openai_whisper-small/TextDecoder.mlmodelc/model.mil

program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
{
    func main<ios18>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, state<tensor<fp16, [12, 768, 1, 1536]>> encoder_attn_key_cache, state<tensor<fp16, [1, 1536]>> encoder_attn_key_padding_mask, state<tensor<fp16, [12, 768, 1, 1536]>> encoder_attn_value_cache, tensor<int32, [1]> input_ids, tensor<fp16, [1, 448]> kv_cache_update_mask, state<tensor<fp16, [12, 768, 1, 448]>> self_attn_key_cache, state<tensor<fp16, [12, 768, 1, 448]>> self_attn_value_cache) {
            int32 var_42_axis_0 = const()[name = string("op_42_axis_0"), val = int32(0)];
            int32 var_42_batch_dims_0 = const()[name = string("op_42_batch_dims_0"), val = int32(0)];
            bool var_42_validate_indices_0 = const()[name = string("op_42_validate_indices_0"), val = bool(false)];
            tensor<fp16, [51865, 768]> embed_tokens_weight_to_fp16 = const()[name = string("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51865, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [1, 768]> var_42_cast_fp16 = gather(axis = var_42_axis_0, batch_dims = var_42_batch_dims_0, indices = input_ids, validate_indices = var_42_validate_indices_0, x = embed_tokens_weight_to_fp16)[name = string("op_42_cast_fp16")];
            int32 var_46_axis_0 = const()[name = string("op_46_axis_0"), val = int32(0)];
            int32 var_46_batch_dims_0 = const()[name = string("op_46_batch_dims_0"), val = int32(0)];
            bool var_46_validate_indices_0 = const()[name = string("op_46_validate_indices_0"), val = bool(false)];
            tensor<fp16, [448, 768]> embed_positions_weight_to_fp16 = const()[name = string("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79664768)))];
            string cache_length_to_uint16_dtype_0 = const()[name = string("cache_length_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> cache_length_to_uint16 = cast(dtype = cache_length_to_uint16_dtype_0, x = cache_length)[name = string("cast_183")];
            tensor<fp16, [1, 768]> var_46_cast_fp16_cast_uint16 = gather(axis = var_46_axis_0, batch_dims = var_46_batch_dims_0, indices = cache_length_to_uint16, validate_indices = var_46_validate_indices_0, x = embed_positions_weight_to_fp16)[name = string("op_46_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> hidden_states_1_cast_fp16 = add(x = var_42_cast_fp16, y = var_46_cast_fp16_cast_uint16)[name = string("hidden_states_1_cast_fp16")];
            tensor<int32, [1]> var_60_axes_0 = const()[name = string("op_60_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 768, 1]> var_60_cast_fp16 = expand_dims(axes = var_60_axes_0, x = hidden_states_1_cast_fp16)[name = string("op_60_cast_fp16")];
            tensor<int32, [1]> inputs_1_axes_0 = const()[name = string("inputs_1_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 768, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_60_cast_fp16)[name = string("inputs_1_cast_fp16")];
            tensor<fp16, [12, 768, 1, 448]> read_state_0 = read_state(input = self_attn_key_cache)[name = string("read_state_0")];
            tensor<int32, [12]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [12]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80352960)))];
            int32 var_65_axis_0 = const()[name = string("op_65_axis_0"), val = int32(0)];
            tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_0, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_1, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_2, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_3, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_4, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_5, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_6, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_7, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_8, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_9, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_10, tensor<fp16, [1, 768, 1, 448]> var_65_cast_fp16_11 = split(axis = var_65_axis_0, split_sizes = tile_0, x = read_state_0)[name = string("op_65_cast_fp16")];
            tensor<fp16, [12, 768, 1, 448]> read_state_1 = read_state(input = self_attn_value_cache)[name = string("read_state_1")];
            tensor<int32, [12]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [12]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80353088)))];
            int32 var_80_axis_0 = const()[name = string("op_80_axis_0"), val = int32(0)];
            tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_0, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_1, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_2, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_3, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_4, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_5, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_6, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_7, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_8, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_9, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_10, tensor<fp16, [1, 768, 1, 448]> var_80_cast_fp16_11 = split(axis = var_80_axis_0, split_sizes = tile_1, x = read_state_1)[name = string("op_80_cast_fp16")];
            tensor<fp16, [12, 768, 1, 1536]> read_state_2 = read_state(input = encoder_attn_key_cache)[name = string("read_state_2")];
            tensor<int32, [4]> obj_17_begin_0 = const()[name = string("obj_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> obj_17_end_0 = const()[name = string("obj_17_end_0"), val = tensor<int32, [4]>([1, 768, 1, 1536])];
            tensor<bool, [4]> obj_17_end_mask_0 = const()[name = string("obj_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_17_cast_fp16 = slice_by_index(begin = obj_17_begin_0, end = obj_17_end_0, end_mask = obj_17_end_mask_0, x = read_state_2)[name = string("obj_17_cast_fp16")];
            tensor<fp16, [12, 768, 1, 1536]> read_state_3 = read_state(input = encoder_attn_value_cache)[name = string("read_state_3")];
            tensor<int32, [4]> obj_19_begin_0 = const()[name = string("obj_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> obj_19_end_0 = const()[name = string("obj_19_end_0"), val = tensor<int32, [4]>([1, 768, 1, 1536])];
            tensor<bool, [4]> obj_19_end_mask_0 = const()[name = string("obj_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_19_cast_fp16 = slice_by_index(begin = obj_19_begin_0, end = obj_19_end_0, end_mask = obj_19_end_mask_0, x = read_state_3)[name = string("obj_19_cast_fp16")];
            int32 var_108 = const()[name = string("op_108"), val = int32(3)];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_133_to_fp16 = const()[name = string("op_133_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_133_to_fp16, x = inputs_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [768]> obj_5_mean_0_to_fp16 = const()[name = string("obj_5_mean_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80353216)))];
            tensor<fp16, [768]> obj_5_variance_0_to_fp16 = const()[name = string("obj_5_variance_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80354816)))];
            tensor<fp16, [768]> obj_5_gamma_0_to_fp16 = const()[name = string("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80356416)))];
            tensor<fp16, [768]> obj_5_beta_0_to_fp16 = const()[name = string("obj_5_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80358016)))];
            fp16 obj_5_epsilon_0_to_fp16 = const()[name = string("obj_5_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_1_cast_fp16)[name = string("obj_5_cast_fp16")];
            string query_1_pad_type_0 = const()[name = string("query_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_1_strides_0 = const()[name = string("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_1_pad_0 = const()[name = string("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_1_dilations_0 = const()[name = string("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_1_groups_0 = const()[name = string("query_1_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80359616)))];
            tensor<fp16, [768]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81539328)))];
            tensor<fp16, [1, 768, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("query_1_cast_fp16")];
            string current_key_1_pad_type_0 = const()[name = string("current_key_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_1_strides_0 = const()[name = string("current_key_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_1_pad_0 = const()[name = string("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_1_dilations_0 = const()[name = string("current_key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_1_groups_0 = const()[name = string("current_key_1_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81540928)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("current_key_1_cast_fp16")];
            string current_value_1_pad_type_0 = const()[name = string("current_value_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_1_strides_0 = const()[name = string("current_value_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_1_pad_0 = const()[name = string("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_1_dilations_0 = const()[name = string("current_value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_1_groups_0 = const()[name = string("current_value_1_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82720640)))];
            tensor<fp16, [768]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83900352)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("current_value_1_cast_fp16")];
            tensor<int32, [1]> var_168_axes_0 = const()[name = string("op_168_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 448]> var_168_cast_fp16 = expand_dims(axes = var_168_axes_0, x = kv_cache_update_mask)[name = string("op_168_cast_fp16")];
            tensor<int32, [1]> var_169_axes_0 = const()[name = string("op_169_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 448]> var_169_cast_fp16 = expand_dims(axes = var_169_axes_0, x = var_168_cast_fp16)[name = string("op_169_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_171_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_169_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_1_cast_fp16 = add(x = var_65_cast_fp16_0, y = var_171_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_173_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_169_cast_fp16)[name = string("op_173_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_1_cast_fp16 = add(x = var_80_cast_fp16_0, y = var_173_cast_fp16)[name = string("value_1_cast_fp16")];
            tensor<int32, [4]> var_176 = const()[name = string("op_176"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_176, x = query_1_cast_fp16)[name = string("mh_q_1_cast_fp16")];
            fp16 var_178_to_fp16 = const()[name = string("op_178_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_179_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_178_to_fp16)[name = string("op_179_cast_fp16")];
            tensor<int32, [4]> var_180 = const()[name = string("op_180"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_181_cast_fp16 = reshape(shape = var_180, x = key_1_cast_fp16)[name = string("op_181_cast_fp16")];
            bool mh_w_1_transpose_x_0 = const()[name = string("mh_w_1_transpose_x_0"), val = bool(true)];
            bool mh_w_1_transpose_y_0 = const()[name = string("mh_w_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_179_cast_fp16, y = var_181_cast_fp16)[name = string("mh_w_1_cast_fp16")];
            tensor<int32, [1]> var_185_axes_0 = const()[name = string("op_185_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 448]> var_185_cast_fp16 = expand_dims(axes = var_185_axes_0, x = decoder_key_padding_mask)[name = string("op_185_cast_fp16")];
            tensor<int32, [1]> var_186_axes_0 = const()[name = string("op_186_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 448]> var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_185_cast_fp16)[name = string("op_186_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_3_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_189_cast_fp16 = softmax(axis = var_108, x = mh_w_3_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<int32, [4]> var_190 = const()[name = string("op_190"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_191_cast_fp16 = reshape(shape = var_190, x = value_1_cast_fp16)[name = string("op_191_cast_fp16")];
            bool attn_1_transpose_x_0 = const()[name = string("attn_1_transpose_x_0"), val = bool(false)];
            bool attn_1_transpose_y_0 = const()[name = string("attn_1_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_191_cast_fp16, y = var_189_cast_fp16)[name = string("attn_1_cast_fp16")];
            tensor<int32, [4]> var_194 = const()[name = string("op_194"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_1_cast_fp16 = reshape(shape = var_194, x = attn_1_cast_fp16)[name = string("input_1_cast_fp16")];
            string obj_11_pad_type_0 = const()[name = string("obj_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_11_strides_0 = const()[name = string("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_11_pad_0 = const()[name = string("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_11_dilations_0 = const()[name = string("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_11_groups_0 = const()[name = string("obj_11_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83901952)))];
            tensor<fp16, [768]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85081664)))];
            tensor<fp16, [1, 768, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("obj_11_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_11_cast_fp16)[name = string("inputs_3_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_216_to_fp16 = const()[name = string("op_216_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_216_to_fp16, x = inputs_3_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [768]> obj_13_gamma_0_to_fp16 = const()[name = string("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85083264)))];
            tensor<fp16, [768]> obj_13_beta_0_to_fp16 = const()[name = string("obj_13_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85084864)))];
            fp16 obj_13_epsilon_0_to_fp16 = const()[name = string("obj_13_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_3_cast_fp16)[name = string("obj_13_cast_fp16")];
            string query_3_pad_type_0 = const()[name = string("query_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_3_strides_0 = const()[name = string("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_3_pad_0 = const()[name = string("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_3_dilations_0 = const()[name = string("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_3_groups_0 = const()[name = string("query_3_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85086464)))];
            tensor<fp16, [768]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86266176)))];
            tensor<fp16, [1, 768, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<int32, [4]> var_236 = const()[name = string("op_236"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_236, x = query_3_cast_fp16)[name = string("mh_q_3_cast_fp16")];
            fp16 var_238_to_fp16 = const()[name = string("op_238_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_239_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_238_to_fp16)[name = string("op_239_cast_fp16")];
            tensor<int32, [4]> var_240 = const()[name = string("op_240"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_241_cast_fp16 = reshape(shape = var_240, x = obj_17_cast_fp16)[name = string("op_241_cast_fp16")];
            bool mh_w_5_transpose_x_0 = const()[name = string("mh_w_5_transpose_x_0"), val = bool(true)];
            bool mh_w_5_transpose_y_0 = const()[name = string("mh_w_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_239_cast_fp16, y = var_241_cast_fp16)[name = string("mh_w_5_cast_fp16")];
            tensor<fp16, [1, 1536]> read_state_4 = read_state(input = encoder_attn_key_padding_mask)[name = string("read_state_4")];
            tensor<int32, [1]> var_245_axes_0 = const()[name = string("op_245_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1536]> var_245_cast_fp16 = expand_dims(axes = var_245_axes_0, x = read_state_4)[name = string("op_245_cast_fp16")];
            tensor<int32, [1]> var_246_axes_0 = const()[name = string("op_246_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 1536]> var_246_cast_fp16 = expand_dims(axes = var_246_axes_0, x = var_245_cast_fp16)[name = string("op_246_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_7_cast_fp16 = add(x = mh_w_5_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_7_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_23_cast_fp16 = softmax(axis = var_108, x = mh_w_7_cast_fp16)[name = string("obj_23_cast_fp16")];
            tensor<int32, [4]> var_250 = const()[name = string("op_250"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_251_cast_fp16 = reshape(shape = var_250, x = obj_19_cast_fp16)[name = string("op_251_cast_fp16")];
            bool attn_3_transpose_x_0 = const()[name = string("attn_3_transpose_x_0"), val = bool(false)];
            bool attn_3_transpose_y_0 = const()[name = string("attn_3_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_251_cast_fp16, y = obj_23_cast_fp16)[name = string("attn_3_cast_fp16")];
            tensor<int32, [4]> var_254 = const()[name = string("op_254"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_3_cast_fp16 = reshape(shape = var_254, x = attn_3_cast_fp16)[name = string("input_3_cast_fp16")];
            string obj_21_pad_type_0 = const()[name = string("obj_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_21_strides_0 = const()[name = string("obj_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_21_pad_0 = const()[name = string("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_21_dilations_0 = const()[name = string("obj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_21_groups_0 = const()[name = string("obj_21_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86267776)))];
            tensor<fp16, [768]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87447488)))];
            tensor<fp16, [1, 768, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = string("obj_21_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_21_cast_fp16)[name = string("inputs_5_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_272_to_fp16 = const()[name = string("op_272_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_272_to_fp16, x = inputs_5_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [768]> input_5_gamma_0_to_fp16 = const()[name = string("input_5_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87449088)))];
            tensor<fp16, [768]> input_5_beta_0_to_fp16 = const()[name = string("input_5_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87450688)))];
            fp16 input_5_epsilon_0_to_fp16 = const()[name = string("input_5_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_5_cast_fp16)[name = string("input_5_cast_fp16")];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = string("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87452288)))];
            tensor<fp16, [3072]> layers_0_fc1_bias_to_fp16 = const()[name = string("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92170944)))];
            tensor<fp16, [1, 3072, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
            string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = string("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92177152)))];
            tensor<fp16, [768]> layers_0_fc2_bias_to_fp16 = const()[name = string("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96895808)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = string("inputs_7_cast_fp16")];
            tensor<int32, [4]> obj_35_begin_0 = const()[name = string("obj_35_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> obj_35_end_0 = const()[name = string("obj_35_end_0"), val = tensor<int32, [4]>([2, 768, 1, 1536])];
            tensor<bool, [4]> obj_35_end_mask_0 = const()[name = string("obj_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_35_cast_fp16 = slice_by_index(begin = obj_35_begin_0, end = obj_35_end_0, end_mask = obj_35_end_mask_0, x = read_state_2)[name = string("obj_35_cast_fp16")];
            tensor<int32, [4]> obj_37_begin_0 = const()[name = string("obj_37_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> obj_37_end_0 = const()[name = string("obj_37_end_0"), val = tensor<int32, [4]>([2, 768, 1, 1536])];
            tensor<bool, [4]> obj_37_end_mask_0 = const()[name = string("obj_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_37_cast_fp16 = slice_by_index(begin = obj_37_begin_0, end = obj_37_end_0, end_mask = obj_37_end_mask_0, x = read_state_3)[name = string("obj_37_cast_fp16")];
            int32 var_317 = const()[name = string("op_317"), val = int32(3)];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_342_to_fp16 = const()[name = string("op_342_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_342_to_fp16, x = inputs_7_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [768]> obj_25_gamma_0_to_fp16 = const()[name = string("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96897408)))];
            tensor<fp16, [768]> obj_25_beta_0_to_fp16 = const()[name = string("obj_25_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96899008)))];
            fp16 obj_25_epsilon_0_to_fp16 = const()[name = string("obj_25_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_7_cast_fp16)[name = string("obj_25_cast_fp16")];
            string query_5_pad_type_0 = const()[name = string("query_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_5_strides_0 = const()[name = string("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_5_pad_0 = const()[name = string("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_5_dilations_0 = const()[name = string("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_5_groups_0 = const()[name = string("query_5_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96900608)))];
            tensor<fp16, [768]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98080320)))];
            tensor<fp16, [1, 768, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("query_5_cast_fp16")];
            string current_key_3_pad_type_0 = const()[name = string("current_key_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_3_strides_0 = const()[name = string("current_key_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_3_pad_0 = const()[name = string("current_key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_3_dilations_0 = const()[name = string("current_key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_3_groups_0 = const()[name = string("current_key_3_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98081920)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_3_cast_fp16 = conv(dilations = current_key_3_dilations_0, groups = current_key_3_groups_0, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = current_key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("current_key_3_cast_fp16")];
            string current_value_3_pad_type_0 = const()[name = string("current_value_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_3_strides_0 = const()[name = string("current_value_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_3_pad_0 = const()[name = string("current_value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_3_dilations_0 = const()[name = string("current_value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_3_groups_0 = const()[name = string("current_value_3_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99261632)))];
            tensor<fp16, [768]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100441344)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_3_dilations_0, groups = current_value_3_groups_0, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = current_value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("current_value_3_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_380_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_169_cast_fp16)[name = string("op_380_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_3_cast_fp16 = add(x = var_65_cast_fp16_1, y = var_380_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_382_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_169_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_3_cast_fp16 = add(x = var_80_cast_fp16_1, y = var_382_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<int32, [4]> var_385 = const()[name = string("op_385"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_385, x = query_5_cast_fp16)[name = string("mh_q_5_cast_fp16")];
            fp16 var_387_to_fp16 = const()[name = string("op_387_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_388_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_387_to_fp16)[name = string("op_388_cast_fp16")];
            tensor<int32, [4]> var_389 = const()[name = string("op_389"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_390_cast_fp16 = reshape(shape = var_389, x = key_3_cast_fp16)[name = string("op_390_cast_fp16")];
            bool mh_w_9_transpose_x_0 = const()[name = string("mh_w_9_transpose_x_0"), val = bool(true)];
            bool mh_w_9_transpose_y_0 = const()[name = string("mh_w_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_388_cast_fp16, y = var_390_cast_fp16)[name = string("mh_w_9_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_11_cast_fp16 = add(x = mh_w_9_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_11_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_398_cast_fp16 = softmax(axis = var_317, x = mh_w_11_cast_fp16)[name = string("op_398_cast_fp16")];
            tensor<int32, [4]> var_399 = const()[name = string("op_399"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_400_cast_fp16 = reshape(shape = var_399, x = value_3_cast_fp16)[name = string("op_400_cast_fp16")];
            bool attn_5_transpose_x_0 = const()[name = string("attn_5_transpose_x_0"), val = bool(false)];
            bool attn_5_transpose_y_0 = const()[name = string("attn_5_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_400_cast_fp16, y = var_398_cast_fp16)[name = string("attn_5_cast_fp16")];
            tensor<int32, [4]> var_403 = const()[name = string("op_403"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_11_cast_fp16 = reshape(shape = var_403, x = attn_5_cast_fp16)[name = string("input_11_cast_fp16")];
            string obj_31_pad_type_0 = const()[name = string("obj_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_31_strides_0 = const()[name = string("obj_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_31_pad_0 = const()[name = string("obj_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_31_dilations_0 = const()[name = string("obj_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_31_groups_0 = const()[name = string("obj_31_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100442944)))];
            tensor<fp16, [768]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101622656)))];
            tensor<fp16, [1, 768, 1, 1]> obj_31_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_31_dilations_0, groups = obj_31_groups_0, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = obj_31_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("obj_31_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_31_cast_fp16)[name = string("inputs_9_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_425_to_fp16 = const()[name = string("op_425_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_425_to_fp16, x = inputs_9_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [768]> obj_33_gamma_0_to_fp16 = const()[name = string("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101624256)))];
            tensor<fp16, [768]> obj_33_beta_0_to_fp16 = const()[name = string("obj_33_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101625856)))];
            fp16 obj_33_epsilon_0_to_fp16 = const()[name = string("obj_33_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_9_cast_fp16)[name = string("obj_33_cast_fp16")];
            string query_7_pad_type_0 = const()[name = string("query_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_7_strides_0 = const()[name = string("query_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_7_pad_0 = const()[name = string("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_7_dilations_0 = const()[name = string("query_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_7_groups_0 = const()[name = string("query_7_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101627456)))];
            tensor<fp16, [768]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102807168)))];
            tensor<fp16, [1, 768, 1, 1]> query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<int32, [4]> var_445 = const()[name = string("op_445"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_7_cast_fp16 = reshape(shape = var_445, x = query_7_cast_fp16)[name = string("mh_q_7_cast_fp16")];
            fp16 var_447_to_fp16 = const()[name = string("op_447_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_448_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_447_to_fp16)[name = string("op_448_cast_fp16")];
            tensor<int32, [4]> var_449 = const()[name = string("op_449"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_450_cast_fp16 = reshape(shape = var_449, x = obj_35_cast_fp16)[name = string("op_450_cast_fp16")];
            bool mh_w_13_transpose_x_0 = const()[name = string("mh_w_13_transpose_x_0"), val = bool(true)];
            bool mh_w_13_transpose_y_0 = const()[name = string("mh_w_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_448_cast_fp16, y = var_450_cast_fp16)[name = string("mh_w_13_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_15_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_41_cast_fp16 = softmax(axis = var_317, x = mh_w_15_cast_fp16)[name = string("obj_41_cast_fp16")];
            tensor<int32, [4]> var_459 = const()[name = string("op_459"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_460_cast_fp16 = reshape(shape = var_459, x = obj_37_cast_fp16)[name = string("op_460_cast_fp16")];
            bool attn_7_transpose_x_0 = const()[name = string("attn_7_transpose_x_0"), val = bool(false)];
            bool attn_7_transpose_y_0 = const()[name = string("attn_7_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_460_cast_fp16, y = obj_41_cast_fp16)[name = string("attn_7_cast_fp16")];
            tensor<int32, [4]> var_463 = const()[name = string("op_463"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_13_cast_fp16 = reshape(shape = var_463, x = attn_7_cast_fp16)[name = string("input_13_cast_fp16")];
            string obj_39_pad_type_0 = const()[name = string("obj_39_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_39_strides_0 = const()[name = string("obj_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_39_pad_0 = const()[name = string("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_39_dilations_0 = const()[name = string("obj_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_39_groups_0 = const()[name = string("obj_39_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102808768)))];
            tensor<fp16, [768]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103988480)))];
            tensor<fp16, [1, 768, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = string("obj_39_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_39_cast_fp16)[name = string("inputs_11_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_481_to_fp16 = const()[name = string("op_481_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_481_to_fp16, x = inputs_11_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [768]> input_15_gamma_0_to_fp16 = const()[name = string("input_15_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103990080)))];
            tensor<fp16, [768]> input_15_beta_0_to_fp16 = const()[name = string("input_15_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103991680)))];
            fp16 input_15_epsilon_0_to_fp16 = const()[name = string("input_15_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_11_cast_fp16)[name = string("input_15_cast_fp16")];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = string("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103993280)))];
            tensor<fp16, [3072]> layers_1_fc1_bias_to_fp16 = const()[name = string("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108711936)))];
            tensor<fp16, [1, 3072, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
            string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = string("input_19_cast_fp16")];
            string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = string("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108718144)))];
            tensor<fp16, [768]> layers_1_fc2_bias_to_fp16 = const()[name = string("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113436800)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = string("inputs_13_cast_fp16")];
            tensor<int32, [4]> obj_53_begin_0 = const()[name = string("obj_53_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> obj_53_end_0 = const()[name = string("obj_53_end_0"), val = tensor<int32, [4]>([3, 768, 1, 1536])];
            tensor<bool, [4]> obj_53_end_mask_0 = const()[name = string("obj_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_53_cast_fp16 = slice_by_index(begin = obj_53_begin_0, end = obj_53_end_0, end_mask = obj_53_end_mask_0, x = read_state_2)[name = string("obj_53_cast_fp16")];
            tensor<int32, [4]> obj_55_begin_0 = const()[name = string("obj_55_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> obj_55_end_0 = const()[name = string("obj_55_end_0"), val = tensor<int32, [4]>([3, 768, 1, 1536])];
            tensor<bool, [4]> obj_55_end_mask_0 = const()[name = string("obj_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_55_cast_fp16 = slice_by_index(begin = obj_55_begin_0, end = obj_55_end_0, end_mask = obj_55_end_mask_0, x = read_state_3)[name = string("obj_55_cast_fp16")];
            int32 var_526 = const()[name = string("op_526"), val = int32(3)];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_551_to_fp16 = const()[name = string("op_551_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_551_to_fp16, x = inputs_13_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [768]> obj_43_gamma_0_to_fp16 = const()[name = string("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113438400)))];
            tensor<fp16, [768]> obj_43_beta_0_to_fp16 = const()[name = string("obj_43_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113440000)))];
            fp16 obj_43_epsilon_0_to_fp16 = const()[name = string("obj_43_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_13_cast_fp16)[name = string("obj_43_cast_fp16")];
            string query_9_pad_type_0 = const()[name = string("query_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_9_strides_0 = const()[name = string("query_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_9_pad_0 = const()[name = string("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_9_dilations_0 = const()[name = string("query_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_9_groups_0 = const()[name = string("query_9_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113441600)))];
            tensor<fp16, [768]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114621312)))];
            tensor<fp16, [1, 768, 1, 1]> query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = string("query_9_cast_fp16")];
            string current_key_5_pad_type_0 = const()[name = string("current_key_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_5_strides_0 = const()[name = string("current_key_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_5_pad_0 = const()[name = string("current_key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_5_dilations_0 = const()[name = string("current_key_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_5_groups_0 = const()[name = string("current_key_5_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114622912)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_5_cast_fp16 = conv(dilations = current_key_5_dilations_0, groups = current_key_5_groups_0, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = current_key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = string("current_key_5_cast_fp16")];
            string current_value_5_pad_type_0 = const()[name = string("current_value_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_5_strides_0 = const()[name = string("current_value_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_5_pad_0 = const()[name = string("current_value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_5_dilations_0 = const()[name = string("current_value_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_5_groups_0 = const()[name = string("current_value_5_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115802624)))];
            tensor<fp16, [768]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116982336)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = current_value_5_dilations_0, groups = current_value_5_groups_0, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = current_value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = string("current_value_5_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_589_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_169_cast_fp16)[name = string("op_589_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_5_cast_fp16 = add(x = var_65_cast_fp16_2, y = var_589_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_591_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_169_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_5_cast_fp16 = add(x = var_80_cast_fp16_2, y = var_591_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<int32, [4]> var_594 = const()[name = string("op_594"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_9_cast_fp16 = reshape(shape = var_594, x = query_9_cast_fp16)[name = string("mh_q_9_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_597_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<int32, [4]> var_598 = const()[name = string("op_598"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_599_cast_fp16 = reshape(shape = var_598, x = key_5_cast_fp16)[name = string("op_599_cast_fp16")];
            bool mh_w_17_transpose_x_0 = const()[name = string("mh_w_17_transpose_x_0"), val = bool(true)];
            bool mh_w_17_transpose_y_0 = const()[name = string("mh_w_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_597_cast_fp16, y = var_599_cast_fp16)[name = string("mh_w_17_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_19_cast_fp16 = add(x = mh_w_17_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_19_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_607_cast_fp16 = softmax(axis = var_526, x = mh_w_19_cast_fp16)[name = string("op_607_cast_fp16")];
            tensor<int32, [4]> var_608 = const()[name = string("op_608"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_609_cast_fp16 = reshape(shape = var_608, x = value_5_cast_fp16)[name = string("op_609_cast_fp16")];
            bool attn_9_transpose_x_0 = const()[name = string("attn_9_transpose_x_0"), val = bool(false)];
            bool attn_9_transpose_y_0 = const()[name = string("attn_9_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_609_cast_fp16, y = var_607_cast_fp16)[name = string("attn_9_cast_fp16")];
            tensor<int32, [4]> var_612 = const()[name = string("op_612"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_21_cast_fp16 = reshape(shape = var_612, x = attn_9_cast_fp16)[name = string("input_21_cast_fp16")];
            string obj_49_pad_type_0 = const()[name = string("obj_49_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_49_strides_0 = const()[name = string("obj_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_49_pad_0 = const()[name = string("obj_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_49_dilations_0 = const()[name = string("obj_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_49_groups_0 = const()[name = string("obj_49_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116983936)))];
            tensor<fp16, [768]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118163648)))];
            tensor<fp16, [1, 768, 1, 1]> obj_49_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = string("obj_49_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_49_cast_fp16)[name = string("inputs_15_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_634_to_fp16 = const()[name = string("op_634_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_634_to_fp16, x = inputs_15_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [768]> obj_51_gamma_0_to_fp16 = const()[name = string("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118165248)))];
            tensor<fp16, [768]> obj_51_beta_0_to_fp16 = const()[name = string("obj_51_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118166848)))];
            fp16 obj_51_epsilon_0_to_fp16 = const()[name = string("obj_51_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_15_cast_fp16)[name = string("obj_51_cast_fp16")];
            string query_11_pad_type_0 = const()[name = string("query_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_11_strides_0 = const()[name = string("query_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_11_pad_0 = const()[name = string("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_11_dilations_0 = const()[name = string("query_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_11_groups_0 = const()[name = string("query_11_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118168448)))];
            tensor<fp16, [768]> layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119348160)))];
            tensor<fp16, [1, 768, 1, 1]> query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<int32, [4]> var_654 = const()[name = string("op_654"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_11_cast_fp16 = reshape(shape = var_654, x = query_11_cast_fp16)[name = string("mh_q_11_cast_fp16")];
            fp16 var_656_to_fp16 = const()[name = string("op_656_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_657_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_656_to_fp16)[name = string("op_657_cast_fp16")];
            tensor<int32, [4]> var_658 = const()[name = string("op_658"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_659_cast_fp16 = reshape(shape = var_658, x = obj_53_cast_fp16)[name = string("op_659_cast_fp16")];
            bool mh_w_21_transpose_x_0 = const()[name = string("mh_w_21_transpose_x_0"), val = bool(true)];
            bool mh_w_21_transpose_y_0 = const()[name = string("mh_w_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_657_cast_fp16, y = var_659_cast_fp16)[name = string("mh_w_21_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_23_cast_fp16 = add(x = mh_w_21_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_23_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_59_cast_fp16 = softmax(axis = var_526, x = mh_w_23_cast_fp16)[name = string("obj_59_cast_fp16")];
            tensor<int32, [4]> var_668 = const()[name = string("op_668"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_669_cast_fp16 = reshape(shape = var_668, x = obj_55_cast_fp16)[name = string("op_669_cast_fp16")];
            bool attn_11_transpose_x_0 = const()[name = string("attn_11_transpose_x_0"), val = bool(false)];
            bool attn_11_transpose_y_0 = const()[name = string("attn_11_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_669_cast_fp16, y = obj_59_cast_fp16)[name = string("attn_11_cast_fp16")];
            tensor<int32, [4]> var_672 = const()[name = string("op_672"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_23_cast_fp16 = reshape(shape = var_672, x = attn_11_cast_fp16)[name = string("input_23_cast_fp16")];
            string obj_57_pad_type_0 = const()[name = string("obj_57_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_57_strides_0 = const()[name = string("obj_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_57_pad_0 = const()[name = string("obj_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_57_dilations_0 = const()[name = string("obj_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_57_groups_0 = const()[name = string("obj_57_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119349760)))];
            tensor<fp16, [768]> layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120529472)))];
            tensor<fp16, [1, 768, 1, 1]> obj_57_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = obj_57_dilations_0, groups = obj_57_groups_0, pad = obj_57_pad_0, pad_type = obj_57_pad_type_0, strides = obj_57_strides_0, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = string("obj_57_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_57_cast_fp16)[name = string("inputs_17_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_690_to_fp16 = const()[name = string("op_690_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_690_to_fp16, x = inputs_17_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [768]> input_25_gamma_0_to_fp16 = const()[name = string("input_25_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120531072)))];
            tensor<fp16, [768]> input_25_beta_0_to_fp16 = const()[name = string("input_25_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120532672)))];
            fp16 input_25_epsilon_0_to_fp16 = const()[name = string("input_25_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_17_cast_fp16)[name = string("input_25_cast_fp16")];
            string input_27_pad_type_0 = const()[name = string("input_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_27_strides_0 = const()[name = string("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_27_pad_0 = const()[name = string("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_27_dilations_0 = const()[name = string("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_27_groups_0 = const()[name = string("input_27_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_2_fc1_weight_to_fp16 = const()[name = string("layers_2_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120534272)))];
            tensor<fp16, [3072]> layers_2_fc1_bias_to_fp16 = const()[name = string("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125252928)))];
            tensor<fp16, [1, 3072, 1, 1]> input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = string("input_27_cast_fp16")];
            string input_29_mode_0 = const()[name = string("input_29_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_2_fc2_weight_to_fp16 = const()[name = string("layers_2_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125259136)))];
            tensor<fp16, [768]> layers_2_fc2_bias_to_fp16 = const()[name = string("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129977792)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("inputs_19_cast_fp16")];
            tensor<int32, [4]> obj_71_begin_0 = const()[name = string("obj_71_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> obj_71_end_0 = const()[name = string("obj_71_end_0"), val = tensor<int32, [4]>([4, 768, 1, 1536])];
            tensor<bool, [4]> obj_71_end_mask_0 = const()[name = string("obj_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_71_cast_fp16 = slice_by_index(begin = obj_71_begin_0, end = obj_71_end_0, end_mask = obj_71_end_mask_0, x = read_state_2)[name = string("obj_71_cast_fp16")];
            tensor<int32, [4]> obj_73_begin_0 = const()[name = string("obj_73_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> obj_73_end_0 = const()[name = string("obj_73_end_0"), val = tensor<int32, [4]>([4, 768, 1, 1536])];
            tensor<bool, [4]> obj_73_end_mask_0 = const()[name = string("obj_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_73_cast_fp16 = slice_by_index(begin = obj_73_begin_0, end = obj_73_end_0, end_mask = obj_73_end_mask_0, x = read_state_3)[name = string("obj_73_cast_fp16")];
            int32 var_735 = const()[name = string("op_735"), val = int32(3)];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_760_to_fp16 = const()[name = string("op_760_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_760_to_fp16, x = inputs_19_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [768]> obj_61_gamma_0_to_fp16 = const()[name = string("obj_61_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129979392)))];
            tensor<fp16, [768]> obj_61_beta_0_to_fp16 = const()[name = string("obj_61_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129980992)))];
            fp16 obj_61_epsilon_0_to_fp16 = const()[name = string("obj_61_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_19_cast_fp16)[name = string("obj_61_cast_fp16")];
            string query_13_pad_type_0 = const()[name = string("query_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_13_strides_0 = const()[name = string("query_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_13_pad_0 = const()[name = string("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_13_dilations_0 = const()[name = string("query_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_13_groups_0 = const()[name = string("query_13_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129982592)))];
            tensor<fp16, [768]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131162304)))];
            tensor<fp16, [1, 768, 1, 1]> query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = string("query_13_cast_fp16")];
            string current_key_7_pad_type_0 = const()[name = string("current_key_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_7_strides_0 = const()[name = string("current_key_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_7_pad_0 = const()[name = string("current_key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_7_dilations_0 = const()[name = string("current_key_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_7_groups_0 = const()[name = string("current_key_7_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131163904)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_7_cast_fp16 = conv(dilations = current_key_7_dilations_0, groups = current_key_7_groups_0, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = current_key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = string("current_key_7_cast_fp16")];
            string current_value_7_pad_type_0 = const()[name = string("current_value_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_7_strides_0 = const()[name = string("current_value_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_7_pad_0 = const()[name = string("current_value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_7_dilations_0 = const()[name = string("current_value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_7_groups_0 = const()[name = string("current_value_7_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132343616)))];
            tensor<fp16, [768]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133523328)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = current_value_7_dilations_0, groups = current_value_7_groups_0, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = current_value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = string("current_value_7_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_798_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_169_cast_fp16)[name = string("op_798_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_7_cast_fp16 = add(x = var_65_cast_fp16_3, y = var_798_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_800_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_169_cast_fp16)[name = string("op_800_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_7_cast_fp16 = add(x = var_80_cast_fp16_3, y = var_800_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<int32, [4]> var_803 = const()[name = string("op_803"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_13_cast_fp16 = reshape(shape = var_803, x = query_13_cast_fp16)[name = string("mh_q_13_cast_fp16")];
            fp16 var_805_to_fp16 = const()[name = string("op_805_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_806_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_805_to_fp16)[name = string("op_806_cast_fp16")];
            tensor<int32, [4]> var_807 = const()[name = string("op_807"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_808_cast_fp16 = reshape(shape = var_807, x = key_7_cast_fp16)[name = string("op_808_cast_fp16")];
            bool mh_w_25_transpose_x_0 = const()[name = string("mh_w_25_transpose_x_0"), val = bool(true)];
            bool mh_w_25_transpose_y_0 = const()[name = string("mh_w_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_806_cast_fp16, y = var_808_cast_fp16)[name = string("mh_w_25_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_27_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_816_cast_fp16 = softmax(axis = var_735, x = mh_w_27_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<int32, [4]> var_817 = const()[name = string("op_817"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_818_cast_fp16 = reshape(shape = var_817, x = value_7_cast_fp16)[name = string("op_818_cast_fp16")];
            bool attn_13_transpose_x_0 = const()[name = string("attn_13_transpose_x_0"), val = bool(false)];
            bool attn_13_transpose_y_0 = const()[name = string("attn_13_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_818_cast_fp16, y = var_816_cast_fp16)[name = string("attn_13_cast_fp16")];
            tensor<int32, [4]> var_821 = const()[name = string("op_821"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_31_cast_fp16 = reshape(shape = var_821, x = attn_13_cast_fp16)[name = string("input_31_cast_fp16")];
            string obj_67_pad_type_0 = const()[name = string("obj_67_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_67_strides_0 = const()[name = string("obj_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_67_pad_0 = const()[name = string("obj_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_67_dilations_0 = const()[name = string("obj_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_67_groups_0 = const()[name = string("obj_67_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133524928)))];
            tensor<fp16, [768]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134704640)))];
            tensor<fp16, [1, 768, 1, 1]> obj_67_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = string("obj_67_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_67_cast_fp16)[name = string("inputs_21_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_843_to_fp16 = const()[name = string("op_843_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_843_to_fp16, x = inputs_21_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [768]> obj_69_gamma_0_to_fp16 = const()[name = string("obj_69_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134706240)))];
            tensor<fp16, [768]> obj_69_beta_0_to_fp16 = const()[name = string("obj_69_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134707840)))];
            fp16 obj_69_epsilon_0_to_fp16 = const()[name = string("obj_69_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_21_cast_fp16)[name = string("obj_69_cast_fp16")];
            string query_15_pad_type_0 = const()[name = string("query_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_15_strides_0 = const()[name = string("query_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_15_pad_0 = const()[name = string("query_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_15_dilations_0 = const()[name = string("query_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_15_groups_0 = const()[name = string("query_15_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134709440)))];
            tensor<fp16, [768]> layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135889152)))];
            tensor<fp16, [1, 768, 1, 1]> query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<int32, [4]> var_863 = const()[name = string("op_863"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_15_cast_fp16 = reshape(shape = var_863, x = query_15_cast_fp16)[name = string("mh_q_15_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_866_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<int32, [4]> var_867 = const()[name = string("op_867"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_868_cast_fp16 = reshape(shape = var_867, x = obj_71_cast_fp16)[name = string("op_868_cast_fp16")];
            bool mh_w_29_transpose_x_0 = const()[name = string("mh_w_29_transpose_x_0"), val = bool(true)];
            bool mh_w_29_transpose_y_0 = const()[name = string("mh_w_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_866_cast_fp16, y = var_868_cast_fp16)[name = string("mh_w_29_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_31_cast_fp16 = add(x = mh_w_29_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_31_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_77_cast_fp16 = softmax(axis = var_735, x = mh_w_31_cast_fp16)[name = string("obj_77_cast_fp16")];
            tensor<int32, [4]> var_877 = const()[name = string("op_877"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_878_cast_fp16 = reshape(shape = var_877, x = obj_73_cast_fp16)[name = string("op_878_cast_fp16")];
            bool attn_15_transpose_x_0 = const()[name = string("attn_15_transpose_x_0"), val = bool(false)];
            bool attn_15_transpose_y_0 = const()[name = string("attn_15_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_878_cast_fp16, y = obj_77_cast_fp16)[name = string("attn_15_cast_fp16")];
            tensor<int32, [4]> var_881 = const()[name = string("op_881"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_33_cast_fp16 = reshape(shape = var_881, x = attn_15_cast_fp16)[name = string("input_33_cast_fp16")];
            string obj_75_pad_type_0 = const()[name = string("obj_75_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_75_strides_0 = const()[name = string("obj_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_75_pad_0 = const()[name = string("obj_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_75_dilations_0 = const()[name = string("obj_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_75_groups_0 = const()[name = string("obj_75_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135890752)))];
            tensor<fp16, [768]> layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137070464)))];
            tensor<fp16, [1, 768, 1, 1]> obj_75_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = obj_75_dilations_0, groups = obj_75_groups_0, pad = obj_75_pad_0, pad_type = obj_75_pad_type_0, strides = obj_75_strides_0, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = string("obj_75_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_75_cast_fp16)[name = string("inputs_23_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_899_to_fp16 = const()[name = string("op_899_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_899_to_fp16, x = inputs_23_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [768]> input_35_gamma_0_to_fp16 = const()[name = string("input_35_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137072064)))];
            tensor<fp16, [768]> input_35_beta_0_to_fp16 = const()[name = string("input_35_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137073664)))];
            fp16 input_35_epsilon_0_to_fp16 = const()[name = string("input_35_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_23_cast_fp16)[name = string("input_35_cast_fp16")];
            string input_37_pad_type_0 = const()[name = string("input_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_37_strides_0 = const()[name = string("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_37_pad_0 = const()[name = string("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_37_dilations_0 = const()[name = string("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_37_groups_0 = const()[name = string("input_37_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_3_fc1_weight_to_fp16 = const()[name = string("layers_3_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137075264)))];
            tensor<fp16, [3072]> layers_3_fc1_bias_to_fp16 = const()[name = string("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141793920)))];
            tensor<fp16, [1, 3072, 1, 1]> input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = string("input_37_cast_fp16")];
            string input_39_mode_0 = const()[name = string("input_39_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = string("input_39_cast_fp16")];
            string hidden_states_9_pad_type_0 = const()[name = string("hidden_states_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = string("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = string("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = string("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_9_groups_0 = const()[name = string("hidden_states_9_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = string("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141800128)))];
            tensor<fp16, [768]> layers_3_fc2_bias_to_fp16 = const()[name = string("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146518784)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = string("inputs_25_cast_fp16")];
            tensor<int32, [4]> obj_89_begin_0 = const()[name = string("obj_89_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> obj_89_end_0 = const()[name = string("obj_89_end_0"), val = tensor<int32, [4]>([5, 768, 1, 1536])];
            tensor<bool, [4]> obj_89_end_mask_0 = const()[name = string("obj_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_89_cast_fp16 = slice_by_index(begin = obj_89_begin_0, end = obj_89_end_0, end_mask = obj_89_end_mask_0, x = read_state_2)[name = string("obj_89_cast_fp16")];
            tensor<int32, [4]> obj_91_begin_0 = const()[name = string("obj_91_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> obj_91_end_0 = const()[name = string("obj_91_end_0"), val = tensor<int32, [4]>([5, 768, 1, 1536])];
            tensor<bool, [4]> obj_91_end_mask_0 = const()[name = string("obj_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_91_cast_fp16 = slice_by_index(begin = obj_91_begin_0, end = obj_91_end_0, end_mask = obj_91_end_mask_0, x = read_state_3)[name = string("obj_91_cast_fp16")];
            int32 var_944 = const()[name = string("op_944"), val = int32(3)];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_969_to_fp16 = const()[name = string("op_969_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_969_to_fp16, x = inputs_25_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [768]> obj_79_gamma_0_to_fp16 = const()[name = string("obj_79_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146520384)))];
            tensor<fp16, [768]> obj_79_beta_0_to_fp16 = const()[name = string("obj_79_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146521984)))];
            fp16 obj_79_epsilon_0_to_fp16 = const()[name = string("obj_79_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_25_cast_fp16)[name = string("obj_79_cast_fp16")];
            string query_17_pad_type_0 = const()[name = string("query_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_17_strides_0 = const()[name = string("query_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_17_pad_0 = const()[name = string("query_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_17_dilations_0 = const()[name = string("query_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_17_groups_0 = const()[name = string("query_17_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146523584)))];
            tensor<fp16, [768]> layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147703296)))];
            tensor<fp16, [1, 768, 1, 1]> query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = string("query_17_cast_fp16")];
            string current_key_9_pad_type_0 = const()[name = string("current_key_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_9_strides_0 = const()[name = string("current_key_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_9_pad_0 = const()[name = string("current_key_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_9_dilations_0 = const()[name = string("current_key_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_9_groups_0 = const()[name = string("current_key_9_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147704896)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_9_cast_fp16 = conv(dilations = current_key_9_dilations_0, groups = current_key_9_groups_0, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = current_key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = string("current_key_9_cast_fp16")];
            string current_value_9_pad_type_0 = const()[name = string("current_value_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_9_strides_0 = const()[name = string("current_value_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_9_pad_0 = const()[name = string("current_value_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_9_dilations_0 = const()[name = string("current_value_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_9_groups_0 = const()[name = string("current_value_9_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148884608)))];
            tensor<fp16, [768]> layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150064320)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = current_value_9_dilations_0, groups = current_value_9_groups_0, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = current_value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = string("current_value_9_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1007_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_169_cast_fp16)[name = string("op_1007_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_9_cast_fp16 = add(x = var_65_cast_fp16_4, y = var_1007_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1009_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_169_cast_fp16)[name = string("op_1009_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_9_cast_fp16 = add(x = var_80_cast_fp16_4, y = var_1009_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<int32, [4]> var_1012 = const()[name = string("op_1012"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_17_cast_fp16 = reshape(shape = var_1012, x = query_17_cast_fp16)[name = string("mh_q_17_cast_fp16")];
            fp16 var_1014_to_fp16 = const()[name = string("op_1014_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1015_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1014_to_fp16)[name = string("op_1015_cast_fp16")];
            tensor<int32, [4]> var_1016 = const()[name = string("op_1016"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1017_cast_fp16 = reshape(shape = var_1016, x = key_9_cast_fp16)[name = string("op_1017_cast_fp16")];
            bool mh_w_33_transpose_x_0 = const()[name = string("mh_w_33_transpose_x_0"), val = bool(true)];
            bool mh_w_33_transpose_y_0 = const()[name = string("mh_w_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_1015_cast_fp16, y = var_1017_cast_fp16)[name = string("mh_w_33_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_35_cast_fp16 = add(x = mh_w_33_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_35_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_1025_cast_fp16 = softmax(axis = var_944, x = mh_w_35_cast_fp16)[name = string("op_1025_cast_fp16")];
            tensor<int32, [4]> var_1026 = const()[name = string("op_1026"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1027_cast_fp16 = reshape(shape = var_1026, x = value_9_cast_fp16)[name = string("op_1027_cast_fp16")];
            bool attn_17_transpose_x_0 = const()[name = string("attn_17_transpose_x_0"), val = bool(false)];
            bool attn_17_transpose_y_0 = const()[name = string("attn_17_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1027_cast_fp16, y = var_1025_cast_fp16)[name = string("attn_17_cast_fp16")];
            tensor<int32, [4]> var_1030 = const()[name = string("op_1030"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_41_cast_fp16 = reshape(shape = var_1030, x = attn_17_cast_fp16)[name = string("input_41_cast_fp16")];
            string obj_85_pad_type_0 = const()[name = string("obj_85_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_85_strides_0 = const()[name = string("obj_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_85_pad_0 = const()[name = string("obj_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_85_dilations_0 = const()[name = string("obj_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_85_groups_0 = const()[name = string("obj_85_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150065920)))];
            tensor<fp16, [768]> layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151245632)))];
            tensor<fp16, [1, 768, 1, 1]> obj_85_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_85_dilations_0, groups = obj_85_groups_0, pad = obj_85_pad_0, pad_type = obj_85_pad_type_0, strides = obj_85_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = string("obj_85_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_85_cast_fp16)[name = string("inputs_27_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1052_to_fp16 = const()[name = string("op_1052_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1052_to_fp16, x = inputs_27_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [768]> obj_87_gamma_0_to_fp16 = const()[name = string("obj_87_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151247232)))];
            tensor<fp16, [768]> obj_87_beta_0_to_fp16 = const()[name = string("obj_87_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151248832)))];
            fp16 obj_87_epsilon_0_to_fp16 = const()[name = string("obj_87_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_87_cast_fp16 = batch_norm(beta = obj_87_beta_0_to_fp16, epsilon = obj_87_epsilon_0_to_fp16, gamma = obj_87_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_27_cast_fp16)[name = string("obj_87_cast_fp16")];
            string query_19_pad_type_0 = const()[name = string("query_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_19_strides_0 = const()[name = string("query_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_19_pad_0 = const()[name = string("query_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_19_dilations_0 = const()[name = string("query_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_19_groups_0 = const()[name = string("query_19_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151250432)))];
            tensor<fp16, [768]> layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152430144)))];
            tensor<fp16, [1, 768, 1, 1]> query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_87_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<int32, [4]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_19_cast_fp16 = reshape(shape = var_1072, x = query_19_cast_fp16)[name = string("mh_q_19_cast_fp16")];
            fp16 var_1074_to_fp16 = const()[name = string("op_1074_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1075_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1074_to_fp16)[name = string("op_1075_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1077_cast_fp16 = reshape(shape = var_1076, x = obj_89_cast_fp16)[name = string("op_1077_cast_fp16")];
            bool mh_w_37_transpose_x_0 = const()[name = string("mh_w_37_transpose_x_0"), val = bool(true)];
            bool mh_w_37_transpose_y_0 = const()[name = string("mh_w_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1075_cast_fp16, y = var_1077_cast_fp16)[name = string("mh_w_37_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_39_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_95_cast_fp16 = softmax(axis = var_944, x = mh_w_39_cast_fp16)[name = string("obj_95_cast_fp16")];
            tensor<int32, [4]> var_1086 = const()[name = string("op_1086"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1087_cast_fp16 = reshape(shape = var_1086, x = obj_91_cast_fp16)[name = string("op_1087_cast_fp16")];
            bool attn_19_transpose_x_0 = const()[name = string("attn_19_transpose_x_0"), val = bool(false)];
            bool attn_19_transpose_y_0 = const()[name = string("attn_19_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1087_cast_fp16, y = obj_95_cast_fp16)[name = string("attn_19_cast_fp16")];
            tensor<int32, [4]> var_1090 = const()[name = string("op_1090"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_43_cast_fp16 = reshape(shape = var_1090, x = attn_19_cast_fp16)[name = string("input_43_cast_fp16")];
            string obj_93_pad_type_0 = const()[name = string("obj_93_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_93_strides_0 = const()[name = string("obj_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_93_pad_0 = const()[name = string("obj_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_93_dilations_0 = const()[name = string("obj_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_93_groups_0 = const()[name = string("obj_93_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_4_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_4_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152431744)))];
            tensor<fp16, [768]> layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153611456)))];
            tensor<fp16, [1, 768, 1, 1]> obj_93_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = obj_93_dilations_0, groups = obj_93_groups_0, pad = obj_93_pad_0, pad_type = obj_93_pad_type_0, strides = obj_93_strides_0, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = string("obj_93_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_93_cast_fp16)[name = string("inputs_29_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1108_to_fp16 = const()[name = string("op_1108_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1108_to_fp16, x = inputs_29_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [768]> input_45_gamma_0_to_fp16 = const()[name = string("input_45_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153613056)))];
            tensor<fp16, [768]> input_45_beta_0_to_fp16 = const()[name = string("input_45_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153614656)))];
            fp16 input_45_epsilon_0_to_fp16 = const()[name = string("input_45_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_29_cast_fp16)[name = string("input_45_cast_fp16")];
            string input_47_pad_type_0 = const()[name = string("input_47_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_47_strides_0 = const()[name = string("input_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_47_pad_0 = const()[name = string("input_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_47_dilations_0 = const()[name = string("input_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_47_groups_0 = const()[name = string("input_47_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_4_fc1_weight_to_fp16 = const()[name = string("layers_4_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153616256)))];
            tensor<fp16, [3072]> layers_4_fc1_bias_to_fp16 = const()[name = string("layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158334912)))];
            tensor<fp16, [1, 3072, 1, 1]> input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            string input_49_mode_0 = const()[name = string("input_49_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = string("input_49_cast_fp16")];
            string hidden_states_11_pad_type_0 = const()[name = string("hidden_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_11_strides_0 = const()[name = string("hidden_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = string("hidden_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_11_dilations_0 = const()[name = string("hidden_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_11_groups_0 = const()[name = string("hidden_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_4_fc2_weight_to_fp16 = const()[name = string("layers_4_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158341120)))];
            tensor<fp16, [768]> layers_4_fc2_bias_to_fp16 = const()[name = string("layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163059776)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = string("hidden_states_11_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = string("inputs_31_cast_fp16")];
            tensor<int32, [4]> obj_107_begin_0 = const()[name = string("obj_107_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> obj_107_end_0 = const()[name = string("obj_107_end_0"), val = tensor<int32, [4]>([6, 768, 1, 1536])];
            tensor<bool, [4]> obj_107_end_mask_0 = const()[name = string("obj_107_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_107_cast_fp16 = slice_by_index(begin = obj_107_begin_0, end = obj_107_end_0, end_mask = obj_107_end_mask_0, x = read_state_2)[name = string("obj_107_cast_fp16")];
            tensor<int32, [4]> obj_109_begin_0 = const()[name = string("obj_109_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> obj_109_end_0 = const()[name = string("obj_109_end_0"), val = tensor<int32, [4]>([6, 768, 1, 1536])];
            tensor<bool, [4]> obj_109_end_mask_0 = const()[name = string("obj_109_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_109_cast_fp16 = slice_by_index(begin = obj_109_begin_0, end = obj_109_end_0, end_mask = obj_109_end_mask_0, x = read_state_3)[name = string("obj_109_cast_fp16")];
            int32 var_1153 = const()[name = string("op_1153"), val = int32(3)];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1178_to_fp16 = const()[name = string("op_1178_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1178_to_fp16, x = inputs_31_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [768]> obj_97_gamma_0_to_fp16 = const()[name = string("obj_97_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163061376)))];
            tensor<fp16, [768]> obj_97_beta_0_to_fp16 = const()[name = string("obj_97_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163062976)))];
            fp16 obj_97_epsilon_0_to_fp16 = const()[name = string("obj_97_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_31_cast_fp16)[name = string("obj_97_cast_fp16")];
            string query_21_pad_type_0 = const()[name = string("query_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_21_strides_0 = const()[name = string("query_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_21_pad_0 = const()[name = string("query_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_21_dilations_0 = const()[name = string("query_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_21_groups_0 = const()[name = string("query_21_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163064576)))];
            tensor<fp16, [768]> layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164244288)))];
            tensor<fp16, [1, 768, 1, 1]> query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = string("query_21_cast_fp16")];
            string current_key_11_pad_type_0 = const()[name = string("current_key_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_11_strides_0 = const()[name = string("current_key_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_11_pad_0 = const()[name = string("current_key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_11_dilations_0 = const()[name = string("current_key_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_11_groups_0 = const()[name = string("current_key_11_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164245888)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_11_cast_fp16 = conv(dilations = current_key_11_dilations_0, groups = current_key_11_groups_0, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = current_key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = string("current_key_11_cast_fp16")];
            string current_value_11_pad_type_0 = const()[name = string("current_value_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_11_strides_0 = const()[name = string("current_value_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_11_pad_0 = const()[name = string("current_value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_11_dilations_0 = const()[name = string("current_value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_11_groups_0 = const()[name = string("current_value_11_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165425600)))];
            tensor<fp16, [768]> layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166605312)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = current_value_11_dilations_0, groups = current_value_11_groups_0, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = current_value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = string("current_value_11_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1216_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_169_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_11_cast_fp16 = add(x = var_65_cast_fp16_5, y = var_1216_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1218_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_169_cast_fp16)[name = string("op_1218_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_11_cast_fp16 = add(x = var_80_cast_fp16_5, y = var_1218_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_21_cast_fp16 = reshape(shape = var_1221, x = query_21_cast_fp16)[name = string("mh_q_21_cast_fp16")];
            fp16 var_1223_to_fp16 = const()[name = string("op_1223_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1224_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1223_to_fp16)[name = string("op_1224_cast_fp16")];
            tensor<int32, [4]> var_1225 = const()[name = string("op_1225"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1226_cast_fp16 = reshape(shape = var_1225, x = key_11_cast_fp16)[name = string("op_1226_cast_fp16")];
            bool mh_w_41_transpose_x_0 = const()[name = string("mh_w_41_transpose_x_0"), val = bool(true)];
            bool mh_w_41_transpose_y_0 = const()[name = string("mh_w_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1224_cast_fp16, y = var_1226_cast_fp16)[name = string("mh_w_41_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_43_cast_fp16 = add(x = mh_w_41_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_43_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_1234_cast_fp16 = softmax(axis = var_1153, x = mh_w_43_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<int32, [4]> var_1235 = const()[name = string("op_1235"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1236_cast_fp16 = reshape(shape = var_1235, x = value_11_cast_fp16)[name = string("op_1236_cast_fp16")];
            bool attn_21_transpose_x_0 = const()[name = string("attn_21_transpose_x_0"), val = bool(false)];
            bool attn_21_transpose_y_0 = const()[name = string("attn_21_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1236_cast_fp16, y = var_1234_cast_fp16)[name = string("attn_21_cast_fp16")];
            tensor<int32, [4]> var_1239 = const()[name = string("op_1239"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_51_cast_fp16 = reshape(shape = var_1239, x = attn_21_cast_fp16)[name = string("input_51_cast_fp16")];
            string obj_103_pad_type_0 = const()[name = string("obj_103_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_103_strides_0 = const()[name = string("obj_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_103_pad_0 = const()[name = string("obj_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_103_dilations_0 = const()[name = string("obj_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_103_groups_0 = const()[name = string("obj_103_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166606912)))];
            tensor<fp16, [768]> layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167786624)))];
            tensor<fp16, [1, 768, 1, 1]> obj_103_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_103_dilations_0, groups = obj_103_groups_0, pad = obj_103_pad_0, pad_type = obj_103_pad_type_0, strides = obj_103_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = string("obj_103_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_103_cast_fp16)[name = string("inputs_33_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1261_to_fp16 = const()[name = string("op_1261_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1261_to_fp16, x = inputs_33_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [768]> obj_105_gamma_0_to_fp16 = const()[name = string("obj_105_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167788224)))];
            tensor<fp16, [768]> obj_105_beta_0_to_fp16 = const()[name = string("obj_105_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167789824)))];
            fp16 obj_105_epsilon_0_to_fp16 = const()[name = string("obj_105_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_33_cast_fp16)[name = string("obj_105_cast_fp16")];
            string query_23_pad_type_0 = const()[name = string("query_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_23_strides_0 = const()[name = string("query_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_23_pad_0 = const()[name = string("query_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_23_dilations_0 = const()[name = string("query_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_23_groups_0 = const()[name = string("query_23_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167791424)))];
            tensor<fp16, [768]> layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168971136)))];
            tensor<fp16, [1, 768, 1, 1]> query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<int32, [4]> var_1281 = const()[name = string("op_1281"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_23_cast_fp16 = reshape(shape = var_1281, x = query_23_cast_fp16)[name = string("mh_q_23_cast_fp16")];
            fp16 var_1283_to_fp16 = const()[name = string("op_1283_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1284_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1283_to_fp16)[name = string("op_1284_cast_fp16")];
            tensor<int32, [4]> var_1285 = const()[name = string("op_1285"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1286_cast_fp16 = reshape(shape = var_1285, x = obj_107_cast_fp16)[name = string("op_1286_cast_fp16")];
            bool mh_w_45_transpose_x_0 = const()[name = string("mh_w_45_transpose_x_0"), val = bool(true)];
            bool mh_w_45_transpose_y_0 = const()[name = string("mh_w_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_1284_cast_fp16, y = var_1286_cast_fp16)[name = string("mh_w_45_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_47_cast_fp16 = add(x = mh_w_45_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_47_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_113_cast_fp16 = softmax(axis = var_1153, x = mh_w_47_cast_fp16)[name = string("obj_113_cast_fp16")];
            tensor<int32, [4]> var_1295 = const()[name = string("op_1295"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1296_cast_fp16 = reshape(shape = var_1295, x = obj_109_cast_fp16)[name = string("op_1296_cast_fp16")];
            bool attn_23_transpose_x_0 = const()[name = string("attn_23_transpose_x_0"), val = bool(false)];
            bool attn_23_transpose_y_0 = const()[name = string("attn_23_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1296_cast_fp16, y = obj_113_cast_fp16)[name = string("attn_23_cast_fp16")];
            tensor<int32, [4]> var_1299 = const()[name = string("op_1299"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_53_cast_fp16 = reshape(shape = var_1299, x = attn_23_cast_fp16)[name = string("input_53_cast_fp16")];
            string obj_111_pad_type_0 = const()[name = string("obj_111_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_111_strides_0 = const()[name = string("obj_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_111_pad_0 = const()[name = string("obj_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_111_dilations_0 = const()[name = string("obj_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_111_groups_0 = const()[name = string("obj_111_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168972736)))];
            tensor<fp16, [768]> layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170152448)))];
            tensor<fp16, [1, 768, 1, 1]> obj_111_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = obj_111_dilations_0, groups = obj_111_groups_0, pad = obj_111_pad_0, pad_type = obj_111_pad_type_0, strides = obj_111_strides_0, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("obj_111_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_111_cast_fp16)[name = string("inputs_35_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1320_to_fp16 = const()[name = string("op_1320_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1320_to_fp16, x = inputs_35_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [768]> input_55_gamma_0_to_fp16 = const()[name = string("input_55_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170154048)))];
            tensor<fp16, [768]> input_55_beta_0_to_fp16 = const()[name = string("input_55_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170155648)))];
            fp16 input_55_epsilon_0_to_fp16 = const()[name = string("input_55_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_55_cast_fp16 = batch_norm(beta = input_55_beta_0_to_fp16, epsilon = input_55_epsilon_0_to_fp16, gamma = input_55_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_35_cast_fp16)[name = string("input_55_cast_fp16")];
            string input_57_pad_type_0 = const()[name = string("input_57_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_57_strides_0 = const()[name = string("input_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_57_pad_0 = const()[name = string("input_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_57_dilations_0 = const()[name = string("input_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_57_groups_0 = const()[name = string("input_57_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_5_fc1_weight_to_fp16 = const()[name = string("layers_5_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170157248)))];
            tensor<fp16, [3072]> layers_5_fc1_bias_to_fp16 = const()[name = string("layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174875904)))];
            tensor<fp16, [1, 3072, 1, 1]> input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = string("input_57_cast_fp16")];
            string input_59_mode_0 = const()[name = string("input_59_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_5_fc2_weight_to_fp16 = const()[name = string("layers_5_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174882112)))];
            tensor<fp16, [768]> layers_5_fc2_bias_to_fp16 = const()[name = string("layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179600768)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("inputs_37_cast_fp16")];
            tensor<int32, [4]> obj_125_begin_0 = const()[name = string("obj_125_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> obj_125_end_0 = const()[name = string("obj_125_end_0"), val = tensor<int32, [4]>([7, 768, 1, 1536])];
            tensor<bool, [4]> obj_125_end_mask_0 = const()[name = string("obj_125_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_125_cast_fp16 = slice_by_index(begin = obj_125_begin_0, end = obj_125_end_0, end_mask = obj_125_end_mask_0, x = read_state_2)[name = string("obj_125_cast_fp16")];
            tensor<int32, [4]> obj_127_begin_0 = const()[name = string("obj_127_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> obj_127_end_0 = const()[name = string("obj_127_end_0"), val = tensor<int32, [4]>([7, 768, 1, 1536])];
            tensor<bool, [4]> obj_127_end_mask_0 = const()[name = string("obj_127_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_127_cast_fp16 = slice_by_index(begin = obj_127_begin_0, end = obj_127_end_0, end_mask = obj_127_end_mask_0, x = read_state_3)[name = string("obj_127_cast_fp16")];
            int32 var_1366 = const()[name = string("op_1366"), val = int32(3)];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1391_to_fp16 = const()[name = string("op_1391_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1391_to_fp16, x = inputs_37_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [768]> obj_115_gamma_0_to_fp16 = const()[name = string("obj_115_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179602368)))];
            tensor<fp16, [768]> obj_115_beta_0_to_fp16 = const()[name = string("obj_115_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179603968)))];
            fp16 obj_115_epsilon_0_to_fp16 = const()[name = string("obj_115_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_115_cast_fp16 = batch_norm(beta = obj_115_beta_0_to_fp16, epsilon = obj_115_epsilon_0_to_fp16, gamma = obj_115_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_37_cast_fp16)[name = string("obj_115_cast_fp16")];
            string query_25_pad_type_0 = const()[name = string("query_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_25_strides_0 = const()[name = string("query_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_25_pad_0 = const()[name = string("query_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_25_dilations_0 = const()[name = string("query_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_25_groups_0 = const()[name = string("query_25_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179605568)))];
            tensor<fp16, [768]> layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180785280)))];
            tensor<fp16, [1, 768, 1, 1]> query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_115_cast_fp16)[name = string("query_25_cast_fp16")];
            string current_key_13_pad_type_0 = const()[name = string("current_key_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_13_strides_0 = const()[name = string("current_key_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_13_pad_0 = const()[name = string("current_key_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_13_dilations_0 = const()[name = string("current_key_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_13_groups_0 = const()[name = string("current_key_13_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180786880)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_13_cast_fp16 = conv(dilations = current_key_13_dilations_0, groups = current_key_13_groups_0, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = current_key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_115_cast_fp16)[name = string("current_key_13_cast_fp16")];
            string current_value_13_pad_type_0 = const()[name = string("current_value_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_13_strides_0 = const()[name = string("current_value_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_13_pad_0 = const()[name = string("current_value_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_13_dilations_0 = const()[name = string("current_value_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_13_groups_0 = const()[name = string("current_value_13_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181966592)))];
            tensor<fp16, [768]> layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183146304)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = current_value_13_dilations_0, groups = current_value_13_groups_0, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = current_value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_115_cast_fp16)[name = string("current_value_13_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1429_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_169_cast_fp16)[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_13_cast_fp16 = add(x = var_65_cast_fp16_6, y = var_1429_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1431_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_169_cast_fp16)[name = string("op_1431_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_13_cast_fp16 = add(x = var_80_cast_fp16_6, y = var_1431_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<int32, [4]> var_1434 = const()[name = string("op_1434"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_25_cast_fp16 = reshape(shape = var_1434, x = query_25_cast_fp16)[name = string("mh_q_25_cast_fp16")];
            fp16 var_1436_to_fp16 = const()[name = string("op_1436_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1437_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1436_to_fp16)[name = string("op_1437_cast_fp16")];
            tensor<int32, [4]> var_1438 = const()[name = string("op_1438"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1439_cast_fp16 = reshape(shape = var_1438, x = key_13_cast_fp16)[name = string("op_1439_cast_fp16")];
            bool mh_w_49_transpose_x_0 = const()[name = string("mh_w_49_transpose_x_0"), val = bool(true)];
            bool mh_w_49_transpose_y_0 = const()[name = string("mh_w_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_1437_cast_fp16, y = var_1439_cast_fp16)[name = string("mh_w_49_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_51_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_1447_cast_fp16 = softmax(axis = var_1366, x = mh_w_51_cast_fp16)[name = string("op_1447_cast_fp16")];
            tensor<int32, [4]> var_1448 = const()[name = string("op_1448"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1449_cast_fp16 = reshape(shape = var_1448, x = value_13_cast_fp16)[name = string("op_1449_cast_fp16")];
            bool attn_25_transpose_x_0 = const()[name = string("attn_25_transpose_x_0"), val = bool(false)];
            bool attn_25_transpose_y_0 = const()[name = string("attn_25_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1449_cast_fp16, y = var_1447_cast_fp16)[name = string("attn_25_cast_fp16")];
            tensor<int32, [4]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_61_cast_fp16 = reshape(shape = var_1452, x = attn_25_cast_fp16)[name = string("input_61_cast_fp16")];
            string obj_121_pad_type_0 = const()[name = string("obj_121_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_121_strides_0 = const()[name = string("obj_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_121_pad_0 = const()[name = string("obj_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_121_dilations_0 = const()[name = string("obj_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_121_groups_0 = const()[name = string("obj_121_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183147904)))];
            tensor<fp16, [768]> layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184327616)))];
            tensor<fp16, [1, 768, 1, 1]> obj_121_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_121_dilations_0, groups = obj_121_groups_0, pad = obj_121_pad_0, pad_type = obj_121_pad_type_0, strides = obj_121_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = string("obj_121_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_121_cast_fp16)[name = string("inputs_39_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1474_to_fp16 = const()[name = string("op_1474_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1474_to_fp16, x = inputs_39_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [768]> obj_123_gamma_0_to_fp16 = const()[name = string("obj_123_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184329216)))];
            tensor<fp16, [768]> obj_123_beta_0_to_fp16 = const()[name = string("obj_123_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184330816)))];
            fp16 obj_123_epsilon_0_to_fp16 = const()[name = string("obj_123_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_123_cast_fp16 = batch_norm(beta = obj_123_beta_0_to_fp16, epsilon = obj_123_epsilon_0_to_fp16, gamma = obj_123_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_39_cast_fp16)[name = string("obj_123_cast_fp16")];
            string query_27_pad_type_0 = const()[name = string("query_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_27_strides_0 = const()[name = string("query_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_27_pad_0 = const()[name = string("query_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_27_dilations_0 = const()[name = string("query_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_27_groups_0 = const()[name = string("query_27_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184332416)))];
            tensor<fp16, [768]> layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185512128)))];
            tensor<fp16, [1, 768, 1, 1]> query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_123_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<int32, [4]> var_1494 = const()[name = string("op_1494"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_27_cast_fp16 = reshape(shape = var_1494, x = query_27_cast_fp16)[name = string("mh_q_27_cast_fp16")];
            fp16 var_1496_to_fp16 = const()[name = string("op_1496_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1497_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1496_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [4]> var_1498 = const()[name = string("op_1498"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1499_cast_fp16 = reshape(shape = var_1498, x = obj_125_cast_fp16)[name = string("op_1499_cast_fp16")];
            bool mh_w_53_transpose_x_0 = const()[name = string("mh_w_53_transpose_x_0"), val = bool(true)];
            bool mh_w_53_transpose_y_0 = const()[name = string("mh_w_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_1497_cast_fp16, y = var_1499_cast_fp16)[name = string("mh_w_53_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_55_cast_fp16 = add(x = mh_w_53_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_55_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_131_cast_fp16 = softmax(axis = var_1366, x = mh_w_55_cast_fp16)[name = string("obj_131_cast_fp16")];
            tensor<int32, [4]> var_1508 = const()[name = string("op_1508"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1509_cast_fp16 = reshape(shape = var_1508, x = obj_127_cast_fp16)[name = string("op_1509_cast_fp16")];
            bool attn_27_transpose_x_0 = const()[name = string("attn_27_transpose_x_0"), val = bool(false)];
            bool attn_27_transpose_y_0 = const()[name = string("attn_27_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1509_cast_fp16, y = obj_131_cast_fp16)[name = string("attn_27_cast_fp16")];
            tensor<int32, [4]> var_1512 = const()[name = string("op_1512"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_63_cast_fp16 = reshape(shape = var_1512, x = attn_27_cast_fp16)[name = string("input_63_cast_fp16")];
            string obj_129_pad_type_0 = const()[name = string("obj_129_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_129_strides_0 = const()[name = string("obj_129_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_129_pad_0 = const()[name = string("obj_129_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_129_dilations_0 = const()[name = string("obj_129_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_129_groups_0 = const()[name = string("obj_129_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_6_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_6_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185513728)))];
            tensor<fp16, [768]> layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186693440)))];
            tensor<fp16, [1, 768, 1, 1]> obj_129_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = obj_129_dilations_0, groups = obj_129_groups_0, pad = obj_129_pad_0, pad_type = obj_129_pad_type_0, strides = obj_129_strides_0, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = string("obj_129_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_129_cast_fp16)[name = string("inputs_41_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1530_to_fp16 = const()[name = string("op_1530_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1530_to_fp16, x = inputs_41_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [768]> input_65_gamma_0_to_fp16 = const()[name = string("input_65_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186695040)))];
            tensor<fp16, [768]> input_65_beta_0_to_fp16 = const()[name = string("input_65_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186696640)))];
            fp16 input_65_epsilon_0_to_fp16 = const()[name = string("input_65_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_65_cast_fp16 = batch_norm(beta = input_65_beta_0_to_fp16, epsilon = input_65_epsilon_0_to_fp16, gamma = input_65_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_41_cast_fp16)[name = string("input_65_cast_fp16")];
            string input_67_pad_type_0 = const()[name = string("input_67_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_67_strides_0 = const()[name = string("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_67_pad_0 = const()[name = string("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_67_dilations_0 = const()[name = string("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_67_groups_0 = const()[name = string("input_67_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_6_fc1_weight_to_fp16 = const()[name = string("layers_6_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186698240)))];
            tensor<fp16, [3072]> layers_6_fc1_bias_to_fp16 = const()[name = string("layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191416896)))];
            tensor<fp16, [1, 3072, 1, 1]> input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
            string input_69_mode_0 = const()[name = string("input_69_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = string("input_69_cast_fp16")];
            string hidden_states_15_pad_type_0 = const()[name = string("hidden_states_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_15_strides_0 = const()[name = string("hidden_states_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = string("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_15_dilations_0 = const()[name = string("hidden_states_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_15_groups_0 = const()[name = string("hidden_states_15_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_6_fc2_weight_to_fp16 = const()[name = string("layers_6_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191423104)))];
            tensor<fp16, [768]> layers_6_fc2_bias_to_fp16 = const()[name = string("layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196141760)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = string("inputs_43_cast_fp16")];
            tensor<int32, [4]> obj_143_begin_0 = const()[name = string("obj_143_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> obj_143_end_0 = const()[name = string("obj_143_end_0"), val = tensor<int32, [4]>([8, 768, 1, 1536])];
            tensor<bool, [4]> obj_143_end_mask_0 = const()[name = string("obj_143_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_143_cast_fp16 = slice_by_index(begin = obj_143_begin_0, end = obj_143_end_0, end_mask = obj_143_end_mask_0, x = read_state_2)[name = string("obj_143_cast_fp16")];
            tensor<int32, [4]> obj_145_begin_0 = const()[name = string("obj_145_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> obj_145_end_0 = const()[name = string("obj_145_end_0"), val = tensor<int32, [4]>([8, 768, 1, 1536])];
            tensor<bool, [4]> obj_145_end_mask_0 = const()[name = string("obj_145_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_145_cast_fp16 = slice_by_index(begin = obj_145_begin_0, end = obj_145_end_0, end_mask = obj_145_end_mask_0, x = read_state_3)[name = string("obj_145_cast_fp16")];
            int32 var_1575 = const()[name = string("op_1575"), val = int32(3)];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1600_to_fp16 = const()[name = string("op_1600_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1600_to_fp16, x = inputs_43_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [768]> obj_133_gamma_0_to_fp16 = const()[name = string("obj_133_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196143360)))];
            tensor<fp16, [768]> obj_133_beta_0_to_fp16 = const()[name = string("obj_133_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196144960)))];
            fp16 obj_133_epsilon_0_to_fp16 = const()[name = string("obj_133_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_133_cast_fp16 = batch_norm(beta = obj_133_beta_0_to_fp16, epsilon = obj_133_epsilon_0_to_fp16, gamma = obj_133_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_43_cast_fp16)[name = string("obj_133_cast_fp16")];
            string query_29_pad_type_0 = const()[name = string("query_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_29_strides_0 = const()[name = string("query_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_29_pad_0 = const()[name = string("query_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_29_dilations_0 = const()[name = string("query_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_29_groups_0 = const()[name = string("query_29_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196146560)))];
            tensor<fp16, [768]> layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197326272)))];
            tensor<fp16, [1, 768, 1, 1]> query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_133_cast_fp16)[name = string("query_29_cast_fp16")];
            string current_key_15_pad_type_0 = const()[name = string("current_key_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_15_strides_0 = const()[name = string("current_key_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_15_pad_0 = const()[name = string("current_key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_15_dilations_0 = const()[name = string("current_key_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_15_groups_0 = const()[name = string("current_key_15_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197327872)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_15_cast_fp16 = conv(dilations = current_key_15_dilations_0, groups = current_key_15_groups_0, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = current_key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_133_cast_fp16)[name = string("current_key_15_cast_fp16")];
            string current_value_15_pad_type_0 = const()[name = string("current_value_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_15_strides_0 = const()[name = string("current_value_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_15_pad_0 = const()[name = string("current_value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_15_dilations_0 = const()[name = string("current_value_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_15_groups_0 = const()[name = string("current_value_15_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198507584)))];
            tensor<fp16, [768]> layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199687296)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = current_value_15_dilations_0, groups = current_value_15_groups_0, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = current_value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_133_cast_fp16)[name = string("current_value_15_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1638_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_169_cast_fp16)[name = string("op_1638_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_15_cast_fp16 = add(x = var_65_cast_fp16_7, y = var_1638_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1640_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_169_cast_fp16)[name = string("op_1640_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_15_cast_fp16 = add(x = var_80_cast_fp16_7, y = var_1640_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<int32, [4]> var_1643 = const()[name = string("op_1643"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_29_cast_fp16 = reshape(shape = var_1643, x = query_29_cast_fp16)[name = string("mh_q_29_cast_fp16")];
            fp16 var_1645_to_fp16 = const()[name = string("op_1645_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1646_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1645_to_fp16)[name = string("op_1646_cast_fp16")];
            tensor<int32, [4]> var_1647 = const()[name = string("op_1647"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1648_cast_fp16 = reshape(shape = var_1647, x = key_15_cast_fp16)[name = string("op_1648_cast_fp16")];
            bool mh_w_57_transpose_x_0 = const()[name = string("mh_w_57_transpose_x_0"), val = bool(true)];
            bool mh_w_57_transpose_y_0 = const()[name = string("mh_w_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_1646_cast_fp16, y = var_1648_cast_fp16)[name = string("mh_w_57_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_59_cast_fp16 = add(x = mh_w_57_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_59_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_1656_cast_fp16 = softmax(axis = var_1575, x = mh_w_59_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<int32, [4]> var_1657 = const()[name = string("op_1657"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1658_cast_fp16 = reshape(shape = var_1657, x = value_15_cast_fp16)[name = string("op_1658_cast_fp16")];
            bool attn_29_transpose_x_0 = const()[name = string("attn_29_transpose_x_0"), val = bool(false)];
            bool attn_29_transpose_y_0 = const()[name = string("attn_29_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1658_cast_fp16, y = var_1656_cast_fp16)[name = string("attn_29_cast_fp16")];
            tensor<int32, [4]> var_1661 = const()[name = string("op_1661"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_71_cast_fp16 = reshape(shape = var_1661, x = attn_29_cast_fp16)[name = string("input_71_cast_fp16")];
            string obj_139_pad_type_0 = const()[name = string("obj_139_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_139_strides_0 = const()[name = string("obj_139_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_139_pad_0 = const()[name = string("obj_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_139_dilations_0 = const()[name = string("obj_139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_139_groups_0 = const()[name = string("obj_139_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199688896)))];
            tensor<fp16, [768]> layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200868608)))];
            tensor<fp16, [1, 768, 1, 1]> obj_139_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_139_dilations_0, groups = obj_139_groups_0, pad = obj_139_pad_0, pad_type = obj_139_pad_type_0, strides = obj_139_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = string("obj_139_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_139_cast_fp16)[name = string("inputs_45_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1683_to_fp16 = const()[name = string("op_1683_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1683_to_fp16, x = inputs_45_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [768]> obj_141_gamma_0_to_fp16 = const()[name = string("obj_141_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200870208)))];
            tensor<fp16, [768]> obj_141_beta_0_to_fp16 = const()[name = string("obj_141_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200871808)))];
            fp16 obj_141_epsilon_0_to_fp16 = const()[name = string("obj_141_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_45_cast_fp16)[name = string("obj_141_cast_fp16")];
            string query_31_pad_type_0 = const()[name = string("query_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_31_strides_0 = const()[name = string("query_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_31_pad_0 = const()[name = string("query_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_31_dilations_0 = const()[name = string("query_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_31_groups_0 = const()[name = string("query_31_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200873408)))];
            tensor<fp16, [768]> layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202053120)))];
            tensor<fp16, [1, 768, 1, 1]> query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<int32, [4]> var_1703 = const()[name = string("op_1703"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_31_cast_fp16 = reshape(shape = var_1703, x = query_31_cast_fp16)[name = string("mh_q_31_cast_fp16")];
            fp16 var_1705_to_fp16 = const()[name = string("op_1705_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1706_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_1705_to_fp16)[name = string("op_1706_cast_fp16")];
            tensor<int32, [4]> var_1707 = const()[name = string("op_1707"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1708_cast_fp16 = reshape(shape = var_1707, x = obj_143_cast_fp16)[name = string("op_1708_cast_fp16")];
            bool mh_w_61_transpose_x_0 = const()[name = string("mh_w_61_transpose_x_0"), val = bool(true)];
            bool mh_w_61_transpose_y_0 = const()[name = string("mh_w_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_1706_cast_fp16, y = var_1708_cast_fp16)[name = string("mh_w_61_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_63_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_149_cast_fp16 = softmax(axis = var_1575, x = mh_w_63_cast_fp16)[name = string("obj_149_cast_fp16")];
            tensor<int32, [4]> var_1717 = const()[name = string("op_1717"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1718_cast_fp16 = reshape(shape = var_1717, x = obj_145_cast_fp16)[name = string("op_1718_cast_fp16")];
            bool attn_31_transpose_x_0 = const()[name = string("attn_31_transpose_x_0"), val = bool(false)];
            bool attn_31_transpose_y_0 = const()[name = string("attn_31_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_1718_cast_fp16, y = obj_149_cast_fp16)[name = string("attn_31_cast_fp16")];
            tensor<int32, [4]> var_1721 = const()[name = string("op_1721"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_73_cast_fp16 = reshape(shape = var_1721, x = attn_31_cast_fp16)[name = string("input_73_cast_fp16")];
            string obj_147_pad_type_0 = const()[name = string("obj_147_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_147_strides_0 = const()[name = string("obj_147_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_147_pad_0 = const()[name = string("obj_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_147_dilations_0 = const()[name = string("obj_147_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_147_groups_0 = const()[name = string("obj_147_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_7_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_7_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202054720)))];
            tensor<fp16, [768]> layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203234432)))];
            tensor<fp16, [1, 768, 1, 1]> obj_147_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = obj_147_dilations_0, groups = obj_147_groups_0, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = obj_147_strides_0, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = string("obj_147_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_147_cast_fp16)[name = string("inputs_47_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1739_to_fp16 = const()[name = string("op_1739_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1739_to_fp16, x = inputs_47_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [768]> input_75_gamma_0_to_fp16 = const()[name = string("input_75_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203236032)))];
            tensor<fp16, [768]> input_75_beta_0_to_fp16 = const()[name = string("input_75_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203237632)))];
            fp16 input_75_epsilon_0_to_fp16 = const()[name = string("input_75_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_47_cast_fp16)[name = string("input_75_cast_fp16")];
            string input_77_pad_type_0 = const()[name = string("input_77_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_77_strides_0 = const()[name = string("input_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_77_pad_0 = const()[name = string("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_77_dilations_0 = const()[name = string("input_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_77_groups_0 = const()[name = string("input_77_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_7_fc1_weight_to_fp16 = const()[name = string("layers_7_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203239232)))];
            tensor<fp16, [3072]> layers_7_fc1_bias_to_fp16 = const()[name = string("layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207957888)))];
            tensor<fp16, [1, 3072, 1, 1]> input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
            string input_79_mode_0 = const()[name = string("input_79_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_7_fc2_weight_to_fp16 = const()[name = string("layers_7_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207964096)))];
            tensor<fp16, [768]> layers_7_fc2_bias_to_fp16 = const()[name = string("layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212682752)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("inputs_49_cast_fp16")];
            tensor<int32, [4]> obj_161_begin_0 = const()[name = string("obj_161_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> obj_161_end_0 = const()[name = string("obj_161_end_0"), val = tensor<int32, [4]>([9, 768, 1, 1536])];
            tensor<bool, [4]> obj_161_end_mask_0 = const()[name = string("obj_161_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_161_cast_fp16 = slice_by_index(begin = obj_161_begin_0, end = obj_161_end_0, end_mask = obj_161_end_mask_0, x = read_state_2)[name = string("obj_161_cast_fp16")];
            tensor<int32, [4]> obj_163_begin_0 = const()[name = string("obj_163_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> obj_163_end_0 = const()[name = string("obj_163_end_0"), val = tensor<int32, [4]>([9, 768, 1, 1536])];
            tensor<bool, [4]> obj_163_end_mask_0 = const()[name = string("obj_163_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_163_cast_fp16 = slice_by_index(begin = obj_163_begin_0, end = obj_163_end_0, end_mask = obj_163_end_mask_0, x = read_state_3)[name = string("obj_163_cast_fp16")];
            int32 var_1784 = const()[name = string("op_1784"), val = int32(3)];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1809_to_fp16 = const()[name = string("op_1809_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1809_to_fp16, x = inputs_49_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [768]> obj_151_gamma_0_to_fp16 = const()[name = string("obj_151_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212684352)))];
            tensor<fp16, [768]> obj_151_beta_0_to_fp16 = const()[name = string("obj_151_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212685952)))];
            fp16 obj_151_epsilon_0_to_fp16 = const()[name = string("obj_151_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_151_cast_fp16 = batch_norm(beta = obj_151_beta_0_to_fp16, epsilon = obj_151_epsilon_0_to_fp16, gamma = obj_151_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_49_cast_fp16)[name = string("obj_151_cast_fp16")];
            string query_33_pad_type_0 = const()[name = string("query_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_33_strides_0 = const()[name = string("query_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_33_pad_0 = const()[name = string("query_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_33_dilations_0 = const()[name = string("query_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_33_groups_0 = const()[name = string("query_33_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212687552)))];
            tensor<fp16, [768]> layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213867264)))];
            tensor<fp16, [1, 768, 1, 1]> query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_151_cast_fp16)[name = string("query_33_cast_fp16")];
            string current_key_17_pad_type_0 = const()[name = string("current_key_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_17_strides_0 = const()[name = string("current_key_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_17_pad_0 = const()[name = string("current_key_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_17_dilations_0 = const()[name = string("current_key_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_17_groups_0 = const()[name = string("current_key_17_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213868864)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_17_cast_fp16 = conv(dilations = current_key_17_dilations_0, groups = current_key_17_groups_0, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = current_key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_151_cast_fp16)[name = string("current_key_17_cast_fp16")];
            string current_value_17_pad_type_0 = const()[name = string("current_value_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_17_strides_0 = const()[name = string("current_value_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_17_pad_0 = const()[name = string("current_value_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_17_dilations_0 = const()[name = string("current_value_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_17_groups_0 = const()[name = string("current_value_17_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215048576)))];
            tensor<fp16, [768]> layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216228288)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = current_value_17_dilations_0, groups = current_value_17_groups_0, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = current_value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_151_cast_fp16)[name = string("current_value_17_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1847_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_169_cast_fp16)[name = string("op_1847_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_17_cast_fp16 = add(x = var_65_cast_fp16_8, y = var_1847_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_1849_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_169_cast_fp16)[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_17_cast_fp16 = add(x = var_80_cast_fp16_8, y = var_1849_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<int32, [4]> var_1852 = const()[name = string("op_1852"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_33_cast_fp16 = reshape(shape = var_1852, x = query_33_cast_fp16)[name = string("mh_q_33_cast_fp16")];
            fp16 var_1854_to_fp16 = const()[name = string("op_1854_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1855_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_1854_to_fp16)[name = string("op_1855_cast_fp16")];
            tensor<int32, [4]> var_1856 = const()[name = string("op_1856"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1857_cast_fp16 = reshape(shape = var_1856, x = key_17_cast_fp16)[name = string("op_1857_cast_fp16")];
            bool mh_w_65_transpose_x_0 = const()[name = string("mh_w_65_transpose_x_0"), val = bool(true)];
            bool mh_w_65_transpose_y_0 = const()[name = string("mh_w_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_1855_cast_fp16, y = var_1857_cast_fp16)[name = string("mh_w_65_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_67_cast_fp16 = add(x = mh_w_65_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_67_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_1865_cast_fp16 = softmax(axis = var_1784, x = mh_w_67_cast_fp16)[name = string("op_1865_cast_fp16")];
            tensor<int32, [4]> var_1866 = const()[name = string("op_1866"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_1867_cast_fp16 = reshape(shape = var_1866, x = value_17_cast_fp16)[name = string("op_1867_cast_fp16")];
            bool attn_33_transpose_x_0 = const()[name = string("attn_33_transpose_x_0"), val = bool(false)];
            bool attn_33_transpose_y_0 = const()[name = string("attn_33_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_1867_cast_fp16, y = var_1865_cast_fp16)[name = string("attn_33_cast_fp16")];
            tensor<int32, [4]> var_1870 = const()[name = string("op_1870"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_81_cast_fp16 = reshape(shape = var_1870, x = attn_33_cast_fp16)[name = string("input_81_cast_fp16")];
            string obj_157_pad_type_0 = const()[name = string("obj_157_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_157_strides_0 = const()[name = string("obj_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_157_pad_0 = const()[name = string("obj_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_157_dilations_0 = const()[name = string("obj_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_157_groups_0 = const()[name = string("obj_157_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216229888)))];
            tensor<fp16, [768]> layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217409600)))];
            tensor<fp16, [1, 768, 1, 1]> obj_157_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_157_dilations_0, groups = obj_157_groups_0, pad = obj_157_pad_0, pad_type = obj_157_pad_type_0, strides = obj_157_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = string("obj_157_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_157_cast_fp16)[name = string("inputs_51_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1892_to_fp16 = const()[name = string("op_1892_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1892_to_fp16, x = inputs_51_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [768]> obj_159_gamma_0_to_fp16 = const()[name = string("obj_159_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217411200)))];
            tensor<fp16, [768]> obj_159_beta_0_to_fp16 = const()[name = string("obj_159_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217412800)))];
            fp16 obj_159_epsilon_0_to_fp16 = const()[name = string("obj_159_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_159_cast_fp16 = batch_norm(beta = obj_159_beta_0_to_fp16, epsilon = obj_159_epsilon_0_to_fp16, gamma = obj_159_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_51_cast_fp16)[name = string("obj_159_cast_fp16")];
            string query_35_pad_type_0 = const()[name = string("query_35_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_35_strides_0 = const()[name = string("query_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_35_pad_0 = const()[name = string("query_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_35_dilations_0 = const()[name = string("query_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_35_groups_0 = const()[name = string("query_35_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217414400)))];
            tensor<fp16, [768]> layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218594112)))];
            tensor<fp16, [1, 768, 1, 1]> query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_159_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<int32, [4]> var_1912 = const()[name = string("op_1912"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_35_cast_fp16 = reshape(shape = var_1912, x = query_35_cast_fp16)[name = string("mh_q_35_cast_fp16")];
            fp16 var_1914_to_fp16 = const()[name = string("op_1914_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_1915_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_1914_to_fp16)[name = string("op_1915_cast_fp16")];
            tensor<int32, [4]> var_1916 = const()[name = string("op_1916"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1917_cast_fp16 = reshape(shape = var_1916, x = obj_161_cast_fp16)[name = string("op_1917_cast_fp16")];
            bool mh_w_69_transpose_x_0 = const()[name = string("mh_w_69_transpose_x_0"), val = bool(true)];
            bool mh_w_69_transpose_y_0 = const()[name = string("mh_w_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_69_cast_fp16 = matmul(transpose_x = mh_w_69_transpose_x_0, transpose_y = mh_w_69_transpose_y_0, x = var_1915_cast_fp16, y = var_1917_cast_fp16)[name = string("mh_w_69_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_71_cast_fp16 = add(x = mh_w_69_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_71_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_167_cast_fp16 = softmax(axis = var_1784, x = mh_w_71_cast_fp16)[name = string("obj_167_cast_fp16")];
            tensor<int32, [4]> var_1926 = const()[name = string("op_1926"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_1927_cast_fp16 = reshape(shape = var_1926, x = obj_163_cast_fp16)[name = string("op_1927_cast_fp16")];
            bool attn_35_transpose_x_0 = const()[name = string("attn_35_transpose_x_0"), val = bool(false)];
            bool attn_35_transpose_y_0 = const()[name = string("attn_35_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_1927_cast_fp16, y = obj_167_cast_fp16)[name = string("attn_35_cast_fp16")];
            tensor<int32, [4]> var_1930 = const()[name = string("op_1930"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_83_cast_fp16 = reshape(shape = var_1930, x = attn_35_cast_fp16)[name = string("input_83_cast_fp16")];
            string obj_165_pad_type_0 = const()[name = string("obj_165_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_165_strides_0 = const()[name = string("obj_165_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_165_pad_0 = const()[name = string("obj_165_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_165_dilations_0 = const()[name = string("obj_165_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_165_groups_0 = const()[name = string("obj_165_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_8_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218595712)))];
            tensor<fp16, [768]> layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219775424)))];
            tensor<fp16, [1, 768, 1, 1]> obj_165_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = obj_165_dilations_0, groups = obj_165_groups_0, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = obj_165_strides_0, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = string("obj_165_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_165_cast_fp16)[name = string("inputs_53_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1951_to_fp16 = const()[name = string("op_1951_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1951_to_fp16, x = inputs_53_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [768]> input_85_gamma_0_to_fp16 = const()[name = string("input_85_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219777024)))];
            tensor<fp16, [768]> input_85_beta_0_to_fp16 = const()[name = string("input_85_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219778624)))];
            fp16 input_85_epsilon_0_to_fp16 = const()[name = string("input_85_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_85_cast_fp16 = batch_norm(beta = input_85_beta_0_to_fp16, epsilon = input_85_epsilon_0_to_fp16, gamma = input_85_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_53_cast_fp16)[name = string("input_85_cast_fp16")];
            string input_87_pad_type_0 = const()[name = string("input_87_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_87_strides_0 = const()[name = string("input_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_87_pad_0 = const()[name = string("input_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_87_dilations_0 = const()[name = string("input_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_87_groups_0 = const()[name = string("input_87_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_8_fc1_weight_to_fp16 = const()[name = string("layers_8_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219780224)))];
            tensor<fp16, [3072]> layers_8_fc1_bias_to_fp16 = const()[name = string("layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224498880)))];
            tensor<fp16, [1, 3072, 1, 1]> input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = string("input_87_cast_fp16")];
            string input_89_mode_0 = const()[name = string("input_89_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = string("input_89_cast_fp16")];
            string hidden_states_19_pad_type_0 = const()[name = string("hidden_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_19_strides_0 = const()[name = string("hidden_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_19_pad_0 = const()[name = string("hidden_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_19_dilations_0 = const()[name = string("hidden_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_19_groups_0 = const()[name = string("hidden_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_8_fc2_weight_to_fp16 = const()[name = string("layers_8_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224505088)))];
            tensor<fp16, [768]> layers_8_fc2_bias_to_fp16 = const()[name = string("layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229223744)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = string("inputs_55_cast_fp16")];
            tensor<int32, [4]> obj_179_begin_0 = const()[name = string("obj_179_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> obj_179_end_0 = const()[name = string("obj_179_end_0"), val = tensor<int32, [4]>([10, 768, 1, 1536])];
            tensor<bool, [4]> obj_179_end_mask_0 = const()[name = string("obj_179_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_179_cast_fp16 = slice_by_index(begin = obj_179_begin_0, end = obj_179_end_0, end_mask = obj_179_end_mask_0, x = read_state_2)[name = string("obj_179_cast_fp16")];
            tensor<int32, [4]> obj_181_begin_0 = const()[name = string("obj_181_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> obj_181_end_0 = const()[name = string("obj_181_end_0"), val = tensor<int32, [4]>([10, 768, 1, 1536])];
            tensor<bool, [4]> obj_181_end_mask_0 = const()[name = string("obj_181_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_181_cast_fp16 = slice_by_index(begin = obj_181_begin_0, end = obj_181_end_0, end_mask = obj_181_end_mask_0, x = read_state_3)[name = string("obj_181_cast_fp16")];
            int32 var_1997 = const()[name = string("op_1997"), val = int32(3)];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2022_to_fp16 = const()[name = string("op_2022_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_2022_to_fp16, x = inputs_55_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [768]> obj_169_gamma_0_to_fp16 = const()[name = string("obj_169_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229225344)))];
            tensor<fp16, [768]> obj_169_beta_0_to_fp16 = const()[name = string("obj_169_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229226944)))];
            fp16 obj_169_epsilon_0_to_fp16 = const()[name = string("obj_169_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_169_cast_fp16 = batch_norm(beta = obj_169_beta_0_to_fp16, epsilon = obj_169_epsilon_0_to_fp16, gamma = obj_169_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_55_cast_fp16)[name = string("obj_169_cast_fp16")];
            string query_37_pad_type_0 = const()[name = string("query_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_37_strides_0 = const()[name = string("query_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_37_pad_0 = const()[name = string("query_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_37_dilations_0 = const()[name = string("query_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_37_groups_0 = const()[name = string("query_37_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229228544)))];
            tensor<fp16, [768]> layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230408256)))];
            tensor<fp16, [1, 768, 1, 1]> query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = string("query_37_cast_fp16")];
            string current_key_19_pad_type_0 = const()[name = string("current_key_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_19_strides_0 = const()[name = string("current_key_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_19_pad_0 = const()[name = string("current_key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_19_dilations_0 = const()[name = string("current_key_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_19_groups_0 = const()[name = string("current_key_19_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230409856)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_19_cast_fp16 = conv(dilations = current_key_19_dilations_0, groups = current_key_19_groups_0, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = current_key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = string("current_key_19_cast_fp16")];
            string current_value_19_pad_type_0 = const()[name = string("current_value_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_19_strides_0 = const()[name = string("current_value_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_19_pad_0 = const()[name = string("current_value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_19_dilations_0 = const()[name = string("current_value_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_19_groups_0 = const()[name = string("current_value_19_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231589568)))];
            tensor<fp16, [768]> layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232769280)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = current_value_19_dilations_0, groups = current_value_19_groups_0, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = current_value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = string("current_value_19_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_2060_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_169_cast_fp16)[name = string("op_2060_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_19_cast_fp16 = add(x = var_65_cast_fp16_9, y = var_2060_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_2062_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_169_cast_fp16)[name = string("op_2062_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_19_cast_fp16 = add(x = var_80_cast_fp16_9, y = var_2062_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<int32, [4]> var_2065 = const()[name = string("op_2065"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_37_cast_fp16 = reshape(shape = var_2065, x = query_37_cast_fp16)[name = string("mh_q_37_cast_fp16")];
            fp16 var_2067_to_fp16 = const()[name = string("op_2067_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_2068_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2067_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [4]> var_2069 = const()[name = string("op_2069"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_2070_cast_fp16 = reshape(shape = var_2069, x = key_19_cast_fp16)[name = string("op_2070_cast_fp16")];
            bool mh_w_73_transpose_x_0 = const()[name = string("mh_w_73_transpose_x_0"), val = bool(true)];
            bool mh_w_73_transpose_y_0 = const()[name = string("mh_w_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_2068_cast_fp16, y = var_2070_cast_fp16)[name = string("mh_w_73_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_75_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_2078_cast_fp16 = softmax(axis = var_1997, x = mh_w_75_cast_fp16)[name = string("op_2078_cast_fp16")];
            tensor<int32, [4]> var_2079 = const()[name = string("op_2079"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_2080_cast_fp16 = reshape(shape = var_2079, x = value_19_cast_fp16)[name = string("op_2080_cast_fp16")];
            bool attn_37_transpose_x_0 = const()[name = string("attn_37_transpose_x_0"), val = bool(false)];
            bool attn_37_transpose_y_0 = const()[name = string("attn_37_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2080_cast_fp16, y = var_2078_cast_fp16)[name = string("attn_37_cast_fp16")];
            tensor<int32, [4]> var_2083 = const()[name = string("op_2083"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_91_cast_fp16 = reshape(shape = var_2083, x = attn_37_cast_fp16)[name = string("input_91_cast_fp16")];
            string obj_175_pad_type_0 = const()[name = string("obj_175_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_175_strides_0 = const()[name = string("obj_175_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_175_pad_0 = const()[name = string("obj_175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_175_dilations_0 = const()[name = string("obj_175_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_175_groups_0 = const()[name = string("obj_175_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232770880)))];
            tensor<fp16, [768]> layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233950592)))];
            tensor<fp16, [1, 768, 1, 1]> obj_175_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_175_dilations_0, groups = obj_175_groups_0, pad = obj_175_pad_0, pad_type = obj_175_pad_type_0, strides = obj_175_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = string("obj_175_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_175_cast_fp16)[name = string("inputs_57_cast_fp16")];
            tensor<int32, [1]> out_57_axes_0 = const()[name = string("out_57_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2105_to_fp16 = const()[name = string("op_2105_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_2105_to_fp16, x = inputs_57_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [768]> obj_177_gamma_0_to_fp16 = const()[name = string("obj_177_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233952192)))];
            tensor<fp16, [768]> obj_177_beta_0_to_fp16 = const()[name = string("obj_177_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233953792)))];
            fp16 obj_177_epsilon_0_to_fp16 = const()[name = string("obj_177_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_177_cast_fp16 = batch_norm(beta = obj_177_beta_0_to_fp16, epsilon = obj_177_epsilon_0_to_fp16, gamma = obj_177_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_57_cast_fp16)[name = string("obj_177_cast_fp16")];
            string query_39_pad_type_0 = const()[name = string("query_39_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_39_strides_0 = const()[name = string("query_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_39_pad_0 = const()[name = string("query_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_39_dilations_0 = const()[name = string("query_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_39_groups_0 = const()[name = string("query_39_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233955392)))];
            tensor<fp16, [768]> layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235135104)))];
            tensor<fp16, [1, 768, 1, 1]> query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_177_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<int32, [4]> var_2125 = const()[name = string("op_2125"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_39_cast_fp16 = reshape(shape = var_2125, x = query_39_cast_fp16)[name = string("mh_q_39_cast_fp16")];
            fp16 var_2127_to_fp16 = const()[name = string("op_2127_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_2128_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2127_to_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> var_2129 = const()[name = string("op_2129"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_2130_cast_fp16 = reshape(shape = var_2129, x = obj_179_cast_fp16)[name = string("op_2130_cast_fp16")];
            bool mh_w_77_transpose_x_0 = const()[name = string("mh_w_77_transpose_x_0"), val = bool(true)];
            bool mh_w_77_transpose_y_0 = const()[name = string("mh_w_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_2128_cast_fp16, y = var_2130_cast_fp16)[name = string("mh_w_77_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_79_cast_fp16 = add(x = mh_w_77_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_79_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_185_cast_fp16 = softmax(axis = var_1997, x = mh_w_79_cast_fp16)[name = string("obj_185_cast_fp16")];
            tensor<int32, [4]> var_2139 = const()[name = string("op_2139"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_2140_cast_fp16 = reshape(shape = var_2139, x = obj_181_cast_fp16)[name = string("op_2140_cast_fp16")];
            bool attn_39_transpose_x_0 = const()[name = string("attn_39_transpose_x_0"), val = bool(false)];
            bool attn_39_transpose_y_0 = const()[name = string("attn_39_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2140_cast_fp16, y = obj_185_cast_fp16)[name = string("attn_39_cast_fp16")];
            tensor<int32, [4]> var_2143 = const()[name = string("op_2143"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_93_cast_fp16 = reshape(shape = var_2143, x = attn_39_cast_fp16)[name = string("input_93_cast_fp16")];
            string obj_183_pad_type_0 = const()[name = string("obj_183_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_183_strides_0 = const()[name = string("obj_183_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_183_pad_0 = const()[name = string("obj_183_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_183_dilations_0 = const()[name = string("obj_183_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_183_groups_0 = const()[name = string("obj_183_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_9_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_9_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235136704)))];
            tensor<fp16, [768]> layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236316416)))];
            tensor<fp16, [1, 768, 1, 1]> obj_183_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = obj_183_dilations_0, groups = obj_183_groups_0, pad = obj_183_pad_0, pad_type = obj_183_pad_type_0, strides = obj_183_strides_0, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = string("obj_183_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_183_cast_fp16)[name = string("inputs_59_cast_fp16")];
            tensor<int32, [1]> out_59_axes_0 = const()[name = string("out_59_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2164_to_fp16 = const()[name = string("op_2164_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_2164_to_fp16, x = inputs_59_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [768]> input_95_gamma_0_to_fp16 = const()[name = string("input_95_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236318016)))];
            tensor<fp16, [768]> input_95_beta_0_to_fp16 = const()[name = string("input_95_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236319616)))];
            fp16 input_95_epsilon_0_to_fp16 = const()[name = string("input_95_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_59_cast_fp16)[name = string("input_95_cast_fp16")];
            string input_97_pad_type_0 = const()[name = string("input_97_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_97_strides_0 = const()[name = string("input_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_97_pad_0 = const()[name = string("input_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_97_dilations_0 = const()[name = string("input_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_97_groups_0 = const()[name = string("input_97_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_9_fc1_weight_to_fp16 = const()[name = string("layers_9_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236321216)))];
            tensor<fp16, [3072]> layers_9_fc1_bias_to_fp16 = const()[name = string("layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241039872)))];
            tensor<fp16, [1, 3072, 1, 1]> input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = string("input_97_cast_fp16")];
            string input_99_mode_0 = const()[name = string("input_99_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = string("input_99_cast_fp16")];
            string hidden_states_21_pad_type_0 = const()[name = string("hidden_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_21_strides_0 = const()[name = string("hidden_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_21_pad_0 = const()[name = string("hidden_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_21_dilations_0 = const()[name = string("hidden_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_21_groups_0 = const()[name = string("hidden_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_9_fc2_weight_to_fp16 = const()[name = string("layers_9_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241046080)))];
            tensor<fp16, [768]> layers_9_fc2_bias_to_fp16 = const()[name = string("layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245764736)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = string("hidden_states_21_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = string("inputs_61_cast_fp16")];
            tensor<int32, [4]> obj_197_begin_0 = const()[name = string("obj_197_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> obj_197_end_0 = const()[name = string("obj_197_end_0"), val = tensor<int32, [4]>([11, 768, 1, 1536])];
            tensor<bool, [4]> obj_197_end_mask_0 = const()[name = string("obj_197_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_197_cast_fp16 = slice_by_index(begin = obj_197_begin_0, end = obj_197_end_0, end_mask = obj_197_end_mask_0, x = read_state_2)[name = string("obj_197_cast_fp16")];
            tensor<int32, [4]> obj_199_begin_0 = const()[name = string("obj_199_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> obj_199_end_0 = const()[name = string("obj_199_end_0"), val = tensor<int32, [4]>([11, 768, 1, 1536])];
            tensor<bool, [4]> obj_199_end_mask_0 = const()[name = string("obj_199_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_199_cast_fp16 = slice_by_index(begin = obj_199_begin_0, end = obj_199_end_0, end_mask = obj_199_end_mask_0, x = read_state_3)[name = string("obj_199_cast_fp16")];
            int32 var_2210 = const()[name = string("op_2210"), val = int32(3)];
            tensor<int32, [1]> out_61_axes_0 = const()[name = string("out_61_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2235_to_fp16 = const()[name = string("op_2235_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_2235_to_fp16, x = inputs_61_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [768]> obj_187_gamma_0_to_fp16 = const()[name = string("obj_187_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245766336)))];
            tensor<fp16, [768]> obj_187_beta_0_to_fp16 = const()[name = string("obj_187_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245767936)))];
            fp16 obj_187_epsilon_0_to_fp16 = const()[name = string("obj_187_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_187_cast_fp16 = batch_norm(beta = obj_187_beta_0_to_fp16, epsilon = obj_187_epsilon_0_to_fp16, gamma = obj_187_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_61_cast_fp16)[name = string("obj_187_cast_fp16")];
            string query_41_pad_type_0 = const()[name = string("query_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_41_strides_0 = const()[name = string("query_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_41_pad_0 = const()[name = string("query_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_41_dilations_0 = const()[name = string("query_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_41_groups_0 = const()[name = string("query_41_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245769536)))];
            tensor<fp16, [768]> layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246949248)))];
            tensor<fp16, [1, 768, 1, 1]> query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_187_cast_fp16)[name = string("query_41_cast_fp16")];
            string current_key_21_pad_type_0 = const()[name = string("current_key_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_21_strides_0 = const()[name = string("current_key_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_21_pad_0 = const()[name = string("current_key_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_21_dilations_0 = const()[name = string("current_key_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_21_groups_0 = const()[name = string("current_key_21_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246950848)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_21_cast_fp16 = conv(dilations = current_key_21_dilations_0, groups = current_key_21_groups_0, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = current_key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_187_cast_fp16)[name = string("current_key_21_cast_fp16")];
            string current_value_21_pad_type_0 = const()[name = string("current_value_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_21_strides_0 = const()[name = string("current_value_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_21_pad_0 = const()[name = string("current_value_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_21_dilations_0 = const()[name = string("current_value_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_21_groups_0 = const()[name = string("current_value_21_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248130560)))];
            tensor<fp16, [768]> layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249310272)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = current_value_21_dilations_0, groups = current_value_21_groups_0, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = current_value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_187_cast_fp16)[name = string("current_value_21_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_2273_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_169_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_21_cast_fp16 = add(x = var_65_cast_fp16_10, y = var_2273_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_2275_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_169_cast_fp16)[name = string("op_2275_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_21_cast_fp16 = add(x = var_80_cast_fp16_10, y = var_2275_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_41_cast_fp16 = reshape(shape = var_2278, x = query_41_cast_fp16)[name = string("mh_q_41_cast_fp16")];
            fp16 var_2280_to_fp16 = const()[name = string("op_2280_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_2281_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2280_to_fp16)[name = string("op_2281_cast_fp16")];
            tensor<int32, [4]> var_2282 = const()[name = string("op_2282"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_2283_cast_fp16 = reshape(shape = var_2282, x = key_21_cast_fp16)[name = string("op_2283_cast_fp16")];
            bool mh_w_81_transpose_x_0 = const()[name = string("mh_w_81_transpose_x_0"), val = bool(true)];
            bool mh_w_81_transpose_y_0 = const()[name = string("mh_w_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_81_cast_fp16 = matmul(transpose_x = mh_w_81_transpose_x_0, transpose_y = mh_w_81_transpose_y_0, x = var_2281_cast_fp16, y = var_2283_cast_fp16)[name = string("mh_w_81_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_83_cast_fp16 = add(x = mh_w_81_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_83_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_2291_cast_fp16 = softmax(axis = var_2210, x = mh_w_83_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<int32, [4]> var_2292 = const()[name = string("op_2292"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_2293_cast_fp16 = reshape(shape = var_2292, x = value_21_cast_fp16)[name = string("op_2293_cast_fp16")];
            bool attn_41_transpose_x_0 = const()[name = string("attn_41_transpose_x_0"), val = bool(false)];
            bool attn_41_transpose_y_0 = const()[name = string("attn_41_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2293_cast_fp16, y = var_2291_cast_fp16)[name = string("attn_41_cast_fp16")];
            tensor<int32, [4]> var_2296 = const()[name = string("op_2296"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_101_cast_fp16 = reshape(shape = var_2296, x = attn_41_cast_fp16)[name = string("input_101_cast_fp16")];
            string obj_193_pad_type_0 = const()[name = string("obj_193_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_193_strides_0 = const()[name = string("obj_193_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_193_pad_0 = const()[name = string("obj_193_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_193_dilations_0 = const()[name = string("obj_193_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_193_groups_0 = const()[name = string("obj_193_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249311872)))];
            tensor<fp16, [768]> layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250491584)))];
            tensor<fp16, [1, 768, 1, 1]> obj_193_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_193_dilations_0, groups = obj_193_groups_0, pad = obj_193_pad_0, pad_type = obj_193_pad_type_0, strides = obj_193_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("obj_193_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_193_cast_fp16)[name = string("inputs_63_cast_fp16")];
            tensor<int32, [1]> out_63_axes_0 = const()[name = string("out_63_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2318_to_fp16, x = inputs_63_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [768]> obj_195_gamma_0_to_fp16 = const()[name = string("obj_195_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250493184)))];
            tensor<fp16, [768]> obj_195_beta_0_to_fp16 = const()[name = string("obj_195_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250494784)))];
            fp16 obj_195_epsilon_0_to_fp16 = const()[name = string("obj_195_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_195_cast_fp16 = batch_norm(beta = obj_195_beta_0_to_fp16, epsilon = obj_195_epsilon_0_to_fp16, gamma = obj_195_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_63_cast_fp16)[name = string("obj_195_cast_fp16")];
            string query_43_pad_type_0 = const()[name = string("query_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_43_strides_0 = const()[name = string("query_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_43_pad_0 = const()[name = string("query_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_43_dilations_0 = const()[name = string("query_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_43_groups_0 = const()[name = string("query_43_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250496384)))];
            tensor<fp16, [768]> layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251676096)))];
            tensor<fp16, [1, 768, 1, 1]> query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_195_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<int32, [4]> var_2338 = const()[name = string("op_2338"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_43_cast_fp16 = reshape(shape = var_2338, x = query_43_cast_fp16)[name = string("mh_q_43_cast_fp16")];
            fp16 var_2340_to_fp16 = const()[name = string("op_2340_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_2341_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2340_to_fp16)[name = string("op_2341_cast_fp16")];
            tensor<int32, [4]> var_2342 = const()[name = string("op_2342"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_2343_cast_fp16 = reshape(shape = var_2342, x = obj_197_cast_fp16)[name = string("op_2343_cast_fp16")];
            bool mh_w_85_transpose_x_0 = const()[name = string("mh_w_85_transpose_x_0"), val = bool(true)];
            bool mh_w_85_transpose_y_0 = const()[name = string("mh_w_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_2341_cast_fp16, y = var_2343_cast_fp16)[name = string("mh_w_85_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_87_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_203_cast_fp16 = softmax(axis = var_2210, x = mh_w_87_cast_fp16)[name = string("obj_203_cast_fp16")];
            tensor<int32, [4]> var_2352 = const()[name = string("op_2352"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_2353_cast_fp16 = reshape(shape = var_2352, x = obj_199_cast_fp16)[name = string("op_2353_cast_fp16")];
            bool attn_43_transpose_x_0 = const()[name = string("attn_43_transpose_x_0"), val = bool(false)];
            bool attn_43_transpose_y_0 = const()[name = string("attn_43_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2353_cast_fp16, y = obj_203_cast_fp16)[name = string("attn_43_cast_fp16")];
            tensor<int32, [4]> var_2356 = const()[name = string("op_2356"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_103_cast_fp16 = reshape(shape = var_2356, x = attn_43_cast_fp16)[name = string("input_103_cast_fp16")];
            string obj_201_pad_type_0 = const()[name = string("obj_201_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_201_strides_0 = const()[name = string("obj_201_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_201_pad_0 = const()[name = string("obj_201_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_201_dilations_0 = const()[name = string("obj_201_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_201_groups_0 = const()[name = string("obj_201_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_10_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_10_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251677696)))];
            tensor<fp16, [768]> layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252857408)))];
            tensor<fp16, [1, 768, 1, 1]> obj_201_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = obj_201_dilations_0, groups = obj_201_groups_0, pad = obj_201_pad_0, pad_type = obj_201_pad_type_0, strides = obj_201_strides_0, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = string("obj_201_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_201_cast_fp16)[name = string("inputs_65_cast_fp16")];
            tensor<int32, [1]> out_65_axes_0 = const()[name = string("out_65_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2377_to_fp16 = const()[name = string("op_2377_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2377_to_fp16, x = inputs_65_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [768]> input_105_gamma_0_to_fp16 = const()[name = string("input_105_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252859008)))];
            tensor<fp16, [768]> input_105_beta_0_to_fp16 = const()[name = string("input_105_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252860608)))];
            fp16 input_105_epsilon_0_to_fp16 = const()[name = string("input_105_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_105_cast_fp16 = batch_norm(beta = input_105_beta_0_to_fp16, epsilon = input_105_epsilon_0_to_fp16, gamma = input_105_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_65_cast_fp16)[name = string("input_105_cast_fp16")];
            string input_107_pad_type_0 = const()[name = string("input_107_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_107_strides_0 = const()[name = string("input_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_107_pad_0 = const()[name = string("input_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_107_dilations_0 = const()[name = string("input_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_107_groups_0 = const()[name = string("input_107_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_10_fc1_weight_to_fp16 = const()[name = string("layers_10_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252862208)))];
            tensor<fp16, [3072]> layers_10_fc1_bias_to_fp16 = const()[name = string("layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257580864)))];
            tensor<fp16, [1, 3072, 1, 1]> input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
            string input_109_mode_0 = const()[name = string("input_109_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = string("input_109_cast_fp16")];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_10_fc2_weight_to_fp16 = const()[name = string("layers_10_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257587072)))];
            tensor<fp16, [768]> layers_10_fc2_bias_to_fp16 = const()[name = string("layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262305728)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("inputs_67_cast_fp16")];
            tensor<int32, [4]> obj_215_begin_0 = const()[name = string("obj_215_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> obj_215_end_0 = const()[name = string("obj_215_end_0"), val = tensor<int32, [4]>([12, 768, 1, 1536])];
            tensor<bool, [4]> obj_215_end_mask_0 = const()[name = string("obj_215_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_215_cast_fp16 = slice_by_index(begin = obj_215_begin_0, end = obj_215_end_0, end_mask = obj_215_end_mask_0, x = read_state_2)[name = string("obj_215_cast_fp16")];
            tensor<int32, [4]> obj_217_begin_0 = const()[name = string("obj_217_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> obj_217_end_0 = const()[name = string("obj_217_end_0"), val = tensor<int32, [4]>([12, 768, 1, 1536])];
            tensor<bool, [4]> obj_217_end_mask_0 = const()[name = string("obj_217_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 768, 1, 1536]> obj_217_cast_fp16 = slice_by_index(begin = obj_217_begin_0, end = obj_217_end_0, end_mask = obj_217_end_mask_0, x = read_state_3)[name = string("obj_217_cast_fp16")];
            int32 var_2423 = const()[name = string("op_2423"), val = int32(3)];
            tensor<int32, [1]> out_67_axes_0 = const()[name = string("out_67_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2448_to_fp16 = const()[name = string("op_2448_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2448_to_fp16, x = inputs_67_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [768]> obj_205_gamma_0_to_fp16 = const()[name = string("obj_205_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262307328)))];
            tensor<fp16, [768]> obj_205_beta_0_to_fp16 = const()[name = string("obj_205_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262308928)))];
            fp16 obj_205_epsilon_0_to_fp16 = const()[name = string("obj_205_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_205_cast_fp16 = batch_norm(beta = obj_205_beta_0_to_fp16, epsilon = obj_205_epsilon_0_to_fp16, gamma = obj_205_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_67_cast_fp16)[name = string("obj_205_cast_fp16")];
            string query_45_pad_type_0 = const()[name = string("query_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_45_strides_0 = const()[name = string("query_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_45_pad_0 = const()[name = string("query_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_45_dilations_0 = const()[name = string("query_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_45_groups_0 = const()[name = string("query_45_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262310528)))];
            tensor<fp16, [768]> layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263490240)))];
            tensor<fp16, [1, 768, 1, 1]> query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = string("query_45_cast_fp16")];
            string current_key_pad_type_0 = const()[name = string("current_key_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_key_strides_0 = const()[name = string("current_key_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_pad_0 = const()[name = string("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_dilations_0 = const()[name = string("current_key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_key_groups_0 = const()[name = string("current_key_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263491840)))];
            tensor<fp16, [1, 768, 1, 1]> current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = string("current_key_cast_fp16")];
            string current_value_pad_type_0 = const()[name = string("current_value_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> current_value_strides_0 = const()[name = string("current_value_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_pad_0 = const()[name = string("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_dilations_0 = const()[name = string("current_value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 current_value_groups_0 = const()[name = string("current_value_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264671552)))];
            tensor<fp16, [768]> layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265851264)))];
            tensor<fp16, [1, 768, 1, 1]> current_value_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = string("current_value_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_2486_cast_fp16 = mul(x = current_key_cast_fp16, y = var_169_cast_fp16)[name = string("op_2486_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> key_cast_fp16 = add(x = var_65_cast_fp16_11, y = var_2486_cast_fp16)[name = string("key_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> var_2488_cast_fp16 = mul(x = current_value_cast_fp16, y = var_169_cast_fp16)[name = string("op_2488_cast_fp16")];
            tensor<fp16, [1, 768, 1, 448]> value_cast_fp16 = add(x = var_80_cast_fp16_11, y = var_2488_cast_fp16)[name = string("value_cast_fp16")];
            tensor<int32, [4]> var_2491 = const()[name = string("op_2491"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_45_cast_fp16 = reshape(shape = var_2491, x = query_45_cast_fp16)[name = string("mh_q_45_cast_fp16")];
            fp16 var_2493_to_fp16 = const()[name = string("op_2493_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_2494_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2493_to_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> var_2495 = const()[name = string("op_2495"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_2496_cast_fp16 = reshape(shape = var_2495, x = key_cast_fp16)[name = string("op_2496_cast_fp16")];
            bool mh_w_89_transpose_x_0 = const()[name = string("mh_w_89_transpose_x_0"), val = bool(true)];
            bool mh_w_89_transpose_y_0 = const()[name = string("mh_w_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 448]> mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_2494_cast_fp16, y = var_2496_cast_fp16)[name = string("mh_w_89_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> mh_w_91_cast_fp16 = add(x = mh_w_89_cast_fp16, y = var_186_cast_fp16)[name = string("mh_w_91_cast_fp16")];
            tensor<fp16, [1, 12, 1, 448]> var_2504_cast_fp16 = softmax(axis = var_2423, x = mh_w_91_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [4]> var_2505 = const()[name = string("op_2505"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 448]> var_2506_cast_fp16 = reshape(shape = var_2505, x = value_cast_fp16)[name = string("op_2506_cast_fp16")];
            bool attn_45_transpose_x_0 = const()[name = string("attn_45_transpose_x_0"), val = bool(false)];
            bool attn_45_transpose_y_0 = const()[name = string("attn_45_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2506_cast_fp16, y = var_2504_cast_fp16)[name = string("attn_45_cast_fp16")];
            tensor<int32, [4]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_111_cast_fp16 = reshape(shape = var_2509, x = attn_45_cast_fp16)[name = string("input_111_cast_fp16")];
            string obj_211_pad_type_0 = const()[name = string("obj_211_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_211_strides_0 = const()[name = string("obj_211_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_211_pad_0 = const()[name = string("obj_211_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_211_dilations_0 = const()[name = string("obj_211_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_211_groups_0 = const()[name = string("obj_211_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265852864)))];
            tensor<fp16, [768]> layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267032576)))];
            tensor<fp16, [1, 768, 1, 1]> obj_211_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_211_dilations_0, groups = obj_211_groups_0, pad = obj_211_pad_0, pad_type = obj_211_pad_type_0, strides = obj_211_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = string("obj_211_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_211_cast_fp16)[name = string("inputs_69_cast_fp16")];
            tensor<int32, [1]> out_69_axes_0 = const()[name = string("out_69_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2531_to_fp16 = const()[name = string("op_2531_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2531_to_fp16, x = inputs_69_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [768]> obj_213_gamma_0_to_fp16 = const()[name = string("obj_213_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267034176)))];
            tensor<fp16, [768]> obj_213_beta_0_to_fp16 = const()[name = string("obj_213_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267035776)))];
            fp16 obj_213_epsilon_0_to_fp16 = const()[name = string("obj_213_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> obj_213_cast_fp16 = batch_norm(beta = obj_213_beta_0_to_fp16, epsilon = obj_213_epsilon_0_to_fp16, gamma = obj_213_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_69_cast_fp16)[name = string("obj_213_cast_fp16")];
            string query_pad_type_0 = const()[name = string("query_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_strides_0 = const()[name = string("query_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_pad_0 = const()[name = string("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_dilations_0 = const()[name = string("query_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_groups_0 = const()[name = string("query_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_11_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267037376)))];
            tensor<fp16, [768]> layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268217088)))];
            tensor<fp16, [1, 768, 1, 1]> query_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_213_cast_fp16)[name = string("query_cast_fp16")];
            tensor<int32, [4]> var_2551 = const()[name = string("op_2551"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_2551, x = query_cast_fp16)[name = string("mh_q_cast_fp16")];
            fp16 var_2553_to_fp16 = const()[name = string("op_2553_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 64, 1]> var_2554_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_2553_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [4]> var_2555 = const()[name = string("op_2555"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_2556_cast_fp16 = reshape(shape = var_2555, x = obj_215_cast_fp16)[name = string("op_2556_cast_fp16")];
            bool mh_w_93_transpose_x_0 = const()[name = string("mh_w_93_transpose_x_0"), val = bool(true)];
            bool mh_w_93_transpose_y_0 = const()[name = string("mh_w_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_93_cast_fp16 = matmul(transpose_x = mh_w_93_transpose_x_0, transpose_y = mh_w_93_transpose_y_0, x = var_2554_cast_fp16, y = var_2556_cast_fp16)[name = string("mh_w_93_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> mh_w_cast_fp16 = add(x = mh_w_93_cast_fp16, y = var_246_cast_fp16)[name = string("mh_w_cast_fp16")];
            tensor<fp16, [1, 12, 1, 1536]> obj_221_cast_fp16 = softmax(axis = var_2423, x = mh_w_cast_fp16)[name = string("obj_221_cast_fp16")];
            tensor<int32, [4]> var_2565 = const()[name = string("op_2565"), val = tensor<int32, [4]>([1, 12, 64, -1])];
            tensor<fp16, [1, 12, 64, 1536]> var_2566_cast_fp16 = reshape(shape = var_2565, x = obj_217_cast_fp16)[name = string("op_2566_cast_fp16")];
            bool attn_transpose_x_0 = const()[name = string("attn_transpose_x_0"), val = bool(false)];
            bool attn_transpose_y_0 = const()[name = string("attn_transpose_y_0"), val = bool(true)];
            tensor<fp16, [1, 12, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_2566_cast_fp16, y = obj_221_cast_fp16)[name = string("attn_cast_fp16")];
            tensor<int32, [4]> var_2569 = const()[name = string("op_2569"), val = tensor<int32, [4]>([1, 768, 1, -1])];
            tensor<fp16, [1, 768, 1, 1]> input_113_cast_fp16 = reshape(shape = var_2569, x = attn_cast_fp16)[name = string("input_113_cast_fp16")];
            string obj_219_pad_type_0 = const()[name = string("obj_219_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> obj_219_strides_0 = const()[name = string("obj_219_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_219_pad_0 = const()[name = string("obj_219_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_219_dilations_0 = const()[name = string("obj_219_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 obj_219_groups_0 = const()[name = string("obj_219_groups_0"), val = int32(1)];
            tensor<fp16, [768, 768, 1, 1]> layers_11_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_11_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268218688)))];
            tensor<fp16, [768]> layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269398400)))];
            tensor<fp16, [1, 768, 1, 1]> obj_219_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = obj_219_dilations_0, groups = obj_219_groups_0, pad = obj_219_pad_0, pad_type = obj_219_pad_type_0, strides = obj_219_strides_0, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = string("obj_219_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_219_cast_fp16)[name = string("inputs_71_cast_fp16")];
            tensor<int32, [1]> out_71_axes_0 = const()[name = string("out_71_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2587_to_fp16 = const()[name = string("op_2587_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2587_to_fp16, x = inputs_71_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [768]> input_115_gamma_0_to_fp16 = const()[name = string("input_115_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269400000)))];
            tensor<fp16, [768]> input_115_beta_0_to_fp16 = const()[name = string("input_115_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269401600)))];
            fp16 input_115_epsilon_0_to_fp16 = const()[name = string("input_115_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_71_cast_fp16)[name = string("input_115_cast_fp16")];
            string input_117_pad_type_0 = const()[name = string("input_117_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_117_strides_0 = const()[name = string("input_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_117_pad_0 = const()[name = string("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_117_dilations_0 = const()[name = string("input_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_117_groups_0 = const()[name = string("input_117_groups_0"), val = int32(1)];
            tensor<fp16, [3072, 768, 1, 1]> layers_11_fc1_weight_to_fp16 = const()[name = string("layers_11_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269403200)))];
            tensor<fp16, [3072]> layers_11_fc1_bias_to_fp16 = const()[name = string("layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274121856)))];
            tensor<fp16, [1, 3072, 1, 1]> input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")];
            string input_mode_0 = const()[name = string("input_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 3072, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_117_cast_fp16)[name = string("input_cast_fp16")];
            string hidden_states_25_pad_type_0 = const()[name = string("hidden_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_25_strides_0 = const()[name = string("hidden_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_25_pad_0 = const()[name = string("hidden_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_25_dilations_0 = const()[name = string("hidden_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_25_groups_0 = const()[name = string("hidden_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [768, 3072, 1, 1]> layers_11_fc2_weight_to_fp16 = const()[name = string("layers_11_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274128064)))];
            tensor<fp16, [768]> layers_11_fc2_bias_to_fp16 = const()[name = string("layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278846720)))];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            tensor<fp16, [1, 768, 1, 1]> inputs_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = string("inputs_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2629_to_fp16 = const()[name = string("op_2629_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_2629_to_fp16, x = inputs_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [768]> hidden_states_gamma_0_to_fp16 = const()[name = string("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278848320)))];
            tensor<fp16, [768]> hidden_states_beta_0_to_fp16 = const()[name = string("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278849920)))];
            fp16 hidden_states_epsilon_0_to_fp16 = const()[name = string("hidden_states_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 768, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<int32, [1]> var_2640_axes_0 = const()[name = string("op_2640_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 768, 1]> var_2640_cast_fp16 = squeeze(axes = var_2640_axes_0, x = hidden_states_cast_fp16)[name = string("op_2640_cast_fp16")];
            tensor<int32, [3]> var_2643_perm_0 = const()[name = string("op_2643_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [51865]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51865]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278851520)))];
            tensor<fp16, [1, 1, 768]> var_2643_cast_fp16 = transpose(perm = var_2643_perm_0, x = var_2640_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 1, 51865]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_2643_cast_fp16)[name = string("linear_0_cast_fp16")];
            int32 var_2647 = const()[name = string("op_2647"), val = int32(1)];
            bool obj_225_interleave_0 = const()[name = string("obj_225_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 9216, 1, 1]> key_cache_updates = concat(axis = var_2647, interleave = obj_225_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_11_cast_fp16, current_key_13_cast_fp16, current_key_15_cast_fp16, current_key_17_cast_fp16, current_key_19_cast_fp16, current_key_21_cast_fp16, current_key_cast_fp16))[name = string("obj_225_cast_fp16")];
            int32 var_2650 = const()[name = string("op_2650"), val = int32(1)];
            bool obj_227_interleave_0 = const()[name = string("obj_227_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 9216, 1, 1]> value_cache_updates = concat(axis = var_2650, interleave = obj_227_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_11_cast_fp16, current_value_13_cast_fp16, current_value_15_cast_fp16, current_value_17_cast_fp16, current_value_19_cast_fp16, current_value_21_cast_fp16, current_value_cast_fp16))[name = string("obj_227_cast_fp16")];
            tensor<int32, [4]> var_2661_begin_0 = const()[name = string("op_2661_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])];
            tensor<int32, [4]> var_2661_end_0 = const()[name = string("op_2661_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1536])];
            tensor<bool, [4]> var_2661_end_mask_0 = const()[name = string("op_2661_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2661_cast_fp16 = slice_by_index(begin = var_2661_begin_0, end = var_2661_end_0, end_mask = var_2661_end_mask_0, x = obj_113_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<int32, [4]> var_2664_begin_0 = const()[name = string("op_2664_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2664_end_0 = const()[name = string("op_2664_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2664_end_mask_0 = const()[name = string("op_2664_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2664_squeeze_mask_0 = const()[name = string("op_2664_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2664_cast_fp16 = slice_by_index(begin = var_2664_begin_0, end = var_2664_end_0, end_mask = var_2664_end_mask_0, squeeze_mask = var_2664_squeeze_mask_0, x = var_2661_cast_fp16)[name = string("op_2664_cast_fp16")];
            tensor<int32, [4]> var_2679_begin_0 = const()[name = string("op_2679_begin_0"), val = tensor<int32, [4]>([0, 9, 0, 0])];
            tensor<int32, [4]> var_2679_end_0 = const()[name = string("op_2679_end_0"), val = tensor<int32, [4]>([1, 10, 1, 1536])];
            tensor<bool, [4]> var_2679_end_mask_0 = const()[name = string("op_2679_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2679_cast_fp16 = slice_by_index(begin = var_2679_begin_0, end = var_2679_end_0, end_mask = var_2679_end_mask_0, x = obj_113_cast_fp16)[name = string("op_2679_cast_fp16")];
            tensor<int32, [4]> var_2682_begin_0 = const()[name = string("op_2682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2682_end_0 = const()[name = string("op_2682_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2682_end_mask_0 = const()[name = string("op_2682_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2682_squeeze_mask_0 = const()[name = string("op_2682_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2682_cast_fp16 = slice_by_index(begin = var_2682_begin_0, end = var_2682_end_0, end_mask = var_2682_end_mask_0, squeeze_mask = var_2682_squeeze_mask_0, x = var_2679_cast_fp16)[name = string("op_2682_cast_fp16")];
            tensor<int32, [4]> var_2697_begin_0 = const()[name = string("op_2697_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2697_end_0 = const()[name = string("op_2697_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2697_end_mask_0 = const()[name = string("op_2697_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2697_cast_fp16 = slice_by_index(begin = var_2697_begin_0, end = var_2697_end_0, end_mask = var_2697_end_mask_0, x = obj_167_cast_fp16)[name = string("op_2697_cast_fp16")];
            tensor<int32, [4]> var_2700_begin_0 = const()[name = string("op_2700_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2700_end_0 = const()[name = string("op_2700_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2700_end_mask_0 = const()[name = string("op_2700_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2700_squeeze_mask_0 = const()[name = string("op_2700_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2700_cast_fp16 = slice_by_index(begin = var_2700_begin_0, end = var_2700_end_0, end_mask = var_2700_end_mask_0, squeeze_mask = var_2700_squeeze_mask_0, x = var_2697_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [4]> var_2715_begin_0 = const()[name = string("op_2715_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
            tensor<int32, [4]> var_2715_end_0 = const()[name = string("op_2715_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1536])];
            tensor<bool, [4]> var_2715_end_mask_0 = const()[name = string("op_2715_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2715_cast_fp16 = slice_by_index(begin = var_2715_begin_0, end = var_2715_end_0, end_mask = var_2715_end_mask_0, x = obj_167_cast_fp16)[name = string("op_2715_cast_fp16")];
            tensor<int32, [4]> var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2718_squeeze_mask_0 = const()[name = string("op_2718_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, squeeze_mask = var_2718_squeeze_mask_0, x = var_2715_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<int32, [4]> var_2733_begin_0 = const()[name = string("op_2733_begin_0"), val = tensor<int32, [4]>([0, 7, 0, 0])];
            tensor<int32, [4]> var_2733_end_0 = const()[name = string("op_2733_end_0"), val = tensor<int32, [4]>([1, 8, 1, 1536])];
            tensor<bool, [4]> var_2733_end_mask_0 = const()[name = string("op_2733_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2733_cast_fp16 = slice_by_index(begin = var_2733_begin_0, end = var_2733_end_0, end_mask = var_2733_end_mask_0, x = obj_167_cast_fp16)[name = string("op_2733_cast_fp16")];
            tensor<int32, [4]> var_2736_begin_0 = const()[name = string("op_2736_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2736_end_0 = const()[name = string("op_2736_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2736_end_mask_0 = const()[name = string("op_2736_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2736_squeeze_mask_0 = const()[name = string("op_2736_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2736_cast_fp16 = slice_by_index(begin = var_2736_begin_0, end = var_2736_end_0, end_mask = var_2736_end_mask_0, squeeze_mask = var_2736_squeeze_mask_0, x = var_2733_cast_fp16)[name = string("op_2736_cast_fp16")];
            tensor<int32, [4]> var_2751_begin_0 = const()[name = string("op_2751_begin_0"), val = tensor<int32, [4]>([0, 8, 0, 0])];
            tensor<int32, [4]> var_2751_end_0 = const()[name = string("op_2751_end_0"), val = tensor<int32, [4]>([1, 9, 1, 1536])];
            tensor<bool, [4]> var_2751_end_mask_0 = const()[name = string("op_2751_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2751_cast_fp16 = slice_by_index(begin = var_2751_begin_0, end = var_2751_end_0, end_mask = var_2751_end_mask_0, x = obj_167_cast_fp16)[name = string("op_2751_cast_fp16")];
            tensor<int32, [4]> var_2754_begin_0 = const()[name = string("op_2754_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2754_end_0 = const()[name = string("op_2754_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2754_end_mask_0 = const()[name = string("op_2754_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2754_squeeze_mask_0 = const()[name = string("op_2754_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2754_cast_fp16 = slice_by_index(begin = var_2754_begin_0, end = var_2754_end_0, end_mask = var_2754_end_mask_0, squeeze_mask = var_2754_squeeze_mask_0, x = var_2751_cast_fp16)[name = string("op_2754_cast_fp16")];
            tensor<int32, [4]> var_2769_begin_0 = const()[name = string("op_2769_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2769_end_0 = const()[name = string("op_2769_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2769_end_mask_0 = const()[name = string("op_2769_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2769_cast_fp16 = slice_by_index(begin = var_2769_begin_0, end = var_2769_end_0, end_mask = var_2769_end_mask_0, x = obj_185_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [4]> var_2772_begin_0 = const()[name = string("op_2772_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2772_end_0 = const()[name = string("op_2772_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2772_end_mask_0 = const()[name = string("op_2772_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2772_squeeze_mask_0 = const()[name = string("op_2772_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2772_cast_fp16 = slice_by_index(begin = var_2772_begin_0, end = var_2772_end_0, end_mask = var_2772_end_mask_0, squeeze_mask = var_2772_squeeze_mask_0, x = var_2769_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<int32, [4]> var_2787_begin_0 = const()[name = string("op_2787_begin_0"), val = tensor<int32, [4]>([0, 7, 0, 0])];
            tensor<int32, [4]> var_2787_end_0 = const()[name = string("op_2787_end_0"), val = tensor<int32, [4]>([1, 8, 1, 1536])];
            tensor<bool, [4]> var_2787_end_mask_0 = const()[name = string("op_2787_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2787_cast_fp16 = slice_by_index(begin = var_2787_begin_0, end = var_2787_end_0, end_mask = var_2787_end_mask_0, x = obj_185_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<int32, [4]> var_2790_begin_0 = const()[name = string("op_2790_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2790_end_0 = const()[name = string("op_2790_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2790_end_mask_0 = const()[name = string("op_2790_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2790_squeeze_mask_0 = const()[name = string("op_2790_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2790_cast_fp16 = slice_by_index(begin = var_2790_begin_0, end = var_2790_end_0, end_mask = var_2790_end_mask_0, squeeze_mask = var_2790_squeeze_mask_0, x = var_2787_cast_fp16)[name = string("op_2790_cast_fp16")];
            tensor<int32, [4]> var_2805_begin_0 = const()[name = string("op_2805_begin_0"), val = tensor<int32, [4]>([0, 9, 0, 0])];
            tensor<int32, [4]> var_2805_end_0 = const()[name = string("op_2805_end_0"), val = tensor<int32, [4]>([1, 10, 1, 1536])];
            tensor<bool, [4]> var_2805_end_mask_0 = const()[name = string("op_2805_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2805_cast_fp16 = slice_by_index(begin = var_2805_begin_0, end = var_2805_end_0, end_mask = var_2805_end_mask_0, x = obj_185_cast_fp16)[name = string("op_2805_cast_fp16")];
            tensor<int32, [4]> var_2808_begin_0 = const()[name = string("op_2808_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2808_end_0 = const()[name = string("op_2808_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2808_end_mask_0 = const()[name = string("op_2808_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2808_squeeze_mask_0 = const()[name = string("op_2808_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2808_cast_fp16 = slice_by_index(begin = var_2808_begin_0, end = var_2808_end_0, end_mask = var_2808_end_mask_0, squeeze_mask = var_2808_squeeze_mask_0, x = var_2805_cast_fp16)[name = string("op_2808_cast_fp16")];
            tensor<int32, [4]> var_2823_begin_0 = const()[name = string("op_2823_begin_0"), val = tensor<int32, [4]>([0, 5, 0, 0])];
            tensor<int32, [4]> var_2823_end_0 = const()[name = string("op_2823_end_0"), val = tensor<int32, [4]>([1, 6, 1, 1536])];
            tensor<bool, [4]> var_2823_end_mask_0 = const()[name = string("op_2823_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1536]> var_2823_cast_fp16 = slice_by_index(begin = var_2823_begin_0, end = var_2823_end_0, end_mask = var_2823_end_mask_0, x = obj_203_cast_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [4]> var_2826_begin_0 = const()[name = string("op_2826_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2826_end_0 = const()[name = string("op_2826_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
            tensor<bool, [4]> var_2826_end_mask_0 = const()[name = string("op_2826_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2826_squeeze_mask_0 = const()[name = string("op_2826_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1536]> var_2826_cast_fp16 = slice_by_index(begin = var_2826_begin_0, end = var_2826_end_0, end_mask = var_2826_end_mask_0, squeeze_mask = var_2826_squeeze_mask_0, x = var_2823_cast_fp16)[name = string("op_2826_cast_fp16")];
            int32 var_2833 = const()[name = string("op_2833"), val = int32(1)];
            bool var_2834_interleave_0 = const()[name = string("op_2834_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 10, 1536]> var_2834_cast_fp16 = concat(axis = var_2833, interleave = var_2834_interleave_0, values = (var_2664_cast_fp16, var_2682_cast_fp16, var_2700_cast_fp16, var_2718_cast_fp16, var_2736_cast_fp16, var_2754_cast_fp16, var_2772_cast_fp16, var_2790_cast_fp16, var_2808_cast_fp16, var_2826_cast_fp16))[name = string("op_2834_cast_fp16")];
            bool var_2837 = const()[name = string("op_2837"), val = bool(false)];
            tensor<int32, [1]> obj_axes_0 = const()[name = string("obj_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1536]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_2837, x = var_2834_cast_fp16)[name = string("obj_cast_fp16")];
        } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights);
}