_49_transpose_y_0"), val = 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_4503_cast_fp16, y = var_4501_cast_fp16)[name = string("attn_49_cast_fp16")];
            tensor<int32, [4]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_193_cast_fp16 = reshape(shape = var_4506, x = attn_49_cast_fp16)[name = string("input_193_cast_fp16")];
            string var_4516_pad_type_0 = const()[name = string("op_4516_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4516_strides_0 = const()[name = string("op_4516_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4516_pad_0 = const()[name = string("op_4516_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4516_dilations_0 = const()[name = string("op_4516_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4516_groups_0 = const()[name = string("op_4516_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325026944))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325846208))))[name = string("layers_24_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_24_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_24_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325846336)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4516_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_4516_dilations_0, groups = var_4516_groups_0, pad = var_4516_pad_0, pad_type = var_4516_pad_type_0, strides = var_4516_strides_0, weight = layers_24_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_193_cast_fp16)[name = string("op_4516_cast_fp16")];
            string var_4522_pad_type_0 = const()[name = string("op_4522_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4522_strides_0 = const()[name = string("op_4522_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4522_pad_0 = const()[name = string("op_4522_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4522_dilations_0 = const()[name = string("op_4522_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4522_groups_0 = const()[name = string("op_4522_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325859712))), nonzero_data = tensor<fp16, [5325]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325848960))))[name = string("layers_24_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4522_cast_fp16 = conv(dilations = var_4522_dilations_0, groups = var_4522_groups_0, pad = var_4522_pad_0, pad_type = var_4522_pad_type_0, strides = var_4522_strides_0, weight = layers_24_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_193_cast_fp16)[name = string("op_4522_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_99_cast_fp16 = add(x = var_4516_cast_fp16, y = var_4522_cast_fp16)[name = 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 = string("inputs_99_cast_fp16")];
            tensor<int32, [1]> out_99_axes_0 = const()[name = string("out_99_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4533_to_fp16 = const()[name = string("op_4533_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_4533_to_fp16, x = inputs_99_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [1280]> input_195_gamma_0_to_fp16 = const()[name = string("input_195_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326064576)))];
            tensor<fp16, [1280]> input_195_beta_0_to_fp16 = const()[name = string("input_195_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326067200)))];
            fp16 input_195_epsilon_0_to_fp16 = const()[name = string("input_195_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = string("input_195_cast_fp16")];
            string var_4551_pad_type_0 = const()[name = string("op_4551_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4551_strides_0 = const()[name = string("op_4551_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4551_pad_0 = const()[name = string("op_4551_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4551_dilations_0 = const()[name = string("op_4551_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4551_groups_0 = const()[name = string("op_4551_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_24_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326069824))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329346688))))[name = string("layers_24_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_24_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_24_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329346816)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_4551_cast_fp16 = conv(bias = layers_24_fc1_inlier_module_bias_to_fp16, dilations = var_4551_dilations_0, groups = var_4551_groups_0, pad = var_4551_pad_0, pad_type = var_4551_pad_type_0, strides = var_4551_strides_0, weight = layers_24_fc1_inlier_module_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("op_4551_cast_fp16")];
            string var_4557_pad_type_0 = const()[name = string("op_4557_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4557_strides_0 = const()[name = string("op_4557_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4557_pad_0 = const()[name = string("op_4557_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4557_dilations_0 = const()[name = string("op_4557_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4557_groups_0 = const()[name = string("op_4557_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_24_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329417984))), nonzero_data = tensor<fp16, [30384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329357120))))[name = string("layers_24_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_4557_cast_fp16 = conv(dilations = var_4557_dilations_0, groups = var_4557_groups_0, pad = var_4557_pad_0, pad_type = var_4557_pad_type_0, strides = var_4557_strides_0, weight = layers_24_fc1_outlier_module_weight_to_fp16_sparsified, x = input_195_cast_fp16)[name = string("op_4557_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_197_cast_fp16 = add(x = var_4551_cast_fp16, y = var_4557_cast_fp16)[name = string("input_197_cast_fp16")];
            string input_199_mode_0 = const()[name = string("input_199_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = string("input_199_cast_fp16")];
            string var_4568_pad_type_0 = const()[name = string("op_4568_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4568_strides_0 = const()[name = string("op_4568_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4568_pad_0 = const()[name = string("op_4568_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4568_dilations_0 = const()[name = string("op_4568_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4568_groups_0 = const()[name = string("op_4568_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_24_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330237248))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333514112))))[name = string("layers_24_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_24_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_24_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333514240)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4568_cast_fp16 = conv(bias = layers_24_fc2_inlier_module_bias_to_fp16, dilations = var_4568_dilations_0, groups = var_4568_groups_0, pad = var_4568_pad_0, pad_type = var_4568_pad_type_0, strides = var_4568_strides_0, weight = layers_24_fc2_inlier_module_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("op_4568_cast_fp16")];
            string var_4574_pad_type_0 = const()[name = string("op_4574_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4574_strides_0 = const()[name = string("op_4574_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4574_pad_0 = const()[name = string("op_4574_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4574_dilations_0 = const()[name = string("op_4574_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4574_groups_0 = const()[name = string("op_4574_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_24_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333569216))), nonzero_data = tensor<fp16, [26123]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333516864))))[name = string("layers_24_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4574_cast_fp16 = conv(dilations = var_4574_dilations_0, groups = var_4574_groups_0, pad = var_4574_pad_0, pad_type = var_4574_pad_type_0, strides = var_4574_strides_0, weight = layers_24_fc2_outlier_module_weight_to_fp16_sparsified, x = input_199_cast_fp16)[name = string("op_4574_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_53_cast_fp16 = add(x = var_4568_cast_fp16, y = var_4574_cast_fp16)[name = 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 = string("inputs_101_cast_fp16")];
            int32 var_4584 = const()[name = string("op_4584"), val = int32(3)];
            tensor<int32, [1]> out_101_axes_0 = const()[name = string("out_101_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4603_to_fp16 = const()[name = string("op_4603_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_4603_to_fp16, x = inputs_101_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [1280]> obj_101_gamma_0_to_fp16 = const()[name = string("obj_101_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334388480)))];
            tensor<fp16, [1280]> obj_101_beta_0_to_fp16 = const()[name = string("obj_101_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334391104)))];
            fp16 obj_101_epsilon_0_to_fp16 = const()[name = string("obj_101_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = string("obj_101_cast_fp16")];
            string var_4625_pad_type_0 = const()[name = string("op_4625_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4625_strides_0 = const()[name = string("op_4625_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4625_pad_0 = const()[name = string("op_4625_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4625_dilations_0 = const()[name = string("op_4625_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4625_groups_0 = const()[name = string("op_4625_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334393728))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335212992))))[name = string("layers_25_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_25_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_25_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335213120)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4625_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_4625_dilations_0, groups = var_4625_groups_0, pad = var_4625_pad_0, pad_type = var_4625_pad_type_0, strides = var_4625_strides_0, weight = layers_25_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = string("op_4625_cast_fp16")];
            string var_4631_pad_type_0 = const()[name = string("op_4631_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4631_strides_0 = const()[name = string("op_4631_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4631_pad_0 = const()[name = string("op_4631_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4631_dilations_0 = const()[name = string("op_4631_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4631_groups_0 = const()[name = string("op_4631_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335253376))), nonzero_data = tensor<fp16, [18769]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335215744))))[name = string("layers_25_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4631_cast_fp16 = conv(dilations = var_4631_dilations_0, groups = var_4631_groups_0, pad = var_4631_pad_0, pad_type = var_4631_pad_type_0, strides = var_4631_strides_0, weight = layers_25_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_101_cast_fp16)[name = string("op_4631_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_51_cast_fp16 = add(x = var_4625_cast_fp16, y = var_4631_cast_fp16)[name = string("query_51_cast_fp16")];
            string var_4640_pad_type_0 = const()[name = string("op_4640_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4640_strides_0 = const()[name = string("op_4640_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4640_pad_0 = const()[name = string("op_4640_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4640_dilations_0 = const()[name = string("op_4640_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4640_groups_0 = const()[name = string("op_4640_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335458240))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336277504))))[name = string("layers_25_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4640_cast_fp16 = conv(dilations = var_4640_dilations_0, groups = var_4640_groups_0, pad = var_4640_pad_0, pad_type = var_4640_pad_type_0, strides = var_4640_strides_0, weight = layers_25_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = string("op_4640_cast_fp16")];
            string var_4646_pad_type_0 = const()[name = string("op_4646_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4646_strides_0 = const()[name = string("op_4646_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4646_pad_0 = const()[name = string("op_4646_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4646_dilations_0 = const()[name = string("op_4646_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4646_groups_0 = const()[name = string("op_4646_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336302912))), nonzero_data = tensor<fp16, [12599]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336277632))))[name = string("layers_25_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4646_cast_fp16 = conv(dilations = var_4646_dilations_0, groups = var_4646_groups_0, pad = var_4646_pad_0, pad_type = var_4646_pad_type_0, strides = var_4646_strides_0, weight = layers_25_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_101_cast_fp16)[name = string("op_4646_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_51_cast_fp16 = add(x = var_4640_cast_fp16, y = var_4646_cast_fp16)[name = string("key_51_cast_fp16")];
            string var_4656_pad_type_0 = const()[name = string("op_4656_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4656_strides_0 = const()[name = string("op_4656_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4656_pad_0 = const()[name = string("op_4656_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4656_dilations_0 = const()[name = string("op_4656_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4656_groups_0 = const()[name = string("op_4656_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336507776))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337327040))))[name = string("layers_25_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_25_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_25_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337327168)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4656_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_4656_dilations_0, groups = var_4656_groups_0, pad = var_4656_pad_0, pad_type = var_4656_pad_type_0, strides = var_4656_strides_0, weight = layers_25_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = string("op_4656_cast_fp16")];
            string var_4662_pad_type_0 = const()[name = string("op_4662_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4662_strides_0 = const()[name = string("op_4662_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4662_pad_0 = const()[name = string("op_4662_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4662_dilations_0 = const()[name = string("op_4662_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4662_groups_0 = const()[name = string("op_4662_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337341760))), nonzero_data = tensor<fp16, [5951]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337329792))))[name = string("layers_25_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4662_cast_fp16 = conv(dilations = var_4662_dilations_0, groups = var_4662_groups_0, pad = var_4662_pad_0, pad_type = var_4662_pad_type_0, strides = var_4662_strides_0, weight = layers_25_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_101_cast_fp16)[name = string("op_4662_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_51_cast_fp16 = add(x = var_4656_cast_fp16, y = var_4662_cast_fp16)[name = string("value_51_cast_fp16")];
            tensor<int32, [4]> var_4665 = const()[name = string("op_4665"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_51_cast_fp16 = reshape(shape = var_4665, x = query_51_cast_fp16)[name = string("mh_q_51_cast_fp16")];
            fp16 var_4667_to_fp16 = const()[name = string("op_4667_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_4668_cast_fp16 = mul(x = mh_q_51_cast_fp16, y = var_4667_to_fp16)[name = string("op_4668_cast_fp16")];
            tensor<int32, [4]> var_4669 = const()[name = string("op_4669"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4670_cast_fp16 = reshape(shape = var_4669, x = key_51_cast_fp16)[name = string("op_4670_cast_fp16")];
            bool mh_w_51_transpose_x_0 = const()[name = string("mh_w_51_transpose_x_0"), val = bool(true)];
            bool mh_w_51_transpose_y_0 = const()[name = string("mh_w_51_transpose_y_0"), val = 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_4668_cast_fp16, y = var_4670_cast_fp16)[name = string("mh_w_51_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_4673_cast_fp16 = softmax(axis = var_4584, x = mh_w_51_cast_fp16)[name = string("op_4673_cast_fp16")];
            tensor<int32, [4]> var_4674 = const()[name = string("op_4674"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4675_cast_fp16 = reshape(shape = var_4674, x = value_51_cast_fp16)[name = string("op_4675_cast_fp16")];
            bool attn_51_transpose_x_0 = const()[name = string("attn_51_transpose_x_0"), val = bool(false)];
            bool attn_51_transpose_y_0 = const()[name = string("attn_51_transpose_y_0"), val = 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_4675_cast_fp16, y = var_4673_cast_fp16)[name = string("attn_51_cast_fp16")];
            tensor<int32, [4]> var_4678 = const()[name = string("op_4678"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_201_cast_fp16 = reshape(shape = var_4678, x = attn_51_cast_fp16)[name = string("input_201_cast_fp16")];
            string var_4688_pad_type_0 = const()[name = string("op_4688_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4688_strides_0 = const()[name = string("op_4688_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4688_pad_0 = const()[name = string("op_4688_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4688_dilations_0 = const()[name = string("op_4688_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4688_groups_0 = const()[name = string("op_4688_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337546624))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338365888))))[name = string("layers_25_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_25_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_25_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338366016)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4688_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_4688_dilations_0, groups = var_4688_groups_0, pad = var_4688_pad_0, pad_type = var_4688_pad_type_0, strides = var_4688_strides_0, weight = layers_25_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_201_cast_fp16)[name = string("op_4688_cast_fp16")];
            string var_4694_pad_type_0 = const()[name = string("op_4694_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4694_strides_0 = const()[name = string("op_4694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4694_pad_0 = const()[name = string("op_4694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4694_dilations_0 = const()[name = string("op_4694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4694_groups_0 = const()[name = string("op_4694_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338379584))), nonzero_data = tensor<fp16, [5419]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338368640))))[name = string("layers_25_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4694_cast_fp16 = conv(dilations = var_4694_dilations_0, groups = var_4694_groups_0, pad = var_4694_pad_0, pad_type = var_4694_pad_type_0, strides = var_4694_strides_0, weight = layers_25_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_201_cast_fp16)[name = string("op_4694_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_103_cast_fp16 = add(x = var_4688_cast_fp16, y = var_4694_cast_fp16)[name = 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 = string("inputs_103_cast_fp16")];
            tensor<int32, [1]> out_103_axes_0 = const()[name = string("out_103_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4705_to_fp16 = const()[name = string("op_4705_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_4705_to_fp16, x = inputs_103_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [1280]> input_203_gamma_0_to_fp16 = const()[name = string("input_203_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338584448)))];
            tensor<fp16, [1280]> input_203_beta_0_to_fp16 = const()[name = string("input_203_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338587072)))];
            fp16 input_203_epsilon_0_to_fp16 = const()[name = string("input_203_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = string("input_203_cast_fp16")];
            string var_4723_pad_type_0 = const()[name = string("op_4723_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4723_strides_0 = const()[name = string("op_4723_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4723_pad_0 = const()[name = string("op_4723_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4723_dilations_0 = const()[name = string("op_4723_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4723_groups_0 = const()[name = string("op_4723_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_25_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338589696))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341866560))))[name = string("layers_25_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_25_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_25_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341866688)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_4723_cast_fp16 = conv(bias = layers_25_fc1_inlier_module_bias_to_fp16, dilations = var_4723_dilations_0, groups = var_4723_groups_0, pad = var_4723_pad_0, pad_type = var_4723_pad_type_0, strides = var_4723_strides_0, weight = layers_25_fc1_inlier_module_weight_to_fp16_palettized, x = input_203_cast_fp16)[name = string("op_4723_cast_fp16")];
            string var_4729_pad_type_0 = const()[name = string("op_4729_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4729_strides_0 = const()[name = string("op_4729_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4729_pad_0 = const()[name = string("op_4729_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4729_dilations_0 = const()[name = string("op_4729_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4729_groups_0 = const()[name = string("op_4729_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_25_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341933760))), nonzero_data = tensor<fp16, [28333]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341876992))))[name = string("layers_25_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_4729_cast_fp16 = conv(dilations = var_4729_dilations_0, groups = var_4729_groups_0, pad = var_4729_pad_0, pad_type = var_4729_pad_type_0, strides = var_4729_strides_0, weight = layers_25_fc1_outlier_module_weight_to_fp16_sparsified, x = input_203_cast_fp16)[name = string("op_4729_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_205_cast_fp16 = add(x = var_4723_cast_fp16, y = var_4729_cast_fp16)[name = string("input_205_cast_fp16")];
            string input_207_mode_0 = const()[name = string("input_207_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_207_cast_fp16 = gelu(mode = input_207_mode_0, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")];
            string var_4740_pad_type_0 = const()[name = string("op_4740_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4740_strides_0 = const()[name = string("op_4740_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4740_pad_0 = const()[name = string("op_4740_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4740_dilations_0 = const()[name = string("op_4740_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4740_groups_0 = const()[name = string("op_4740_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_25_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342753024))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346029888))))[name = string("layers_25_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_25_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_25_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346030016)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4740_cast_fp16 = conv(bias = layers_25_fc2_inlier_module_bias_to_fp16, dilations = var_4740_dilations_0, groups = var_4740_groups_0, pad = var_4740_pad_0, pad_type = var_4740_pad_type_0, strides = var_4740_strides_0, weight = layers_25_fc2_inlier_module_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = string("op_4740_cast_fp16")];
            string var_4746_pad_type_0 = const()[name = string("op_4746_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4746_strides_0 = const()[name = string("op_4746_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4746_pad_0 = const()[name = string("op_4746_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4746_dilations_0 = const()[name = string("op_4746_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4746_groups_0 = const()[name = string("op_4746_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_25_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346081152))), nonzero_data = tensor<fp16, [24223]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346032640))))[name = string("layers_25_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4746_cast_fp16 = conv(dilations = var_4746_dilations_0, groups = var_4746_groups_0, pad = var_4746_pad_0, pad_type = var_4746_pad_type_0, strides = var_4746_strides_0, weight = layers_25_fc2_outlier_module_weight_to_fp16_sparsified, x = input_207_cast_fp16)[name = string("op_4746_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_55_cast_fp16 = add(x = var_4740_cast_fp16, y = var_4746_cast_fp16)[name = 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 = string("inputs_105_cast_fp16")];
            int32 var_4756 = const()[name = string("op_4756"), val = int32(3)];
            tensor<int32, [1]> out_105_axes_0 = const()[name = string("out_105_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4775_to_fp16 = const()[name = string("op_4775_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_4775_to_fp16, x = inputs_105_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [1280]> obj_105_gamma_0_to_fp16 = const()[name = string("obj_105_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346900416)))];
            tensor<fp16, [1280]> obj_105_beta_0_to_fp16 = const()[name = string("obj_105_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346903040)))];
            fp16 obj_105_epsilon_0_to_fp16 = const()[name = string("obj_105_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = string("obj_105_cast_fp16")];
            string var_4797_pad_type_0 = const()[name = string("op_4797_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4797_strides_0 = const()[name = string("op_4797_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4797_pad_0 = const()[name = string("op_4797_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4797_dilations_0 = const()[name = string("op_4797_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4797_groups_0 = const()[name = string("op_4797_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346905664))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347724928))))[name = string("layers_26_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_26_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_26_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347725056)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4797_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_4797_dilations_0, groups = var_4797_groups_0, pad = var_4797_pad_0, pad_type = var_4797_pad_type_0, strides = var_4797_strides_0, weight = layers_26_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = string("op_4797_cast_fp16")];
            string var_4803_pad_type_0 = const()[name = string("op_4803_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4803_strides_0 = const()[name = string("op_4803_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4803_pad_0 = const()[name = string("op_4803_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4803_dilations_0 = const()[name = string("op_4803_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4803_groups_0 = const()[name = string("op_4803_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347761152))), nonzero_data = tensor<fp16, [16685]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347727680))))[name = string("layers_26_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4803_cast_fp16 = conv(dilations = var_4803_dilations_0, groups = var_4803_groups_0, pad = var_4803_pad_0, pad_type = var_4803_pad_type_0, strides = var_4803_strides_0, weight = layers_26_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_105_cast_fp16)[name = string("op_4803_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_53_cast_fp16 = add(x = var_4797_cast_fp16, y = var_4803_cast_fp16)[name = string("query_53_cast_fp16")];
            string var_4812_pad_type_0 = const()[name = string("op_4812_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4812_strides_0 = const()[name = string("op_4812_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4812_pad_0 = const()[name = string("op_4812_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4812_dilations_0 = const()[name = string("op_4812_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4812_groups_0 = const()[name = string("op_4812_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347966016))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348785280))))[name = string("layers_26_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4812_cast_fp16 = conv(dilations = var_4812_dilations_0, groups = var_4812_groups_0, pad = var_4812_pad_0, pad_type = var_4812_pad_type_0, strides = var_4812_strides_0, weight = layers_26_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = string("op_4812_cast_fp16")];
            string var_4818_pad_type_0 = const()[name = string("op_4818_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4818_strides_0 = const()[name = string("op_4818_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4818_pad_0 = const()[name = string("op_4818_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4818_dilations_0 = const()[name = string("op_4818_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4818_groups_0 = const()[name = string("op_4818_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348811904))), nonzero_data = tensor<fp16, [13198]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348785408))))[name = string("layers_26_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4818_cast_fp16 = conv(dilations = var_4818_dilations_0, groups = var_4818_groups_0, pad = var_4818_pad_0, pad_type = var_4818_pad_type_0, strides = var_4818_strides_0, weight = layers_26_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_105_cast_fp16)[name = string("op_4818_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_53_cast_fp16 = add(x = var_4812_cast_fp16, y = var_4818_cast_fp16)[name = string("key_53_cast_fp16")];
            string var_4828_pad_type_0 = const()[name = string("op_4828_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4828_strides_0 = const()[name = string("op_4828_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4828_pad_0 = const()[name = string("op_4828_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4828_dilations_0 = const()[name = string("op_4828_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4828_groups_0 = const()[name = string("op_4828_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349016768))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349836032))))[name = string("layers_26_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_26_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_26_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349836160)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4828_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_4828_dilations_0, groups = var_4828_groups_0, pad = var_4828_pad_0, pad_type = var_4828_pad_type_0, strides = var_4828_strides_0, weight = layers_26_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = string("op_4828_cast_fp16")];
            string var_4834_pad_type_0 = const()[name = string("op_4834_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4834_strides_0 = const()[name = string("op_4834_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4834_pad_0 = const()[name = string("op_4834_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4834_dilations_0 = const()[name = string("op_4834_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4834_groups_0 = const()[name = string("op_4834_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349851264))), nonzero_data = tensor<fp16, [6184]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349838784))))[name = string("layers_26_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4834_cast_fp16 = conv(dilations = var_4834_dilations_0, groups = var_4834_groups_0, pad = var_4834_pad_0, pad_type = var_4834_pad_type_0, strides = var_4834_strides_0, weight = layers_26_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_105_cast_fp16)[name = string("op_4834_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_53_cast_fp16 = add(x = var_4828_cast_fp16, y = var_4834_cast_fp16)[name = string("value_53_cast_fp16")];
            tensor<int32, [4]> var_4837 = const()[name = string("op_4837"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_53_cast_fp16 = reshape(shape = var_4837, x = query_53_cast_fp16)[name = string("mh_q_53_cast_fp16")];
            fp16 var_4839_to_fp16 = const()[name = string("op_4839_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_4840_cast_fp16 = mul(x = mh_q_53_cast_fp16, y = var_4839_to_fp16)[name = string("op_4840_cast_fp16")];
            tensor<int32, [4]> var_4841 = const()[name = string("op_4841"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4842_cast_fp16 = reshape(shape = var_4841, x = key_53_cast_fp16)[name = string("op_4842_cast_fp16")];
            bool mh_w_53_transpose_x_0 = const()[name = string("mh_w_53_transpose_x_0"), val = bool(true)];
            bool mh_w_53_transpose_y_0 = const()[name = string("mh_w_53_transpose_y_0"), val = 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_4840_cast_fp16, y = var_4842_cast_fp16)[name = string("mh_w_53_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_4845_cast_fp16 = softmax(axis = var_4756, x = mh_w_53_cast_fp16)[name = string("op_4845_cast_fp16")];
            tensor<int32, [4]> var_4846 = const()[name = string("op_4846"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4847_cast_fp16 = reshape(shape = var_4846, x = value_53_cast_fp16)[name = string("op_4847_cast_fp16")];
            bool attn_53_transpose_x_0 = const()[name = string("attn_53_transpose_x_0"), val = bool(false)];
            bool attn_53_transpose_y_0 = const()[name = string("attn_53_transpose_y_0"), val = 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_4847_cast_fp16, y = var_4845_cast_fp16)[name = string("attn_53_cast_fp16")];
            tensor<int32, [4]> var_4850 = const()[name = string("op_4850"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_209_cast_fp16 = reshape(shape = var_4850, x = attn_53_cast_fp16)[name = string("input_209_cast_fp16")];
            string var_4860_pad_type_0 = const()[name = string("op_4860_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4860_strides_0 = const()[name = string("op_4860_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4860_pad_0 = const()[name = string("op_4860_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4860_dilations_0 = const()[name = string("op_4860_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4860_groups_0 = const()[name = string("op_4860_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350056128))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350875392))))[name = string("layers_26_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_26_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_26_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350875520)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4860_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_4860_dilations_0, groups = var_4860_groups_0, pad = var_4860_pad_0, pad_type = var_4860_pad_type_0, strides = var_4860_strides_0, weight = layers_26_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_209_cast_fp16)[name = string("op_4860_cast_fp16")];
            string var_4866_pad_type_0 = const()[name = string("op_4866_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4866_strides_0 = const()[name = string("op_4866_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4866_pad_0 = const()[name = string("op_4866_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4866_dilations_0 = const()[name = string("op_4866_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4866_groups_0 = const()[name = string("op_4866_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350889984))), nonzero_data = tensor<fp16, [5866]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350878144))))[name = string("layers_26_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4866_cast_fp16 = conv(dilations = var_4866_dilations_0, groups = var_4866_groups_0, pad = var_4866_pad_0, pad_type = var_4866_pad_type_0, strides = var_4866_strides_0, weight = layers_26_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_209_cast_fp16)[name = string("op_4866_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_107_cast_fp16 = add(x = var_4860_cast_fp16, y = var_4866_cast_fp16)[name = 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 = string("inputs_107_cast_fp16")];
            tensor<int32, [1]> out_107_axes_0 = const()[name = string("out_107_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4877_to_fp16 = const()[name = string("op_4877_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_4877_to_fp16, x = inputs_107_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [1280]> input_211_gamma_0_to_fp16 = const()[name = string("input_211_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351094848)))];
            tensor<fp16, [1280]> input_211_beta_0_to_fp16 = const()[name = string("input_211_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351097472)))];
            fp16 input_211_epsilon_0_to_fp16 = const()[name = string("input_211_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = string("input_211_cast_fp16")];
            string var_4895_pad_type_0 = const()[name = string("op_4895_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4895_strides_0 = const()[name = string("op_4895_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4895_pad_0 = const()[name = string("op_4895_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4895_dilations_0 = const()[name = string("op_4895_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4895_groups_0 = const()[name = string("op_4895_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_26_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351100096))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354376960))))[name = string("layers_26_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_26_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_26_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354377088)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_4895_cast_fp16 = conv(bias = layers_26_fc1_inlier_module_bias_to_fp16, dilations = var_4895_dilations_0, groups = var_4895_groups_0, pad = var_4895_pad_0, pad_type = var_4895_pad_type_0, strides = var_4895_strides_0, weight = layers_26_fc1_inlier_module_weight_to_fp16_palettized, x = input_211_cast_fp16)[name = string("op_4895_cast_fp16")];
            string var_4901_pad_type_0 = const()[name = string("op_4901_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4901_strides_0 = const()[name = string("op_4901_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4901_pad_0 = const()[name = string("op_4901_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4901_dilations_0 = const()[name = string("op_4901_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4901_groups_0 = const()[name = string("op_4901_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_26_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354440128))), nonzero_data = tensor<fp16, [26314]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354387392))))[name = string("layers_26_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_4901_cast_fp16 = conv(dilations = var_4901_dilations_0, groups = var_4901_groups_0, pad = var_4901_pad_0, pad_type = var_4901_pad_type_0, strides = var_4901_strides_0, weight = layers_26_fc1_outlier_module_weight_to_fp16_sparsified, x = input_211_cast_fp16)[name = string("op_4901_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_213_cast_fp16 = add(x = var_4895_cast_fp16, y = var_4901_cast_fp16)[name = string("input_213_cast_fp16")];
            string input_215_mode_0 = const()[name = string("input_215_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
            string var_4912_pad_type_0 = const()[name = string("op_4912_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4912_strides_0 = const()[name = string("op_4912_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4912_pad_0 = const()[name = string("op_4912_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4912_dilations_0 = const()[name = string("op_4912_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4912_groups_0 = const()[name = string("op_4912_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_26_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355259392))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358536256))))[name = string("layers_26_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_26_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_26_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358536384)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4912_cast_fp16 = conv(bias = layers_26_fc2_inlier_module_bias_to_fp16, dilations = var_4912_dilations_0, groups = var_4912_groups_0, pad = var_4912_pad_0, pad_type = var_4912_pad_type_0, strides = var_4912_strides_0, weight = layers_26_fc2_inlier_module_weight_to_fp16_palettized, x = input_215_cast_fp16)[name = string("op_4912_cast_fp16")];
            string var_4918_pad_type_0 = const()[name = string("op_4918_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4918_strides_0 = const()[name = string("op_4918_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4918_pad_0 = const()[name = string("op_4918_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4918_dilations_0 = const()[name = string("op_4918_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4918_groups_0 = const()[name = string("op_4918_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_26_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358588352))), nonzero_data = tensor<fp16, [24636]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358539008))))[name = string("layers_26_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4918_cast_fp16 = conv(dilations = var_4918_dilations_0, groups = var_4918_groups_0, pad = var_4918_pad_0, pad_type = var_4918_pad_type_0, strides = var_4918_strides_0, weight = layers_26_fc2_outlier_module_weight_to_fp16_sparsified, x = input_215_cast_fp16)[name = string("op_4918_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_57_cast_fp16 = add(x = var_4912_cast_fp16, y = var_4918_cast_fp16)[name = 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 = string("inputs_109_cast_fp16")];
            int32 var_4928 = const()[name = string("op_4928"), val = int32(3)];
            tensor<int32, [1]> out_109_axes_0 = const()[name = string("out_109_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_4947_to_fp16 = const()[name = string("op_4947_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_4947_to_fp16, x = inputs_109_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [1280]> obj_109_gamma_0_to_fp16 = const()[name = string("obj_109_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359407616)))];
            tensor<fp16, [1280]> obj_109_beta_0_to_fp16 = const()[name = string("obj_109_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359410240)))];
            fp16 obj_109_epsilon_0_to_fp16 = const()[name = string("obj_109_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = string("obj_109_cast_fp16")];
            string var_4969_pad_type_0 = const()[name = string("op_4969_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4969_strides_0 = const()[name = string("op_4969_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4969_pad_0 = const()[name = string("op_4969_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4969_dilations_0 = const()[name = string("op_4969_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4969_groups_0 = const()[name = string("op_4969_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359412864))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360232128))))[name = string("layers_27_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_27_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_27_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360232256)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_4969_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_4969_dilations_0, groups = var_4969_groups_0, pad = var_4969_pad_0, pad_type = var_4969_pad_type_0, strides = var_4969_strides_0, weight = layers_27_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = string("op_4969_cast_fp16")];
            string var_4975_pad_type_0 = const()[name = string("op_4975_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4975_strides_0 = const()[name = string("op_4975_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4975_pad_0 = const()[name = string("op_4975_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4975_dilations_0 = const()[name = string("op_4975_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4975_groups_0 = const()[name = string("op_4975_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360271616))), nonzero_data = tensor<fp16, [18336]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360234880))))[name = string("layers_27_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4975_cast_fp16 = conv(dilations = var_4975_dilations_0, groups = var_4975_groups_0, pad = var_4975_pad_0, pad_type = var_4975_pad_type_0, strides = var_4975_strides_0, weight = layers_27_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_109_cast_fp16)[name = string("op_4975_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_55_cast_fp16 = add(x = var_4969_cast_fp16, y = var_4975_cast_fp16)[name = string("query_55_cast_fp16")];
            string var_4984_pad_type_0 = const()[name = string("op_4984_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4984_strides_0 = const()[name = string("op_4984_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4984_pad_0 = const()[name = string("op_4984_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4984_dilations_0 = const()[name = string("op_4984_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4984_groups_0 = const()[name = string("op_4984_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360476480))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361295744))))[name = string("layers_27_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4984_cast_fp16 = conv(dilations = var_4984_dilations_0, groups = var_4984_groups_0, pad = var_4984_pad_0, pad_type = var_4984_pad_type_0, strides = var_4984_strides_0, weight = layers_27_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = string("op_4984_cast_fp16")];
            string var_4990_pad_type_0 = const()[name = string("op_4990_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4990_strides_0 = const()[name = string("op_4990_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4990_pad_0 = const()[name = string("op_4990_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4990_dilations_0 = const()[name = string("op_4990_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4990_groups_0 = const()[name = string("op_4990_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361321984))), nonzero_data = tensor<fp16, [12996]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361295872))))[name = string("layers_27_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_4990_cast_fp16 = conv(dilations = var_4990_dilations_0, groups = var_4990_groups_0, pad = var_4990_pad_0, pad_type = var_4990_pad_type_0, strides = var_4990_strides_0, weight = layers_27_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_109_cast_fp16)[name = string("op_4990_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_55_cast_fp16 = add(x = var_4984_cast_fp16, y = var_4990_cast_fp16)[name = string("key_55_cast_fp16")];
            string var_5000_pad_type_0 = const()[name = string("op_5000_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5000_strides_0 = const()[name = string("op_5000_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5000_pad_0 = const()[name = string("op_5000_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5000_dilations_0 = const()[name = string("op_5000_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5000_groups_0 = const()[name = string("op_5000_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361526848))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362346112))))[name = string("layers_27_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_27_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_27_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362346240)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5000_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5000_dilations_0, groups = var_5000_groups_0, pad = var_5000_pad_0, pad_type = var_5000_pad_type_0, strides = var_5000_strides_0, weight = layers_27_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = string("op_5000_cast_fp16")];
            string var_5006_pad_type_0 = const()[name = string("op_5006_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5006_strides_0 = const()[name = string("op_5006_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5006_pad_0 = const()[name = string("op_5006_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5006_dilations_0 = const()[name = string("op_5006_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5006_groups_0 = const()[name = string("op_5006_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362359552))), nonzero_data = tensor<fp16, [5306]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362348864))))[name = string("layers_27_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5006_cast_fp16 = conv(dilations = var_5006_dilations_0, groups = var_5006_groups_0, pad = var_5006_pad_0, pad_type = var_5006_pad_type_0, strides = var_5006_strides_0, weight = layers_27_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_109_cast_fp16)[name = string("op_5006_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_55_cast_fp16 = add(x = var_5000_cast_fp16, y = var_5006_cast_fp16)[name = string("value_55_cast_fp16")];
            tensor<int32, [4]> var_5009 = const()[name = string("op_5009"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_55_cast_fp16 = reshape(shape = var_5009, x = query_55_cast_fp16)[name = string("mh_q_55_cast_fp16")];
            fp16 var_5011_to_fp16 = const()[name = string("op_5011_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_5012_cast_fp16 = mul(x = mh_q_55_cast_fp16, y = var_5011_to_fp16)[name = string("op_5012_cast_fp16")];
            tensor<int32, [4]> var_5013 = const()[name = string("op_5013"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5014_cast_fp16 = reshape(shape = var_5013, x = key_55_cast_fp16)[name = string("op_5014_cast_fp16")];
            bool mh_w_55_transpose_x_0 = const()[name = string("mh_w_55_transpose_x_0"), val = bool(true)];
            bool mh_w_55_transpose_y_0 = const()[name = string("mh_w_55_transpose_y_0"), val = 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_5012_cast_fp16, y = var_5014_cast_fp16)[name = string("mh_w_55_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_5017_cast_fp16 = softmax(axis = var_4928, x = mh_w_55_cast_fp16)[name = string("op_5017_cast_fp16")];
            tensor<int32, [4]> var_5018 = const()[name = string("op_5018"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5019_cast_fp16 = reshape(shape = var_5018, x = value_55_cast_fp16)[name = string("op_5019_cast_fp16")];
            bool attn_55_transpose_x_0 = const()[name = string("attn_55_transpose_x_0"), val = bool(false)];
            bool attn_55_transpose_y_0 = const()[name = string("attn_55_transpose_y_0"), val = 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_5019_cast_fp16, y = var_5017_cast_fp16)[name = string("attn_55_cast_fp16")];
            tensor<int32, [4]> var_5022 = const()[name = string("op_5022"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_217_cast_fp16 = reshape(shape = var_5022, x = attn_55_cast_fp16)[name = string("input_217_cast_fp16")];
            string var_5032_pad_type_0 = const()[name = string("op_5032_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5032_strides_0 = const()[name = string("op_5032_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5032_pad_0 = const()[name = string("op_5032_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5032_dilations_0 = const()[name = string("op_5032_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5032_groups_0 = const()[name = string("op_5032_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362564416))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363383680))))[name = string("layers_27_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_27_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_27_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363383808)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5032_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_5032_dilations_0, groups = var_5032_groups_0, pad = var_5032_pad_0, pad_type = var_5032_pad_type_0, strides = var_5032_strides_0, weight = layers_27_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_217_cast_fp16)[name = string("op_5032_cast_fp16")];
            string var_5038_pad_type_0 = const()[name = string("op_5038_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5038_strides_0 = const()[name = string("op_5038_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5038_pad_0 = const()[name = string("op_5038_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5038_dilations_0 = const()[name = string("op_5038_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5038_groups_0 = const()[name = string("op_5038_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363396544))), nonzero_data = tensor<fp16, [4994]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363386432))))[name = string("layers_27_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5038_cast_fp16 = conv(dilations = var_5038_dilations_0, groups = var_5038_groups_0, pad = var_5038_pad_0, pad_type = var_5038_pad_type_0, strides = var_5038_strides_0, weight = layers_27_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_217_cast_fp16)[name = string("op_5038_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_111_cast_fp16 = add(x = var_5032_cast_fp16, y = var_5038_cast_fp16)[name = 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 = string("inputs_111_cast_fp16")];
            tensor<int32, [1]> out_111_axes_0 = const()[name = string("out_111_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5049_to_fp16 = const()[name = string("op_5049_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_5049_to_fp16, x = inputs_111_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [1280]> input_219_gamma_0_to_fp16 = const()[name = string("input_219_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363601408)))];
            tensor<fp16, [1280]> input_219_beta_0_to_fp16 = const()[name = string("input_219_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363604032)))];
            fp16 input_219_epsilon_0_to_fp16 = const()[name = string("input_219_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = string("input_219_cast_fp16")];
            string var_5067_pad_type_0 = const()[name = string("op_5067_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5067_strides_0 = const()[name = string("op_5067_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5067_pad_0 = const()[name = string("op_5067_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5067_dilations_0 = const()[name = string("op_5067_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5067_groups_0 = const()[name = string("op_5067_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_27_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363606656))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366883520))))[name = string("layers_27_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_27_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_27_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366883648)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_5067_cast_fp16 = conv(bias = layers_27_fc1_inlier_module_bias_to_fp16, dilations = var_5067_dilations_0, groups = var_5067_groups_0, pad = var_5067_pad_0, pad_type = var_5067_pad_type_0, strides = var_5067_strides_0, weight = layers_27_fc1_inlier_module_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = string("op_5067_cast_fp16")];
            string var_5073_pad_type_0 = const()[name = string("op_5073_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5073_strides_0 = const()[name = string("op_5073_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5073_pad_0 = const()[name = string("op_5073_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5073_dilations_0 = const()[name = string("op_5073_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5073_groups_0 = const()[name = string("op_5073_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_27_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366948672))), nonzero_data = tensor<fp16, [27310]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366893952))))[name = string("layers_27_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_5073_cast_fp16 = conv(dilations = var_5073_dilations_0, groups = var_5073_groups_0, pad = var_5073_pad_0, pad_type = var_5073_pad_type_0, strides = var_5073_strides_0, weight = layers_27_fc1_outlier_module_weight_to_fp16_sparsified, x = input_219_cast_fp16)[name = string("op_5073_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_221_cast_fp16 = add(x = var_5067_cast_fp16, y = var_5073_cast_fp16)[name = string("input_221_cast_fp16")];
            string input_223_mode_0 = const()[name = string("input_223_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_223_cast_fp16 = gelu(mode = input_223_mode_0, x = input_221_cast_fp16)[name = string("input_223_cast_fp16")];
            string var_5084_pad_type_0 = const()[name = string("op_5084_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5084_strides_0 = const()[name = string("op_5084_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5084_pad_0 = const()[name = string("op_5084_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5084_dilations_0 = const()[name = string("op_5084_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5084_groups_0 = const()[name = string("op_5084_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_27_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367767936))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371044800))))[name = string("layers_27_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_27_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_27_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371044928)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5084_cast_fp16 = conv(bias = layers_27_fc2_inlier_module_bias_to_fp16, dilations = var_5084_dilations_0, groups = var_5084_groups_0, pad = var_5084_pad_0, pad_type = var_5084_pad_type_0, strides = var_5084_strides_0, weight = layers_27_fc2_inlier_module_weight_to_fp16_palettized, x = input_223_cast_fp16)[name = string("op_5084_cast_fp16")];
            string var_5090_pad_type_0 = const()[name = string("op_5090_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5090_strides_0 = const()[name = string("op_5090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5090_pad_0 = const()[name = string("op_5090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5090_dilations_0 = const()[name = string("op_5090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5090_groups_0 = const()[name = string("op_5090_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_27_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371098048))), nonzero_data = tensor<fp16, [25203]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371047552))))[name = string("layers_27_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5090_cast_fp16 = conv(dilations = var_5090_dilations_0, groups = var_5090_groups_0, pad = var_5090_pad_0, pad_type = var_5090_pad_type_0, strides = var_5090_strides_0, weight = layers_27_fc2_outlier_module_weight_to_fp16_sparsified, x = input_223_cast_fp16)[name = string("op_5090_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_59_cast_fp16 = add(x = var_5084_cast_fp16, y = var_5090_cast_fp16)[name = 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 = string("inputs_113_cast_fp16")];
            int32 var_5100 = const()[name = string("op_5100"), val = int32(3)];
            tensor<int32, [1]> out_113_axes_0 = const()[name = string("out_113_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5119_to_fp16 = const()[name = string("op_5119_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_5119_to_fp16, x = inputs_113_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [1280]> obj_113_gamma_0_to_fp16 = const()[name = string("obj_113_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371917312)))];
            tensor<fp16, [1280]> obj_113_beta_0_to_fp16 = const()[name = string("obj_113_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371919936)))];
            fp16 obj_113_epsilon_0_to_fp16 = const()[name = string("obj_113_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = string("obj_113_cast_fp16")];
            string var_5141_pad_type_0 = const()[name = string("op_5141_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5141_strides_0 = const()[name = string("op_5141_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5141_pad_0 = const()[name = string("op_5141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5141_dilations_0 = const()[name = string("op_5141_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5141_groups_0 = const()[name = string("op_5141_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371922560))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372741824))))[name = string("layers_28_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_28_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_28_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372741952)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5141_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_5141_dilations_0, groups = var_5141_groups_0, pad = var_5141_pad_0, pad_type = var_5141_pad_type_0, strides = var_5141_strides_0, weight = layers_28_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = string("op_5141_cast_fp16")];
            string var_5147_pad_type_0 = const()[name = string("op_5147_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5147_strides_0 = const()[name = string("op_5147_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5147_pad_0 = const()[name = string("op_5147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5147_dilations_0 = const()[name = string("op_5147_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5147_groups_0 = const()[name = string("op_5147_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372772992))), nonzero_data = tensor<fp16, [14146]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372744576))))[name = string("layers_28_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5147_cast_fp16 = conv(dilations = var_5147_dilations_0, groups = var_5147_groups_0, pad = var_5147_pad_0, pad_type = var_5147_pad_type_0, strides = var_5147_strides_0, weight = layers_28_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_113_cast_fp16)[name = string("op_5147_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_57_cast_fp16 = add(x = var_5141_cast_fp16, y = var_5147_cast_fp16)[name = string("query_57_cast_fp16")];
            string var_5156_pad_type_0 = const()[name = string("op_5156_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5156_strides_0 = const()[name = string("op_5156_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5156_pad_0 = const()[name = string("op_5156_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5156_dilations_0 = const()[name = string("op_5156_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5156_groups_0 = const()[name = string("op_5156_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372977856))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373797120))))[name = string("layers_28_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5156_cast_fp16 = conv(dilations = var_5156_dilations_0, groups = var_5156_groups_0, pad = var_5156_pad_0, pad_type = var_5156_pad_type_0, strides = var_5156_strides_0, weight = layers_28_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = string("op_5156_cast_fp16")];
            string var_5162_pad_type_0 = const()[name = string("op_5162_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5162_strides_0 = const()[name = string("op_5162_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5162_pad_0 = const()[name = string("op_5162_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5162_dilations_0 = const()[name = string("op_5162_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5162_groups_0 = const()[name = string("op_5162_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373819328))), nonzero_data = tensor<fp16, [10981]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373797248))))[name = string("layers_28_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5162_cast_fp16 = conv(dilations = var_5162_dilations_0, groups = var_5162_groups_0, pad = var_5162_pad_0, pad_type = var_5162_pad_type_0, strides = var_5162_strides_0, weight = layers_28_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_113_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_57_cast_fp16 = add(x = var_5156_cast_fp16, y = var_5162_cast_fp16)[name = string("key_57_cast_fp16")];
            string var_5172_pad_type_0 = const()[name = string("op_5172_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5172_strides_0 = const()[name = string("op_5172_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5172_pad_0 = const()[name = string("op_5172_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5172_dilations_0 = const()[name = string("op_5172_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5172_groups_0 = const()[name = string("op_5172_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374024192))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374843456))))[name = string("layers_28_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_28_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_28_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374843584)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5172_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5172_dilations_0, groups = var_5172_groups_0, pad = var_5172_pad_0, pad_type = var_5172_pad_type_0, strides = var_5172_strides_0, weight = layers_28_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = string("op_5172_cast_fp16")];
            string var_5178_pad_type_0 = const()[name = string("op_5178_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5178_strides_0 = const()[name = string("op_5178_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5178_pad_0 = const()[name = string("op_5178_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5178_dilations_0 = const()[name = string("op_5178_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5178_groups_0 = const()[name = string("op_5178_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374858048))), nonzero_data = tensor<fp16, [5866]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374846208))))[name = string("layers_28_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5178_cast_fp16 = conv(dilations = var_5178_dilations_0, groups = var_5178_groups_0, pad = var_5178_pad_0, pad_type = var_5178_pad_type_0, strides = var_5178_strides_0, weight = layers_28_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_113_cast_fp16)[name = string("op_5178_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_57_cast_fp16 = add(x = var_5172_cast_fp16, y = var_5178_cast_fp16)[name = string("value_57_cast_fp16")];
            tensor<int32, [4]> var_5181 = const()[name = string("op_5181"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_57_cast_fp16 = reshape(shape = var_5181, x = query_57_cast_fp16)[name = string("mh_q_57_cast_fp16")];
            fp16 var_5183_to_fp16 = const()[name = string("op_5183_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_5184_cast_fp16 = mul(x = mh_q_57_cast_fp16, y = var_5183_to_fp16)[name = string("op_5184_cast_fp16")];
            tensor<int32, [4]> var_5185 = const()[name = string("op_5185"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5186_cast_fp16 = reshape(shape = var_5185, x = key_57_cast_fp16)[name = string("op_5186_cast_fp16")];
            bool mh_w_57_transpose_x_0 = const()[name = string("mh_w_57_transpose_x_0"), val = bool(true)];
            bool mh_w_57_transpose_y_0 = const()[name = string("mh_w_57_transpose_y_0"), val = 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_5184_cast_fp16, y = var_5186_cast_fp16)[name = string("mh_w_57_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_5189_cast_fp16 = softmax(axis = var_5100, x = mh_w_57_cast_fp16)[name = string("op_5189_cast_fp16")];
            tensor<int32, [4]> var_5190 = const()[name = string("op_5190"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5191_cast_fp16 = reshape(shape = var_5190, x = value_57_cast_fp16)[name = string("op_5191_cast_fp16")];
            bool attn_57_transpose_x_0 = const()[name = string("attn_57_transpose_x_0"), val = bool(false)];
            bool attn_57_transpose_y_0 = const()[name = string("attn_57_transpose_y_0"), val = 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_5191_cast_fp16, y = var_5189_cast_fp16)[name = string("attn_57_cast_fp16")];
            tensor<int32, [4]> var_5194 = const()[name = string("op_5194"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_225_cast_fp16 = reshape(shape = var_5194, x = attn_57_cast_fp16)[name = string("input_225_cast_fp16")];
            string var_5204_pad_type_0 = const()[name = string("op_5204_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5204_strides_0 = const()[name = string("op_5204_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5204_pad_0 = const()[name = string("op_5204_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5204_dilations_0 = const()[name = string("op_5204_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5204_groups_0 = const()[name = string("op_5204_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375062912))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375882176))))[name = string("layers_28_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_28_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_28_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375882304)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5204_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_5204_dilations_0, groups = var_5204_groups_0, pad = var_5204_pad_0, pad_type = var_5204_pad_type_0, strides = var_5204_strides_0, weight = layers_28_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("op_5204_cast_fp16")];
            string var_5210_pad_type_0 = const()[name = string("op_5210_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5210_strides_0 = const()[name = string("op_5210_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5210_pad_0 = const()[name = string("op_5210_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5210_dilations_0 = const()[name = string("op_5210_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5210_groups_0 = const()[name = string("op_5210_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375896192))), nonzero_data = tensor<fp16, [5591]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375884928))))[name = string("layers_28_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5210_cast_fp16 = conv(dilations = var_5210_dilations_0, groups = var_5210_groups_0, pad = var_5210_pad_0, pad_type = var_5210_pad_type_0, strides = var_5210_strides_0, weight = layers_28_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_225_cast_fp16)[name = string("op_5210_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_115_cast_fp16 = add(x = var_5204_cast_fp16, y = var_5210_cast_fp16)[name = 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 = string("inputs_115_cast_fp16")];
            tensor<int32, [1]> out_115_axes_0 = const()[name = string("out_115_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5221_to_fp16 = const()[name = string("op_5221_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_5221_to_fp16, x = inputs_115_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [1280]> input_227_gamma_0_to_fp16 = const()[name = string("input_227_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376101056)))];
            tensor<fp16, [1280]> input_227_beta_0_to_fp16 = const()[name = string("input_227_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376103680)))];
            fp16 input_227_epsilon_0_to_fp16 = const()[name = string("input_227_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = string("input_227_cast_fp16")];
            string var_5239_pad_type_0 = const()[name = string("op_5239_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5239_strides_0 = const()[name = string("op_5239_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5239_pad_0 = const()[name = string("op_5239_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5239_dilations_0 = const()[name = string("op_5239_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5239_groups_0 = const()[name = string("op_5239_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_28_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376106304))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379383168))))[name = string("layers_28_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_28_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_28_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379383296)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_5239_cast_fp16 = conv(bias = layers_28_fc1_inlier_module_bias_to_fp16, dilations = var_5239_dilations_0, groups = var_5239_groups_0, pad = var_5239_pad_0, pad_type = var_5239_pad_type_0, strides = var_5239_strides_0, weight = layers_28_fc1_inlier_module_weight_to_fp16_palettized, x = input_227_cast_fp16)[name = string("op_5239_cast_fp16")];
            string var_5245_pad_type_0 = const()[name = string("op_5245_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5245_strides_0 = const()[name = string("op_5245_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5245_pad_0 = const()[name = string("op_5245_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5245_dilations_0 = const()[name = string("op_5245_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5245_groups_0 = const()[name = string("op_5245_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_28_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379450816))), nonzero_data = tensor<fp16, [28562]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379393600))))[name = string("layers_28_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_5245_cast_fp16 = conv(dilations = var_5245_dilations_0, groups = var_5245_groups_0, pad = var_5245_pad_0, pad_type = var_5245_pad_type_0, strides = var_5245_strides_0, weight = layers_28_fc1_outlier_module_weight_to_fp16_sparsified, x = input_227_cast_fp16)[name = string("op_5245_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_229_cast_fp16 = add(x = var_5239_cast_fp16, y = var_5245_cast_fp16)[name = string("input_229_cast_fp16")];
            string input_231_mode_0 = const()[name = string("input_231_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = string("input_231_cast_fp16")];
            string var_5256_pad_type_0 = const()[name = string("op_5256_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5256_strides_0 = const()[name = string("op_5256_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5256_pad_0 = const()[name = string("op_5256_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5256_dilations_0 = const()[name = string("op_5256_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5256_groups_0 = const()[name = string("op_5256_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_28_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380270080))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383546944))))[name = string("layers_28_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_28_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_28_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383547072)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5256_cast_fp16 = conv(bias = layers_28_fc2_inlier_module_bias_to_fp16, dilations = var_5256_dilations_0, groups = var_5256_groups_0, pad = var_5256_pad_0, pad_type = var_5256_pad_type_0, strides = var_5256_strides_0, weight = layers_28_fc2_inlier_module_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = string("op_5256_cast_fp16")];
            string var_5262_pad_type_0 = const()[name = string("op_5262_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5262_strides_0 = const()[name = string("op_5262_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5262_pad_0 = const()[name = string("op_5262_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5262_dilations_0 = const()[name = string("op_5262_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5262_groups_0 = const()[name = string("op_5262_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_28_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383608320))), nonzero_data = tensor<fp16, [29259]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383549696))))[name = string("layers_28_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5262_cast_fp16 = conv(dilations = var_5262_dilations_0, groups = var_5262_groups_0, pad = var_5262_pad_0, pad_type = var_5262_pad_type_0, strides = var_5262_strides_0, weight = layers_28_fc2_outlier_module_weight_to_fp16_sparsified, x = input_231_cast_fp16)[name = string("op_5262_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_61_cast_fp16 = add(x = var_5256_cast_fp16, y = var_5262_cast_fp16)[name = 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 = string("inputs_117_cast_fp16")];
            int32 var_5272 = const()[name = string("op_5272"), val = int32(3)];
            tensor<int32, [1]> out_117_axes_0 = const()[name = string("out_117_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5291_to_fp16 = const()[name = string("op_5291_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_5291_to_fp16, x = inputs_117_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [1280]> obj_117_gamma_0_to_fp16 = const()[name = string("obj_117_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384427584)))];
            tensor<fp16, [1280]> obj_117_beta_0_to_fp16 = const()[name = string("obj_117_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384430208)))];
            fp16 obj_117_epsilon_0_to_fp16 = const()[name = string("obj_117_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = string("obj_117_cast_fp16")];
            string var_5313_pad_type_0 = const()[name = string("op_5313_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5313_strides_0 = const()[name = string("op_5313_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5313_pad_0 = const()[name = string("op_5313_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5313_dilations_0 = const()[name = string("op_5313_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5313_groups_0 = const()[name = string("op_5313_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384432832))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385252096))))[name = string("layers_29_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_29_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_29_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385252224)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5313_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_5313_dilations_0, groups = var_5313_groups_0, pad = var_5313_pad_0, pad_type = var_5313_pad_type_0, strides = var_5313_strides_0, weight = layers_29_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = string("op_5313_cast_fp16")];
            string var_5319_pad_type_0 = const()[name = string("op_5319_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5319_strides_0 = const()[name = string("op_5319_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5319_pad_0 = const()[name = string("op_5319_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5319_dilations_0 = const()[name = string("op_5319_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5319_groups_0 = const()[name = string("op_5319_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385277312))), nonzero_data = tensor<fp16, [11179]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385254848))))[name = string("layers_29_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5319_cast_fp16 = conv(dilations = var_5319_dilations_0, groups = var_5319_groups_0, pad = var_5319_pad_0, pad_type = var_5319_pad_type_0, strides = var_5319_strides_0, weight = layers_29_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_117_cast_fp16)[name = string("op_5319_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_59_cast_fp16 = add(x = var_5313_cast_fp16, y = var_5319_cast_fp16)[name = string("query_59_cast_fp16")];
            string var_5328_pad_type_0 = const()[name = string("op_5328_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5328_strides_0 = const()[name = string("op_5328_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5328_pad_0 = const()[name = string("op_5328_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5328_dilations_0 = const()[name = string("op_5328_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5328_groups_0 = const()[name = string("op_5328_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385482176))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386301440))))[name = string("layers_29_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5328_cast_fp16 = conv(dilations = var_5328_dilations_0, groups = var_5328_groups_0, pad = var_5328_pad_0, pad_type = var_5328_pad_type_0, strides = var_5328_strides_0, weight = layers_29_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = string("op_5328_cast_fp16")];
            string var_5334_pad_type_0 = const()[name = string("op_5334_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5334_strides_0 = const()[name = string("op_5334_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5334_pad_0 = const()[name = string("op_5334_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5334_dilations_0 = const()[name = string("op_5334_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5334_groups_0 = const()[name = string("op_5334_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386323840))), nonzero_data = tensor<fp16, [11076]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386301568))))[name = string("layers_29_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5334_cast_fp16 = conv(dilations = var_5334_dilations_0, groups = var_5334_groups_0, pad = var_5334_pad_0, pad_type = var_5334_pad_type_0, strides = var_5334_strides_0, weight = layers_29_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_117_cast_fp16)[name = string("op_5334_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_59_cast_fp16 = add(x = var_5328_cast_fp16, y = var_5334_cast_fp16)[name = string("key_59_cast_fp16")];
            string var_5344_pad_type_0 = const()[name = string("op_5344_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5344_strides_0 = const()[name = string("op_5344_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5344_pad_0 = const()[name = string("op_5344_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5344_dilations_0 = const()[name = string("op_5344_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5344_groups_0 = const()[name = string("op_5344_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386528704))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387347968))))[name = string("layers_29_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_29_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_29_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387348096)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5344_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5344_dilations_0, groups = var_5344_groups_0, pad = var_5344_pad_0, pad_type = var_5344_pad_type_0, strides = var_5344_strides_0, weight = layers_29_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = string("op_5344_cast_fp16")];
            string var_5350_pad_type_0 = const()[name = string("op_5350_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5350_strides_0 = const()[name = string("op_5350_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5350_pad_0 = const()[name = string("op_5350_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5350_dilations_0 = const()[name = string("op_5350_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5350_groups_0 = const()[name = string("op_5350_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387363392))), nonzero_data = tensor<fp16, [6296]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387350720))))[name = string("layers_29_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5350_cast_fp16 = conv(dilations = var_5350_dilations_0, groups = var_5350_groups_0, pad = var_5350_pad_0, pad_type = var_5350_pad_type_0, strides = var_5350_strides_0, weight = layers_29_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_117_cast_fp16)[name = string("op_5350_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_59_cast_fp16 = add(x = var_5344_cast_fp16, y = var_5350_cast_fp16)[name = string("value_59_cast_fp16")];
            tensor<int32, [4]> var_5353 = const()[name = string("op_5353"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_59_cast_fp16 = reshape(shape = var_5353, x = query_59_cast_fp16)[name = string("mh_q_59_cast_fp16")];
            fp16 var_5355_to_fp16 = const()[name = string("op_5355_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_5356_cast_fp16 = mul(x = mh_q_59_cast_fp16, y = var_5355_to_fp16)[name = string("op_5356_cast_fp16")];
            tensor<int32, [4]> var_5357 = const()[name = string("op_5357"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5358_cast_fp16 = reshape(shape = var_5357, x = key_59_cast_fp16)[name = string("op_5358_cast_fp16")];
            bool mh_w_59_transpose_x_0 = const()[name = string("mh_w_59_transpose_x_0"), val = bool(true)];
            bool mh_w_59_transpose_y_0 = const()[name = string("mh_w_59_transpose_y_0"), val = 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_5356_cast_fp16, y = var_5358_cast_fp16)[name = string("mh_w_59_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_5361_cast_fp16 = softmax(axis = var_5272, x = mh_w_59_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<int32, [4]> var_5362 = const()[name = string("op_5362"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5363_cast_fp16 = reshape(shape = var_5362, x = value_59_cast_fp16)[name = string("op_5363_cast_fp16")];
            bool attn_59_transpose_x_0 = const()[name = string("attn_59_transpose_x_0"), val = bool(false)];
            bool attn_59_transpose_y_0 = const()[name = string("attn_59_transpose_y_0"), val = 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_5363_cast_fp16, y = var_5361_cast_fp16)[name = string("attn_59_cast_fp16")];
            tensor<int32, [4]> var_5366 = const()[name = string("op_5366"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_233_cast_fp16 = reshape(shape = var_5366, x = attn_59_cast_fp16)[name = string("input_233_cast_fp16")];
            string var_5376_pad_type_0 = const()[name = string("op_5376_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5376_strides_0 = const()[name = string("op_5376_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5376_pad_0 = const()[name = string("op_5376_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5376_dilations_0 = const()[name = string("op_5376_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5376_groups_0 = const()[name = string("op_5376_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387568256))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388387520))))[name = string("layers_29_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_29_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_29_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388387648)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5376_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_5376_dilations_0, groups = var_5376_groups_0, pad = var_5376_pad_0, pad_type = var_5376_pad_type_0, strides = var_5376_strides_0, weight = layers_29_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_233_cast_fp16)[name = string("op_5376_cast_fp16")];
            string var_5382_pad_type_0 = const()[name = string("op_5382_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5382_strides_0 = const()[name = string("op_5382_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5382_pad_0 = const()[name = string("op_5382_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5382_dilations_0 = const()[name = string("op_5382_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5382_groups_0 = const()[name = string("op_5382_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388402304))), nonzero_data = tensor<fp16, [5965]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388390272))))[name = string("layers_29_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5382_cast_fp16 = conv(dilations = var_5382_dilations_0, groups = var_5382_groups_0, pad = var_5382_pad_0, pad_type = var_5382_pad_type_0, strides = var_5382_strides_0, weight = layers_29_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_233_cast_fp16)[name = string("op_5382_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_119_cast_fp16 = add(x = var_5376_cast_fp16, y = var_5382_cast_fp16)[name = 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 = string("inputs_119_cast_fp16")];
            tensor<int32, [1]> out_119_axes_0 = const()[name = string("out_119_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5393_to_fp16 = const()[name = string("op_5393_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_5393_to_fp16, x = inputs_119_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [1280]> input_235_gamma_0_to_fp16 = const()[name = string("input_235_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388607168)))];
            tensor<fp16, [1280]> input_235_beta_0_to_fp16 = const()[name = string("input_235_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388609792)))];
            fp16 input_235_epsilon_0_to_fp16 = const()[name = string("input_235_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = string("input_235_cast_fp16")];
            string var_5411_pad_type_0 = const()[name = string("op_5411_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5411_strides_0 = const()[name = string("op_5411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5411_pad_0 = const()[name = string("op_5411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5411_dilations_0 = const()[name = string("op_5411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5411_groups_0 = const()[name = string("op_5411_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_29_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388612416))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391889280))))[name = string("layers_29_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_29_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_29_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391889408)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_5411_cast_fp16 = conv(bias = layers_29_fc1_inlier_module_bias_to_fp16, dilations = var_5411_dilations_0, groups = var_5411_groups_0, pad = var_5411_pad_0, pad_type = var_5411_pad_type_0, strides = var_5411_strides_0, weight = layers_29_fc1_inlier_module_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("op_5411_cast_fp16")];
            string var_5417_pad_type_0 = const()[name = string("op_5417_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5417_strides_0 = const()[name = string("op_5417_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5417_pad_0 = const()[name = string("op_5417_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5417_dilations_0 = const()[name = string("op_5417_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5417_groups_0 = const()[name = string("op_5417_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_29_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391957888))), nonzero_data = tensor<fp16, [29027]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391899712))))[name = string("layers_29_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_5417_cast_fp16 = conv(dilations = var_5417_dilations_0, groups = var_5417_groups_0, pad = var_5417_pad_0, pad_type = var_5417_pad_type_0, strides = var_5417_strides_0, weight = layers_29_fc1_outlier_module_weight_to_fp16_sparsified, x = input_235_cast_fp16)[name = string("op_5417_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_237_cast_fp16 = add(x = var_5411_cast_fp16, y = var_5417_cast_fp16)[name = string("input_237_cast_fp16")];
            string input_239_mode_0 = const()[name = string("input_239_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = input_237_cast_fp16)[name = string("input_239_cast_fp16")];
            string var_5428_pad_type_0 = const()[name = string("op_5428_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5428_strides_0 = const()[name = string("op_5428_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5428_pad_0 = const()[name = string("op_5428_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5428_dilations_0 = const()[name = string("op_5428_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5428_groups_0 = const()[name = string("op_5428_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_29_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392777152))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396054016))))[name = string("layers_29_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_29_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_29_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396054144)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5428_cast_fp16 = conv(bias = layers_29_fc2_inlier_module_bias_to_fp16, dilations = var_5428_dilations_0, groups = var_5428_groups_0, pad = var_5428_pad_0, pad_type = var_5428_pad_type_0, strides = var_5428_strides_0, weight = layers_29_fc2_inlier_module_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("op_5428_cast_fp16")];
            string var_5434_pad_type_0 = const()[name = string("op_5434_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5434_strides_0 = const()[name = string("op_5434_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5434_pad_0 = const()[name = string("op_5434_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5434_dilations_0 = const()[name = string("op_5434_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5434_groups_0 = const()[name = string("op_5434_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_29_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396118272))), nonzero_data = tensor<fp16, [30702]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396056768))))[name = string("layers_29_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5434_cast_fp16 = conv(dilations = var_5434_dilations_0, groups = var_5434_groups_0, pad = var_5434_pad_0, pad_type = var_5434_pad_type_0, strides = var_5434_strides_0, weight = layers_29_fc2_outlier_module_weight_to_fp16_sparsified, x = input_239_cast_fp16)[name = string("op_5434_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_63_cast_fp16 = add(x = var_5428_cast_fp16, y = var_5434_cast_fp16)[name = 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 = string("inputs_121_cast_fp16")];
            int32 var_5444 = const()[name = string("op_5444"), val = int32(3)];
            tensor<int32, [1]> out_121_axes_0 = const()[name = string("out_121_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5463_to_fp16 = const()[name = string("op_5463_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_5463_to_fp16, x = inputs_121_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [1280]> obj_121_gamma_0_to_fp16 = const()[name = string("obj_121_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396937536)))];
            tensor<fp16, [1280]> obj_121_beta_0_to_fp16 = const()[name = string("obj_121_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396940160)))];
            fp16 obj_121_epsilon_0_to_fp16 = const()[name = string("obj_121_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = string("obj_121_cast_fp16")];
            string var_5485_pad_type_0 = const()[name = string("op_5485_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5485_strides_0 = const()[name = string("op_5485_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5485_pad_0 = const()[name = string("op_5485_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5485_dilations_0 = const()[name = string("op_5485_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5485_groups_0 = const()[name = string("op_5485_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396942784))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397762048))))[name = string("layers_30_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_30_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_30_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397762176)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5485_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_5485_dilations_0, groups = var_5485_groups_0, pad = var_5485_pad_0, pad_type = var_5485_pad_type_0, strides = var_5485_strides_0, weight = layers_30_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = string("op_5485_cast_fp16")];
            string var_5491_pad_type_0 = const()[name = string("op_5491_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5491_strides_0 = const()[name = string("op_5491_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5491_pad_0 = const()[name = string("op_5491_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5491_dilations_0 = const()[name = string("op_5491_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5491_groups_0 = const()[name = string("op_5491_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397784960))), nonzero_data = tensor<fp16, [10020]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397764800))))[name = string("layers_30_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5491_cast_fp16 = conv(dilations = var_5491_dilations_0, groups = var_5491_groups_0, pad = var_5491_pad_0, pad_type = var_5491_pad_type_0, strides = var_5491_strides_0, weight = layers_30_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_121_cast_fp16)[name = string("op_5491_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_61_cast_fp16 = add(x = var_5485_cast_fp16, y = var_5491_cast_fp16)[name = string("query_61_cast_fp16")];
            string var_5500_pad_type_0 = const()[name = string("op_5500_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5500_strides_0 = const()[name = string("op_5500_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5500_pad_0 = const()[name = string("op_5500_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5500_dilations_0 = const()[name = string("op_5500_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5500_groups_0 = const()[name = string("op_5500_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397989824))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398809088))))[name = string("layers_30_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5500_cast_fp16 = conv(dilations = var_5500_dilations_0, groups = var_5500_groups_0, pad = var_5500_pad_0, pad_type = var_5500_pad_type_0, strides = var_5500_strides_0, weight = layers_30_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = string("op_5500_cast_fp16")];
            string var_5506_pad_type_0 = const()[name = string("op_5506_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5506_strides_0 = const()[name = string("op_5506_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5506_pad_0 = const()[name = string("op_5506_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5506_dilations_0 = const()[name = string("op_5506_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5506_groups_0 = const()[name = string("op_5506_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398830400))), nonzero_data = tensor<fp16, [10558]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398809216))))[name = string("layers_30_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5506_cast_fp16 = conv(dilations = var_5506_dilations_0, groups = var_5506_groups_0, pad = var_5506_pad_0, pad_type = var_5506_pad_type_0, strides = var_5506_strides_0, weight = layers_30_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_121_cast_fp16)[name = string("op_5506_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_61_cast_fp16 = add(x = var_5500_cast_fp16, y = var_5506_cast_fp16)[name = string("key_61_cast_fp16")];
            string var_5516_pad_type_0 = const()[name = string("op_5516_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5516_strides_0 = const()[name = string("op_5516_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5516_pad_0 = const()[name = string("op_5516_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5516_dilations_0 = const()[name = string("op_5516_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5516_groups_0 = const()[name = string("op_5516_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399035264))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399854528))))[name = string("layers_30_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_30_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_30_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399854656)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5516_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5516_dilations_0, groups = var_5516_groups_0, pad = var_5516_pad_0, pad_type = var_5516_pad_type_0, strides = var_5516_strides_0, weight = layers_30_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = string("op_5516_cast_fp16")];
            string var_5522_pad_type_0 = const()[name = string("op_5522_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5522_strides_0 = const()[name = string("op_5522_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5522_pad_0 = const()[name = string("op_5522_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5522_dilations_0 = const()[name = string("op_5522_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5522_groups_0 = const()[name = string("op_5522_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399869696))), nonzero_data = tensor<fp16, [6176]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399857280))))[name = string("layers_30_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5522_cast_fp16 = conv(dilations = var_5522_dilations_0, groups = var_5522_groups_0, pad = var_5522_pad_0, pad_type = var_5522_pad_type_0, strides = var_5522_strides_0, weight = layers_30_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_121_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_61_cast_fp16 = add(x = var_5516_cast_fp16, y = var_5522_cast_fp16)[name = string("value_61_cast_fp16")];
            tensor<int32, [4]> var_5525 = const()[name = string("op_5525"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_61_cast_fp16 = reshape(shape = var_5525, x = query_61_cast_fp16)[name = string("mh_q_61_cast_fp16")];
            fp16 var_5527_to_fp16 = const()[name = string("op_5527_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_5528_cast_fp16 = mul(x = mh_q_61_cast_fp16, y = var_5527_to_fp16)[name = string("op_5528_cast_fp16")];
            tensor<int32, [4]> var_5529 = const()[name = string("op_5529"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5530_cast_fp16 = reshape(shape = var_5529, x = key_61_cast_fp16)[name = string("op_5530_cast_fp16")];
            bool mh_w_61_transpose_x_0 = const()[name = string("mh_w_61_transpose_x_0"), val = bool(true)];
            bool mh_w_61_transpose_y_0 = const()[name = string("mh_w_61_transpose_y_0"), val = 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_5528_cast_fp16, y = var_5530_cast_fp16)[name = string("mh_w_61_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_5533_cast_fp16 = softmax(axis = var_5444, x = mh_w_61_cast_fp16)[name = string("op_5533_cast_fp16")];
            tensor<int32, [4]> var_5534 = const()[name = string("op_5534"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5535_cast_fp16 = reshape(shape = var_5534, x = value_61_cast_fp16)[name = string("op_5535_cast_fp16")];
            bool attn_61_transpose_x_0 = const()[name = string("attn_61_transpose_x_0"), val = bool(false)];
            bool attn_61_transpose_y_0 = const()[name = string("attn_61_transpose_y_0"), val = 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_5535_cast_fp16, y = var_5533_cast_fp16)[name = string("attn_61_cast_fp16")];
            tensor<int32, [4]> var_5538 = const()[name = string("op_5538"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_241_cast_fp16 = reshape(shape = var_5538, x = attn_61_cast_fp16)[name = string("input_241_cast_fp16")];
            string var_5548_pad_type_0 = const()[name = string("op_5548_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5548_strides_0 = const()[name = string("op_5548_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5548_pad_0 = const()[name = string("op_5548_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5548_dilations_0 = const()[name = string("op_5548_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5548_groups_0 = const()[name = string("op_5548_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400074560))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400893824))))[name = string("layers_30_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_30_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_30_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400893952)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5548_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_5548_dilations_0, groups = var_5548_groups_0, pad = var_5548_pad_0, pad_type = var_5548_pad_type_0, strides = var_5548_strides_0, weight = layers_30_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_241_cast_fp16)[name = string("op_5548_cast_fp16")];
            string var_5554_pad_type_0 = const()[name = string("op_5554_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5554_strides_0 = const()[name = string("op_5554_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5554_pad_0 = const()[name = string("op_5554_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5554_dilations_0 = const()[name = string("op_5554_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5554_groups_0 = const()[name = string("op_5554_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400908032))), nonzero_data = tensor<fp16, [5673]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400896576))))[name = string("layers_30_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5554_cast_fp16 = conv(dilations = var_5554_dilations_0, groups = var_5554_groups_0, pad = var_5554_pad_0, pad_type = var_5554_pad_type_0, strides = var_5554_strides_0, weight = layers_30_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_241_cast_fp16)[name = string("op_5554_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_123_cast_fp16 = add(x = var_5548_cast_fp16, y = var_5554_cast_fp16)[name = 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 = string("inputs_123_cast_fp16")];
            tensor<int32, [1]> out_123_axes_0 = const()[name = string("out_123_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5565_to_fp16 = const()[name = string("op_5565_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_5565_to_fp16, x = inputs_123_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [1280]> input_243_gamma_0_to_fp16 = const()[name = string("input_243_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401112896)))];
            tensor<fp16, [1280]> input_243_beta_0_to_fp16 = const()[name = string("input_243_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401115520)))];
            fp16 input_243_epsilon_0_to_fp16 = const()[name = string("input_243_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = string("input_243_cast_fp16")];
            string var_5583_pad_type_0 = const()[name = string("op_5583_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5583_strides_0 = const()[name = string("op_5583_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5583_pad_0 = const()[name = string("op_5583_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5583_dilations_0 = const()[name = string("op_5583_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5583_groups_0 = const()[name = string("op_5583_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_30_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401118144))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404395008))))[name = string("layers_30_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_30_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_30_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404395136)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_5583_cast_fp16 = conv(bias = layers_30_fc1_inlier_module_bias_to_fp16, dilations = var_5583_dilations_0, groups = var_5583_groups_0, pad = var_5583_pad_0, pad_type = var_5583_pad_type_0, strides = var_5583_strides_0, weight = layers_30_fc1_inlier_module_weight_to_fp16_palettized, x = input_243_cast_fp16)[name = string("op_5583_cast_fp16")];
            string var_5589_pad_type_0 = const()[name = string("op_5589_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5589_strides_0 = const()[name = string("op_5589_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5589_pad_0 = const()[name = string("op_5589_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5589_dilations_0 = const()[name = string("op_5589_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5589_groups_0 = const()[name = string("op_5589_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_30_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404464960))), nonzero_data = tensor<fp16, [29723]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404405440))))[name = string("layers_30_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_5589_cast_fp16 = conv(dilations = var_5589_dilations_0, groups = var_5589_groups_0, pad = var_5589_pad_0, pad_type = var_5589_pad_type_0, strides = var_5589_strides_0, weight = layers_30_fc1_outlier_module_weight_to_fp16_sparsified, x = input_243_cast_fp16)[name = string("op_5589_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_245_cast_fp16 = add(x = var_5583_cast_fp16, y = var_5589_cast_fp16)[name = string("input_245_cast_fp16")];
            string input_247_mode_0 = const()[name = string("input_247_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_247_cast_fp16 = gelu(mode = input_247_mode_0, x = input_245_cast_fp16)[name = string("input_247_cast_fp16")];
            string var_5600_pad_type_0 = const()[name = string("op_5600_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5600_strides_0 = const()[name = string("op_5600_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5600_pad_0 = const()[name = string("op_5600_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5600_dilations_0 = const()[name = string("op_5600_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5600_groups_0 = const()[name = string("op_5600_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_30_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405284224))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408561088))))[name = string("layers_30_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_30_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_30_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408561216)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5600_cast_fp16 = conv(bias = layers_30_fc2_inlier_module_bias_to_fp16, dilations = var_5600_dilations_0, groups = var_5600_groups_0, pad = var_5600_pad_0, pad_type = var_5600_pad_type_0, strides = var_5600_strides_0, weight = layers_30_fc2_inlier_module_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("op_5600_cast_fp16")];
            string var_5606_pad_type_0 = const()[name = string("op_5606_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5606_strides_0 = const()[name = string("op_5606_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5606_pad_0 = const()[name = string("op_5606_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5606_dilations_0 = const()[name = string("op_5606_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5606_groups_0 = const()[name = string("op_5606_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_30_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408624832))), nonzero_data = tensor<fp16, [30449]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408563840))))[name = string("layers_30_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5606_cast_fp16 = conv(dilations = var_5606_dilations_0, groups = var_5606_groups_0, pad = var_5606_pad_0, pad_type = var_5606_pad_type_0, strides = var_5606_strides_0, weight = layers_30_fc2_outlier_module_weight_to_fp16_sparsified, x = input_247_cast_fp16)[name = string("op_5606_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_65_cast_fp16 = add(x = var_5600_cast_fp16, y = var_5606_cast_fp16)[name = 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 = string("inputs_125_cast_fp16")];
            int32 var_5616 = const()[name = string("op_5616"), val = int32(3)];
            tensor<int32, [1]> out_125_axes_0 = const()[name = string("out_125_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5635_to_fp16 = const()[name = string("op_5635_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_5635_to_fp16, x = inputs_125_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [1280]> obj_125_gamma_0_to_fp16 = const()[name = string("obj_125_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409444096)))];
            tensor<fp16, [1280]> obj_125_beta_0_to_fp16 = const()[name = string("obj_125_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409446720)))];
            fp16 obj_125_epsilon_0_to_fp16 = const()[name = string("obj_125_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = string("obj_125_cast_fp16")];
            string var_5657_pad_type_0 = const()[name = string("op_5657_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5657_strides_0 = const()[name = string("op_5657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5657_pad_0 = const()[name = string("op_5657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5657_dilations_0 = const()[name = string("op_5657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5657_groups_0 = const()[name = string("op_5657_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409449344))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410268608))))[name = string("layers_31_self_attn_q_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_31_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_31_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410268736)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5657_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_5657_dilations_0, groups = var_5657_groups_0, pad = var_5657_pad_0, pad_type = var_5657_pad_type_0, strides = var_5657_strides_0, weight = layers_31_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = string("op_5657_cast_fp16")];
            string var_5663_pad_type_0 = const()[name = string("op_5663_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5663_strides_0 = const()[name = string("op_5663_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5663_pad_0 = const()[name = string("op_5663_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5663_dilations_0 = const()[name = string("op_5663_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5663_groups_0 = const()[name = string("op_5663_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410290240))), nonzero_data = tensor<fp16, [9377]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410271360))))[name = string("layers_31_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5663_cast_fp16 = conv(dilations = var_5663_dilations_0, groups = var_5663_groups_0, pad = var_5663_pad_0, pad_type = var_5663_pad_type_0, strides = var_5663_strides_0, weight = layers_31_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_125_cast_fp16)[name = string("op_5663_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> query_cast_fp16 = add(x = var_5657_cast_fp16, y = var_5663_cast_fp16)[name = string("query_cast_fp16")];
            string var_5672_pad_type_0 = const()[name = string("op_5672_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5672_strides_0 = const()[name = string("op_5672_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5672_pad_0 = const()[name = string("op_5672_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5672_dilations_0 = const()[name = string("op_5672_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5672_groups_0 = const()[name = string("op_5672_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410495104))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411314368))))[name = string("layers_31_self_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5672_cast_fp16 = conv(dilations = var_5672_dilations_0, groups = var_5672_groups_0, pad = var_5672_pad_0, pad_type = var_5672_pad_type_0, strides = var_5672_strides_0, weight = layers_31_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = string("op_5672_cast_fp16")];
            string var_5678_pad_type_0 = const()[name = string("op_5678_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5678_strides_0 = const()[name = string("op_5678_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5678_pad_0 = const()[name = string("op_5678_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5678_dilations_0 = const()[name = string("op_5678_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5678_groups_0 = const()[name = string("op_5678_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411336768))), nonzero_data = tensor<fp16, [11093]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411314496))))[name = string("layers_31_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5678_cast_fp16 = conv(dilations = var_5678_dilations_0, groups = var_5678_groups_0, pad = var_5678_pad_0, pad_type = var_5678_pad_type_0, strides = var_5678_strides_0, weight = layers_31_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_125_cast_fp16)[name = string("op_5678_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = add(x = var_5672_cast_fp16, y = var_5678_cast_fp16)[name = string("key_cast_fp16")];
            string var_5688_pad_type_0 = const()[name = string("op_5688_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5688_strides_0 = const()[name = string("op_5688_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5688_pad_0 = const()[name = string("op_5688_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5688_dilations_0 = const()[name = string("op_5688_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5688_groups_0 = const()[name = string("op_5688_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411541632))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412360896))))[name = string("layers_31_self_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_31_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_31_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412361024)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5688_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5688_dilations_0, groups = var_5688_groups_0, pad = var_5688_pad_0, pad_type = var_5688_pad_type_0, strides = var_5688_strides_0, weight = layers_31_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = string("op_5688_cast_fp16")];
            string var_5694_pad_type_0 = const()[name = string("op_5694_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5694_strides_0 = const()[name = string("op_5694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5694_pad_0 = const()[name = string("op_5694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5694_dilations_0 = const()[name = string("op_5694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5694_groups_0 = const()[name = string("op_5694_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412376512))), nonzero_data = tensor<fp16, [6384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412363648))))[name = string("layers_31_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5694_cast_fp16 = conv(dilations = var_5694_dilations_0, groups = var_5694_groups_0, pad = var_5694_pad_0, pad_type = var_5694_pad_type_0, strides = var_5694_strides_0, weight = layers_31_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_125_cast_fp16)[name = string("op_5694_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = add(x = var_5688_cast_fp16, y = var_5694_cast_fp16)[name = string("value_cast_fp16")];
            tensor<int32, [4]> var_5697 = const()[name = string("op_5697"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> mh_q_cast_fp16 = reshape(shape = var_5697, x = query_cast_fp16)[name = string("mh_q_cast_fp16")];
            fp16 var_5699_to_fp16 = const()[name = string("op_5699_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1500]> var_5700_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_5699_to_fp16)[name = string("op_5700_cast_fp16")];
            tensor<int32, [4]> var_5701 = const()[name = string("op_5701"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5702_cast_fp16 = reshape(shape = var_5701, x = key_cast_fp16)[name = string("op_5702_cast_fp16")];
            bool mh_w_transpose_x_0 = const()[name = string("mh_w_transpose_x_0"), val = bool(true)];
            bool mh_w_transpose_y_0 = const()[name = string("mh_w_transpose_y_0"), val = 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_5700_cast_fp16, y = var_5702_cast_fp16)[name = string("mh_w_cast_fp16")];
            tensor<fp16, [1, 20, 1500, 1500]> var_5705_cast_fp16 = softmax(axis = var_5616, x = mh_w_cast_fp16)[name = string("op_5705_cast_fp16")];
            tensor<int32, [4]> var_5706 = const()[name = string("op_5706"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5707_cast_fp16 = reshape(shape = var_5706, x = value_cast_fp16)[name = string("op_5707_cast_fp16")];
            bool attn_transpose_x_0 = const()[name = string("attn_transpose_x_0"), val = bool(false)];
            bool attn_transpose_y_0 = const()[name = string("attn_transpose_y_0"), val = 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_5707_cast_fp16, y = var_5705_cast_fp16)[name = string("attn_cast_fp16")];
            tensor<int32, [4]> var_5710 = const()[name = string("op_5710"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1500]> input_249_cast_fp16 = reshape(shape = var_5710, x = attn_cast_fp16)[name = string("input_249_cast_fp16")];
            string var_5720_pad_type_0 = const()[name = string("op_5720_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5720_strides_0 = const()[name = string("op_5720_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5720_pad_0 = const()[name = string("op_5720_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5720_dilations_0 = const()[name = string("op_5720_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5720_groups_0 = const()[name = string("op_5720_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412581376))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413400640))))[name = string("layers_31_self_attn_o_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_31_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = string("layers_31_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413400768)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5720_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_5720_dilations_0, groups = var_5720_groups_0, pad = var_5720_pad_0, pad_type = var_5720_pad_type_0, strides = var_5720_strides_0, weight = layers_31_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_249_cast_fp16)[name = string("op_5720_cast_fp16")];
            string var_5726_pad_type_0 = const()[name = string("op_5726_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5726_strides_0 = const()[name = string("op_5726_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5726_pad_0 = const()[name = string("op_5726_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5726_dilations_0 = const()[name = string("op_5726_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5726_groups_0 = const()[name = string("op_5726_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413416000))), nonzero_data = tensor<fp16, [6256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413403392))))[name = string("layers_31_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5726_cast_fp16 = conv(dilations = var_5726_dilations_0, groups = var_5726_groups_0, pad = var_5726_pad_0, pad_type = var_5726_pad_type_0, strides = var_5726_strides_0, weight = layers_31_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_249_cast_fp16)[name = string("op_5726_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> obj_cast_fp16 = add(x = var_5720_cast_fp16, y = var_5726_cast_fp16)[name = 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 = string("inputs_127_cast_fp16")];
            tensor<int32, [1]> out_127_axes_0 = const()[name = string("out_127_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5737_to_fp16 = const()[name = string("op_5737_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_5737_to_fp16, x = inputs_127_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [1280]> input_251_gamma_0_to_fp16 = const()[name = string("input_251_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413620864)))];
            tensor<fp16, [1280]> input_251_beta_0_to_fp16 = const()[name = string("input_251_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413623488)))];
            fp16 input_251_epsilon_0_to_fp16 = const()[name = string("input_251_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = string("input_251_cast_fp16")];
            string var_5755_pad_type_0 = const()[name = string("op_5755_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5755_strides_0 = const()[name = string("op_5755_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5755_pad_0 = const()[name = string("op_5755_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5755_dilations_0 = const()[name = string("op_5755_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5755_groups_0 = const()[name = string("op_5755_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_31_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413626112))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416902976))))[name = string("layers_31_fc1_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [5120]> layers_31_fc1_inlier_module_bias_to_fp16 = const()[name = string("layers_31_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416903104)))];
            tensor<fp16, [1, 5120, 1, 1500]> var_5755_cast_fp16 = conv(bias = layers_31_fc1_inlier_module_bias_to_fp16, dilations = var_5755_dilations_0, groups = var_5755_groups_0, pad = var_5755_pad_0, pad_type = var_5755_pad_type_0, strides = var_5755_strides_0, weight = layers_31_fc1_inlier_module_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("op_5755_cast_fp16")];
            string var_5761_pad_type_0 = const()[name = string("op_5761_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5761_strides_0 = const()[name = string("op_5761_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5761_pad_0 = const()[name = string("op_5761_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5761_dilations_0 = const()[name = string("op_5761_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5761_groups_0 = const()[name = string("op_5761_groups_0"), val = int32(1)];
            tensor<fp16, [5120, 1280, 1, 1]> layers_31_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416975104))), nonzero_data = tensor<fp16, [30796]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416913408))))[name = string("layers_31_fc1_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 5120, 1, 1500]> var_5761_cast_fp16 = conv(dilations = var_5761_dilations_0, groups = var_5761_groups_0, pad = var_5761_pad_0, pad_type = var_5761_pad_type_0, strides = var_5761_strides_0, weight = layers_31_fc1_outlier_module_weight_to_fp16_sparsified, x = input_251_cast_fp16)[name = string("op_5761_cast_fp16")];
            tensor<fp16, [1, 5120, 1, 1500]> input_253_cast_fp16 = add(x = var_5755_cast_fp16, y = var_5761_cast_fp16)[name = string("input_253_cast_fp16")];
            string input_255_mode_0 = const()[name = string("input_255_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 5120, 1, 1500]> input_255_cast_fp16 = gelu(mode = input_255_mode_0, x = input_253_cast_fp16)[name = string("input_255_cast_fp16")];
            string var_5772_pad_type_0 = const()[name = string("op_5772_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5772_strides_0 = const()[name = string("op_5772_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5772_pad_0 = const()[name = string("op_5772_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5772_dilations_0 = const()[name = string("op_5772_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5772_groups_0 = const()[name = string("op_5772_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_31_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417794368))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421071232))))[name = string("layers_31_fc2_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> layers_31_fc2_inlier_module_bias_to_fp16 = const()[name = string("layers_31_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421071360)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5772_cast_fp16 = conv(bias = layers_31_fc2_inlier_module_bias_to_fp16, dilations = var_5772_dilations_0, groups = var_5772_groups_0, pad = var_5772_pad_0, pad_type = var_5772_pad_type_0, strides = var_5772_strides_0, weight = layers_31_fc2_inlier_module_weight_to_fp16_palettized, x = input_255_cast_fp16)[name = string("op_5772_cast_fp16")];
            string var_5778_pad_type_0 = const()[name = string("op_5778_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5778_strides_0 = const()[name = string("op_5778_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5778_pad_0 = const()[name = string("op_5778_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5778_dilations_0 = const()[name = string("op_5778_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5778_groups_0 = const()[name = string("op_5778_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 5120, 1, 1]> layers_31_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421144256))), nonzero_data = tensor<fp16, [35103]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421073984))))[name = string("layers_31_fc2_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5778_cast_fp16 = conv(dilations = var_5778_dilations_0, groups = var_5778_groups_0, pad = var_5778_pad_0, pad_type = var_5778_pad_type_0, strides = var_5778_strides_0, weight = layers_31_fc2_outlier_module_weight_to_fp16_sparsified, x = input_255_cast_fp16)[name = string("op_5778_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_cast_fp16 = add(x = var_5772_cast_fp16, y = var_5778_cast_fp16)[name = 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 = string("inputs_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_5793_to_fp16 = const()[name = string("op_5793_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1500]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5793_to_fp16, x = inputs_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [1280]> encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = string("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421963520)))];
            tensor<fp16, [1280]> encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = string("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421966144)))];
            fp16 encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = string("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = 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 = var_105_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = string("encoder_output_embeds_type_fp32_cast_fp16")];
            string var_5819_pad_type_0 = const()[name = string("op_5819_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5819_strides_0 = const()[name = string("op_5819_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5819_pad_0 = const()[name = string("op_5819_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5819_dilations_0 = const()[name = string("op_5819_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5819_groups_0 = const()[name = string("op_5819_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421968768))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422788032))))[name = string("decoder_kv_cache_prep_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5819_cast_fp16 = conv(dilations = var_5819_dilations_0, groups = var_5819_groups_0, pad = var_5819_pad_0, pad_type = var_5819_pad_type_0, strides = var_5819_strides_0, weight = decoder_kv_cache_prep_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = string("op_5819_cast_fp16")];
            string var_5825_pad_type_0 = const()[name = string("op_5825_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5825_strides_0 = const()[name = string("op_5825_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5825_pad_0 = const()[name = string("op_5825_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5825_dilations_0 = const()[name = string("op_5825_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5825_groups_0 = const()[name = string("op_5825_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422848128))), nonzero_data = tensor<fp16, [29949]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422788160))))[name = string("decoder_kv_cache_prep_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5825_cast_fp16 = conv(dilations = var_5825_dilations_0, groups = var_5825_groups_0, pad = var_5825_pad_0, pad_type = var_5825_pad_type_0, strides = var_5825_strides_0, weight = decoder_kv_cache_prep_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = string("op_5825_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5826_cast_fp16 = add(x = var_5819_cast_fp16, y = var_5825_cast_fp16)[name = string("op_5826_cast_fp16")];
            string var_5835_pad_type_0 = const()[name = string("op_5835_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5835_strides_0 = const()[name = string("op_5835_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5835_pad_0 = const()[name = string("op_5835_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5835_dilations_0 = const()[name = string("op_5835_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5835_groups_0 = const()[name = string("op_5835_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423052992))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423872256))))[name = string("decoder_kv_cache_prep_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> decoder_kv_cache_prep_0_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("decoder_kv_cache_prep_0_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423872384)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5835_cast_fp16 = conv(bias = decoder_kv_cache_prep_0_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5835_dilations_0, groups = var_5835_groups_0, pad = var_5835_pad_0, pad_type = var_5835_pad_type_0, strides = var_5835_strides_0, weight = decoder_kv_cache_prep_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = string("op_5835_cast_fp16")];
            string var_5841_pad_type_0 = const()[name = string("op_5841_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5841_strides_0 = const()[name = string("op_5841_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5841_pad_0 = const()[name = string("op_5841_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5841_dilations_0 = const()[name = string("op_5841_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5841_groups_0 = const()[name = string("op_5841_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423886272))), nonzero_data = tensor<fp16, [5596]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423875008))))[name = string("decoder_kv_cache_prep_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5841_cast_fp16 = conv(dilations = var_5841_dilations_0, groups = var_5841_groups_0, pad = var_5841_pad_0, pad_type = var_5841_pad_type_0, strides = var_5841_strides_0, weight = decoder_kv_cache_prep_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = string("op_5841_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5842_cast_fp16 = add(x = var_5835_cast_fp16, y = var_5841_cast_fp16)[name = string("op_5842_cast_fp16")];
            string var_5862_pad_type_0 = const()[name = string("op_5862_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5862_strides_0 = const()[name = string("op_5862_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5862_pad_0 = const()[name = string("op_5862_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5862_dilations_0 = const()[name = string("op_5862_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5862_groups_0 = const()[name = string("op_5862_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424091136))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424910400))))[name = string("decoder_kv_cache_prep_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5862_cast_fp16 = conv(dilations = var_5862_dilations_0, groups = var_5862_groups_0, pad = var_5862_pad_0, pad_type = var_5862_pad_type_0, strides = var_5862_strides_0, weight = decoder_kv_cache_prep_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = string("op_5862_cast_fp16")];
            string var_5868_pad_type_0 = const()[name = string("op_5868_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5868_strides_0 = const()[name = string("op_5868_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5868_pad_0 = const()[name = string("op_5868_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5868_dilations_0 = const()[name = string("op_5868_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5868_groups_0 = const()[name = string("op_5868_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424953984))), nonzero_data = tensor<fp16, [21667]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424910528))))[name = string("decoder_kv_cache_prep_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5868_cast_fp16 = conv(dilations = var_5868_dilations_0, groups = var_5868_groups_0, pad = var_5868_pad_0, pad_type = var_5868_pad_type_0, strides = var_5868_strides_0, weight = decoder_kv_cache_prep_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = string("op_5868_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5869_cast_fp16 = add(x = var_5862_cast_fp16, y = var_5868_cast_fp16)[name = string("op_5869_cast_fp16")];
            string var_5878_pad_type_0 = const()[name = string("op_5878_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5878_strides_0 = const()[name = string("op_5878_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5878_pad_0 = const()[name = string("op_5878_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5878_dilations_0 = const()[name = string("op_5878_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5878_groups_0 = const()[name = string("op_5878_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425158848))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425978112))))[name = string("decoder_kv_cache_prep_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> decoder_kv_cache_prep_1_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("decoder_kv_cache_prep_1_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425978240)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5878_cast_fp16 = conv(bias = decoder_kv_cache_prep_1_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5878_dilations_0, groups = var_5878_groups_0, pad = var_5878_pad_0, pad_type = var_5878_pad_type_0, strides = var_5878_strides_0, weight = decoder_kv_cache_prep_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = string("op_5878_cast_fp16")];
            string var_5884_pad_type_0 = const()[name = string("op_5884_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5884_strides_0 = const()[name = string("op_5884_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5884_pad_0 = const()[name = string("op_5884_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5884_dilations_0 = const()[name = string("op_5884_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5884_groups_0 = const()[name = string("op_5884_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425992064))), nonzero_data = tensor<fp16, [5557]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425980864))))[name = string("decoder_kv_cache_prep_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5884_cast_fp16 = conv(dilations = var_5884_dilations_0, groups = var_5884_groups_0, pad = var_5884_pad_0, pad_type = var_5884_pad_type_0, strides = var_5884_strides_0, weight = decoder_kv_cache_prep_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = string("op_5884_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5885_cast_fp16 = add(x = var_5878_cast_fp16, y = var_5884_cast_fp16)[name = string("op_5885_cast_fp16")];
            string var_5905_pad_type_0 = const()[name = string("op_5905_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5905_strides_0 = const()[name = string("op_5905_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5905_pad_0 = const()[name = string("op_5905_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5905_dilations_0 = const()[name = string("op_5905_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5905_groups_0 = const()[name = string("op_5905_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426196928))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427016192))))[name = string("decoder_kv_cache_prep_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5905_cast_fp16 = conv(dilations = var_5905_dilations_0, groups = var_5905_groups_0, pad = var_5905_pad_0, pad_type = var_5905_pad_type_0, strides = var_5905_strides_0, weight = decoder_kv_cache_prep_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = string("op_5905_cast_fp16")];
            string var_5911_pad_type_0 = const()[name = string("op_5911_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5911_strides_0 = const()[name = string("op_5911_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5911_pad_0 = const()[name = string("op_5911_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5911_dilations_0 = const()[name = string("op_5911_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5911_groups_0 = const()[name = string("op_5911_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427044160))), nonzero_data = tensor<fp16, [13879]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427016320))))[name = string("decoder_kv_cache_prep_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5911_cast_fp16 = conv(dilations = var_5911_dilations_0, groups = var_5911_groups_0, pad = var_5911_pad_0, pad_type = var_5911_pad_type_0, strides = var_5911_strides_0, weight = decoder_kv_cache_prep_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = string("op_5911_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5912_cast_fp16 = add(x = var_5905_cast_fp16, y = var_5911_cast_fp16)[name = string("op_5912_cast_fp16")];
            string var_5921_pad_type_0 = const()[name = string("op_5921_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5921_strides_0 = const()[name = string("op_5921_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5921_pad_0 = const()[name = string("op_5921_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5921_dilations_0 = const()[name = string("op_5921_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5921_groups_0 = const()[name = string("op_5921_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427249024))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428068288))))[name = string("decoder_kv_cache_prep_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> decoder_kv_cache_prep_2_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("decoder_kv_cache_prep_2_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428068416)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5921_cast_fp16 = conv(bias = decoder_kv_cache_prep_2_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5921_dilations_0, groups = var_5921_groups_0, pad = var_5921_pad_0, pad_type = var_5921_pad_type_0, strides = var_5921_strides_0, weight = decoder_kv_cache_prep_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = string("op_5921_cast_fp16")];
            string var_5927_pad_type_0 = const()[name = string("op_5927_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5927_strides_0 = const()[name = string("op_5927_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5927_pad_0 = const()[name = string("op_5927_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5927_dilations_0 = const()[name = string("op_5927_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5927_groups_0 = const()[name = string("op_5927_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428082624))), nonzero_data = tensor<fp16, [5756]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428071040))))[name = string("decoder_kv_cache_prep_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5927_cast_fp16 = conv(dilations = var_5927_dilations_0, groups = var_5927_groups_0, pad = var_5927_pad_0, pad_type = var_5927_pad_type_0, strides = var_5927_strides_0, weight = decoder_kv_cache_prep_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = string("op_5927_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5928_cast_fp16 = add(x = var_5921_cast_fp16, y = var_5927_cast_fp16)[name = string("op_5928_cast_fp16")];
            string var_5948_pad_type_0 = const()[name = string("op_5948_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5948_strides_0 = const()[name = string("op_5948_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5948_pad_0 = const()[name = string("op_5948_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5948_dilations_0 = const()[name = string("op_5948_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5948_groups_0 = const()[name = string("op_5948_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428287488))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429106752))))[name = string("decoder_kv_cache_prep_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5948_cast_fp16 = conv(dilations = var_5948_dilations_0, groups = var_5948_groups_0, pad = var_5948_pad_0, pad_type = var_5948_pad_type_0, strides = var_5948_strides_0, weight = decoder_kv_cache_prep_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = string("op_5948_cast_fp16")];
            string var_5954_pad_type_0 = const()[name = string("op_5954_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5954_strides_0 = const()[name = string("op_5954_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5954_pad_0 = const()[name = string("op_5954_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5954_dilations_0 = const()[name = string("op_5954_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5954_groups_0 = const()[name = string("op_5954_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429129600))), nonzero_data = tensor<fp16, [11308]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429106880))))[name = string("decoder_kv_cache_prep_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5954_cast_fp16 = conv(dilations = var_5954_dilations_0, groups = var_5954_groups_0, pad = var_5954_pad_0, pad_type = var_5954_pad_type_0, strides = var_5954_strides_0, weight = decoder_kv_cache_prep_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = string("op_5954_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> k_cast_fp16 = add(x = var_5948_cast_fp16, y = var_5954_cast_fp16)[name = string("k_cast_fp16")];
            string var_5964_pad_type_0 = const()[name = string("op_5964_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5964_strides_0 = const()[name = string("op_5964_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5964_pad_0 = const()[name = string("op_5964_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5964_dilations_0 = const()[name = string("op_5964_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5964_groups_0 = const()[name = string("op_5964_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429334464))), lut = tensor<fp16, [1, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430153728))))[name = string("decoder_kv_cache_prep_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized")];
            tensor<fp16, [1280]> decoder_kv_cache_prep_3_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = string("decoder_kv_cache_prep_3_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430153856)))];
            tensor<fp16, [1, 1280, 1, 1500]> var_5964_cast_fp16 = conv(bias = decoder_kv_cache_prep_3_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_5964_dilations_0, groups = var_5964_groups_0, pad = var_5964_pad_0, pad_type = var_5964_pad_type_0, strides = var_5964_strides_0, weight = decoder_kv_cache_prep_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = string("op_5964_cast_fp16")];
            string var_5970_pad_type_0 = const()[name = string("op_5970_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5970_strides_0 = const()[name = string("op_5970_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5970_pad_0 = const()[name = string("op_5970_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5970_dilations_0 = const()[name = string("op_5970_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5970_groups_0 = const()[name = string("op_5970_groups_0"), val = int32(1)];
            tensor<fp16, [1280, 1280, 1, 1]> decoder_kv_cache_prep_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430170304))), nonzero_data = tensor<fp16, [6870]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430156480))))[name = string("decoder_kv_cache_prep_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1280, 1, 1500]> var_5970_cast_fp16 = conv(dilations = var_5970_dilations_0, groups = var_5970_groups_0, pad = var_5970_pad_0, pad_type = var_5970_pad_type_0, strides = var_5970_strides_0, weight = decoder_kv_cache_prep_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = string("op_5970_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1500]> v_cast_fp16 = add(x = var_5964_cast_fp16, y = var_5970_cast_fp16)[name = string("v_cast_fp16")];
            int32 var_5976 = const()[name = string("op_5976"), val = int32(0)];
            bool input_259_interleave_0 = const()[name = string("input_259_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 1280, 1, 1500]> input_259_cast_fp16 = concat(axis = var_5976, interleave = input_259_interleave_0, values = (var_5826_cast_fp16, var_5869_cast_fp16, var_5912_cast_fp16, k_cast_fp16))[name = string("input_259_cast_fp16")];
            int32 var_5979 = const()[name = string("op_5979"), val = int32(0)];
            bool input_interleave_0 = const()[name = string("input_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 1280, 1, 1500]> input_cast_fp16 = concat(axis = var_5979, interleave = input_interleave_0, values = (var_5842_cast_fp16, var_5885_cast_fp16, var_5928_cast_fp16, v_cast_fp16))[name = string("input_cast_fp16")];
            tensor<int32, [8]> var_5986_pad_0 = const()[name = string("op_5986_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 36])];
            string var_5986_mode_0 = const()[name = string("op_5986_mode_0"), val = string("constant")];
            fp16 const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [4, 1280, 1, 1536]> encoder_attn_key_cache = pad(constant_val = const_33_to_fp16, mode = var_5986_mode_0, pad = var_5986_pad_0, x = input_259_cast_fp16)[name = string("op_5986_cast_fp16")];
            tensor<int32, [8]> var_5992_pad_0 = const()[name = string("op_5992_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 36])];
            string var_5992_mode_0 = const()[name = string("op_5992_mode_0"), val = string("constant")];
            fp16 const_34_to_fp16 = const()[name = string("const_34_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [4, 1280, 1, 1536]> encoder_attn_value_cache = pad(constant_val = const_34_to_fp16, mode = var_5992_mode_0, pad = var_5992_pad_0, x = input_cast_fp16)[name = string("op_5992_cast_fp16")];
        } -> (encoder_output_embeds, encoder_attn_key_cache, encoder_attn_value_cache);
}