Blame for openai_whisper-small_216MB/AudioEncoder.mlmodelc/model.mil

program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3404.16.1"}, {"coremlc-version", "3404.23.1"}})]
{
    func main<ios16>(tensor<fp16, [1, 80, 1, 3000]> melspectrogram_features) {
            tensor<string, []> var_76_pad_type_0 = const()[name = tensor<string, []>("op_76_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_76_pad_0 = const()[name = tensor<string, []>("op_76_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> var_76_strides_0 = const()[name = tensor<string, []>("op_76_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_76_dilations_0 = const()[name = tensor<string, []>("op_76_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_76_groups_0 = const()[name = tensor<string, []>("op_76_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 80, 1, 3]> var_45_to_fp16 = const()[name = tensor<string, []>("op_45_to_fp16"), val = tensor<fp16, [768, 80, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [768]> var_57_to_fp16 = const()[name = tensor<string, []>("op_57_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368768)))];
            tensor<fp16, [1, 768, 1, 3000]> var_76_cast_fp16 = conv(bias = var_57_to_fp16, dilations = var_76_dilations_0, groups = var_76_groups_0, pad = var_76_pad_0, pad_type = var_76_pad_type_0, strides = var_76_strides_0, weight = var_45_to_fp16, x = melspectrogram_features)[name = tensor<string, []>("op_76_cast_fp16")];
            tensor<string, []> var_114_pad_type_0 = const()[name = tensor<string, []>("op_114_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_114_pad_0 = const()[name = tensor<string, []>("op_114_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> var_114_strides_0 = const()[name = tensor<string, []>("op_114_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_114_dilations_0 = const()[name = tensor<string, []>("op_114_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_114_groups_0 = const()[name = tensor<string, []>("op_114_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 80, 1, 3]> op_89_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [92160]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370368))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462592))), name = tensor<string, []>("op_89_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 80, 1, 3])];
            tensor<fp16, [768]> var_95_to_fp16 = const()[name = tensor<string, []>("op_95_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462720)))];
            tensor<fp16, [1, 768, 1, 3000]> var_114_cast_fp16 = conv(bias = var_95_to_fp16, dilations = var_114_dilations_0, groups = var_114_groups_0, pad = var_114_pad_0, pad_type = var_114_pad_type_0, strides = var_114_strides_0, weight = op_89_to_fp16_palettized, x = melspectrogram_features)[name = tensor<string, []>("op_114_cast_fp16")];
            tensor<fp16, [1, 768, 1, 3000]> var_116_cast_fp16 = add(x = var_76_cast_fp16, y = var_114_cast_fp16)[name = tensor<string, []>("op_116_cast_fp16")];
            tensor<string, []> hidden_states_1_mode_0 = const()[name = tensor<string, []>("hidden_states_1_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 768, 1, 3000]> hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_116_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
            tensor<string, []> var_162_pad_type_0 = const()[name = tensor<string, []>("op_162_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_162_pad_0 = const()[name = tensor<string, []>("op_162_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> var_162_strides_0 = const()[name = tensor<string, []>("op_162_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_162_dilations_0 = const()[name = tensor<string, []>("op_162_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_162_groups_0 = const()[name = tensor<string, []>("op_162_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 3]> var_131_to_fp16 = const()[name = tensor<string, []>("op_131_to_fp16"), val = tensor<fp16, [768, 768, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464320)))];
            tensor<fp16, [1, 768, 1, 1500]> var_162_cast_fp16 = conv(bias = var_57_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 = var_131_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_162_cast_fp16")];
            tensor<string, []> var_200_pad_type_0 = const()[name = tensor<string, []>("op_200_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_200_pad_0 = const()[name = tensor<string, []>("op_200_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> var_200_strides_0 = const()[name = tensor<string, []>("op_200_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_200_dilations_0 = const()[name = tensor<string, []>("op_200_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_200_groups_0 = const()[name = tensor<string, []>("op_200_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 3]> op_175_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [884736]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4003328))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4888128))), name = tensor<string, []>("op_175_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 3])];
            tensor<fp16, [768]> var_181_to_fp16 = const()[name = tensor<string, []>("op_181_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4888256)))];
            tensor<fp16, [1, 768, 1, 1500]> var_200_cast_fp16 = conv(bias = var_181_to_fp16, dilations = var_200_dilations_0, groups = var_200_groups_0, pad = var_200_pad_0, pad_type = var_200_pad_type_0, strides = var_200_strides_0, weight = op_175_to_fp16_palettized, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_200_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> var_202_cast_fp16 = add(x = var_162_cast_fp16, y = var_200_cast_fp16)[name = tensor<string, []>("op_202_cast_fp16")];
            tensor<string, []> hidden_states_3_mode_0 = const()[name = tensor<string, []>("hidden_states_3_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_202_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> var_222_to_fp16 = const()[name = tensor<string, []>("op_222_to_fp16"), val = tensor<fp16, [1, 768, 1, 1500]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4889856)))];
            tensor<fp16, [1, 768, 1, 1500]> inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_222_to_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
            tensor<int32, []> var_232 = const()[name = tensor<string, []>("op_232"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_254_to_fp16 = const()[name = tensor<string, []>("op_254_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_254_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
            tensor<fp16, [768]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7193920)))];
            tensor<fp16, [768]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7195520)))];
            tensor<fp16, [768]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7197120)))];
            tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
            tensor<string, []> var_276_pad_type_0 = const()[name = tensor<string, []>("op_276_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_276_strides_0 = const()[name = tensor<string, []>("op_276_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_276_pad_0 = const()[name = tensor<string, []>("op_276_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_276_dilations_0 = const()[name = tensor<string, []>("op_276_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_276_groups_0 = const()[name = tensor<string, []>("op_276_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7198720))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7493696))), name = tensor<string, []>("layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_0_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7493824)))];
            tensor<fp16, [1, 768, 1, 1500]> var_276_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_276_dilations_0, groups = var_276_groups_0, pad = var_276_pad_0, pad_type = var_276_pad_type_0, strides = var_276_strides_0, weight = layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
            tensor<string, []> var_282_pad_type_0 = const()[name = tensor<string, []>("op_282_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_282_strides_0 = const()[name = tensor<string, []>("op_282_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_282_pad_0 = const()[name = tensor<string, []>("op_282_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_282_dilations_0 = const()[name = tensor<string, []>("op_282_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_282_groups_0 = const()[name = tensor<string, []>("op_282_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7520576))), name = tensor<string, []>("layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [12535]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7495424))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_282_cast_fp16 = conv(dilations = var_282_dilations_0, groups = var_282_groups_0, pad = var_282_pad_0, pad_type = var_282_pad_type_0, strides = var_282_strides_0, weight = layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor<string, []>("op_282_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_1_cast_fp16 = add(x = var_276_cast_fp16, y = var_282_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
            tensor<string, []> var_291_pad_type_0 = const()[name = tensor<string, []>("op_291_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_291_strides_0 = const()[name = tensor<string, []>("op_291_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_291_pad_0 = const()[name = tensor<string, []>("op_291_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_291_dilations_0 = const()[name = tensor<string, []>("op_291_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_291_groups_0 = const()[name = tensor<string, []>("op_291_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7594368))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7889344))), name = tensor<string, []>("layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_291_cast_fp16 = conv(dilations = var_291_dilations_0, groups = var_291_groups_0, pad = var_291_pad_0, pad_type = var_291_pad_type_0, strides = var_291_strides_0, weight = layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("op_291_cast_fp16")];
            tensor<string, []> var_297_pad_type_0 = const()[name = tensor<string, []>("op_297_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_297_strides_0 = const()[name = tensor<string, []>("op_297_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_297_pad_0 = const()[name = tensor<string, []>("op_297_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_297_dilations_0 = const()[name = tensor<string, []>("op_297_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_297_groups_0 = const()[name = tensor<string, []>("op_297_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7913600))), name = tensor<string, []>("layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [12023]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7889472))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_297_cast_fp16 = conv(dilations = var_297_dilations_0, groups = var_297_groups_0, pad = var_297_pad_0, pad_type = var_297_pad_type_0, strides = var_297_strides_0, weight = layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor<string, []>("op_297_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_1_cast_fp16 = add(x = var_291_cast_fp16, y = var_297_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
            tensor<string, []> var_307_pad_type_0 = const()[name = tensor<string, []>("op_307_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_307_strides_0 = const()[name = tensor<string, []>("op_307_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_307_pad_0 = const()[name = tensor<string, []>("op_307_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_307_dilations_0 = const()[name = tensor<string, []>("op_307_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_307_groups_0 = const()[name = tensor<string, []>("op_307_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7987392))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8282368))), name = tensor<string, []>("layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_0_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8282496)))];
            tensor<fp16, [1, 768, 1, 1500]> var_307_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_307_dilations_0, groups = var_307_groups_0, pad = var_307_pad_0, pad_type = var_307_pad_type_0, strides = var_307_strides_0, weight = layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("op_307_cast_fp16")];
            tensor<string, []> var_313_pad_type_0 = const()[name = tensor<string, []>("op_313_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_313_strides_0 = const()[name = tensor<string, []>("op_313_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_313_pad_0 = const()[name = tensor<string, []>("op_313_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_313_dilations_0 = const()[name = tensor<string, []>("op_313_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_313_groups_0 = const()[name = tensor<string, []>("op_313_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8308864))), name = tensor<string, []>("layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [12322]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8284096))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_313_cast_fp16 = conv(dilations = var_313_dilations_0, groups = var_313_groups_0, pad = var_313_pad_0, pad_type = var_313_pad_type_0, strides = var_313_strides_0, weight = layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor<string, []>("op_313_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_1_cast_fp16 = add(x = var_307_cast_fp16, y = var_313_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
            tensor<int32, [4]> var_317 = const()[name = tensor<string, []>("op_317"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_1_cast_fp16 = reshape(shape = var_317, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")];
            tensor<fp16, []> var_319_to_fp16 = const()[name = tensor<string, []>("op_319_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_320_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_319_to_fp16)[name = tensor<string, []>("op_320_cast_fp16")];
            tensor<int32, [4]> var_323 = const()[name = tensor<string, []>("op_323"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_324_cast_fp16 = reshape(shape = var_323, x = key_1_cast_fp16)[name = tensor<string, []>("op_324_cast_fp16")];
            tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_320_cast_fp16, y = var_324_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_327_cast_fp16 = softmax(axis = var_232, x = mh_w_1_cast_fp16)[name = tensor<string, []>("op_327_cast_fp16")];
            tensor<int32, [4]> var_328 = const()[name = tensor<string, []>("op_328"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_329_cast_fp16 = reshape(shape = var_328, x = value_1_cast_fp16)[name = tensor<string, []>("op_329_cast_fp16")];
            tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_329_cast_fp16, y = var_327_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
            tensor<int32, [4]> var_332 = const()[name = tensor<string, []>("op_332"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_1_cast_fp16 = reshape(shape = var_332, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<string, []> var_342_pad_type_0 = const()[name = tensor<string, []>("op_342_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_342_strides_0 = const()[name = tensor<string, []>("op_342_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_342_pad_0 = const()[name = tensor<string, []>("op_342_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_342_dilations_0 = const()[name = tensor<string, []>("op_342_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_342_groups_0 = const()[name = tensor<string, []>("op_342_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8382656))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8677632))), name = tensor<string, []>("layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_0_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8677760)))];
            tensor<fp16, [1, 768, 1, 1500]> var_342_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_342_dilations_0, groups = var_342_groups_0, pad = var_342_pad_0, pad_type = var_342_pad_type_0, strides = var_342_strides_0, weight = layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor<string, []>("op_342_cast_fp16")];
            tensor<string, []> var_348_pad_type_0 = const()[name = tensor<string, []>("op_348_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_348_strides_0 = const()[name = tensor<string, []>("op_348_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_348_pad_0 = const()[name = tensor<string, []>("op_348_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_348_dilations_0 = const()[name = tensor<string, []>("op_348_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_348_groups_0 = const()[name = tensor<string, []>("op_348_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8699776))), name = tensor<string, []>("layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [10175]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8679360))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_348_cast_fp16 = conv(dilations = var_348_dilations_0, groups = var_348_groups_0, pad = var_348_pad_0, pad_type = var_348_pad_type_0, strides = var_348_strides_0, weight = layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_1_cast_fp16)[name = tensor<string, []>("op_348_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_3_cast_fp16 = add(x = var_342_cast_fp16, y = var_348_cast_fp16)[name = tensor<string, []>("obj_3_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_359_to_fp16 = const()[name = tensor<string, []>("op_359_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_359_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
            tensor<fp16, [768]> input_3_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_3_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8773568)))];
            tensor<fp16, [768]> input_3_beta_0_to_fp16 = const()[name = tensor<string, []>("input_3_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8775168)))];
            tensor<fp16, []> input_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<string, []> var_377_pad_type_0 = const()[name = tensor<string, []>("op_377_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_377_strides_0 = const()[name = tensor<string, []>("op_377_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_377_pad_0 = const()[name = tensor<string, []>("op_377_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_377_dilations_0 = const()[name = tensor<string, []>("op_377_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_377_groups_0 = const()[name = tensor<string, []>("op_377_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_0_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8776768))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9956480))), name = tensor<string, []>("layers_0_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_0_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9956608)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_377_cast_fp16 = conv(bias = layers_0_fc1_inlier_module_bias_to_fp16, dilations = var_377_dilations_0, groups = var_377_groups_0, pad = var_377_pad_0, pad_type = var_377_pad_type_0, strides = var_377_strides_0, weight = layers_0_fc1_inlier_module_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = tensor<string, []>("op_377_cast_fp16")];
            tensor<string, []> var_383_pad_type_0 = const()[name = tensor<string, []>("op_383_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_383_strides_0 = const()[name = tensor<string, []>("op_383_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_383_pad_0 = const()[name = tensor<string, []>("op_383_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_383_dilations_0 = const()[name = tensor<string, []>("op_383_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_383_groups_0 = const()[name = tensor<string, []>("op_383_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_0_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10057408))), name = tensor<string, []>("layers_0_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [47244]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9962816))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_383_cast_fp16 = conv(dilations = var_383_dilations_0, groups = var_383_groups_0, pad = var_383_pad_0, pad_type = var_383_pad_type_0, strides = var_383_strides_0, weight = layers_0_fc1_outlier_module_weight_to_fp16_sparsified, x = input_3_cast_fp16)[name = tensor<string, []>("op_383_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_5_cast_fp16 = add(x = var_377_cast_fp16, y = var_383_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<string, []> input_7_mode_0 = const()[name = tensor<string, []>("input_7_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<string, []> var_394_pad_type_0 = const()[name = tensor<string, []>("op_394_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_394_strides_0 = const()[name = tensor<string, []>("op_394_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_394_pad_0 = const()[name = tensor<string, []>("op_394_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_394_dilations_0 = const()[name = tensor<string, []>("op_394_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_394_groups_0 = const()[name = tensor<string, []>("op_394_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_0_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10352384))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11532096))), name = tensor<string, []>("layers_0_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_0_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11532224)))];
            tensor<fp16, [1, 768, 1, 1500]> var_394_cast_fp16 = conv(bias = layers_0_fc2_inlier_module_bias_to_fp16, 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_fc2_inlier_module_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor<string, []>("op_394_cast_fp16")];
            tensor<string, []> var_400_pad_type_0 = const()[name = tensor<string, []>("op_400_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_400_strides_0 = const()[name = tensor<string, []>("op_400_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_400_pad_0 = const()[name = tensor<string, []>("op_400_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_400_dilations_0 = const()[name = tensor<string, []>("op_400_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_400_groups_0 = const()[name = tensor<string, []>("op_400_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_0_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11613568))), name = tensor<string, []>("layers_0_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [39820]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11533824))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_400_cast_fp16 = conv(dilations = var_400_dilations_0, groups = var_400_groups_0, pad = var_400_pad_0, pad_type = var_400_pad_type_0, strides = var_400_strides_0, weight = layers_0_fc2_outlier_module_weight_to_fp16_sparsified, x = input_7_cast_fp16)[name = tensor<string, []>("op_400_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_5_cast_fp16 = add(x = var_394_cast_fp16, y = var_400_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
            tensor<int32, []> var_406 = const()[name = tensor<string, []>("op_406"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_428_to_fp16 = const()[name = tensor<string, []>("op_428_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_428_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
            tensor<fp16, [768]> obj_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11908544)))];
            tensor<fp16, [768]> obj_5_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_5_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11910144)))];
            tensor<fp16, []> obj_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("obj_5_cast_fp16")];
            tensor<string, []> var_450_pad_type_0 = const()[name = tensor<string, []>("op_450_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_450_strides_0 = const()[name = tensor<string, []>("op_450_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_450_pad_0 = const()[name = tensor<string, []>("op_450_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_450_dilations_0 = const()[name = tensor<string, []>("op_450_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_450_groups_0 = const()[name = tensor<string, []>("op_450_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11911744))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12206720))), name = tensor<string, []>("layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_1_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12206848)))];
            tensor<fp16, [1, 768, 1, 1500]> var_450_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_450_dilations_0, groups = var_450_groups_0, pad = var_450_pad_0, pad_type = var_450_pad_type_0, strides = var_450_strides_0, weight = layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = tensor<string, []>("op_450_cast_fp16")];
            tensor<string, []> var_456_pad_type_0 = const()[name = tensor<string, []>("op_456_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_456_strides_0 = const()[name = tensor<string, []>("op_456_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_456_pad_0 = const()[name = tensor<string, []>("op_456_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_456_dilations_0 = const()[name = tensor<string, []>("op_456_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_456_groups_0 = const()[name = tensor<string, []>("op_456_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12226688))), name = tensor<string, []>("layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [9080]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12208448))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_456_cast_fp16 = conv(dilations = var_456_dilations_0, groups = var_456_groups_0, pad = var_456_pad_0, pad_type = var_456_pad_type_0, strides = var_456_strides_0, weight = layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_5_cast_fp16)[name = tensor<string, []>("op_456_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_3_cast_fp16 = add(x = var_450_cast_fp16, y = var_456_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
            tensor<string, []> var_465_pad_type_0 = const()[name = tensor<string, []>("op_465_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_465_strides_0 = const()[name = tensor<string, []>("op_465_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_465_pad_0 = const()[name = tensor<string, []>("op_465_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_465_dilations_0 = const()[name = tensor<string, []>("op_465_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_465_groups_0 = const()[name = tensor<string, []>("op_465_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12300480))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12595456))), name = tensor<string, []>("layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_465_cast_fp16 = conv(dilations = var_465_dilations_0, groups = var_465_groups_0, pad = var_465_pad_0, pad_type = var_465_pad_type_0, strides = var_465_strides_0, weight = layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = tensor<string, []>("op_465_cast_fp16")];
            tensor<string, []> var_471_pad_type_0 = const()[name = tensor<string, []>("op_471_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_471_strides_0 = const()[name = tensor<string, []>("op_471_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_471_pad_0 = const()[name = tensor<string, []>("op_471_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_471_dilations_0 = const()[name = tensor<string, []>("op_471_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_471_groups_0 = const()[name = tensor<string, []>("op_471_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12614464))), name = tensor<string, []>("layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [9397]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12595584))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_471_cast_fp16 = conv(dilations = var_471_dilations_0, groups = var_471_groups_0, pad = var_471_pad_0, pad_type = var_471_pad_type_0, strides = var_471_strides_0, weight = layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_5_cast_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_3_cast_fp16 = add(x = var_465_cast_fp16, y = var_471_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
            tensor<string, []> var_481_pad_type_0 = const()[name = tensor<string, []>("op_481_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_481_strides_0 = const()[name = tensor<string, []>("op_481_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_481_pad_0 = const()[name = tensor<string, []>("op_481_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_481_dilations_0 = const()[name = tensor<string, []>("op_481_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_481_groups_0 = const()[name = tensor<string, []>("op_481_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12688256))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12983232))), name = tensor<string, []>("layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_1_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12983360)))];
            tensor<fp16, [1, 768, 1, 1500]> var_481_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_481_dilations_0, groups = var_481_groups_0, pad = var_481_pad_0, pad_type = var_481_pad_type_0, strides = var_481_strides_0, weight = layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = tensor<string, []>("op_481_cast_fp16")];
            tensor<string, []> var_487_pad_type_0 = const()[name = tensor<string, []>("op_487_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_487_strides_0 = const()[name = tensor<string, []>("op_487_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_487_pad_0 = const()[name = tensor<string, []>("op_487_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_487_dilations_0 = const()[name = tensor<string, []>("op_487_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_487_groups_0 = const()[name = tensor<string, []>("op_487_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13000768))), name = tensor<string, []>("layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [7866]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12984960))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_487_cast_fp16 = conv(dilations = var_487_dilations_0, groups = var_487_groups_0, pad = var_487_pad_0, pad_type = var_487_pad_type_0, strides = var_487_strides_0, weight = layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_5_cast_fp16)[name = tensor<string, []>("op_487_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_3_cast_fp16 = add(x = var_481_cast_fp16, y = var_487_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")];
            tensor<int32, [4]> var_491 = const()[name = tensor<string, []>("op_491"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_3_cast_fp16 = reshape(shape = var_491, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")];
            tensor<fp16, []> var_493_to_fp16 = const()[name = tensor<string, []>("op_493_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_494_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_493_to_fp16)[name = tensor<string, []>("op_494_cast_fp16")];
            tensor<int32, [4]> var_497 = const()[name = tensor<string, []>("op_497"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_498_cast_fp16 = reshape(shape = var_497, x = key_3_cast_fp16)[name = tensor<string, []>("op_498_cast_fp16")];
            tensor<bool, []> mh_w_3_transpose_x_0 = const()[name = tensor<string, []>("mh_w_3_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_3_transpose_y_0 = const()[name = tensor<string, []>("mh_w_3_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_494_cast_fp16, y = var_498_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_501_cast_fp16 = softmax(axis = var_406, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_501_cast_fp16")];
            tensor<int32, [4]> var_502 = const()[name = tensor<string, []>("op_502"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_503_cast_fp16 = reshape(shape = var_502, x = value_3_cast_fp16)[name = tensor<string, []>("op_503_cast_fp16")];
            tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_503_cast_fp16, y = var_501_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
            tensor<int32, [4]> var_506 = const()[name = tensor<string, []>("op_506"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_9_cast_fp16 = reshape(shape = var_506, x = attn_3_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<string, []> var_516_pad_type_0 = const()[name = tensor<string, []>("op_516_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_516_strides_0 = const()[name = tensor<string, []>("op_516_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_516_pad_0 = const()[name = tensor<string, []>("op_516_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_516_dilations_0 = const()[name = tensor<string, []>("op_516_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_516_groups_0 = const()[name = tensor<string, []>("op_516_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13074560))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13369536))), name = tensor<string, []>("layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_1_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13369664)))];
            tensor<fp16, [1, 768, 1, 1500]> var_516_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_516_dilations_0, groups = var_516_groups_0, pad = var_516_pad_0, pad_type = var_516_pad_type_0, strides = var_516_strides_0, weight = layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = tensor<string, []>("op_516_cast_fp16")];
            tensor<string, []> var_522_pad_type_0 = const()[name = tensor<string, []>("op_522_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_522_strides_0 = const()[name = tensor<string, []>("op_522_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_522_pad_0 = const()[name = tensor<string, []>("op_522_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_522_dilations_0 = const()[name = tensor<string, []>("op_522_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_522_groups_0 = const()[name = tensor<string, []>("op_522_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13384576))), name = tensor<string, []>("layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6617]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13371264))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_522_cast_fp16 = conv(dilations = var_522_dilations_0, groups = var_522_groups_0, pad = var_522_pad_0, pad_type = var_522_pad_type_0, strides = var_522_strides_0, weight = layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_9_cast_fp16)[name = tensor<string, []>("op_522_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_7_cast_fp16 = add(x = var_516_cast_fp16, y = var_522_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_533_to_fp16 = const()[name = tensor<string, []>("op_533_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_533_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
            tensor<fp16, [768]> input_11_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_11_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13458368)))];
            tensor<fp16, [768]> input_11_beta_0_to_fp16 = const()[name = tensor<string, []>("input_11_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13459968)))];
            tensor<fp16, []> input_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<string, []> var_551_pad_type_0 = const()[name = tensor<string, []>("op_551_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_551_strides_0 = const()[name = tensor<string, []>("op_551_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_551_pad_0 = const()[name = tensor<string, []>("op_551_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_551_dilations_0 = const()[name = tensor<string, []>("op_551_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_551_groups_0 = const()[name = tensor<string, []>("op_551_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_1_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13461568))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14641280))), name = tensor<string, []>("layers_1_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_1_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14641408)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_551_cast_fp16 = conv(bias = layers_1_fc1_inlier_module_bias_to_fp16, dilations = var_551_dilations_0, groups = var_551_groups_0, pad = var_551_pad_0, pad_type = var_551_pad_type_0, strides = var_551_strides_0, weight = layers_1_fc1_inlier_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor<string, []>("op_551_cast_fp16")];
            tensor<string, []> var_557_pad_type_0 = const()[name = tensor<string, []>("op_557_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_557_strides_0 = const()[name = tensor<string, []>("op_557_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_557_pad_0 = const()[name = tensor<string, []>("op_557_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_557_dilations_0 = const()[name = tensor<string, []>("op_557_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_557_groups_0 = const()[name = tensor<string, []>("op_557_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_1_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14721024))), name = tensor<string, []>("layers_1_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [36655]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14647616))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_557_cast_fp16 = conv(dilations = var_557_dilations_0, groups = var_557_groups_0, pad = var_557_pad_0, pad_type = var_557_pad_type_0, strides = var_557_strides_0, weight = layers_1_fc1_outlier_module_weight_to_fp16_sparsified, x = input_11_cast_fp16)[name = tensor<string, []>("op_557_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_13_cast_fp16 = add(x = var_551_cast_fp16, y = var_557_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<string, []> input_15_mode_0 = const()[name = tensor<string, []>("input_15_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<string, []> var_568_pad_type_0 = const()[name = tensor<string, []>("op_568_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_568_strides_0 = const()[name = tensor<string, []>("op_568_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_568_pad_0 = const()[name = tensor<string, []>("op_568_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_568_dilations_0 = const()[name = tensor<string, []>("op_568_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_568_groups_0 = const()[name = tensor<string, []>("op_568_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_1_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15016000))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16195712))), name = tensor<string, []>("layers_1_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_1_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16195840)))];
            tensor<fp16, [1, 768, 1, 1500]> var_568_cast_fp16 = conv(bias = layers_1_fc2_inlier_module_bias_to_fp16, dilations = var_568_dilations_0, groups = var_568_groups_0, pad = var_568_pad_0, pad_type = var_568_pad_type_0, strides = var_568_strides_0, weight = layers_1_fc2_inlier_module_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor<string, []>("op_568_cast_fp16")];
            tensor<string, []> var_574_pad_type_0 = const()[name = tensor<string, []>("op_574_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_574_strides_0 = const()[name = tensor<string, []>("op_574_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_574_pad_0 = const()[name = tensor<string, []>("op_574_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_574_dilations_0 = const()[name = tensor<string, []>("op_574_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_574_groups_0 = const()[name = tensor<string, []>("op_574_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_1_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16268160))), name = tensor<string, []>("layers_1_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [35313]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16197440))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_574_cast_fp16 = conv(dilations = var_574_dilations_0, groups = var_574_groups_0, pad = var_574_pad_0, pad_type = var_574_pad_type_0, strides = var_574_strides_0, weight = layers_1_fc2_outlier_module_weight_to_fp16_sparsified, x = input_15_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_7_cast_fp16 = add(x = var_568_cast_fp16, y = var_574_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
            tensor<int32, []> var_580 = const()[name = tensor<string, []>("op_580"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_602_to_fp16 = const()[name = tensor<string, []>("op_602_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_602_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
            tensor<fp16, [768]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16563136)))];
            tensor<fp16, [768]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16564736)))];
            tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
            tensor<string, []> var_624_pad_type_0 = const()[name = tensor<string, []>("op_624_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_624_strides_0 = const()[name = tensor<string, []>("op_624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_624_pad_0 = const()[name = tensor<string, []>("op_624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_624_dilations_0 = const()[name = tensor<string, []>("op_624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_624_groups_0 = const()[name = tensor<string, []>("op_624_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16566336))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16861312))), name = tensor<string, []>("layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_2_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16861440)))];
            tensor<fp16, [1, 768, 1, 1500]> var_624_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_624_dilations_0, groups = var_624_groups_0, pad = var_624_pad_0, pad_type = var_624_pad_type_0, strides = var_624_strides_0, weight = layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("op_624_cast_fp16")];
            tensor<string, []> var_630_pad_type_0 = const()[name = tensor<string, []>("op_630_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_630_strides_0 = const()[name = tensor<string, []>("op_630_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_630_pad_0 = const()[name = tensor<string, []>("op_630_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_630_dilations_0 = const()[name = tensor<string, []>("op_630_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_630_groups_0 = const()[name = tensor<string, []>("op_630_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16878784))), name = tensor<string, []>("layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [7822]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16863040))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_630_cast_fp16 = conv(dilations = var_630_dilations_0, groups = var_630_groups_0, pad = var_630_pad_0, pad_type = var_630_pad_type_0, strides = var_630_strides_0, weight = layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_9_cast_fp16)[name = tensor<string, []>("op_630_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_5_cast_fp16 = add(x = var_624_cast_fp16, y = var_630_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
            tensor<string, []> var_639_pad_type_0 = const()[name = tensor<string, []>("op_639_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_639_strides_0 = const()[name = tensor<string, []>("op_639_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_639_pad_0 = const()[name = tensor<string, []>("op_639_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_639_dilations_0 = const()[name = tensor<string, []>("op_639_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_639_groups_0 = const()[name = tensor<string, []>("op_639_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16952576))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17247552))), name = tensor<string, []>("layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_639_cast_fp16 = conv(dilations = var_639_dilations_0, groups = var_639_groups_0, pad = var_639_pad_0, pad_type = var_639_pad_type_0, strides = var_639_strides_0, weight = layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("op_639_cast_fp16")];
            tensor<string, []> var_645_pad_type_0 = const()[name = tensor<string, []>("op_645_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_645_strides_0 = const()[name = tensor<string, []>("op_645_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_645_pad_0 = const()[name = tensor<string, []>("op_645_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_645_dilations_0 = const()[name = tensor<string, []>("op_645_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_645_groups_0 = const()[name = tensor<string, []>("op_645_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17262400))), name = tensor<string, []>("layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [7313]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17247680))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_645_cast_fp16 = conv(dilations = var_645_dilations_0, groups = var_645_groups_0, pad = var_645_pad_0, pad_type = var_645_pad_type_0, strides = var_645_strides_0, weight = layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_9_cast_fp16)[name = tensor<string, []>("op_645_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_5_cast_fp16 = add(x = var_639_cast_fp16, y = var_645_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
            tensor<string, []> var_655_pad_type_0 = const()[name = tensor<string, []>("op_655_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_655_strides_0 = const()[name = tensor<string, []>("op_655_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_655_pad_0 = const()[name = tensor<string, []>("op_655_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_655_dilations_0 = const()[name = tensor<string, []>("op_655_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_655_groups_0 = const()[name = tensor<string, []>("op_655_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17336192))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17631168))), name = tensor<string, []>("layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_2_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17631296)))];
            tensor<fp16, [1, 768, 1, 1500]> var_655_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_655_dilations_0, groups = var_655_groups_0, pad = var_655_pad_0, pad_type = var_655_pad_type_0, strides = var_655_strides_0, weight = layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("op_655_cast_fp16")];
            tensor<string, []> var_661_pad_type_0 = const()[name = tensor<string, []>("op_661_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_661_strides_0 = const()[name = tensor<string, []>("op_661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_661_pad_0 = const()[name = tensor<string, []>("op_661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_661_dilations_0 = const()[name = tensor<string, []>("op_661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_661_groups_0 = const()[name = tensor<string, []>("op_661_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17644992))), name = tensor<string, []>("layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5996]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17632896))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_661_cast_fp16 = conv(dilations = var_661_dilations_0, groups = var_661_groups_0, pad = var_661_pad_0, pad_type = var_661_pad_type_0, strides = var_661_strides_0, weight = layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_9_cast_fp16)[name = tensor<string, []>("op_661_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_5_cast_fp16 = add(x = var_655_cast_fp16, y = var_661_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
            tensor<int32, [4]> var_665 = const()[name = tensor<string, []>("op_665"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_5_cast_fp16 = reshape(shape = var_665, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")];
            tensor<fp16, []> var_667_to_fp16 = const()[name = tensor<string, []>("op_667_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_668_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_667_to_fp16)[name = tensor<string, []>("op_668_cast_fp16")];
            tensor<int32, [4]> var_671 = const()[name = tensor<string, []>("op_671"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_672_cast_fp16 = reshape(shape = var_671, x = key_5_cast_fp16)[name = tensor<string, []>("op_672_cast_fp16")];
            tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_668_cast_fp16, y = var_672_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_675_cast_fp16 = softmax(axis = var_580, x = mh_w_5_cast_fp16)[name = tensor<string, []>("op_675_cast_fp16")];
            tensor<int32, [4]> var_676 = const()[name = tensor<string, []>("op_676"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_677_cast_fp16 = reshape(shape = var_676, x = value_5_cast_fp16)[name = tensor<string, []>("op_677_cast_fp16")];
            tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_677_cast_fp16, y = var_675_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
            tensor<int32, [4]> var_680 = const()[name = tensor<string, []>("op_680"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_17_cast_fp16 = reshape(shape = var_680, x = attn_5_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<string, []> var_690_pad_type_0 = const()[name = tensor<string, []>("op_690_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_690_strides_0 = const()[name = tensor<string, []>("op_690_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_690_pad_0 = const()[name = tensor<string, []>("op_690_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_690_dilations_0 = const()[name = tensor<string, []>("op_690_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_690_groups_0 = const()[name = tensor<string, []>("op_690_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17718784))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18013760))), name = tensor<string, []>("layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_2_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18013888)))];
            tensor<fp16, [1, 768, 1, 1500]> var_690_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_inlier_module_bias_to_fp16, 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_o_proj_inlier_module_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor<string, []>("op_690_cast_fp16")];
            tensor<string, []> var_696_pad_type_0 = const()[name = tensor<string, []>("op_696_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_696_strides_0 = const()[name = tensor<string, []>("op_696_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_696_pad_0 = const()[name = tensor<string, []>("op_696_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_696_dilations_0 = const()[name = tensor<string, []>("op_696_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_696_groups_0 = const()[name = tensor<string, []>("op_696_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18025408))), name = tensor<string, []>("layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [4911]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18015488))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_696_cast_fp16 = conv(dilations = var_696_dilations_0, groups = var_696_groups_0, pad = var_696_pad_0, pad_type = var_696_pad_type_0, strides = var_696_strides_0, weight = layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_17_cast_fp16)[name = tensor<string, []>("op_696_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_11_cast_fp16 = add(x = var_690_cast_fp16, y = var_696_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_707_to_fp16 = const()[name = tensor<string, []>("op_707_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_707_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
            tensor<fp16, [768]> input_19_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_19_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18099200)))];
            tensor<fp16, [768]> input_19_beta_0_to_fp16 = const()[name = tensor<string, []>("input_19_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18100800)))];
            tensor<fp16, []> input_19_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_19_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<string, []> var_725_pad_type_0 = const()[name = tensor<string, []>("op_725_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_725_strides_0 = const()[name = tensor<string, []>("op_725_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_725_pad_0 = const()[name = tensor<string, []>("op_725_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_725_dilations_0 = const()[name = tensor<string, []>("op_725_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_725_groups_0 = const()[name = tensor<string, []>("op_725_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_2_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18102400))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19282112))), name = tensor<string, []>("layers_2_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_2_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19282240)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_725_cast_fp16 = conv(bias = layers_2_fc1_inlier_module_bias_to_fp16, dilations = var_725_dilations_0, groups = var_725_groups_0, pad = var_725_pad_0, pad_type = var_725_pad_type_0, strides = var_725_strides_0, weight = layers_2_fc1_inlier_module_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = tensor<string, []>("op_725_cast_fp16")];
            tensor<string, []> var_731_pad_type_0 = const()[name = tensor<string, []>("op_731_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_731_strides_0 = const()[name = tensor<string, []>("op_731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_731_pad_0 = const()[name = tensor<string, []>("op_731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_731_dilations_0 = const()[name = tensor<string, []>("op_731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_731_groups_0 = const()[name = tensor<string, []>("op_731_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_2_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19352448))), name = tensor<string, []>("layers_2_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [31950]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19288448))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_731_cast_fp16 = conv(dilations = var_731_dilations_0, groups = var_731_groups_0, pad = var_731_pad_0, pad_type = var_731_pad_type_0, strides = var_731_strides_0, weight = layers_2_fc1_outlier_module_weight_to_fp16_sparsified, x = input_19_cast_fp16)[name = tensor<string, []>("op_731_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_21_cast_fp16 = add(x = var_725_cast_fp16, y = var_731_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<string, []> input_23_mode_0 = const()[name = tensor<string, []>("input_23_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<string, []> var_742_pad_type_0 = const()[name = tensor<string, []>("op_742_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_742_strides_0 = const()[name = tensor<string, []>("op_742_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_742_pad_0 = const()[name = tensor<string, []>("op_742_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_742_dilations_0 = const()[name = tensor<string, []>("op_742_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_742_groups_0 = const()[name = tensor<string, []>("op_742_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_2_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19647424))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20827136))), name = tensor<string, []>("layers_2_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_2_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20827264)))];
            tensor<fp16, [1, 768, 1, 1500]> var_742_cast_fp16 = conv(bias = layers_2_fc2_inlier_module_bias_to_fp16, dilations = var_742_dilations_0, groups = var_742_groups_0, pad = var_742_pad_0, pad_type = var_742_pad_type_0, strides = var_742_strides_0, weight = layers_2_fc2_inlier_module_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor<string, []>("op_742_cast_fp16")];
            tensor<string, []> var_748_pad_type_0 = const()[name = tensor<string, []>("op_748_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_748_strides_0 = const()[name = tensor<string, []>("op_748_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_748_pad_0 = const()[name = tensor<string, []>("op_748_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_748_dilations_0 = const()[name = tensor<string, []>("op_748_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_748_groups_0 = const()[name = tensor<string, []>("op_748_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_2_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20892032))), name = tensor<string, []>("layers_2_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [31546]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20828864))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_748_cast_fp16 = conv(dilations = var_748_dilations_0, groups = var_748_groups_0, pad = var_748_pad_0, pad_type = var_748_pad_type_0, strides = var_748_strides_0, weight = layers_2_fc2_outlier_module_weight_to_fp16_sparsified, x = input_23_cast_fp16)[name = tensor<string, []>("op_748_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_9_cast_fp16 = add(x = var_742_cast_fp16, y = var_748_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
            tensor<int32, []> var_754 = const()[name = tensor<string, []>("op_754"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_776_to_fp16 = const()[name = tensor<string, []>("op_776_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_776_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
            tensor<fp16, [768]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21187008)))];
            tensor<fp16, [768]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21188608)))];
            tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
            tensor<string, []> var_798_pad_type_0 = const()[name = tensor<string, []>("op_798_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_798_strides_0 = const()[name = tensor<string, []>("op_798_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_798_pad_0 = const()[name = tensor<string, []>("op_798_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_798_dilations_0 = const()[name = tensor<string, []>("op_798_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_798_groups_0 = const()[name = tensor<string, []>("op_798_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21190208))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21485184))), name = tensor<string, []>("layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_3_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21485312)))];
            tensor<fp16, [1, 768, 1, 1500]> var_798_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_798_dilations_0, groups = var_798_groups_0, pad = var_798_pad_0, pad_type = var_798_pad_type_0, strides = var_798_strides_0, weight = layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor<string, []>("op_798_cast_fp16")];
            tensor<string, []> var_804_pad_type_0 = const()[name = tensor<string, []>("op_804_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_804_strides_0 = const()[name = tensor<string, []>("op_804_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_804_pad_0 = const()[name = tensor<string, []>("op_804_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_804_dilations_0 = const()[name = tensor<string, []>("op_804_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_804_groups_0 = const()[name = tensor<string, []>("op_804_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21500352))), name = tensor<string, []>("layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6682]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21486912))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_804_cast_fp16 = conv(dilations = var_804_dilations_0, groups = var_804_groups_0, pad = var_804_pad_0, pad_type = var_804_pad_type_0, strides = var_804_strides_0, weight = layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_13_cast_fp16)[name = tensor<string, []>("op_804_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_7_cast_fp16 = add(x = var_798_cast_fp16, y = var_804_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
            tensor<string, []> var_813_pad_type_0 = const()[name = tensor<string, []>("op_813_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_813_strides_0 = const()[name = tensor<string, []>("op_813_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_813_pad_0 = const()[name = tensor<string, []>("op_813_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_813_dilations_0 = const()[name = tensor<string, []>("op_813_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_813_groups_0 = const()[name = tensor<string, []>("op_813_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21574144))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21869120))), name = tensor<string, []>("layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_813_cast_fp16 = conv(dilations = var_813_dilations_0, groups = var_813_groups_0, pad = var_813_pad_0, pad_type = var_813_pad_type_0, strides = var_813_strides_0, weight = layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor<string, []>("op_813_cast_fp16")];
            tensor<string, []> var_819_pad_type_0 = const()[name = tensor<string, []>("op_819_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_819_strides_0 = const()[name = tensor<string, []>("op_819_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_819_pad_0 = const()[name = tensor<string, []>("op_819_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_819_dilations_0 = const()[name = tensor<string, []>("op_819_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_819_groups_0 = const()[name = tensor<string, []>("op_819_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21881920))), name = tensor<string, []>("layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21869248))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_819_cast_fp16 = conv(dilations = var_819_dilations_0, groups = var_819_groups_0, pad = var_819_pad_0, pad_type = var_819_pad_type_0, strides = var_819_strides_0, weight = layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_13_cast_fp16)[name = tensor<string, []>("op_819_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_7_cast_fp16 = add(x = var_813_cast_fp16, y = var_819_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
            tensor<string, []> var_829_pad_type_0 = const()[name = tensor<string, []>("op_829_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_829_strides_0 = const()[name = tensor<string, []>("op_829_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_829_pad_0 = const()[name = tensor<string, []>("op_829_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_829_dilations_0 = const()[name = tensor<string, []>("op_829_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_829_groups_0 = const()[name = tensor<string, []>("op_829_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21955712))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22250688))), name = tensor<string, []>("layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_3_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22250816)))];
            tensor<fp16, [1, 768, 1, 1500]> var_829_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_829_dilations_0, groups = var_829_groups_0, pad = var_829_pad_0, pad_type = var_829_pad_type_0, strides = var_829_strides_0, weight = layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor<string, []>("op_829_cast_fp16")];
            tensor<string, []> var_835_pad_type_0 = const()[name = tensor<string, []>("op_835_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_835_strides_0 = const()[name = tensor<string, []>("op_835_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_835_pad_0 = const()[name = tensor<string, []>("op_835_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_835_dilations_0 = const()[name = tensor<string, []>("op_835_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_835_groups_0 = const()[name = tensor<string, []>("op_835_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22262080))), name = tensor<string, []>("layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [4773]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22252416))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_835_cast_fp16 = conv(dilations = var_835_dilations_0, groups = var_835_groups_0, pad = var_835_pad_0, pad_type = var_835_pad_type_0, strides = var_835_strides_0, weight = layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_13_cast_fp16)[name = tensor<string, []>("op_835_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_7_cast_fp16 = add(x = var_829_cast_fp16, y = var_835_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")];
            tensor<int32, [4]> var_839 = const()[name = tensor<string, []>("op_839"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_7_cast_fp16 = reshape(shape = var_839, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")];
            tensor<fp16, []> var_841_to_fp16 = const()[name = tensor<string, []>("op_841_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_842_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_841_to_fp16)[name = tensor<string, []>("op_842_cast_fp16")];
            tensor<int32, [4]> var_845 = const()[name = tensor<string, []>("op_845"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_846_cast_fp16 = reshape(shape = var_845, x = key_7_cast_fp16)[name = tensor<string, []>("op_846_cast_fp16")];
            tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_842_cast_fp16, y = var_846_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_849_cast_fp16 = softmax(axis = var_754, x = mh_w_7_cast_fp16)[name = tensor<string, []>("op_849_cast_fp16")];
            tensor<int32, [4]> var_850 = const()[name = tensor<string, []>("op_850"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_851_cast_fp16 = reshape(shape = var_850, x = value_7_cast_fp16)[name = tensor<string, []>("op_851_cast_fp16")];
            tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_851_cast_fp16, y = var_849_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
            tensor<int32, [4]> var_854 = const()[name = tensor<string, []>("op_854"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_25_cast_fp16 = reshape(shape = var_854, x = attn_7_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<string, []> var_864_pad_type_0 = const()[name = tensor<string, []>("op_864_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_864_strides_0 = const()[name = tensor<string, []>("op_864_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_864_pad_0 = const()[name = tensor<string, []>("op_864_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_864_dilations_0 = const()[name = tensor<string, []>("op_864_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_864_groups_0 = const()[name = tensor<string, []>("op_864_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22335872))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22630848))), name = tensor<string, []>("layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_3_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22630976)))];
            tensor<fp16, [1, 768, 1, 1500]> var_864_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_864_dilations_0, groups = var_864_groups_0, pad = var_864_pad_0, pad_type = var_864_pad_type_0, strides = var_864_strides_0, weight = layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor<string, []>("op_864_cast_fp16")];
            tensor<string, []> var_870_pad_type_0 = const()[name = tensor<string, []>("op_870_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_870_strides_0 = const()[name = tensor<string, []>("op_870_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_870_pad_0 = const()[name = tensor<string, []>("op_870_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_870_dilations_0 = const()[name = tensor<string, []>("op_870_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_870_groups_0 = const()[name = tensor<string, []>("op_870_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22640640))), name = tensor<string, []>("layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3976]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22632576))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_870_cast_fp16 = conv(dilations = var_870_dilations_0, groups = var_870_groups_0, pad = var_870_pad_0, pad_type = var_870_pad_type_0, strides = var_870_strides_0, weight = layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_25_cast_fp16)[name = tensor<string, []>("op_870_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_15_cast_fp16 = add(x = var_864_cast_fp16, y = var_870_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_15_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_881_to_fp16 = const()[name = tensor<string, []>("op_881_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_881_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
            tensor<fp16, [768]> input_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_27_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22714432)))];
            tensor<fp16, [768]> input_27_beta_0_to_fp16 = const()[name = tensor<string, []>("input_27_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22716032)))];
            tensor<fp16, []> input_27_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_27_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<string, []> var_899_pad_type_0 = const()[name = tensor<string, []>("op_899_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_899_strides_0 = const()[name = tensor<string, []>("op_899_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_899_pad_0 = const()[name = tensor<string, []>("op_899_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_899_dilations_0 = const()[name = tensor<string, []>("op_899_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_899_groups_0 = const()[name = tensor<string, []>("op_899_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_3_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22717632))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23897344))), name = tensor<string, []>("layers_3_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_3_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23897472)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_899_cast_fp16 = conv(bias = layers_3_fc1_inlier_module_bias_to_fp16, dilations = var_899_dilations_0, groups = var_899_groups_0, pad = var_899_pad_0, pad_type = var_899_pad_type_0, strides = var_899_strides_0, weight = layers_3_fc1_inlier_module_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor<string, []>("op_899_cast_fp16")];
            tensor<string, []> var_905_pad_type_0 = const()[name = tensor<string, []>("op_905_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_905_strides_0 = const()[name = tensor<string, []>("op_905_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_905_pad_0 = const()[name = tensor<string, []>("op_905_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_905_dilations_0 = const()[name = tensor<string, []>("op_905_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_905_groups_0 = const()[name = tensor<string, []>("op_905_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_3_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23960640))), name = tensor<string, []>("layers_3_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [28437]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23903680))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_905_cast_fp16 = conv(dilations = var_905_dilations_0, groups = var_905_groups_0, pad = var_905_pad_0, pad_type = var_905_pad_type_0, strides = var_905_strides_0, weight = layers_3_fc1_outlier_module_weight_to_fp16_sparsified, x = input_27_cast_fp16)[name = tensor<string, []>("op_905_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_29_cast_fp16 = add(x = var_899_cast_fp16, y = var_905_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<string, []> input_31_mode_0 = const()[name = tensor<string, []>("input_31_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<string, []> var_916_pad_type_0 = const()[name = tensor<string, []>("op_916_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_916_strides_0 = const()[name = tensor<string, []>("op_916_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_916_pad_0 = const()[name = tensor<string, []>("op_916_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_916_dilations_0 = const()[name = tensor<string, []>("op_916_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_916_groups_0 = const()[name = tensor<string, []>("op_916_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_3_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24255616))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25435328))), name = tensor<string, []>("layers_3_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_3_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25435456)))];
            tensor<fp16, [1, 768, 1, 1500]> var_916_cast_fp16 = conv(bias = layers_3_fc2_inlier_module_bias_to_fp16, dilations = var_916_dilations_0, groups = var_916_groups_0, pad = var_916_pad_0, pad_type = var_916_pad_type_0, strides = var_916_strides_0, weight = layers_3_fc2_inlier_module_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor<string, []>("op_916_cast_fp16")];
            tensor<string, []> var_922_pad_type_0 = const()[name = tensor<string, []>("op_922_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_922_strides_0 = const()[name = tensor<string, []>("op_922_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_922_pad_0 = const()[name = tensor<string, []>("op_922_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_922_dilations_0 = const()[name = tensor<string, []>("op_922_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_922_groups_0 = const()[name = tensor<string, []>("op_922_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_3_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25492672))), name = tensor<string, []>("layers_3_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [27768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25437056))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_922_cast_fp16 = conv(dilations = var_922_dilations_0, groups = var_922_groups_0, pad = var_922_pad_0, pad_type = var_922_pad_type_0, strides = var_922_strides_0, weight = layers_3_fc2_outlier_module_weight_to_fp16_sparsified, x = input_31_cast_fp16)[name = tensor<string, []>("op_922_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_11_cast_fp16 = add(x = var_916_cast_fp16, y = var_922_cast_fp16)[name = tensor<string, []>("hidden_states_11_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
            tensor<int32, []> var_928 = const()[name = tensor<string, []>("op_928"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_950_to_fp16 = const()[name = tensor<string, []>("op_950_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_950_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
            tensor<fp16, [768]> obj_17_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_17_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25787648)))];
            tensor<fp16, [768]> obj_17_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_17_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25789248)))];
            tensor<fp16, []> obj_17_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_17_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("obj_17_cast_fp16")];
            tensor<string, []> var_972_pad_type_0 = const()[name = tensor<string, []>("op_972_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_972_strides_0 = const()[name = tensor<string, []>("op_972_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_972_pad_0 = const()[name = tensor<string, []>("op_972_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_972_dilations_0 = const()[name = tensor<string, []>("op_972_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_972_groups_0 = const()[name = tensor<string, []>("op_972_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25790848))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26085824))), name = tensor<string, []>("layers_4_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_4_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26085952)))];
            tensor<fp16, [1, 768, 1, 1500]> var_972_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_972_dilations_0, groups = var_972_groups_0, pad = var_972_pad_0, pad_type = var_972_pad_type_0, strides = var_972_strides_0, weight = layers_4_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor<string, []>("op_972_cast_fp16")];
            tensor<string, []> var_978_pad_type_0 = const()[name = tensor<string, []>("op_978_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_978_strides_0 = const()[name = tensor<string, []>("op_978_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_978_pad_0 = const()[name = tensor<string, []>("op_978_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_978_dilations_0 = const()[name = tensor<string, []>("op_978_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_978_groups_0 = const()[name = tensor<string, []>("op_978_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26100672))), name = tensor<string, []>("layers_4_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6507]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26087552))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_978_cast_fp16 = conv(dilations = var_978_dilations_0, groups = var_978_groups_0, pad = var_978_pad_0, pad_type = var_978_pad_type_0, strides = var_978_strides_0, weight = layers_4_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_17_cast_fp16)[name = tensor<string, []>("op_978_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_9_cast_fp16 = add(x = var_972_cast_fp16, y = var_978_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
            tensor<string, []> var_987_pad_type_0 = const()[name = tensor<string, []>("op_987_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_987_strides_0 = const()[name = tensor<string, []>("op_987_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_987_pad_0 = const()[name = tensor<string, []>("op_987_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_987_dilations_0 = const()[name = tensor<string, []>("op_987_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_987_groups_0 = const()[name = tensor<string, []>("op_987_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26174464))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26469440))), name = tensor<string, []>("layers_4_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_987_cast_fp16 = conv(dilations = var_987_dilations_0, groups = var_987_groups_0, pad = var_987_pad_0, pad_type = var_987_pad_type_0, strides = var_987_strides_0, weight = layers_4_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor<string, []>("op_987_cast_fp16")];
            tensor<string, []> var_993_pad_type_0 = const()[name = tensor<string, []>("op_993_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_993_strides_0 = const()[name = tensor<string, []>("op_993_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_993_pad_0 = const()[name = tensor<string, []>("op_993_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_993_dilations_0 = const()[name = tensor<string, []>("op_993_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_993_groups_0 = const()[name = tensor<string, []>("op_993_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26482240))), name = tensor<string, []>("layers_4_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6290]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26469568))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_993_cast_fp16 = conv(dilations = var_993_dilations_0, groups = var_993_groups_0, pad = var_993_pad_0, pad_type = var_993_pad_type_0, strides = var_993_strides_0, weight = layers_4_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_17_cast_fp16)[name = tensor<string, []>("op_993_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_9_cast_fp16 = add(x = var_987_cast_fp16, y = var_993_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
            tensor<string, []> var_1003_pad_type_0 = const()[name = tensor<string, []>("op_1003_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1003_strides_0 = const()[name = tensor<string, []>("op_1003_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1003_pad_0 = const()[name = tensor<string, []>("op_1003_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1003_dilations_0 = const()[name = tensor<string, []>("op_1003_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1003_groups_0 = const()[name = tensor<string, []>("op_1003_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26556032))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26851008))), name = tensor<string, []>("layers_4_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_4_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26851136)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1003_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1003_dilations_0, groups = var_1003_groups_0, pad = var_1003_pad_0, pad_type = var_1003_pad_type_0, strides = var_1003_strides_0, weight = layers_4_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor<string, []>("op_1003_cast_fp16")];
            tensor<string, []> var_1009_pad_type_0 = const()[name = tensor<string, []>("op_1009_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1009_strides_0 = const()[name = tensor<string, []>("op_1009_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1009_pad_0 = const()[name = tensor<string, []>("op_1009_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1009_dilations_0 = const()[name = tensor<string, []>("op_1009_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1009_groups_0 = const()[name = tensor<string, []>("op_1009_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26862144))), name = tensor<string, []>("layers_4_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [4665]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26852736))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1009_cast_fp16 = conv(dilations = var_1009_dilations_0, groups = var_1009_groups_0, pad = var_1009_pad_0, pad_type = var_1009_pad_type_0, strides = var_1009_strides_0, weight = layers_4_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_17_cast_fp16)[name = tensor<string, []>("op_1009_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_9_cast_fp16 = add(x = var_1003_cast_fp16, y = var_1009_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
            tensor<int32, [4]> var_1013 = const()[name = tensor<string, []>("op_1013"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_9_cast_fp16 = reshape(shape = var_1013, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")];
            tensor<fp16, []> var_1015_to_fp16 = const()[name = tensor<string, []>("op_1015_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_1016_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_1015_to_fp16)[name = tensor<string, []>("op_1016_cast_fp16")];
            tensor<int32, [4]> var_1019 = const()[name = tensor<string, []>("op_1019"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1020_cast_fp16 = reshape(shape = var_1019, x = key_9_cast_fp16)[name = tensor<string, []>("op_1020_cast_fp16")];
            tensor<bool, []> mh_w_9_transpose_x_0 = const()[name = tensor<string, []>("mh_w_9_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_9_transpose_y_0 = const()[name = tensor<string, []>("mh_w_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_1016_cast_fp16, y = var_1020_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_1023_cast_fp16 = softmax(axis = var_928, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_1023_cast_fp16")];
            tensor<int32, [4]> var_1024 = const()[name = tensor<string, []>("op_1024"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1025_cast_fp16 = reshape(shape = var_1024, x = value_9_cast_fp16)[name = tensor<string, []>("op_1025_cast_fp16")];
            tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_1025_cast_fp16, y = var_1023_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
            tensor<int32, [4]> var_1028 = const()[name = tensor<string, []>("op_1028"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_33_cast_fp16 = reshape(shape = var_1028, x = attn_9_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<string, []> var_1038_pad_type_0 = const()[name = tensor<string, []>("op_1038_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1038_strides_0 = const()[name = tensor<string, []>("op_1038_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1038_pad_0 = const()[name = tensor<string, []>("op_1038_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1038_dilations_0 = const()[name = tensor<string, []>("op_1038_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1038_groups_0 = const()[name = tensor<string, []>("op_1038_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26935936))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27230912))), name = tensor<string, []>("layers_4_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_4_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27231040)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1038_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1038_dilations_0, groups = var_1038_groups_0, pad = var_1038_pad_0, pad_type = var_1038_pad_type_0, strides = var_1038_strides_0, weight = layers_4_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = tensor<string, []>("op_1038_cast_fp16")];
            tensor<string, []> var_1044_pad_type_0 = const()[name = tensor<string, []>("op_1044_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1044_strides_0 = const()[name = tensor<string, []>("op_1044_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1044_pad_0 = const()[name = tensor<string, []>("op_1044_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1044_dilations_0 = const()[name = tensor<string, []>("op_1044_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1044_groups_0 = const()[name = tensor<string, []>("op_1044_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27241344))), name = tensor<string, []>("layers_4_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [4306]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27232640))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1044_cast_fp16 = conv(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_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_33_cast_fp16)[name = tensor<string, []>("op_1044_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_19_cast_fp16 = add(x = var_1038_cast_fp16, y = var_1044_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = obj_19_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1055_to_fp16 = const()[name = tensor<string, []>("op_1055_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1055_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
            tensor<fp16, [768]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27315136)))];
            tensor<fp16, [768]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27316736)))];
            tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<string, []> var_1073_pad_type_0 = const()[name = tensor<string, []>("op_1073_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1073_strides_0 = const()[name = tensor<string, []>("op_1073_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1073_pad_0 = const()[name = tensor<string, []>("op_1073_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1073_dilations_0 = const()[name = tensor<string, []>("op_1073_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1073_groups_0 = const()[name = tensor<string, []>("op_1073_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_4_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27318336))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28498048))), name = tensor<string, []>("layers_4_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_4_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28498176)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_1073_cast_fp16 = conv(bias = layers_4_fc1_inlier_module_bias_to_fp16, dilations = var_1073_dilations_0, groups = var_1073_groups_0, pad = var_1073_pad_0, pad_type = var_1073_pad_type_0, strides = var_1073_strides_0, weight = layers_4_fc1_inlier_module_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor<string, []>("op_1073_cast_fp16")];
            tensor<string, []> var_1079_pad_type_0 = const()[name = tensor<string, []>("op_1079_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1079_strides_0 = const()[name = tensor<string, []>("op_1079_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1079_pad_0 = const()[name = tensor<string, []>("op_1079_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1079_dilations_0 = const()[name = tensor<string, []>("op_1079_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1079_groups_0 = const()[name = tensor<string, []>("op_1079_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_4_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28545344))), name = tensor<string, []>("layers_4_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [20439]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28504384))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_1079_cast_fp16 = conv(dilations = var_1079_dilations_0, groups = var_1079_groups_0, pad = var_1079_pad_0, pad_type = var_1079_pad_type_0, strides = var_1079_strides_0, weight = layers_4_fc1_outlier_module_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = tensor<string, []>("op_1079_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_37_cast_fp16 = add(x = var_1073_cast_fp16, y = var_1079_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<string, []> input_39_mode_0 = const()[name = tensor<string, []>("input_39_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<string, []> var_1090_pad_type_0 = const()[name = tensor<string, []>("op_1090_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1090_strides_0 = const()[name = tensor<string, []>("op_1090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1090_pad_0 = const()[name = tensor<string, []>("op_1090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1090_dilations_0 = const()[name = tensor<string, []>("op_1090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1090_groups_0 = const()[name = tensor<string, []>("op_1090_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_4_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28840320))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30020032))), name = tensor<string, []>("layers_4_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_4_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30020160)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1090_cast_fp16 = conv(bias = layers_4_fc2_inlier_module_bias_to_fp16, dilations = var_1090_dilations_0, groups = var_1090_groups_0, pad = var_1090_pad_0, pad_type = var_1090_pad_type_0, strides = var_1090_strides_0, weight = layers_4_fc2_inlier_module_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = tensor<string, []>("op_1090_cast_fp16")];
            tensor<string, []> var_1096_pad_type_0 = const()[name = tensor<string, []>("op_1096_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1096_strides_0 = const()[name = tensor<string, []>("op_1096_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1096_pad_0 = const()[name = tensor<string, []>("op_1096_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1096_dilations_0 = const()[name = tensor<string, []>("op_1096_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1096_groups_0 = const()[name = tensor<string, []>("op_1096_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_4_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30065024))), name = tensor<string, []>("layers_4_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [21573]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30021760))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1096_cast_fp16 = conv(dilations = var_1096_dilations_0, groups = var_1096_groups_0, pad = var_1096_pad_0, pad_type = var_1096_pad_type_0, strides = var_1096_strides_0, weight = layers_4_fc2_outlier_module_weight_to_fp16_sparsified, x = input_39_cast_fp16)[name = tensor<string, []>("op_1096_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_13_cast_fp16 = add(x = var_1090_cast_fp16, y = var_1096_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
            tensor<int32, []> var_1102 = const()[name = tensor<string, []>("op_1102"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1124_to_fp16 = const()[name = tensor<string, []>("op_1124_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1124_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
            tensor<fp16, [768]> obj_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_21_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30360000)))];
            tensor<fp16, [768]> obj_21_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_21_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30361600)))];
            tensor<fp16, []> obj_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
            tensor<string, []> var_1146_pad_type_0 = const()[name = tensor<string, []>("op_1146_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1146_strides_0 = const()[name = tensor<string, []>("op_1146_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1146_pad_0 = const()[name = tensor<string, []>("op_1146_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1146_dilations_0 = const()[name = tensor<string, []>("op_1146_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1146_groups_0 = const()[name = tensor<string, []>("op_1146_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30363200))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30658176))), name = tensor<string, []>("layers_5_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_5_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30658304)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1146_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1146_dilations_0, groups = var_1146_groups_0, pad = var_1146_pad_0, pad_type = var_1146_pad_type_0, strides = var_1146_strides_0, weight = layers_5_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor<string, []>("op_1146_cast_fp16")];
            tensor<string, []> var_1152_pad_type_0 = const()[name = tensor<string, []>("op_1152_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1152_strides_0 = const()[name = tensor<string, []>("op_1152_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1152_pad_0 = const()[name = tensor<string, []>("op_1152_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1152_dilations_0 = const()[name = tensor<string, []>("op_1152_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1152_groups_0 = const()[name = tensor<string, []>("op_1152_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30670336))), name = tensor<string, []>("layers_5_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30659904))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1152_cast_fp16 = conv(dilations = var_1152_dilations_0, groups = var_1152_groups_0, pad = var_1152_pad_0, pad_type = var_1152_pad_type_0, strides = var_1152_strides_0, weight = layers_5_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_21_cast_fp16)[name = tensor<string, []>("op_1152_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_11_cast_fp16 = add(x = var_1146_cast_fp16, y = var_1152_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
            tensor<string, []> var_1161_pad_type_0 = const()[name = tensor<string, []>("op_1161_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1161_strides_0 = const()[name = tensor<string, []>("op_1161_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1161_pad_0 = const()[name = tensor<string, []>("op_1161_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1161_dilations_0 = const()[name = tensor<string, []>("op_1161_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1161_groups_0 = const()[name = tensor<string, []>("op_1161_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30744128))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31039104))), name = tensor<string, []>("layers_5_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1161_cast_fp16 = conv(dilations = var_1161_dilations_0, groups = var_1161_groups_0, pad = var_1161_pad_0, pad_type = var_1161_pad_type_0, strides = var_1161_strides_0, weight = layers_5_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor<string, []>("op_1161_cast_fp16")];
            tensor<string, []> var_1167_pad_type_0 = const()[name = tensor<string, []>("op_1167_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1167_strides_0 = const()[name = tensor<string, []>("op_1167_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1167_pad_0 = const()[name = tensor<string, []>("op_1167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1167_dilations_0 = const()[name = tensor<string, []>("op_1167_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1167_groups_0 = const()[name = tensor<string, []>("op_1167_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31048960))), name = tensor<string, []>("layers_5_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [4811]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31039232))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1167_cast_fp16 = conv(dilations = var_1167_dilations_0, groups = var_1167_groups_0, pad = var_1167_pad_0, pad_type = var_1167_pad_type_0, strides = var_1167_strides_0, weight = layers_5_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_21_cast_fp16)[name = tensor<string, []>("op_1167_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_11_cast_fp16 = add(x = var_1161_cast_fp16, y = var_1167_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
            tensor<string, []> var_1177_pad_type_0 = const()[name = tensor<string, []>("op_1177_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1177_strides_0 = const()[name = tensor<string, []>("op_1177_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1177_pad_0 = const()[name = tensor<string, []>("op_1177_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1177_dilations_0 = const()[name = tensor<string, []>("op_1177_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1177_groups_0 = const()[name = tensor<string, []>("op_1177_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31122752))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31417728))), name = tensor<string, []>("layers_5_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_5_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31417856)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1177_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1177_dilations_0, groups = var_1177_groups_0, pad = var_1177_pad_0, pad_type = var_1177_pad_type_0, strides = var_1177_strides_0, weight = layers_5_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor<string, []>("op_1177_cast_fp16")];
            tensor<string, []> var_1183_pad_type_0 = const()[name = tensor<string, []>("op_1183_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1183_strides_0 = const()[name = tensor<string, []>("op_1183_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1183_pad_0 = const()[name = tensor<string, []>("op_1183_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1183_dilations_0 = const()[name = tensor<string, []>("op_1183_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1183_groups_0 = const()[name = tensor<string, []>("op_1183_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31426368))), name = tensor<string, []>("layers_5_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3420]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31419456))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1183_cast_fp16 = conv(dilations = var_1183_dilations_0, groups = var_1183_groups_0, pad = var_1183_pad_0, pad_type = var_1183_pad_type_0, strides = var_1183_strides_0, weight = layers_5_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_21_cast_fp16)[name = tensor<string, []>("op_1183_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_11_cast_fp16 = add(x = var_1177_cast_fp16, y = var_1183_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")];
            tensor<int32, [4]> var_1187 = const()[name = tensor<string, []>("op_1187"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_11_cast_fp16 = reshape(shape = var_1187, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")];
            tensor<fp16, []> var_1189_to_fp16 = const()[name = tensor<string, []>("op_1189_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_1190_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1189_to_fp16)[name = tensor<string, []>("op_1190_cast_fp16")];
            tensor<int32, [4]> var_1193 = const()[name = tensor<string, []>("op_1193"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1194_cast_fp16 = reshape(shape = var_1193, x = key_11_cast_fp16)[name = tensor<string, []>("op_1194_cast_fp16")];
            tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_1190_cast_fp16, y = var_1194_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_1197_cast_fp16 = softmax(axis = var_1102, x = mh_w_11_cast_fp16)[name = tensor<string, []>("op_1197_cast_fp16")];
            tensor<int32, [4]> var_1198 = const()[name = tensor<string, []>("op_1198"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1199_cast_fp16 = reshape(shape = var_1198, x = value_11_cast_fp16)[name = tensor<string, []>("op_1199_cast_fp16")];
            tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1199_cast_fp16, y = var_1197_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
            tensor<int32, [4]> var_1202 = const()[name = tensor<string, []>("op_1202"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_41_cast_fp16 = reshape(shape = var_1202, x = attn_11_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<string, []> var_1212_pad_type_0 = const()[name = tensor<string, []>("op_1212_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1212_strides_0 = const()[name = tensor<string, []>("op_1212_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1212_pad_0 = const()[name = tensor<string, []>("op_1212_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1212_dilations_0 = const()[name = tensor<string, []>("op_1212_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1212_groups_0 = const()[name = tensor<string, []>("op_1212_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31500160))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31795136))), name = tensor<string, []>("layers_5_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_5_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31795264)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1212_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1212_dilations_0, groups = var_1212_groups_0, pad = var_1212_pad_0, pad_type = var_1212_pad_type_0, strides = var_1212_strides_0, weight = layers_5_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = tensor<string, []>("op_1212_cast_fp16")];
            tensor<string, []> var_1218_pad_type_0 = const()[name = tensor<string, []>("op_1218_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1218_strides_0 = const()[name = tensor<string, []>("op_1218_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1218_pad_0 = const()[name = tensor<string, []>("op_1218_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1218_dilations_0 = const()[name = tensor<string, []>("op_1218_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1218_groups_0 = const()[name = tensor<string, []>("op_1218_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31804736))), name = tensor<string, []>("layers_5_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3878]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31796864))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1218_cast_fp16 = conv(dilations = var_1218_dilations_0, groups = var_1218_groups_0, pad = var_1218_pad_0, pad_type = var_1218_pad_type_0, strides = var_1218_strides_0, weight = layers_5_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_41_cast_fp16)[name = tensor<string, []>("op_1218_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_23_cast_fp16 = add(x = var_1212_cast_fp16, y = var_1218_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_23_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1229_to_fp16 = const()[name = tensor<string, []>("op_1229_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_1229_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
            tensor<fp16, [768]> input_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_43_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31878528)))];
            tensor<fp16, [768]> input_43_beta_0_to_fp16 = const()[name = tensor<string, []>("input_43_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31880128)))];
            tensor<fp16, []> input_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<string, []> var_1247_pad_type_0 = const()[name = tensor<string, []>("op_1247_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1247_strides_0 = const()[name = tensor<string, []>("op_1247_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1247_pad_0 = const()[name = tensor<string, []>("op_1247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1247_dilations_0 = const()[name = tensor<string, []>("op_1247_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1247_groups_0 = const()[name = tensor<string, []>("op_1247_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_5_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31881728))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33061440))), name = tensor<string, []>("layers_5_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_5_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33061568)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_1247_cast_fp16 = conv(bias = layers_5_fc1_inlier_module_bias_to_fp16, dilations = var_1247_dilations_0, groups = var_1247_groups_0, pad = var_1247_pad_0, pad_type = var_1247_pad_type_0, strides = var_1247_strides_0, weight = layers_5_fc1_inlier_module_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor<string, []>("op_1247_cast_fp16")];
            tensor<string, []> var_1253_pad_type_0 = const()[name = tensor<string, []>("op_1253_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1253_strides_0 = const()[name = tensor<string, []>("op_1253_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1253_pad_0 = const()[name = tensor<string, []>("op_1253_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1253_dilations_0 = const()[name = tensor<string, []>("op_1253_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1253_groups_0 = const()[name = tensor<string, []>("op_1253_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_5_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33101248))), name = tensor<string, []>("layers_5_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [16704]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33067776))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_1253_cast_fp16 = conv(dilations = var_1253_dilations_0, groups = var_1253_groups_0, pad = var_1253_pad_0, pad_type = var_1253_pad_type_0, strides = var_1253_strides_0, weight = layers_5_fc1_outlier_module_weight_to_fp16_sparsified, x = input_43_cast_fp16)[name = tensor<string, []>("op_1253_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_45_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1253_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<string, []> input_47_mode_0 = const()[name = tensor<string, []>("input_47_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<string, []> var_1264_pad_type_0 = const()[name = tensor<string, []>("op_1264_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1264_strides_0 = const()[name = tensor<string, []>("op_1264_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1264_pad_0 = const()[name = tensor<string, []>("op_1264_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1264_dilations_0 = const()[name = tensor<string, []>("op_1264_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1264_groups_0 = const()[name = tensor<string, []>("op_1264_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_5_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33396224))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34575936))), name = tensor<string, []>("layers_5_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_5_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34576064)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1264_cast_fp16 = conv(bias = layers_5_fc2_inlier_module_bias_to_fp16, dilations = var_1264_dilations_0, groups = var_1264_groups_0, pad = var_1264_pad_0, pad_type = var_1264_pad_type_0, strides = var_1264_strides_0, weight = layers_5_fc2_inlier_module_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor<string, []>("op_1264_cast_fp16")];
            tensor<string, []> var_1270_pad_type_0 = const()[name = tensor<string, []>("op_1270_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1270_strides_0 = const()[name = tensor<string, []>("op_1270_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1270_pad_0 = const()[name = tensor<string, []>("op_1270_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1270_dilations_0 = const()[name = tensor<string, []>("op_1270_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1270_groups_0 = const()[name = tensor<string, []>("op_1270_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_5_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34614208))), name = tensor<string, []>("layers_5_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [18230]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34577664))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1270_cast_fp16 = conv(dilations = var_1270_dilations_0, groups = var_1270_groups_0, pad = var_1270_pad_0, pad_type = var_1270_pad_type_0, strides = var_1270_strides_0, weight = layers_5_fc2_outlier_module_weight_to_fp16_sparsified, x = input_47_cast_fp16)[name = tensor<string, []>("op_1270_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_15_cast_fp16 = add(x = var_1264_cast_fp16, y = var_1270_cast_fp16)[name = tensor<string, []>("hidden_states_15_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor<string, []>("inputs_25_cast_fp16")];
            tensor<int32, []> var_1276 = const()[name = tensor<string, []>("op_1276"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_25_axes_0 = const()[name = tensor<string, []>("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1298_to_fp16 = const()[name = tensor<string, []>("op_1298_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_1298_to_fp16, x = inputs_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
            tensor<fp16, [768]> obj_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34909184)))];
            tensor<fp16, [768]> obj_25_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_25_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34910784)))];
            tensor<fp16, []> obj_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
            tensor<string, []> var_1320_pad_type_0 = const()[name = tensor<string, []>("op_1320_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1320_strides_0 = const()[name = tensor<string, []>("op_1320_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1320_pad_0 = const()[name = tensor<string, []>("op_1320_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1320_dilations_0 = const()[name = tensor<string, []>("op_1320_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1320_groups_0 = const()[name = tensor<string, []>("op_1320_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34912384))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35207360))), name = tensor<string, []>("layers_6_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_6_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35207488)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1320_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1320_dilations_0, groups = var_1320_groups_0, pad = var_1320_pad_0, pad_type = var_1320_pad_type_0, strides = var_1320_strides_0, weight = layers_6_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("op_1320_cast_fp16")];
            tensor<string, []> var_1326_pad_type_0 = const()[name = tensor<string, []>("op_1326_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1326_strides_0 = const()[name = tensor<string, []>("op_1326_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1326_pad_0 = const()[name = tensor<string, []>("op_1326_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1326_dilations_0 = const()[name = tensor<string, []>("op_1326_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1326_groups_0 = const()[name = tensor<string, []>("op_1326_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35217280))), name = tensor<string, []>("layers_6_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [4054]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35209088))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1326_cast_fp16 = conv(dilations = var_1326_dilations_0, groups = var_1326_groups_0, pad = var_1326_pad_0, pad_type = var_1326_pad_type_0, strides = var_1326_strides_0, weight = layers_6_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_25_cast_fp16)[name = tensor<string, []>("op_1326_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_13_cast_fp16 = add(x = var_1320_cast_fp16, y = var_1326_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
            tensor<string, []> var_1335_pad_type_0 = const()[name = tensor<string, []>("op_1335_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1335_strides_0 = const()[name = tensor<string, []>("op_1335_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1335_pad_0 = const()[name = tensor<string, []>("op_1335_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1335_dilations_0 = const()[name = tensor<string, []>("op_1335_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1335_groups_0 = const()[name = tensor<string, []>("op_1335_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35291072))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35586048))), name = tensor<string, []>("layers_6_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1335_cast_fp16 = conv(dilations = var_1335_dilations_0, groups = var_1335_groups_0, pad = var_1335_pad_0, pad_type = var_1335_pad_type_0, strides = var_1335_strides_0, weight = layers_6_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("op_1335_cast_fp16")];
            tensor<string, []> var_1341_pad_type_0 = const()[name = tensor<string, []>("op_1341_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1341_strides_0 = const()[name = tensor<string, []>("op_1341_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1341_pad_0 = const()[name = tensor<string, []>("op_1341_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1341_dilations_0 = const()[name = tensor<string, []>("op_1341_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1341_groups_0 = const()[name = tensor<string, []>("op_1341_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35593984))), name = tensor<string, []>("layers_6_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3869]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35586176))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1341_cast_fp16 = conv(dilations = var_1341_dilations_0, groups = var_1341_groups_0, pad = var_1341_pad_0, pad_type = var_1341_pad_type_0, strides = var_1341_strides_0, weight = layers_6_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_25_cast_fp16)[name = tensor<string, []>("op_1341_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_13_cast_fp16 = add(x = var_1335_cast_fp16, y = var_1341_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
            tensor<string, []> var_1351_pad_type_0 = const()[name = tensor<string, []>("op_1351_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1351_strides_0 = const()[name = tensor<string, []>("op_1351_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1351_pad_0 = const()[name = tensor<string, []>("op_1351_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1351_dilations_0 = const()[name = tensor<string, []>("op_1351_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1351_groups_0 = const()[name = tensor<string, []>("op_1351_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35667776))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35962752))), name = tensor<string, []>("layers_6_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_6_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35962880)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1351_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1351_dilations_0, groups = var_1351_groups_0, pad = var_1351_pad_0, pad_type = var_1351_pad_type_0, strides = var_1351_strides_0, weight = layers_6_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("op_1351_cast_fp16")];
            tensor<string, []> var_1357_pad_type_0 = const()[name = tensor<string, []>("op_1357_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1357_strides_0 = const()[name = tensor<string, []>("op_1357_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1357_pad_0 = const()[name = tensor<string, []>("op_1357_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1357_dilations_0 = const()[name = tensor<string, []>("op_1357_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1357_groups_0 = const()[name = tensor<string, []>("op_1357_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35970496))), name = tensor<string, []>("layers_6_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2961]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35964480))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1357_cast_fp16 = conv(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_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_25_cast_fp16)[name = tensor<string, []>("op_1357_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_13_cast_fp16 = add(x = var_1351_cast_fp16, y = var_1357_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
            tensor<int32, [4]> var_1361 = const()[name = tensor<string, []>("op_1361"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_13_cast_fp16 = reshape(shape = var_1361, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")];
            tensor<fp16, []> var_1363_to_fp16 = const()[name = tensor<string, []>("op_1363_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_1364_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1363_to_fp16)[name = tensor<string, []>("op_1364_cast_fp16")];
            tensor<int32, [4]> var_1367 = const()[name = tensor<string, []>("op_1367"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1368_cast_fp16 = reshape(shape = var_1367, x = key_13_cast_fp16)[name = tensor<string, []>("op_1368_cast_fp16")];
            tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_1364_cast_fp16, y = var_1368_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_1371_cast_fp16 = softmax(axis = var_1276, x = mh_w_13_cast_fp16)[name = tensor<string, []>("op_1371_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = tensor<string, []>("op_1372"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1373_cast_fp16 = reshape(shape = var_1372, x = value_13_cast_fp16)[name = tensor<string, []>("op_1373_cast_fp16")];
            tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1373_cast_fp16, y = var_1371_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
            tensor<int32, [4]> var_1376 = const()[name = tensor<string, []>("op_1376"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_49_cast_fp16 = reshape(shape = var_1376, x = attn_13_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<string, []> var_1386_pad_type_0 = const()[name = tensor<string, []>("op_1386_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1386_strides_0 = const()[name = tensor<string, []>("op_1386_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1386_pad_0 = const()[name = tensor<string, []>("op_1386_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1386_dilations_0 = const()[name = tensor<string, []>("op_1386_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1386_groups_0 = const()[name = tensor<string, []>("op_1386_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36044288))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36339264))), name = tensor<string, []>("layers_6_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_6_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36339392)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1386_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1386_dilations_0, groups = var_1386_groups_0, pad = var_1386_pad_0, pad_type = var_1386_pad_type_0, strides = var_1386_strides_0, weight = layers_6_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("op_1386_cast_fp16")];
            tensor<string, []> var_1392_pad_type_0 = const()[name = tensor<string, []>("op_1392_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1392_strides_0 = const()[name = tensor<string, []>("op_1392_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1392_pad_0 = const()[name = tensor<string, []>("op_1392_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1392_dilations_0 = const()[name = tensor<string, []>("op_1392_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1392_groups_0 = const()[name = tensor<string, []>("op_1392_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36346496))), name = tensor<string, []>("layers_6_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2712]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36340992))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1392_cast_fp16 = conv(dilations = var_1392_dilations_0, groups = var_1392_groups_0, pad = var_1392_pad_0, pad_type = var_1392_pad_type_0, strides = var_1392_strides_0, weight = layers_6_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_49_cast_fp16)[name = tensor<string, []>("op_1392_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_27_cast_fp16 = add(x = var_1386_cast_fp16, y = var_1392_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("inputs_27_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = tensor<string, []>("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1403_to_fp16 = const()[name = tensor<string, []>("op_1403_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1403_to_fp16, x = inputs_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
            tensor<fp16, [768]> input_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_51_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36420288)))];
            tensor<fp16, [768]> input_51_beta_0_to_fp16 = const()[name = tensor<string, []>("input_51_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36421888)))];
            tensor<fp16, []> input_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<string, []> var_1421_pad_type_0 = const()[name = tensor<string, []>("op_1421_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1421_strides_0 = const()[name = tensor<string, []>("op_1421_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1421_pad_0 = const()[name = tensor<string, []>("op_1421_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1421_dilations_0 = const()[name = tensor<string, []>("op_1421_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1421_groups_0 = const()[name = tensor<string, []>("op_1421_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_6_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36423488))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37603200))), name = tensor<string, []>("layers_6_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_6_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37603328)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_1421_cast_fp16 = conv(bias = layers_6_fc1_inlier_module_bias_to_fp16, dilations = var_1421_dilations_0, groups = var_1421_groups_0, pad = var_1421_pad_0, pad_type = var_1421_pad_type_0, strides = var_1421_strides_0, weight = layers_6_fc1_inlier_module_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor<string, []>("op_1421_cast_fp16")];
            tensor<string, []> var_1427_pad_type_0 = const()[name = tensor<string, []>("op_1427_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1427_strides_0 = const()[name = tensor<string, []>("op_1427_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1427_pad_0 = const()[name = tensor<string, []>("op_1427_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1427_dilations_0 = const()[name = tensor<string, []>("op_1427_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1427_groups_0 = const()[name = tensor<string, []>("op_1427_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_6_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37638976))), name = tensor<string, []>("layers_6_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [14658]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37609536))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_1427_cast_fp16 = conv(dilations = var_1427_dilations_0, groups = var_1427_groups_0, pad = var_1427_pad_0, pad_type = var_1427_pad_type_0, strides = var_1427_strides_0, weight = layers_6_fc1_outlier_module_weight_to_fp16_sparsified, x = input_51_cast_fp16)[name = tensor<string, []>("op_1427_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_53_cast_fp16 = add(x = var_1421_cast_fp16, y = var_1427_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<string, []> input_55_mode_0 = const()[name = tensor<string, []>("input_55_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = input_53_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<string, []> var_1438_pad_type_0 = const()[name = tensor<string, []>("op_1438_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1438_strides_0 = const()[name = tensor<string, []>("op_1438_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1438_pad_0 = const()[name = tensor<string, []>("op_1438_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1438_dilations_0 = const()[name = tensor<string, []>("op_1438_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1438_groups_0 = const()[name = tensor<string, []>("op_1438_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_6_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37933952))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39113664))), name = tensor<string, []>("layers_6_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_6_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39113792)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1438_cast_fp16 = conv(bias = layers_6_fc2_inlier_module_bias_to_fp16, dilations = var_1438_dilations_0, groups = var_1438_groups_0, pad = var_1438_pad_0, pad_type = var_1438_pad_type_0, strides = var_1438_strides_0, weight = layers_6_fc2_inlier_module_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = tensor<string, []>("op_1438_cast_fp16")];
            tensor<string, []> var_1444_pad_type_0 = const()[name = tensor<string, []>("op_1444_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1444_strides_0 = const()[name = tensor<string, []>("op_1444_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1444_pad_0 = const()[name = tensor<string, []>("op_1444_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1444_dilations_0 = const()[name = tensor<string, []>("op_1444_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1444_groups_0 = const()[name = tensor<string, []>("op_1444_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_6_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39142720))), name = tensor<string, []>("layers_6_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [13601]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39115392))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1444_cast_fp16 = conv(dilations = var_1444_dilations_0, groups = var_1444_groups_0, pad = var_1444_pad_0, pad_type = var_1444_pad_type_0, strides = var_1444_strides_0, weight = layers_6_fc2_outlier_module_weight_to_fp16_sparsified, x = input_55_cast_fp16)[name = tensor<string, []>("op_1444_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_17_cast_fp16 = add(x = var_1438_cast_fp16, y = var_1444_cast_fp16)[name = tensor<string, []>("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor<string, []>("inputs_29_cast_fp16")];
            tensor<int32, []> var_1450 = const()[name = tensor<string, []>("op_1450"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_29_axes_0 = const()[name = tensor<string, []>("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1472_to_fp16 = const()[name = tensor<string, []>("op_1472_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1472_to_fp16, x = inputs_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
            tensor<fp16, [768]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39437696)))];
            tensor<fp16, [768]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39439296)))];
            tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
            tensor<string, []> var_1494_pad_type_0 = const()[name = tensor<string, []>("op_1494_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1494_strides_0 = const()[name = tensor<string, []>("op_1494_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1494_pad_0 = const()[name = tensor<string, []>("op_1494_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1494_dilations_0 = const()[name = tensor<string, []>("op_1494_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1494_groups_0 = const()[name = tensor<string, []>("op_1494_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39440896))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39735872))), name = tensor<string, []>("layers_7_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_7_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39736000)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1494_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1494_dilations_0, groups = var_1494_groups_0, pad = var_1494_pad_0, pad_type = var_1494_pad_type_0, strides = var_1494_strides_0, weight = layers_7_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("op_1494_cast_fp16")];
            tensor<string, []> var_1500_pad_type_0 = const()[name = tensor<string, []>("op_1500_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1500_strides_0 = const()[name = tensor<string, []>("op_1500_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1500_pad_0 = const()[name = tensor<string, []>("op_1500_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1500_dilations_0 = const()[name = tensor<string, []>("op_1500_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1500_groups_0 = const()[name = tensor<string, []>("op_1500_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39745536))), name = tensor<string, []>("layers_7_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3933]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39737600))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1500_cast_fp16 = conv(dilations = var_1500_dilations_0, groups = var_1500_groups_0, pad = var_1500_pad_0, pad_type = var_1500_pad_type_0, strides = var_1500_strides_0, weight = layers_7_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor<string, []>("op_1500_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_15_cast_fp16 = add(x = var_1494_cast_fp16, y = var_1500_cast_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
            tensor<string, []> var_1509_pad_type_0 = const()[name = tensor<string, []>("op_1509_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1509_strides_0 = const()[name = tensor<string, []>("op_1509_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1509_pad_0 = const()[name = tensor<string, []>("op_1509_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1509_dilations_0 = const()[name = tensor<string, []>("op_1509_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1509_groups_0 = const()[name = tensor<string, []>("op_1509_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39819328))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40114304))), name = tensor<string, []>("layers_7_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1509_cast_fp16 = conv(dilations = var_1509_dilations_0, groups = var_1509_groups_0, pad = var_1509_pad_0, pad_type = var_1509_pad_type_0, strides = var_1509_strides_0, weight = layers_7_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("op_1509_cast_fp16")];
            tensor<string, []> var_1515_pad_type_0 = const()[name = tensor<string, []>("op_1515_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1515_strides_0 = const()[name = tensor<string, []>("op_1515_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1515_pad_0 = const()[name = tensor<string, []>("op_1515_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1515_dilations_0 = const()[name = tensor<string, []>("op_1515_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1515_groups_0 = const()[name = tensor<string, []>("op_1515_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40123840))), name = tensor<string, []>("layers_7_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [4657]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40114432))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1515_cast_fp16 = conv(dilations = var_1515_dilations_0, groups = var_1515_groups_0, pad = var_1515_pad_0, pad_type = var_1515_pad_type_0, strides = var_1515_strides_0, weight = layers_7_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor<string, []>("op_1515_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_15_cast_fp16 = add(x = var_1509_cast_fp16, y = var_1515_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
            tensor<string, []> var_1525_pad_type_0 = const()[name = tensor<string, []>("op_1525_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1525_strides_0 = const()[name = tensor<string, []>("op_1525_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1525_pad_0 = const()[name = tensor<string, []>("op_1525_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1525_dilations_0 = const()[name = tensor<string, []>("op_1525_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1525_groups_0 = const()[name = tensor<string, []>("op_1525_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40197632))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40492608))), name = tensor<string, []>("layers_7_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_7_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40492736)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1525_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1525_dilations_0, groups = var_1525_groups_0, pad = var_1525_pad_0, pad_type = var_1525_pad_type_0, strides = var_1525_strides_0, weight = layers_7_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("op_1525_cast_fp16")];
            tensor<string, []> var_1531_pad_type_0 = const()[name = tensor<string, []>("op_1531_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1531_strides_0 = const()[name = tensor<string, []>("op_1531_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1531_pad_0 = const()[name = tensor<string, []>("op_1531_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1531_dilations_0 = const()[name = tensor<string, []>("op_1531_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1531_groups_0 = const()[name = tensor<string, []>("op_1531_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40503360))), name = tensor<string, []>("layers_7_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [4463]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40494336))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1531_cast_fp16 = conv(dilations = var_1531_dilations_0, groups = var_1531_groups_0, pad = var_1531_pad_0, pad_type = var_1531_pad_type_0, strides = var_1531_strides_0, weight = layers_7_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor<string, []>("op_1531_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_15_cast_fp16 = add(x = var_1525_cast_fp16, y = var_1531_cast_fp16)[name = tensor<string, []>("value_15_cast_fp16")];
            tensor<int32, [4]> var_1535 = const()[name = tensor<string, []>("op_1535"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_15_cast_fp16 = reshape(shape = var_1535, x = query_15_cast_fp16)[name = tensor<string, []>("mh_q_15_cast_fp16")];
            tensor<fp16, []> var_1537_to_fp16 = const()[name = tensor<string, []>("op_1537_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_1538_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_1537_to_fp16)[name = tensor<string, []>("op_1538_cast_fp16")];
            tensor<int32, [4]> var_1541 = const()[name = tensor<string, []>("op_1541"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1542_cast_fp16 = reshape(shape = var_1541, x = key_15_cast_fp16)[name = tensor<string, []>("op_1542_cast_fp16")];
            tensor<bool, []> mh_w_15_transpose_x_0 = const()[name = tensor<string, []>("mh_w_15_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_15_transpose_y_0 = const()[name = tensor<string, []>("mh_w_15_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_1538_cast_fp16, y = var_1542_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_1545_cast_fp16 = softmax(axis = var_1450, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_1545_cast_fp16")];
            tensor<int32, [4]> var_1546 = const()[name = tensor<string, []>("op_1546"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1547_cast_fp16 = reshape(shape = var_1546, x = value_15_cast_fp16)[name = tensor<string, []>("op_1547_cast_fp16")];
            tensor<bool, []> attn_15_transpose_x_0 = const()[name = tensor<string, []>("attn_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_15_transpose_y_0 = const()[name = tensor<string, []>("attn_15_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_1547_cast_fp16, y = var_1545_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
            tensor<int32, [4]> var_1550 = const()[name = tensor<string, []>("op_1550"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_57_cast_fp16 = reshape(shape = var_1550, x = attn_15_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<string, []> var_1560_pad_type_0 = const()[name = tensor<string, []>("op_1560_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1560_strides_0 = const()[name = tensor<string, []>("op_1560_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1560_pad_0 = const()[name = tensor<string, []>("op_1560_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1560_dilations_0 = const()[name = tensor<string, []>("op_1560_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1560_groups_0 = const()[name = tensor<string, []>("op_1560_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40577152))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40872128))), name = tensor<string, []>("layers_7_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_7_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40872256)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1560_cast_fp16 = conv(bias = layers_7_self_attn_o_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_o_proj_inlier_module_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor<string, []>("op_1560_cast_fp16")];
            tensor<string, []> var_1566_pad_type_0 = const()[name = tensor<string, []>("op_1566_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1566_strides_0 = const()[name = tensor<string, []>("op_1566_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1566_pad_0 = const()[name = tensor<string, []>("op_1566_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1566_dilations_0 = const()[name = tensor<string, []>("op_1566_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1566_groups_0 = const()[name = tensor<string, []>("op_1566_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40885568))), name = tensor<string, []>("layers_7_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5795]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40873856))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 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_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_57_cast_fp16)[name = tensor<string, []>("op_1566_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_31_cast_fp16 = add(x = var_1560_cast_fp16, y = var_1566_cast_fp16)[name = tensor<string, []>("obj_31_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = obj_31_cast_fp16)[name = tensor<string, []>("inputs_31_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = tensor<string, []>("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1577_to_fp16 = const()[name = tensor<string, []>("op_1577_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1577_to_fp16, x = inputs_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
            tensor<fp16, [768]> input_59_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_59_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40959360)))];
            tensor<fp16, [768]> input_59_beta_0_to_fp16 = const()[name = tensor<string, []>("input_59_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40960960)))];
            tensor<fp16, []> input_59_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_59_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<string, []> var_1595_pad_type_0 = const()[name = tensor<string, []>("op_1595_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1595_strides_0 = const()[name = tensor<string, []>("op_1595_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1595_pad_0 = const()[name = tensor<string, []>("op_1595_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1595_dilations_0 = const()[name = tensor<string, []>("op_1595_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1595_groups_0 = const()[name = tensor<string, []>("op_1595_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_7_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40962560))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42142272))), name = tensor<string, []>("layers_7_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_7_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42142400)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_1595_cast_fp16 = conv(bias = layers_7_fc1_inlier_module_bias_to_fp16, dilations = var_1595_dilations_0, groups = var_1595_groups_0, pad = var_1595_pad_0, pad_type = var_1595_pad_type_0, strides = var_1595_strides_0, weight = layers_7_fc1_inlier_module_weight_to_fp16_palettized, x = input_59_cast_fp16)[name = tensor<string, []>("op_1595_cast_fp16")];
            tensor<string, []> var_1601_pad_type_0 = const()[name = tensor<string, []>("op_1601_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1601_strides_0 = const()[name = tensor<string, []>("op_1601_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1601_pad_0 = const()[name = tensor<string, []>("op_1601_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1601_dilations_0 = const()[name = tensor<string, []>("op_1601_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1601_groups_0 = const()[name = tensor<string, []>("op_1601_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_7_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42172224))), name = tensor<string, []>("layers_7_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [11748]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42148608))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_1601_cast_fp16 = conv(dilations = var_1601_dilations_0, groups = var_1601_groups_0, pad = var_1601_pad_0, pad_type = var_1601_pad_type_0, strides = var_1601_strides_0, weight = layers_7_fc1_outlier_module_weight_to_fp16_sparsified, x = input_59_cast_fp16)[name = tensor<string, []>("op_1601_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_61_cast_fp16 = add(x = var_1595_cast_fp16, y = var_1601_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
            tensor<string, []> input_63_mode_0 = const()[name = tensor<string, []>("input_63_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_63_cast_fp16 = gelu(mode = input_63_mode_0, x = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<string, []> var_1612_pad_type_0 = const()[name = tensor<string, []>("op_1612_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1612_strides_0 = const()[name = tensor<string, []>("op_1612_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1612_pad_0 = const()[name = tensor<string, []>("op_1612_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1612_dilations_0 = const()[name = tensor<string, []>("op_1612_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1612_groups_0 = const()[name = tensor<string, []>("op_1612_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_7_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42467200))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43646912))), name = tensor<string, []>("layers_7_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_7_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43647040)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1612_cast_fp16 = conv(bias = layers_7_fc2_inlier_module_bias_to_fp16, dilations = var_1612_dilations_0, groups = var_1612_groups_0, pad = var_1612_pad_0, pad_type = var_1612_pad_type_0, strides = var_1612_strides_0, weight = layers_7_fc2_inlier_module_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = tensor<string, []>("op_1612_cast_fp16")];
            tensor<string, []> var_1618_pad_type_0 = const()[name = tensor<string, []>("op_1618_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1618_strides_0 = const()[name = tensor<string, []>("op_1618_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1618_pad_0 = const()[name = tensor<string, []>("op_1618_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1618_dilations_0 = const()[name = tensor<string, []>("op_1618_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1618_groups_0 = const()[name = tensor<string, []>("op_1618_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_7_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43671552))), name = tensor<string, []>("layers_7_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [11417]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43648640))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1618_cast_fp16 = conv(dilations = var_1618_dilations_0, groups = var_1618_groups_0, pad = var_1618_pad_0, pad_type = var_1618_pad_type_0, strides = var_1618_strides_0, weight = layers_7_fc2_outlier_module_weight_to_fp16_sparsified, x = input_63_cast_fp16)[name = tensor<string, []>("op_1618_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_19_cast_fp16 = add(x = var_1612_cast_fp16, y = var_1618_cast_fp16)[name = tensor<string, []>("hidden_states_19_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")];
            tensor<int32, []> var_1624 = const()[name = tensor<string, []>("op_1624"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_33_axes_0 = const()[name = tensor<string, []>("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1646_to_fp16 = const()[name = tensor<string, []>("op_1646_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1646_to_fp16, x = inputs_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
            tensor<fp16, [768]> obj_33_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43966528)))];
            tensor<fp16, [768]> obj_33_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_33_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43968128)))];
            tensor<fp16, []> obj_33_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_33_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor<string, []>("obj_33_cast_fp16")];
            tensor<string, []> var_1668_pad_type_0 = const()[name = tensor<string, []>("op_1668_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1668_strides_0 = const()[name = tensor<string, []>("op_1668_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1668_pad_0 = const()[name = tensor<string, []>("op_1668_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1668_dilations_0 = const()[name = tensor<string, []>("op_1668_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1668_groups_0 = const()[name = tensor<string, []>("op_1668_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43969728))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44264704))), name = tensor<string, []>("layers_8_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_8_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44264832)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1668_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1668_dilations_0, groups = var_1668_groups_0, pad = var_1668_pad_0, pad_type = var_1668_pad_type_0, strides = var_1668_strides_0, weight = layers_8_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor<string, []>("op_1668_cast_fp16")];
            tensor<string, []> var_1674_pad_type_0 = const()[name = tensor<string, []>("op_1674_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1674_strides_0 = const()[name = tensor<string, []>("op_1674_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1674_pad_0 = const()[name = tensor<string, []>("op_1674_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1674_dilations_0 = const()[name = tensor<string, []>("op_1674_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1674_groups_0 = const()[name = tensor<string, []>("op_1674_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44272704))), name = tensor<string, []>("layers_8_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3098]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44266432))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1674_cast_fp16 = conv(dilations = var_1674_dilations_0, groups = var_1674_groups_0, pad = var_1674_pad_0, pad_type = var_1674_pad_type_0, strides = var_1674_strides_0, weight = layers_8_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_33_cast_fp16)[name = tensor<string, []>("op_1674_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_17_cast_fp16 = add(x = var_1668_cast_fp16, y = var_1674_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
            tensor<string, []> var_1683_pad_type_0 = const()[name = tensor<string, []>("op_1683_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1683_strides_0 = const()[name = tensor<string, []>("op_1683_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1683_pad_0 = const()[name = tensor<string, []>("op_1683_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1683_dilations_0 = const()[name = tensor<string, []>("op_1683_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1683_groups_0 = const()[name = tensor<string, []>("op_1683_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44346496))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44641472))), name = tensor<string, []>("layers_8_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1683_cast_fp16 = conv(dilations = var_1683_dilations_0, groups = var_1683_groups_0, pad = var_1683_pad_0, pad_type = var_1683_pad_type_0, strides = var_1683_strides_0, weight = layers_8_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor<string, []>("op_1683_cast_fp16")];
            tensor<string, []> var_1689_pad_type_0 = const()[name = tensor<string, []>("op_1689_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1689_strides_0 = const()[name = tensor<string, []>("op_1689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1689_pad_0 = const()[name = tensor<string, []>("op_1689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1689_dilations_0 = const()[name = tensor<string, []>("op_1689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1689_groups_0 = const()[name = tensor<string, []>("op_1689_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44648384))), name = tensor<string, []>("layers_8_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3329]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44641600))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1689_cast_fp16 = conv(dilations = var_1689_dilations_0, groups = var_1689_groups_0, pad = var_1689_pad_0, pad_type = var_1689_pad_type_0, strides = var_1689_strides_0, weight = layers_8_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_33_cast_fp16)[name = tensor<string, []>("op_1689_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_17_cast_fp16 = add(x = var_1683_cast_fp16, y = var_1689_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
            tensor<string, []> var_1699_pad_type_0 = const()[name = tensor<string, []>("op_1699_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1699_strides_0 = const()[name = tensor<string, []>("op_1699_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1699_pad_0 = const()[name = tensor<string, []>("op_1699_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1699_dilations_0 = const()[name = tensor<string, []>("op_1699_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1699_groups_0 = const()[name = tensor<string, []>("op_1699_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44722176))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45017152))), name = tensor<string, []>("layers_8_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_8_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45017280)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1699_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1699_dilations_0, groups = var_1699_groups_0, pad = var_1699_pad_0, pad_type = var_1699_pad_type_0, strides = var_1699_strides_0, weight = layers_8_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor<string, []>("op_1699_cast_fp16")];
            tensor<string, []> var_1705_pad_type_0 = const()[name = tensor<string, []>("op_1705_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1705_strides_0 = const()[name = tensor<string, []>("op_1705_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1705_pad_0 = const()[name = tensor<string, []>("op_1705_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1705_dilations_0 = const()[name = tensor<string, []>("op_1705_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1705_groups_0 = const()[name = tensor<string, []>("op_1705_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45025152))), name = tensor<string, []>("layers_8_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3094]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45018880))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1705_cast_fp16 = conv(dilations = var_1705_dilations_0, groups = var_1705_groups_0, pad = var_1705_pad_0, pad_type = var_1705_pad_type_0, strides = var_1705_strides_0, weight = layers_8_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_33_cast_fp16)[name = tensor<string, []>("op_1705_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_17_cast_fp16 = add(x = var_1699_cast_fp16, y = var_1705_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")];
            tensor<int32, [4]> var_1709 = const()[name = tensor<string, []>("op_1709"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_17_cast_fp16 = reshape(shape = var_1709, x = query_17_cast_fp16)[name = tensor<string, []>("mh_q_17_cast_fp16")];
            tensor<fp16, []> var_1711_to_fp16 = const()[name = tensor<string, []>("op_1711_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_1712_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1711_to_fp16)[name = tensor<string, []>("op_1712_cast_fp16")];
            tensor<int32, [4]> var_1715 = const()[name = tensor<string, []>("op_1715"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1716_cast_fp16 = reshape(shape = var_1715, x = key_17_cast_fp16)[name = tensor<string, []>("op_1716_cast_fp16")];
            tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_1712_cast_fp16, y = var_1716_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_1719_cast_fp16 = softmax(axis = var_1624, x = mh_w_17_cast_fp16)[name = tensor<string, []>("op_1719_cast_fp16")];
            tensor<int32, [4]> var_1720 = const()[name = tensor<string, []>("op_1720"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1721_cast_fp16 = reshape(shape = var_1720, x = value_17_cast_fp16)[name = tensor<string, []>("op_1721_cast_fp16")];
            tensor<bool, []> attn_17_transpose_x_0 = const()[name = tensor<string, []>("attn_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_17_transpose_y_0 = const()[name = tensor<string, []>("attn_17_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1721_cast_fp16, y = var_1719_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
            tensor<int32, [4]> var_1724 = const()[name = tensor<string, []>("op_1724"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_65_cast_fp16 = reshape(shape = var_1724, x = attn_17_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<string, []> var_1734_pad_type_0 = const()[name = tensor<string, []>("op_1734_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1734_strides_0 = const()[name = tensor<string, []>("op_1734_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1734_pad_0 = const()[name = tensor<string, []>("op_1734_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1734_dilations_0 = const()[name = tensor<string, []>("op_1734_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1734_groups_0 = const()[name = tensor<string, []>("op_1734_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45098944))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45393920))), name = tensor<string, []>("layers_8_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_8_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45394048)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1734_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1734_dilations_0, groups = var_1734_groups_0, pad = var_1734_pad_0, pad_type = var_1734_pad_type_0, strides = var_1734_strides_0, weight = layers_8_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = tensor<string, []>("op_1734_cast_fp16")];
            tensor<string, []> var_1740_pad_type_0 = const()[name = tensor<string, []>("op_1740_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1740_strides_0 = const()[name = tensor<string, []>("op_1740_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1740_pad_0 = const()[name = tensor<string, []>("op_1740_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1740_dilations_0 = const()[name = tensor<string, []>("op_1740_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1740_groups_0 = const()[name = tensor<string, []>("op_1740_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45402880))), name = tensor<string, []>("layers_8_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3581]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45395648))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1740_cast_fp16 = conv(dilations = var_1740_dilations_0, groups = var_1740_groups_0, pad = var_1740_pad_0, pad_type = var_1740_pad_type_0, strides = var_1740_strides_0, weight = layers_8_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_65_cast_fp16)[name = tensor<string, []>("op_1740_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_35_cast_fp16 = add(x = var_1734_cast_fp16, y = var_1740_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_35_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = tensor<string, []>("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1751_to_fp16 = const()[name = tensor<string, []>("op_1751_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1751_to_fp16, x = inputs_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
            tensor<fp16, [768]> input_67_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_67_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45476672)))];
            tensor<fp16, [768]> input_67_beta_0_to_fp16 = const()[name = tensor<string, []>("input_67_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45478272)))];
            tensor<fp16, []> input_67_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_67_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<string, []> var_1769_pad_type_0 = const()[name = tensor<string, []>("op_1769_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1769_strides_0 = const()[name = tensor<string, []>("op_1769_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1769_pad_0 = const()[name = tensor<string, []>("op_1769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1769_dilations_0 = const()[name = tensor<string, []>("op_1769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1769_groups_0 = const()[name = tensor<string, []>("op_1769_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_8_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45479872))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46659584))), name = tensor<string, []>("layers_8_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_8_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46659712)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_1769_cast_fp16 = conv(bias = layers_8_fc1_inlier_module_bias_to_fp16, dilations = var_1769_dilations_0, groups = var_1769_groups_0, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1769_strides_0, weight = layers_8_fc1_inlier_module_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor<string, []>("op_1769_cast_fp16")];
            tensor<string, []> var_1775_pad_type_0 = const()[name = tensor<string, []>("op_1775_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1775_strides_0 = const()[name = tensor<string, []>("op_1775_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1775_pad_0 = const()[name = tensor<string, []>("op_1775_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1775_dilations_0 = const()[name = tensor<string, []>("op_1775_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1775_groups_0 = const()[name = tensor<string, []>("op_1775_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_8_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46692096))), name = tensor<string, []>("layers_8_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [13025]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46665920))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_1775_cast_fp16 = conv(dilations = var_1775_dilations_0, groups = var_1775_groups_0, pad = var_1775_pad_0, pad_type = var_1775_pad_type_0, strides = var_1775_strides_0, weight = layers_8_fc1_outlier_module_weight_to_fp16_sparsified, x = input_67_cast_fp16)[name = tensor<string, []>("op_1775_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_69_cast_fp16 = add(x = var_1769_cast_fp16, y = var_1775_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<string, []> input_71_mode_0 = const()[name = tensor<string, []>("input_71_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<string, []> var_1786_pad_type_0 = const()[name = tensor<string, []>("op_1786_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1786_strides_0 = const()[name = tensor<string, []>("op_1786_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1786_pad_0 = const()[name = tensor<string, []>("op_1786_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1786_dilations_0 = const()[name = tensor<string, []>("op_1786_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1786_groups_0 = const()[name = tensor<string, []>("op_1786_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_8_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46987072))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48166784))), name = tensor<string, []>("layers_8_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_8_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48166912)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1786_cast_fp16 = conv(bias = layers_8_fc2_inlier_module_bias_to_fp16, dilations = var_1786_dilations_0, groups = var_1786_groups_0, pad = var_1786_pad_0, pad_type = var_1786_pad_type_0, strides = var_1786_strides_0, weight = layers_8_fc2_inlier_module_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = tensor<string, []>("op_1786_cast_fp16")];
            tensor<string, []> var_1792_pad_type_0 = const()[name = tensor<string, []>("op_1792_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1792_strides_0 = const()[name = tensor<string, []>("op_1792_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1792_pad_0 = const()[name = tensor<string, []>("op_1792_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1792_dilations_0 = const()[name = tensor<string, []>("op_1792_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1792_groups_0 = const()[name = tensor<string, []>("op_1792_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_8_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48189952))), name = tensor<string, []>("layers_8_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [10681]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48168512))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1792_cast_fp16 = conv(dilations = var_1792_dilations_0, groups = var_1792_groups_0, pad = var_1792_pad_0, pad_type = var_1792_pad_type_0, strides = var_1792_strides_0, weight = layers_8_fc2_outlier_module_weight_to_fp16_sparsified, x = input_71_cast_fp16)[name = tensor<string, []>("op_1792_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_21_cast_fp16 = add(x = var_1786_cast_fp16, y = var_1792_cast_fp16)[name = tensor<string, []>("hidden_states_21_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor<string, []>("inputs_37_cast_fp16")];
            tensor<int32, []> var_1798 = const()[name = tensor<string, []>("op_1798"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_37_axes_0 = const()[name = tensor<string, []>("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1820_to_fp16 = const()[name = tensor<string, []>("op_1820_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1820_to_fp16, x = inputs_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
            tensor<fp16, [768]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48484928)))];
            tensor<fp16, [768]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48486528)))];
            tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
            tensor<string, []> var_1842_pad_type_0 = const()[name = tensor<string, []>("op_1842_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1842_strides_0 = const()[name = tensor<string, []>("op_1842_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1842_pad_0 = const()[name = tensor<string, []>("op_1842_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1842_dilations_0 = const()[name = tensor<string, []>("op_1842_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1842_groups_0 = const()[name = tensor<string, []>("op_1842_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48488128))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48783104))), name = tensor<string, []>("layers_9_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_9_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48783232)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1842_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1842_dilations_0, groups = var_1842_groups_0, pad = var_1842_pad_0, pad_type = var_1842_pad_type_0, strides = var_1842_strides_0, weight = layers_9_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("op_1842_cast_fp16")];
            tensor<string, []> var_1848_pad_type_0 = const()[name = tensor<string, []>("op_1848_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1848_strides_0 = const()[name = tensor<string, []>("op_1848_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1848_pad_0 = const()[name = tensor<string, []>("op_1848_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1848_dilations_0 = const()[name = tensor<string, []>("op_1848_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1848_groups_0 = const()[name = tensor<string, []>("op_1848_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48790784))), name = tensor<string, []>("layers_9_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2925]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48784832))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1848_cast_fp16 = conv(dilations = var_1848_dilations_0, groups = var_1848_groups_0, pad = var_1848_pad_0, pad_type = var_1848_pad_type_0, strides = var_1848_strides_0, weight = layers_9_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_37_cast_fp16)[name = tensor<string, []>("op_1848_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_19_cast_fp16 = add(x = var_1842_cast_fp16, y = var_1848_cast_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
            tensor<string, []> var_1857_pad_type_0 = const()[name = tensor<string, []>("op_1857_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1857_strides_0 = const()[name = tensor<string, []>("op_1857_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1857_pad_0 = const()[name = tensor<string, []>("op_1857_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1857_dilations_0 = const()[name = tensor<string, []>("op_1857_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1857_groups_0 = const()[name = tensor<string, []>("op_1857_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48864576))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49159552))), name = tensor<string, []>("layers_9_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1857_cast_fp16 = conv(dilations = var_1857_dilations_0, groups = var_1857_groups_0, pad = var_1857_pad_0, pad_type = var_1857_pad_type_0, strides = var_1857_strides_0, weight = layers_9_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("op_1857_cast_fp16")];
            tensor<string, []> var_1863_pad_type_0 = const()[name = tensor<string, []>("op_1863_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1863_strides_0 = const()[name = tensor<string, []>("op_1863_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1863_pad_0 = const()[name = tensor<string, []>("op_1863_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1863_dilations_0 = const()[name = tensor<string, []>("op_1863_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1863_groups_0 = const()[name = tensor<string, []>("op_1863_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49166528))), name = tensor<string, []>("layers_9_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3376]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49159680))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1863_cast_fp16 = conv(dilations = var_1863_dilations_0, groups = var_1863_groups_0, pad = var_1863_pad_0, pad_type = var_1863_pad_type_0, strides = var_1863_strides_0, weight = layers_9_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_37_cast_fp16)[name = tensor<string, []>("op_1863_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_19_cast_fp16 = add(x = var_1857_cast_fp16, y = var_1863_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
            tensor<string, []> var_1873_pad_type_0 = const()[name = tensor<string, []>("op_1873_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1873_strides_0 = const()[name = tensor<string, []>("op_1873_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1873_pad_0 = const()[name = tensor<string, []>("op_1873_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1873_dilations_0 = const()[name = tensor<string, []>("op_1873_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1873_groups_0 = const()[name = tensor<string, []>("op_1873_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49240320))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49535296))), name = tensor<string, []>("layers_9_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_9_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49535424)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1873_cast_fp16 = conv(bias = layers_9_self_attn_v_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_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("op_1873_cast_fp16")];
            tensor<string, []> var_1879_pad_type_0 = const()[name = tensor<string, []>("op_1879_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1879_strides_0 = const()[name = tensor<string, []>("op_1879_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1879_pad_0 = const()[name = tensor<string, []>("op_1879_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1879_dilations_0 = const()[name = tensor<string, []>("op_1879_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1879_groups_0 = const()[name = tensor<string, []>("op_1879_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49541568))), name = tensor<string, []>("layers_9_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2221]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49537024))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 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_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_37_cast_fp16)[name = tensor<string, []>("op_1879_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_19_cast_fp16 = add(x = var_1873_cast_fp16, y = var_1879_cast_fp16)[name = tensor<string, []>("value_19_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = tensor<string, []>("op_1883"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_19_cast_fp16 = reshape(shape = var_1883, x = query_19_cast_fp16)[name = tensor<string, []>("mh_q_19_cast_fp16")];
            tensor<fp16, []> var_1885_to_fp16 = const()[name = tensor<string, []>("op_1885_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_1886_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1885_to_fp16)[name = tensor<string, []>("op_1886_cast_fp16")];
            tensor<int32, [4]> var_1889 = const()[name = tensor<string, []>("op_1889"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1890_cast_fp16 = reshape(shape = var_1889, x = key_19_cast_fp16)[name = tensor<string, []>("op_1890_cast_fp16")];
            tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_1886_cast_fp16, y = var_1890_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_1893_cast_fp16 = softmax(axis = var_1798, x = mh_w_19_cast_fp16)[name = tensor<string, []>("op_1893_cast_fp16")];
            tensor<int32, [4]> var_1894 = const()[name = tensor<string, []>("op_1894"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_1895_cast_fp16 = reshape(shape = var_1894, x = value_19_cast_fp16)[name = tensor<string, []>("op_1895_cast_fp16")];
            tensor<bool, []> attn_19_transpose_x_0 = const()[name = tensor<string, []>("attn_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_19_transpose_y_0 = const()[name = tensor<string, []>("attn_19_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1895_cast_fp16, y = var_1893_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
            tensor<int32, [4]> var_1898 = const()[name = tensor<string, []>("op_1898"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_73_cast_fp16 = reshape(shape = var_1898, x = attn_19_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<string, []> var_1908_pad_type_0 = const()[name = tensor<string, []>("op_1908_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1908_strides_0 = const()[name = tensor<string, []>("op_1908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1908_pad_0 = const()[name = tensor<string, []>("op_1908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1908_dilations_0 = const()[name = tensor<string, []>("op_1908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1908_groups_0 = const()[name = tensor<string, []>("op_1908_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49615360))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49910336))), name = tensor<string, []>("layers_9_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_9_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49910464)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1908_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1908_dilations_0, groups = var_1908_groups_0, pad = var_1908_pad_0, pad_type = var_1908_pad_type_0, strides = var_1908_strides_0, weight = layers_9_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_73_cast_fp16)[name = tensor<string, []>("op_1908_cast_fp16")];
            tensor<string, []> var_1914_pad_type_0 = const()[name = tensor<string, []>("op_1914_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1914_strides_0 = const()[name = tensor<string, []>("op_1914_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1914_pad_0 = const()[name = tensor<string, []>("op_1914_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1914_dilations_0 = const()[name = tensor<string, []>("op_1914_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1914_groups_0 = const()[name = tensor<string, []>("op_1914_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49916928))), name = tensor<string, []>("layers_9_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2375]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49912064))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1914_cast_fp16 = conv(dilations = var_1914_dilations_0, groups = var_1914_groups_0, pad = var_1914_pad_0, pad_type = var_1914_pad_type_0, strides = var_1914_strides_0, weight = layers_9_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_73_cast_fp16)[name = tensor<string, []>("op_1914_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_39_cast_fp16 = add(x = var_1908_cast_fp16, y = var_1914_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_39_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = tensor<string, []>("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1925_to_fp16 = const()[name = tensor<string, []>("op_1925_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1925_to_fp16, x = inputs_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
            tensor<fp16, [768]> input_75_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_75_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49990720)))];
            tensor<fp16, [768]> input_75_beta_0_to_fp16 = const()[name = tensor<string, []>("input_75_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49992320)))];
            tensor<fp16, []> input_75_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_75_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<string, []> var_1943_pad_type_0 = const()[name = tensor<string, []>("op_1943_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1943_strides_0 = const()[name = tensor<string, []>("op_1943_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1943_pad_0 = const()[name = tensor<string, []>("op_1943_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1943_dilations_0 = const()[name = tensor<string, []>("op_1943_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1943_groups_0 = const()[name = tensor<string, []>("op_1943_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_9_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49993920))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51173632))), name = tensor<string, []>("layers_9_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_9_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51173760)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_1943_cast_fp16 = conv(bias = layers_9_fc1_inlier_module_bias_to_fp16, dilations = var_1943_dilations_0, groups = var_1943_groups_0, pad = var_1943_pad_0, pad_type = var_1943_pad_type_0, strides = var_1943_strides_0, weight = layers_9_fc1_inlier_module_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = tensor<string, []>("op_1943_cast_fp16")];
            tensor<string, []> var_1949_pad_type_0 = const()[name = tensor<string, []>("op_1949_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1949_strides_0 = const()[name = tensor<string, []>("op_1949_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1949_pad_0 = const()[name = tensor<string, []>("op_1949_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1949_dilations_0 = const()[name = tensor<string, []>("op_1949_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1949_groups_0 = const()[name = tensor<string, []>("op_1949_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_9_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51201984))), name = tensor<string, []>("layers_9_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [10959]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51179968))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_1949_cast_fp16 = conv(dilations = var_1949_dilations_0, groups = var_1949_groups_0, pad = var_1949_pad_0, pad_type = var_1949_pad_type_0, strides = var_1949_strides_0, weight = layers_9_fc1_outlier_module_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = tensor<string, []>("op_1949_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_77_cast_fp16 = add(x = var_1943_cast_fp16, y = var_1949_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<string, []> input_79_mode_0 = const()[name = tensor<string, []>("input_79_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<string, []> var_1960_pad_type_0 = const()[name = tensor<string, []>("op_1960_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1960_strides_0 = const()[name = tensor<string, []>("op_1960_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1960_pad_0 = const()[name = tensor<string, []>("op_1960_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1960_dilations_0 = const()[name = tensor<string, []>("op_1960_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1960_groups_0 = const()[name = tensor<string, []>("op_1960_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_9_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51496960))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52676672))), name = tensor<string, []>("layers_9_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_9_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52676800)))];
            tensor<fp16, [1, 768, 1, 1500]> var_1960_cast_fp16 = conv(bias = layers_9_fc2_inlier_module_bias_to_fp16, dilations = var_1960_dilations_0, groups = var_1960_groups_0, pad = var_1960_pad_0, pad_type = var_1960_pad_type_0, strides = var_1960_strides_0, weight = layers_9_fc2_inlier_module_weight_to_fp16_palettized, x = input_79_cast_fp16)[name = tensor<string, []>("op_1960_cast_fp16")];
            tensor<string, []> var_1966_pad_type_0 = const()[name = tensor<string, []>("op_1966_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1966_strides_0 = const()[name = tensor<string, []>("op_1966_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1966_pad_0 = const()[name = tensor<string, []>("op_1966_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1966_dilations_0 = const()[name = tensor<string, []>("op_1966_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1966_groups_0 = const()[name = tensor<string, []>("op_1966_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_9_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52698816))), name = tensor<string, []>("layers_9_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [10150]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52678400))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_1966_cast_fp16 = conv(dilations = var_1966_dilations_0, groups = var_1966_groups_0, pad = var_1966_pad_0, pad_type = var_1966_pad_type_0, strides = var_1966_strides_0, weight = layers_9_fc2_outlier_module_weight_to_fp16_sparsified, x = input_79_cast_fp16)[name = tensor<string, []>("op_1966_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_23_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1966_cast_fp16)[name = tensor<string, []>("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor<string, []>("inputs_41_cast_fp16")];
            tensor<int32, []> var_1972 = const()[name = tensor<string, []>("op_1972"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_41_axes_0 = const()[name = tensor<string, []>("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1994_to_fp16 = const()[name = tensor<string, []>("op_1994_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1994_to_fp16, x = inputs_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
            tensor<fp16, [768]> obj_41_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_41_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52993792)))];
            tensor<fp16, [768]> obj_41_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_41_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52995392)))];
            tensor<fp16, []> obj_41_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_41_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
            tensor<string, []> var_2016_pad_type_0 = const()[name = tensor<string, []>("op_2016_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2016_strides_0 = const()[name = tensor<string, []>("op_2016_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2016_pad_0 = const()[name = tensor<string, []>("op_2016_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2016_dilations_0 = const()[name = tensor<string, []>("op_2016_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2016_groups_0 = const()[name = tensor<string, []>("op_2016_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52996992))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53291968))), name = tensor<string, []>("layers_10_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_10_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53292096)))];
            tensor<fp16, [1, 768, 1, 1500]> var_2016_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_2016_dilations_0, groups = var_2016_groups_0, pad = var_2016_pad_0, pad_type = var_2016_pad_type_0, strides = var_2016_strides_0, weight = layers_10_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor<string, []>("op_2016_cast_fp16")];
            tensor<string, []> var_2022_pad_type_0 = const()[name = tensor<string, []>("op_2022_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2022_strides_0 = const()[name = tensor<string, []>("op_2022_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2022_pad_0 = const()[name = tensor<string, []>("op_2022_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2022_dilations_0 = const()[name = tensor<string, []>("op_2022_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2022_groups_0 = const()[name = tensor<string, []>("op_2022_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53299776))), name = tensor<string, []>("layers_10_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2995]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53293696))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2022_cast_fp16 = conv(dilations = var_2022_dilations_0, groups = var_2022_groups_0, pad = var_2022_pad_0, pad_type = var_2022_pad_type_0, strides = var_2022_strides_0, weight = layers_10_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_41_cast_fp16)[name = tensor<string, []>("op_2022_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_21_cast_fp16 = add(x = var_2016_cast_fp16, y = var_2022_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
            tensor<string, []> var_2031_pad_type_0 = const()[name = tensor<string, []>("op_2031_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2031_strides_0 = const()[name = tensor<string, []>("op_2031_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2031_pad_0 = const()[name = tensor<string, []>("op_2031_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2031_dilations_0 = const()[name = tensor<string, []>("op_2031_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2031_groups_0 = const()[name = tensor<string, []>("op_2031_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53373568))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53668544))), name = tensor<string, []>("layers_10_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2031_cast_fp16 = conv(dilations = var_2031_dilations_0, groups = var_2031_groups_0, pad = var_2031_pad_0, pad_type = var_2031_pad_type_0, strides = var_2031_strides_0, weight = layers_10_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor<string, []>("op_2031_cast_fp16")];
            tensor<string, []> var_2037_pad_type_0 = const()[name = tensor<string, []>("op_2037_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2037_strides_0 = const()[name = tensor<string, []>("op_2037_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2037_pad_0 = const()[name = tensor<string, []>("op_2037_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2037_dilations_0 = const()[name = tensor<string, []>("op_2037_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2037_groups_0 = const()[name = tensor<string, []>("op_2037_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53675072))), name = tensor<string, []>("layers_10_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3150]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53668672))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2037_cast_fp16 = conv(dilations = var_2037_dilations_0, groups = var_2037_groups_0, pad = var_2037_pad_0, pad_type = var_2037_pad_type_0, strides = var_2037_strides_0, weight = layers_10_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_41_cast_fp16)[name = tensor<string, []>("op_2037_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_21_cast_fp16 = add(x = var_2031_cast_fp16, y = var_2037_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
            tensor<string, []> var_2047_pad_type_0 = const()[name = tensor<string, []>("op_2047_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2047_strides_0 = const()[name = tensor<string, []>("op_2047_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2047_pad_0 = const()[name = tensor<string, []>("op_2047_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2047_dilations_0 = const()[name = tensor<string, []>("op_2047_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2047_groups_0 = const()[name = tensor<string, []>("op_2047_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53748864))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54043840))), name = tensor<string, []>("layers_10_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_10_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54043968)))];
            tensor<fp16, [1, 768, 1, 1500]> var_2047_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_2047_dilations_0, groups = var_2047_groups_0, pad = var_2047_pad_0, pad_type = var_2047_pad_type_0, strides = var_2047_strides_0, weight = layers_10_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor<string, []>("op_2047_cast_fp16")];
            tensor<string, []> var_2053_pad_type_0 = const()[name = tensor<string, []>("op_2053_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2053_strides_0 = const()[name = tensor<string, []>("op_2053_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2053_pad_0 = const()[name = tensor<string, []>("op_2053_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2053_dilations_0 = const()[name = tensor<string, []>("op_2053_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2053_groups_0 = const()[name = tensor<string, []>("op_2053_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54050496))), name = tensor<string, []>("layers_10_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2430]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54045568))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2053_cast_fp16 = conv(dilations = var_2053_dilations_0, groups = var_2053_groups_0, pad = var_2053_pad_0, pad_type = var_2053_pad_type_0, strides = var_2053_strides_0, weight = layers_10_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_41_cast_fp16)[name = tensor<string, []>("op_2053_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_21_cast_fp16 = add(x = var_2047_cast_fp16, y = var_2053_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")];
            tensor<int32, [4]> var_2057 = const()[name = tensor<string, []>("op_2057"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_21_cast_fp16 = reshape(shape = var_2057, x = query_21_cast_fp16)[name = tensor<string, []>("mh_q_21_cast_fp16")];
            tensor<fp16, []> var_2059_to_fp16 = const()[name = tensor<string, []>("op_2059_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_2060_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_2059_to_fp16)[name = tensor<string, []>("op_2060_cast_fp16")];
            tensor<int32, [4]> var_2063 = const()[name = tensor<string, []>("op_2063"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_2064_cast_fp16 = reshape(shape = var_2063, x = key_21_cast_fp16)[name = tensor<string, []>("op_2064_cast_fp16")];
            tensor<bool, []> mh_w_21_transpose_x_0 = const()[name = tensor<string, []>("mh_w_21_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_21_transpose_y_0 = const()[name = tensor<string, []>("mh_w_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 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_2060_cast_fp16, y = var_2064_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_2067_cast_fp16 = softmax(axis = var_1972, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_2067_cast_fp16")];
            tensor<int32, [4]> var_2068 = const()[name = tensor<string, []>("op_2068"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_2069_cast_fp16 = reshape(shape = var_2068, x = value_21_cast_fp16)[name = tensor<string, []>("op_2069_cast_fp16")];
            tensor<bool, []> attn_21_transpose_x_0 = const()[name = tensor<string, []>("attn_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_21_transpose_y_0 = const()[name = tensor<string, []>("attn_21_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_2069_cast_fp16, y = var_2067_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
            tensor<int32, [4]> var_2072 = const()[name = tensor<string, []>("op_2072"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_81_cast_fp16 = reshape(shape = var_2072, x = attn_21_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<string, []> var_2082_pad_type_0 = const()[name = tensor<string, []>("op_2082_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2082_strides_0 = const()[name = tensor<string, []>("op_2082_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2082_pad_0 = const()[name = tensor<string, []>("op_2082_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2082_dilations_0 = const()[name = tensor<string, []>("op_2082_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2082_groups_0 = const()[name = tensor<string, []>("op_2082_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54124288))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54419264))), name = tensor<string, []>("layers_10_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_10_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54419392)))];
            tensor<fp16, [1, 768, 1, 1500]> var_2082_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_inlier_module_bias_to_fp16, 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_o_proj_inlier_module_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor<string, []>("op_2082_cast_fp16")];
            tensor<string, []> var_2088_pad_type_0 = const()[name = tensor<string, []>("op_2088_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2088_strides_0 = const()[name = tensor<string, []>("op_2088_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2088_pad_0 = const()[name = tensor<string, []>("op_2088_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2088_dilations_0 = const()[name = tensor<string, []>("op_2088_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2088_groups_0 = const()[name = tensor<string, []>("op_2088_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54426624))), name = tensor<string, []>("layers_10_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2772]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54420992))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2088_cast_fp16 = conv(dilations = var_2088_dilations_0, groups = var_2088_groups_0, pad = var_2088_pad_0, pad_type = var_2088_pad_type_0, strides = var_2088_strides_0, weight = layers_10_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_81_cast_fp16)[name = tensor<string, []>("op_2088_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_43_cast_fp16 = add(x = var_2082_cast_fp16, y = var_2088_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = obj_43_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = tensor<string, []>("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2099_to_fp16 = const()[name = tensor<string, []>("op_2099_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_2099_to_fp16, x = inputs_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
            tensor<fp16, [768]> input_83_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_83_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54500416)))];
            tensor<fp16, [768]> input_83_beta_0_to_fp16 = const()[name = tensor<string, []>("input_83_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54502016)))];
            tensor<fp16, []> input_83_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_83_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<string, []> var_2117_pad_type_0 = const()[name = tensor<string, []>("op_2117_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2117_strides_0 = const()[name = tensor<string, []>("op_2117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2117_pad_0 = const()[name = tensor<string, []>("op_2117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2117_dilations_0 = const()[name = tensor<string, []>("op_2117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2117_groups_0 = const()[name = tensor<string, []>("op_2117_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_10_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54503616))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55683328))), name = tensor<string, []>("layers_10_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_10_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55683456)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_2117_cast_fp16 = conv(bias = layers_10_fc1_inlier_module_bias_to_fp16, dilations = var_2117_dilations_0, groups = var_2117_groups_0, pad = var_2117_pad_0, pad_type = var_2117_pad_type_0, strides = var_2117_strides_0, weight = layers_10_fc1_inlier_module_weight_to_fp16_palettized, x = input_83_cast_fp16)[name = tensor<string, []>("op_2117_cast_fp16")];
            tensor<string, []> var_2123_pad_type_0 = const()[name = tensor<string, []>("op_2123_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2123_strides_0 = const()[name = tensor<string, []>("op_2123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2123_pad_0 = const()[name = tensor<string, []>("op_2123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2123_dilations_0 = const()[name = tensor<string, []>("op_2123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2123_groups_0 = const()[name = tensor<string, []>("op_2123_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_10_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55711488))), name = tensor<string, []>("layers_10_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [10858]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55689664))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_2123_cast_fp16 = conv(dilations = var_2123_dilations_0, groups = var_2123_groups_0, pad = var_2123_pad_0, pad_type = var_2123_pad_type_0, strides = var_2123_strides_0, weight = layers_10_fc1_outlier_module_weight_to_fp16_sparsified, x = input_83_cast_fp16)[name = tensor<string, []>("op_2123_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_85_cast_fp16 = add(x = var_2117_cast_fp16, y = var_2123_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<string, []> input_87_mode_0 = const()[name = tensor<string, []>("input_87_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<string, []> var_2134_pad_type_0 = const()[name = tensor<string, []>("op_2134_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2134_strides_0 = const()[name = tensor<string, []>("op_2134_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2134_pad_0 = const()[name = tensor<string, []>("op_2134_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2134_dilations_0 = const()[name = tensor<string, []>("op_2134_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2134_groups_0 = const()[name = tensor<string, []>("op_2134_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_10_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56006464))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57186176))), name = tensor<string, []>("layers_10_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_10_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57186304)))];
            tensor<fp16, [1, 768, 1, 1500]> var_2134_cast_fp16 = conv(bias = layers_10_fc2_inlier_module_bias_to_fp16, dilations = var_2134_dilations_0, groups = var_2134_groups_0, pad = var_2134_pad_0, pad_type = var_2134_pad_type_0, strides = var_2134_strides_0, weight = layers_10_fc2_inlier_module_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor<string, []>("op_2134_cast_fp16")];
            tensor<string, []> var_2140_pad_type_0 = const()[name = tensor<string, []>("op_2140_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2140_strides_0 = const()[name = tensor<string, []>("op_2140_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2140_pad_0 = const()[name = tensor<string, []>("op_2140_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2140_dilations_0 = const()[name = tensor<string, []>("op_2140_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2140_groups_0 = const()[name = tensor<string, []>("op_2140_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_10_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57213952))), name = tensor<string, []>("layers_10_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [12990]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57187904))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2140_cast_fp16 = conv(dilations = var_2140_dilations_0, groups = var_2140_groups_0, pad = var_2140_pad_0, pad_type = var_2140_pad_type_0, strides = var_2140_strides_0, weight = layers_10_fc2_outlier_module_weight_to_fp16_sparsified, x = input_87_cast_fp16)[name = tensor<string, []>("op_2140_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_25_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2140_cast_fp16)[name = tensor<string, []>("hidden_states_25_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor<string, []>("inputs_45_cast_fp16")];
            tensor<int32, []> var_2146 = const()[name = tensor<string, []>("op_2146"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_45_axes_0 = const()[name = tensor<string, []>("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2168_to_fp16 = const()[name = tensor<string, []>("op_2168_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_2168_to_fp16, x = inputs_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
            tensor<fp16, [768]> obj_45_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_45_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57508928)))];
            tensor<fp16, [768]> obj_45_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_45_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57510528)))];
            tensor<fp16, []> obj_45_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_45_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor<string, []>("obj_45_cast_fp16")];
            tensor<string, []> var_2190_pad_type_0 = const()[name = tensor<string, []>("op_2190_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2190_strides_0 = const()[name = tensor<string, []>("op_2190_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2190_pad_0 = const()[name = tensor<string, []>("op_2190_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2190_dilations_0 = const()[name = tensor<string, []>("op_2190_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2190_groups_0 = const()[name = tensor<string, []>("op_2190_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57512128))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57807104))), name = tensor<string, []>("layers_11_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_11_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57807232)))];
            tensor<fp16, [1, 768, 1, 1500]> var_2190_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_2190_dilations_0, groups = var_2190_groups_0, pad = var_2190_pad_0, pad_type = var_2190_pad_type_0, strides = var_2190_strides_0, weight = layers_11_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor<string, []>("op_2190_cast_fp16")];
            tensor<string, []> var_2196_pad_type_0 = const()[name = tensor<string, []>("op_2196_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2196_strides_0 = const()[name = tensor<string, []>("op_2196_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2196_pad_0 = const()[name = tensor<string, []>("op_2196_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2196_dilations_0 = const()[name = tensor<string, []>("op_2196_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2196_groups_0 = const()[name = tensor<string, []>("op_2196_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57815296))), name = tensor<string, []>("layers_11_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3182]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57808832))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2196_cast_fp16 = conv(dilations = var_2196_dilations_0, groups = var_2196_groups_0, pad = var_2196_pad_0, pad_type = var_2196_pad_type_0, strides = var_2196_strides_0, weight = layers_11_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_45_cast_fp16)[name = tensor<string, []>("op_2196_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> query_cast_fp16 = add(x = var_2190_cast_fp16, y = var_2196_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
            tensor<string, []> var_2205_pad_type_0 = const()[name = tensor<string, []>("op_2205_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2205_strides_0 = const()[name = tensor<string, []>("op_2205_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2205_pad_0 = const()[name = tensor<string, []>("op_2205_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2205_dilations_0 = const()[name = tensor<string, []>("op_2205_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2205_groups_0 = const()[name = tensor<string, []>("op_2205_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57889088))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58184064))), name = tensor<string, []>("layers_11_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2205_cast_fp16 = conv(dilations = var_2205_dilations_0, groups = var_2205_groups_0, pad = var_2205_pad_0, pad_type = var_2205_pad_type_0, strides = var_2205_strides_0, weight = layers_11_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor<string, []>("op_2205_cast_fp16")];
            tensor<string, []> var_2211_pad_type_0 = const()[name = tensor<string, []>("op_2211_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2211_strides_0 = const()[name = tensor<string, []>("op_2211_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2211_pad_0 = const()[name = tensor<string, []>("op_2211_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2211_dilations_0 = const()[name = tensor<string, []>("op_2211_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2211_groups_0 = const()[name = tensor<string, []>("op_2211_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58191872))), name = tensor<string, []>("layers_11_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [3796]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58184192))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2211_cast_fp16 = conv(dilations = var_2211_dilations_0, groups = var_2211_groups_0, pad = var_2211_pad_0, pad_type = var_2211_pad_type_0, strides = var_2211_strides_0, weight = layers_11_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_45_cast_fp16)[name = tensor<string, []>("op_2211_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> key_cast_fp16 = add(x = var_2205_cast_fp16, y = var_2211_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
            tensor<string, []> var_2221_pad_type_0 = const()[name = tensor<string, []>("op_2221_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2221_strides_0 = const()[name = tensor<string, []>("op_2221_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2221_pad_0 = const()[name = tensor<string, []>("op_2221_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2221_dilations_0 = const()[name = tensor<string, []>("op_2221_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2221_groups_0 = const()[name = tensor<string, []>("op_2221_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58265664))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58560640))), name = tensor<string, []>("layers_11_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_11_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58560768)))];
            tensor<fp16, [1, 768, 1, 1500]> var_2221_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_2221_dilations_0, groups = var_2221_groups_0, pad = var_2221_pad_0, pad_type = var_2221_pad_type_0, strides = var_2221_strides_0, weight = layers_11_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor<string, []>("op_2221_cast_fp16")];
            tensor<string, []> var_2227_pad_type_0 = const()[name = tensor<string, []>("op_2227_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2227_strides_0 = const()[name = tensor<string, []>("op_2227_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2227_pad_0 = const()[name = tensor<string, []>("op_2227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2227_dilations_0 = const()[name = tensor<string, []>("op_2227_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2227_groups_0 = const()[name = tensor<string, []>("op_2227_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58567232))), name = tensor<string, []>("layers_11_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2394]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58562368))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2227_cast_fp16 = conv(dilations = var_2227_dilations_0, groups = var_2227_groups_0, pad = var_2227_pad_0, pad_type = var_2227_pad_type_0, strides = var_2227_strides_0, weight = layers_11_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_45_cast_fp16)[name = tensor<string, []>("op_2227_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> value_cast_fp16 = add(x = var_2221_cast_fp16, y = var_2227_cast_fp16)[name = tensor<string, []>("value_cast_fp16")];
            tensor<int32, [4]> var_2231 = const()[name = tensor<string, []>("op_2231"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> mh_q_cast_fp16 = reshape(shape = var_2231, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")];
            tensor<fp16, []> var_2233_to_fp16 = const()[name = tensor<string, []>("op_2233_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1500]> var_2234_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_2233_to_fp16)[name = tensor<string, []>("op_2234_cast_fp16")];
            tensor<int32, [4]> var_2237 = const()[name = tensor<string, []>("op_2237"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_2238_cast_fp16 = reshape(shape = var_2237, x = key_cast_fp16)[name = tensor<string, []>("op_2238_cast_fp16")];
            tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1500, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_2234_cast_fp16, y = var_2238_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
            tensor<fp16, [1, 12, 1500, 1500]> var_2241_cast_fp16 = softmax(axis = var_2146, x = mh_w_cast_fp16)[name = tensor<string, []>("op_2241_cast_fp16")];
            tensor<int32, [4]> var_2242 = const()[name = tensor<string, []>("op_2242"), val = tensor<int32, [4]>([1, 12, 64, 1500])];
            tensor<fp16, [1, 12, 64, 1500]> var_2243_cast_fp16 = reshape(shape = var_2242, x = value_cast_fp16)[name = tensor<string, []>("op_2243_cast_fp16")];
            tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 64, 1500]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_2243_cast_fp16, y = var_2241_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
            tensor<int32, [4]> var_2246 = const()[name = tensor<string, []>("op_2246"), val = tensor<int32, [4]>([1, 768, 1, 1500])];
            tensor<fp16, [1, 768, 1, 1500]> input_89_cast_fp16 = reshape(shape = var_2246, x = attn_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<string, []> var_2256_pad_type_0 = const()[name = tensor<string, []>("op_2256_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2256_strides_0 = const()[name = tensor<string, []>("op_2256_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2256_pad_0 = const()[name = tensor<string, []>("op_2256_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2256_dilations_0 = const()[name = tensor<string, []>("op_2256_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2256_groups_0 = const()[name = tensor<string, []>("op_2256_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58641024))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58936000))), name = tensor<string, []>("layers_11_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [768]> layers_11_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58936128)))];
            tensor<fp16, [1, 768, 1, 1500]> var_2256_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_2256_dilations_0, groups = var_2256_groups_0, pad = var_2256_pad_0, pad_type = var_2256_pad_type_0, strides = var_2256_strides_0, weight = layers_11_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = tensor<string, []>("op_2256_cast_fp16")];
            tensor<string, []> var_2262_pad_type_0 = const()[name = tensor<string, []>("op_2262_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2262_strides_0 = const()[name = tensor<string, []>("op_2262_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2262_pad_0 = const()[name = tensor<string, []>("op_2262_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2262_dilations_0 = const()[name = tensor<string, []>("op_2262_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2262_groups_0 = const()[name = tensor<string, []>("op_2262_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [73728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58943552))), name = tensor<string, []>("layers_11_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [2873]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58937728))), shape = tensor<uint32, [4]>([768, 768, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2262_cast_fp16 = conv(dilations = var_2262_dilations_0, groups = var_2262_groups_0, pad = var_2262_pad_0, pad_type = var_2262_pad_type_0, strides = var_2262_strides_0, weight = layers_11_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_89_cast_fp16)[name = tensor<string, []>("op_2262_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> obj_cast_fp16 = add(x = var_2256_cast_fp16, y = var_2262_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_cast_fp16)[name = tensor<string, []>("inputs_47_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = tensor<string, []>("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2273_to_fp16 = const()[name = tensor<string, []>("op_2273_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_2273_to_fp16, x = inputs_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")];
            tensor<fp16, [768]> input_91_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_91_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59017344)))];
            tensor<fp16, [768]> input_91_beta_0_to_fp16 = const()[name = tensor<string, []>("input_91_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59018944)))];
            tensor<fp16, []> input_91_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_91_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<string, []> var_2291_pad_type_0 = const()[name = tensor<string, []>("op_2291_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2291_strides_0 = const()[name = tensor<string, []>("op_2291_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2291_pad_0 = const()[name = tensor<string, []>("op_2291_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2291_dilations_0 = const()[name = tensor<string, []>("op_2291_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2291_groups_0 = const()[name = tensor<string, []>("op_2291_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_11_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59020544))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60200256))), name = tensor<string, []>("layers_11_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [3072]> layers_11_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60200384)))];
            tensor<fp16, [1, 3072, 1, 1500]> var_2291_cast_fp16 = conv(bias = layers_11_fc1_inlier_module_bias_to_fp16, dilations = var_2291_dilations_0, groups = var_2291_groups_0, pad = var_2291_pad_0, pad_type = var_2291_pad_type_0, strides = var_2291_strides_0, weight = layers_11_fc1_inlier_module_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = tensor<string, []>("op_2291_cast_fp16")];
            tensor<string, []> var_2297_pad_type_0 = const()[name = tensor<string, []>("op_2297_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2297_strides_0 = const()[name = tensor<string, []>("op_2297_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2297_pad_0 = const()[name = tensor<string, []>("op_2297_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2297_dilations_0 = const()[name = tensor<string, []>("op_2297_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2297_groups_0 = const()[name = tensor<string, []>("op_2297_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_11_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60238976))), name = tensor<string, []>("layers_11_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [16136]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60206592))), shape = tensor<uint32, [4]>([3072, 768, 1, 1])];
            tensor<fp16, [1, 3072, 1, 1500]> var_2297_cast_fp16 = conv(dilations = var_2297_dilations_0, groups = var_2297_groups_0, pad = var_2297_pad_0, pad_type = var_2297_pad_type_0, strides = var_2297_strides_0, weight = layers_11_fc1_outlier_module_weight_to_fp16_sparsified, x = input_91_cast_fp16)[name = tensor<string, []>("op_2297_cast_fp16")];
            tensor<fp16, [1, 3072, 1, 1500]> input_93_cast_fp16 = add(x = var_2291_cast_fp16, y = var_2297_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 3072, 1, 1500]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_93_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<string, []> var_2308_pad_type_0 = const()[name = tensor<string, []>("op_2308_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2308_strides_0 = const()[name = tensor<string, []>("op_2308_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2308_pad_0 = const()[name = tensor<string, []>("op_2308_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2308_dilations_0 = const()[name = tensor<string, []>("op_2308_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2308_groups_0 = const()[name = tensor<string, []>("op_2308_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_11_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1179648]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60533952))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61713664))), name = tensor<string, []>("layers_11_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [768]> layers_11_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61713792)))];
            tensor<fp16, [1, 768, 1, 1500]> var_2308_cast_fp16 = conv(bias = layers_11_fc2_inlier_module_bias_to_fp16, dilations = var_2308_dilations_0, groups = var_2308_groups_0, pad = var_2308_pad_0, pad_type = var_2308_pad_type_0, strides = var_2308_strides_0, weight = layers_11_fc2_inlier_module_weight_to_fp16_palettized, x = input_cast_fp16)[name = tensor<string, []>("op_2308_cast_fp16")];
            tensor<string, []> var_2314_pad_type_0 = const()[name = tensor<string, []>("op_2314_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2314_strides_0 = const()[name = tensor<string, []>("op_2314_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2314_pad_0 = const()[name = tensor<string, []>("op_2314_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2314_dilations_0 = const()[name = tensor<string, []>("op_2314_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2314_groups_0 = const()[name = tensor<string, []>("op_2314_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_11_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [294912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61759168))), name = tensor<string, []>("layers_11_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [21830]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61715392))), shape = tensor<uint32, [4]>([768, 3072, 1, 1])];
            tensor<fp16, [1, 768, 1, 1500]> var_2314_cast_fp16 = conv(dilations = var_2314_dilations_0, groups = var_2314_groups_0, pad = var_2314_pad_0, pad_type = var_2314_pad_type_0, strides = var_2314_strides_0, weight = layers_11_fc2_outlier_module_weight_to_fp16_sparsified, x = input_cast_fp16)[name = tensor<string, []>("op_2314_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> hidden_states_cast_fp16 = add(x = var_2308_cast_fp16, y = var_2314_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1500]> inputs_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2329_to_fp16 = const()[name = tensor<string, []>("op_2329_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1500]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_2329_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
            tensor<fp16, [768]> encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62054144)))];
            tensor<fp16, [768]> encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62055744)))];
            tensor<fp16, []> encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768, 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_57_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("encoder_output_embeds_type_fp32_cast_fp16")];
        } -> (encoder_output_embeds);
}