program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}})]
{
    func main<ios17>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, tensor<fp16, [1, 1280, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 40960, 1, 448]> key_cache, tensor<fp16, [1, 448]> kv_cache_update_mask, tensor<fp16, [1, 40960, 1, 448]> value_cache) {
            tensor<int32, []> var_80_axis_0 = const()[name = tensor<string, []>("op_80_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_80_batch_dims_0 = const()[name = tensor<string, []>("op_80_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_80_validate_indices_0 = const()[name = tensor<string, []>("op_80_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [51866, 1280]> embed_tokens_weight_to_fp16 = const()[name = tensor<string, []>("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51866, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [1, 1280]> var_80_cast_fp16 = gather(axis = var_80_axis_0, batch_dims = var_80_batch_dims_0, indices = input_ids, validate_indices = var_80_validate_indices_0, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_80_cast_fp16")];
            tensor<int32, []> var_84_axis_0 = const()[name = tensor<string, []>("op_84_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_84_batch_dims_0 = const()[name = tensor<string, []>("op_84_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_84_validate_indices_0 = const()[name = tensor<string, []>("op_84_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [448, 1280]> embed_positions_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [573440]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132777088))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133350592))), name = tensor<string, []>("embed_positions_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([448, 1280])];
            tensor<string, []> cache_length_to_int16_dtype_0 = const()[name = tensor<string, []>("cache_length_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<int16, [1]> cast_0 = cast(dtype = cache_length_to_int16_dtype_0, x = cache_length)[name = tensor<string, []>("cast_0")];
            tensor<fp16, [1, 1280]> var_84_cast_fp16_cast_int16 = gather(axis = var_84_axis_0, batch_dims = var_84_batch_dims_0, indices = cast_0, validate_indices = var_84_validate_indices_0, x = embed_positions_weight_to_fp16_palettized)[name = tensor<string, []>("op_84_cast_fp16_cast_int16")];
            tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_80_cast_fp16, y = var_84_cast_fp16_cast_int16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
            tensor<int32, [1]> var_98_axes_0 = const()[name = tensor<string, []>("op_98_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1280, 1]> var_98_cast_fp16 = expand_dims(axes = var_98_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_98_cast_fp16")];
            tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_98_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
            tensor<int32, [32]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [32]>([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])];
            tensor<int32, []> var_103_axis_0 = const()[name = tensor<string, []>("op_103_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_3, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_4, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_5, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_6, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_7, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_8, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_9, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_10, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_11, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_12, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_13, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_14, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_15, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_16, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_17, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_18, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_19, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_20, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_21, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_22, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_23, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_24, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_25, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_26, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_27, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_28, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_29, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_30, tensor<fp16, [1, 1280, 1, 448]> var_103_cast_fp16_31 = split(axis = var_103_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_103_cast_fp16")];
            tensor<int32, [32]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [32]>([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])];
            tensor<int32, []> var_138_axis_0 = const()[name = tensor<string, []>("op_138_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_3, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_4, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_5, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_6, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_7, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_8, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_9, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_10, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_11, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_12, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_13, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_14, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_15, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_16, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_17, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_18, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_19, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_20, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_21, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_22, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_23, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_24, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_25, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_26, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_27, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_28, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_29, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_30, tensor<fp16, [1, 1280, 1, 448]> var_138_cast_fp16_31 = split(axis = var_138_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_138_cast_fp16")];
            tensor<int32, []> var_176 = const()[name = tensor<string, []>("op_176"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_183 = const()[name = tensor<string, []>("op_183"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_184 = const()[name = tensor<string, []>("op_184"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_196 = const()[name = tensor<string, []>("op_196"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_1_cast_fp16 = reduce_mean(axes = var_196, keep_dims = var_184, x = inputs_1_cast_fp16)[name = tensor<string, []>("channels_mean_1_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_1_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_sq_1_cast_fp16")];
            tensor<int32, [1]> var_200 = const()[name = tensor<string, []>("op_200"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_201_cast_fp16 = reduce_mean(axes = var_200, keep_dims = var_184, x = zero_mean_sq_1_cast_fp16)[name = tensor<string, []>("op_201_cast_fp16")];
            tensor<fp16, []> var_202_to_fp16 = const()[name = tensor<string, []>("op_202_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_203_cast_fp16 = add(x = var_201_cast_fp16, y = var_202_to_fp16)[name = tensor<string, []>("op_203_cast_fp16")];
            tensor<fp32, []> denom_1_epsilon_0 = const()[name = tensor<string, []>("denom_1_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0, x = var_203_cast_fp16)[name = tensor<string, []>("denom_1_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
            tensor<fp16, [1280]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133351168)))];
            tensor<fp16, [1280]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133353792)))];
            tensor<fp16, [1280]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133356416)))];
            tensor<fp16, [1280]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133359040)))];
            tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
            tensor<int32, [2]> var_218 = const()[name = tensor<string, []>("op_218"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_220 = const()[name = tensor<string, []>("op_220"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133361664)))];
            tensor<fp16, [1280]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136638528)))];
            tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_220, groups = var_183, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_218, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
            tensor<int32, [2]> var_224 = const()[name = tensor<string, []>("op_224"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_226 = const()[name = tensor<string, []>("op_226"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_1_pad_type_0 = const()[name = tensor<string, []>("current_key_1_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_1_pad_0 = const()[name = tensor<string, []>("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136641152)))];
            tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = conv(dilations = var_226, groups = var_183, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_224, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")];
            tensor<int32, [2]> var_231 = const()[name = tensor<string, []>("op_231"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_233 = const()[name = tensor<string, []>("op_233"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_1_pad_type_0 = const()[name = tensor<string, []>("current_value_1_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_1_pad_0 = const()[name = tensor<string, []>("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139918016))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141146880))), name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141147072)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_233, groups = var_183, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_231, weight = layers_0_self_attn_v_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")];
            tensor<int32, [1]> var_237_axes_0 = const()[name = tensor<string, []>("op_237_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 448]> var_237_cast_fp16 = expand_dims(axes = var_237_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_237_cast_fp16")];
            tensor<int32, [1]> var_238_axes_0 = const()[name = tensor<string, []>("op_238_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 448]> var_238_cast_fp16 = expand_dims(axes = var_238_axes_0, x = var_237_cast_fp16)[name = tensor<string, []>("op_238_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_240_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_240_cast_fp16")];
            tensor<fp16, []> var_177_to_fp16 = const()[name = tensor<string, []>("op_177_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 1, 1, 448]> var_241_cast_fp16 = sub(x = var_177_to_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_241_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_242_cast_fp16 = mul(x = var_103_cast_fp16_0, y = var_241_cast_fp16)[name = tensor<string, []>("op_242_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_1_cast_fp16 = add(x = var_240_cast_fp16, y = var_242_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_244_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_244_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_246_cast_fp16 = mul(x = var_138_cast_fp16_0, y = var_241_cast_fp16)[name = tensor<string, []>("op_246_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_1_cast_fp16 = add(x = var_244_cast_fp16, y = var_246_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
            tensor<int32, [4]> var_249 = const()[name = tensor<string, []>("op_249"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_250_cast_fp16 = reshape(shape = var_249, x = query_1_cast_fp16)[name = tensor<string, []>("op_250_cast_fp16")];
            tensor<fp16, []> var_251_to_fp16 = const()[name = tensor<string, []>("op_251_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_252_cast_fp16 = mul(x = var_250_cast_fp16, y = var_251_to_fp16)[name = tensor<string, []>("op_252_cast_fp16")];
            tensor<int32, [4]> var_253 = const()[name = tensor<string, []>("op_253"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_254_cast_fp16 = reshape(shape = var_253, x = key_1_cast_fp16)[name = tensor<string, []>("op_254_cast_fp16")];
            tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 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_252_cast_fp16, y = var_254_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
            tensor<int32, [1]> var_258_axes_0 = const()[name = tensor<string, []>("op_258_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 448]> var_258_cast_fp16 = expand_dims(axes = var_258_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_258_cast_fp16")];
            tensor<int32, [1]> var_259_axes_0 = const()[name = tensor<string, []>("op_259_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 448]> var_259_cast_fp16 = expand_dims(axes = var_259_axes_0, x = var_258_cast_fp16)[name = tensor<string, []>("op_259_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_262_cast_fp16 = softmax(axis = var_176, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_262_cast_fp16")];
            tensor<int32, [4]> var_263 = const()[name = tensor<string, []>("op_263"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_264_cast_fp16 = reshape(shape = var_263, x = value_1_cast_fp16)[name = tensor<string, []>("op_264_cast_fp16")];
            tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_264_cast_fp16, y = var_262_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
            tensor<int32, [4]> var_267 = const()[name = tensor<string, []>("op_267"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_1_cast_fp16 = reshape(shape = var_267, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<int32, [2]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_273 = const()[name = tensor<string, []>("op_273"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141149696)))];
            tensor<fp16, [1280]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144426560)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_273, groups = var_183, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_271, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
            tensor<int32, [1]> var_283 = const()[name = tensor<string, []>("op_283"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_3_cast_fp16 = reduce_mean(axes = var_283, keep_dims = var_184, x = inputs_3_cast_fp16)[name = tensor<string, []>("channels_mean_3_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_3_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_sq_3_cast_fp16")];
            tensor<int32, [1]> var_287 = const()[name = tensor<string, []>("op_287"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_288_cast_fp16 = reduce_mean(axes = var_287, keep_dims = var_184, x = zero_mean_sq_3_cast_fp16)[name = tensor<string, []>("op_288_cast_fp16")];
            tensor<fp16, []> var_289_to_fp16 = const()[name = tensor<string, []>("op_289_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_290_cast_fp16 = add(x = var_288_cast_fp16, y = var_289_to_fp16)[name = tensor<string, []>("op_290_cast_fp16")];
            tensor<fp32, []> denom_3_epsilon_0 = const()[name = tensor<string, []>("denom_3_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0, x = var_290_cast_fp16)[name = tensor<string, []>("denom_3_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
            tensor<fp16, [1280]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144429184)))];
            tensor<fp16, [1280]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144431808)))];
            tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
            tensor<int32, [2]> var_305 = const()[name = tensor<string, []>("op_305"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_307 = const()[name = tensor<string, []>("op_307"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144434432))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145253696))), name = tensor<string, []>("layers_0_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145253824)))];
            tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_307, groups = var_183, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_305, weight = layers_0_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
            tensor<int32, [2]> var_311 = const()[name = tensor<string, []>("op_311"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_313 = const()[name = tensor<string, []>("op_313"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145256448))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146075712))), name = tensor<string, []>("layers_0_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = conv(dilations = var_313, groups = var_183, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_311, weight = layers_0_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_3_cast_fp16")];
            tensor<int32, [2]> var_318 = const()[name = tensor<string, []>("op_318"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_320 = const()[name = tensor<string, []>("op_320"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146075840))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146895104))), name = tensor<string, []>("layers_0_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146895232)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_320, groups = var_183, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_318, weight = layers_0_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_3_cast_fp16")];
            tensor<int32, [4]> var_324 = const()[name = tensor<string, []>("op_324"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_325_cast_fp16 = reshape(shape = var_324, x = query_3_cast_fp16)[name = tensor<string, []>("op_325_cast_fp16")];
            tensor<fp16, []> var_326_to_fp16 = const()[name = tensor<string, []>("op_326_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_327_cast_fp16 = mul(x = var_325_cast_fp16, y = var_326_to_fp16)[name = tensor<string, []>("op_327_cast_fp16")];
            tensor<int32, [4]> var_328 = const()[name = tensor<string, []>("op_328"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_329_cast_fp16 = reshape(shape = var_328, x = key_3_cast_fp16)[name = tensor<string, []>("op_329_cast_fp16")];
            tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_327_cast_fp16, y = var_329_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_332_cast_fp16 = softmax(axis = var_176, x = mh_w_5_cast_fp16)[name = tensor<string, []>("op_332_cast_fp16")];
            tensor<int32, [4]> var_333 = const()[name = tensor<string, []>("op_333"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_334_cast_fp16 = reshape(shape = var_333, x = value_3_cast_fp16)[name = tensor<string, []>("op_334_cast_fp16")];
            tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_334_cast_fp16, y = var_332_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
            tensor<int32, [4]> var_337 = const()[name = tensor<string, []>("op_337"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_3_cast_fp16 = reshape(shape = var_337, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<int32, [2]> var_341 = const()[name = tensor<string, []>("op_341"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_343 = const()[name = tensor<string, []>("op_343"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146897856))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148126720))), name = tensor<string, []>("layers_0_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148126912)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_343, groups = var_183, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_341, weight = layers_0_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
            tensor<int32, [1]> var_349 = const()[name = tensor<string, []>("op_349"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_5_cast_fp16 = reduce_mean(axes = var_349, keep_dims = var_184, x = inputs_5_cast_fp16)[name = tensor<string, []>("channels_mean_5_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_5_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_sq_5_cast_fp16")];
            tensor<int32, [1]> var_353 = const()[name = tensor<string, []>("op_353"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_354_cast_fp16 = reduce_mean(axes = var_353, keep_dims = var_184, x = zero_mean_sq_5_cast_fp16)[name = tensor<string, []>("op_354_cast_fp16")];
            tensor<fp16, []> var_355_to_fp16 = const()[name = tensor<string, []>("op_355_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_356_cast_fp16 = add(x = var_354_cast_fp16, y = var_355_to_fp16)[name = tensor<string, []>("op_356_cast_fp16")];
            tensor<fp32, []> denom_5_epsilon_0 = const()[name = tensor<string, []>("denom_5_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0, x = var_356_cast_fp16)[name = tensor<string, []>("denom_5_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
            tensor<fp16, [1280]> input_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148129536)))];
            tensor<fp16, [1280]> input_5_beta_0_to_fp16 = const()[name = tensor<string, []>("input_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148132160)))];
            tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<int32, [2]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_369 = const()[name = tensor<string, []>("op_369"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148134784))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151411648))), name = tensor<string, []>("layers_0_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151411776)))];
            tensor<fp16, [1, 5120, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_369, groups = var_183, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_367, weight = layers_0_fc1_weight_to_fp16_palettized, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<int32, [2]> var_375 = const()[name = tensor<string, []>("op_375"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_377 = const()[name = tensor<string, []>("op_377"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_3_pad_type_0 = const()[name = tensor<string, []>("hidden_states_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = tensor<string, []>("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151422080))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157975744))), name = tensor<string, []>("layers_0_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157976320)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_377, groups = var_183, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_375, weight = layers_0_fc2_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
            tensor<int32, []> var_390 = const()[name = tensor<string, []>("op_390"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_397 = const()[name = tensor<string, []>("op_397"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_398 = const()[name = tensor<string, []>("op_398"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_410 = const()[name = tensor<string, []>("op_410"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_7_cast_fp16 = reduce_mean(axes = var_410, keep_dims = var_398, x = inputs_7_cast_fp16)[name = tensor<string, []>("channels_mean_7_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_7_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_sq_7_cast_fp16")];
            tensor<int32, [1]> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_415_cast_fp16 = reduce_mean(axes = var_414, keep_dims = var_398, x = zero_mean_sq_7_cast_fp16)[name = tensor<string, []>("op_415_cast_fp16")];
            tensor<fp16, []> var_416_to_fp16 = const()[name = tensor<string, []>("op_416_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_417_cast_fp16 = add(x = var_415_cast_fp16, y = var_416_to_fp16)[name = tensor<string, []>("op_417_cast_fp16")];
            tensor<fp32, []> denom_7_epsilon_0 = const()[name = tensor<string, []>("denom_7_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0, x = var_417_cast_fp16)[name = tensor<string, []>("denom_7_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
            tensor<fp16, [1280]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157978944)))];
            tensor<fp16, [1280]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157981568)))];
            tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
            tensor<int32, [2]> var_432 = const()[name = tensor<string, []>("op_432"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_434 = const()[name = tensor<string, []>("op_434"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157984192))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158803456))), name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158803584)))];
            tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_434, groups = var_397, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_432, weight = layers_1_self_attn_q_proj_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
            tensor<int32, [2]> var_438 = const()[name = tensor<string, []>("op_438"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_440 = const()[name = tensor<string, []>("op_440"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_3_pad_type_0 = const()[name = tensor<string, []>("current_key_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_3_pad_0 = const()[name = tensor<string, []>("current_key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158806208))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159625472))), name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_3_cast_fp16 = conv(dilations = var_440, groups = var_397, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = var_438, weight = layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")];
            tensor<int32, [2]> var_445 = const()[name = tensor<string, []>("op_445"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_447 = const()[name = tensor<string, []>("op_447"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_3_pad_type_0 = const()[name = tensor<string, []>("current_value_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_3_pad_0 = const()[name = tensor<string, []>("current_value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159625600)))];
            tensor<fp16, [1280]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162902464)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_447, groups = var_397, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = var_445, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_454_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_454_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_456_cast_fp16 = mul(x = var_103_cast_fp16_1, y = var_241_cast_fp16)[name = tensor<string, []>("op_456_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_5_cast_fp16 = add(x = var_454_cast_fp16, y = var_456_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_458_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_458_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_460_cast_fp16 = mul(x = var_138_cast_fp16_1, y = var_241_cast_fp16)[name = tensor<string, []>("op_460_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_5_cast_fp16 = add(x = var_458_cast_fp16, y = var_460_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
            tensor<int32, [4]> var_463 = const()[name = tensor<string, []>("op_463"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_464_cast_fp16 = reshape(shape = var_463, x = query_5_cast_fp16)[name = tensor<string, []>("op_464_cast_fp16")];
            tensor<fp16, []> var_465_to_fp16 = const()[name = tensor<string, []>("op_465_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_466_cast_fp16 = mul(x = var_464_cast_fp16, y = var_465_to_fp16)[name = tensor<string, []>("op_466_cast_fp16")];
            tensor<int32, [4]> var_467 = const()[name = tensor<string, []>("op_467"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_468_cast_fp16 = reshape(shape = var_467, x = key_5_cast_fp16)[name = tensor<string, []>("op_468_cast_fp16")];
            tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_466_cast_fp16, y = var_468_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_476_cast_fp16 = softmax(axis = var_390, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_476_cast_fp16")];
            tensor<int32, [4]> var_477 = const()[name = tensor<string, []>("op_477"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_478_cast_fp16 = reshape(shape = var_477, x = value_5_cast_fp16)[name = tensor<string, []>("op_478_cast_fp16")];
            tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_478_cast_fp16, y = var_476_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
            tensor<int32, [4]> var_481 = const()[name = tensor<string, []>("op_481"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_11_cast_fp16 = reshape(shape = var_481, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<int32, [2]> var_485 = const()[name = tensor<string, []>("op_485"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_487 = const()[name = tensor<string, []>("op_487"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_19_pad_type_0 = const()[name = tensor<string, []>("obj_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_19_pad_0 = const()[name = tensor<string, []>("obj_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162905088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164133952))), name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164134144)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_19_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_487, groups = var_397, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = var_485, weight = layers_1_self_attn_o_proj_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_19_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
            tensor<int32, [1]> var_497 = const()[name = tensor<string, []>("op_497"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_9_cast_fp16 = reduce_mean(axes = var_497, keep_dims = var_398, x = inputs_9_cast_fp16)[name = tensor<string, []>("channels_mean_9_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_9_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_sq_9_cast_fp16")];
            tensor<int32, [1]> var_501 = const()[name = tensor<string, []>("op_501"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_502_cast_fp16 = reduce_mean(axes = var_501, keep_dims = var_398, x = zero_mean_sq_9_cast_fp16)[name = tensor<string, []>("op_502_cast_fp16")];
            tensor<fp16, []> var_503_to_fp16 = const()[name = tensor<string, []>("op_503_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_504_cast_fp16 = add(x = var_502_cast_fp16, y = var_503_to_fp16)[name = tensor<string, []>("op_504_cast_fp16")];
            tensor<fp32, []> denom_9_epsilon_0 = const()[name = tensor<string, []>("denom_9_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0, x = var_504_cast_fp16)[name = tensor<string, []>("denom_9_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
            tensor<fp16, [1280]> obj_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_21_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164136768)))];
            tensor<fp16, [1280]> obj_21_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_21_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164139392)))];
            tensor<fp16, []> obj_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
            tensor<int32, [2]> var_519 = const()[name = tensor<string, []>("op_519"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_521 = const()[name = tensor<string, []>("op_521"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_7_pad_type_0 = const()[name = tensor<string, []>("query_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_7_pad_0 = const()[name = tensor<string, []>("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164142016))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164961280))), name = tensor<string, []>("layers_1_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164961408)))];
            tensor<fp16, [1, 1280, 1, 1]> query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_521, groups = var_397, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_519, weight = layers_1_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
            tensor<int32, [2]> var_525 = const()[name = tensor<string, []>("op_525"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_527 = const()[name = tensor<string, []>("op_527"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_7_pad_0 = const()[name = tensor<string, []>("key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164964032))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165783296))), name = tensor<string, []>("layers_1_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_7_cast_fp16 = conv(dilations = var_527, groups = var_397, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = var_525, weight = layers_1_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_7_cast_fp16")];
            tensor<int32, [2]> var_532 = const()[name = tensor<string, []>("op_532"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_534 = const()[name = tensor<string, []>("op_534"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_7_pad_type_0 = const()[name = tensor<string, []>("value_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_7_pad_0 = const()[name = tensor<string, []>("value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165783424))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166602688))), name = tensor<string, []>("layers_1_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166602816)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_534, groups = var_397, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = var_532, weight = layers_1_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_7_cast_fp16")];
            tensor<int32, [4]> var_538 = const()[name = tensor<string, []>("op_538"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_539_cast_fp16 = reshape(shape = var_538, x = query_7_cast_fp16)[name = tensor<string, []>("op_539_cast_fp16")];
            tensor<fp16, []> var_540_to_fp16 = const()[name = tensor<string, []>("op_540_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_541_cast_fp16 = mul(x = var_539_cast_fp16, y = var_540_to_fp16)[name = tensor<string, []>("op_541_cast_fp16")];
            tensor<int32, [4]> var_542 = const()[name = tensor<string, []>("op_542"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_543_cast_fp16 = reshape(shape = var_542, x = key_7_cast_fp16)[name = tensor<string, []>("op_543_cast_fp16")];
            tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_541_cast_fp16, y = var_543_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_546_cast_fp16 = softmax(axis = var_390, x = mh_w_11_cast_fp16)[name = tensor<string, []>("op_546_cast_fp16")];
            tensor<int32, [4]> var_547 = const()[name = tensor<string, []>("op_547"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_548_cast_fp16 = reshape(shape = var_547, x = value_7_cast_fp16)[name = tensor<string, []>("op_548_cast_fp16")];
            tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_548_cast_fp16, y = var_546_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
            tensor<int32, [4]> var_551 = const()[name = tensor<string, []>("op_551"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_13_cast_fp16 = reshape(shape = var_551, x = attn_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<int32, [2]> var_555 = const()[name = tensor<string, []>("op_555"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_557 = const()[name = tensor<string, []>("op_557"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_23_pad_type_0 = const()[name = tensor<string, []>("obj_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_23_pad_0 = const()[name = tensor<string, []>("obj_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166605440))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167834304))), name = tensor<string, []>("layers_1_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167834496)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_23_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_557, groups = var_397, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = var_555, weight = layers_1_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_23_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
            tensor<int32, [1]> var_563 = const()[name = tensor<string, []>("op_563"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_11_cast_fp16 = reduce_mean(axes = var_563, keep_dims = var_398, x = inputs_11_cast_fp16)[name = tensor<string, []>("channels_mean_11_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_11_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_sq_11_cast_fp16")];
            tensor<int32, [1]> var_567 = const()[name = tensor<string, []>("op_567"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_568_cast_fp16 = reduce_mean(axes = var_567, keep_dims = var_398, x = zero_mean_sq_11_cast_fp16)[name = tensor<string, []>("op_568_cast_fp16")];
            tensor<fp16, []> var_569_to_fp16 = const()[name = tensor<string, []>("op_569_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_570_cast_fp16 = add(x = var_568_cast_fp16, y = var_569_to_fp16)[name = tensor<string, []>("op_570_cast_fp16")];
            tensor<fp32, []> denom_11_epsilon_0 = const()[name = tensor<string, []>("denom_11_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0, x = var_570_cast_fp16)[name = tensor<string, []>("denom_11_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
            tensor<fp16, [1280]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167837120)))];
            tensor<fp16, [1280]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167839744)))];
            tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<int32, [2]> var_581 = const()[name = tensor<string, []>("op_581"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_583 = const()[name = tensor<string, []>("op_583"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167842368))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172757632))), name = tensor<string, []>("layers_1_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172757824)))];
            tensor<fp16, [1, 5120, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_583, groups = var_397, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_581, weight = layers_1_fc1_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<int32, [2]> var_589 = const()[name = tensor<string, []>("op_589"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_591 = const()[name = tensor<string, []>("op_591"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172768128))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177683392))), name = tensor<string, []>("layers_1_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177683584)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_591, groups = var_397, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_589, weight = layers_1_fc2_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
            tensor<int32, []> var_604 = const()[name = tensor<string, []>("op_604"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_611 = const()[name = tensor<string, []>("op_611"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_612 = const()[name = tensor<string, []>("op_612"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_624 = const()[name = tensor<string, []>("op_624"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_13_cast_fp16 = reduce_mean(axes = var_624, keep_dims = var_612, x = inputs_13_cast_fp16)[name = tensor<string, []>("channels_mean_13_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_13_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_sq_13_cast_fp16")];
            tensor<int32, [1]> var_628 = const()[name = tensor<string, []>("op_628"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_629_cast_fp16 = reduce_mean(axes = var_628, keep_dims = var_612, x = zero_mean_sq_13_cast_fp16)[name = tensor<string, []>("op_629_cast_fp16")];
            tensor<fp16, []> var_630_to_fp16 = const()[name = tensor<string, []>("op_630_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_631_cast_fp16 = add(x = var_629_cast_fp16, y = var_630_to_fp16)[name = tensor<string, []>("op_631_cast_fp16")];
            tensor<fp32, []> denom_13_epsilon_0 = const()[name = tensor<string, []>("denom_13_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0, x = var_631_cast_fp16)[name = tensor<string, []>("denom_13_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
            tensor<fp16, [1280]> obj_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177686208)))];
            tensor<fp16, [1280]> obj_25_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177688832)))];
            tensor<fp16, []> obj_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
            tensor<int32, [2]> var_646 = const()[name = tensor<string, []>("op_646"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_648 = const()[name = tensor<string, []>("op_648"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177691456))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178510720))), name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178510848)))];
            tensor<fp16, [1, 1280, 1, 1]> query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_648, groups = var_611, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_646, weight = layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
            tensor<int32, [2]> var_652 = const()[name = tensor<string, []>("op_652"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_654 = const()[name = tensor<string, []>("op_654"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_5_pad_type_0 = const()[name = tensor<string, []>("current_key_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_5_pad_0 = const()[name = tensor<string, []>("current_key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178513472))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179332736))), name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_5_cast_fp16 = conv(dilations = var_654, groups = var_611, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_652, weight = layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")];
            tensor<int32, [2]> var_659 = const()[name = tensor<string, []>("op_659"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_661 = const()[name = tensor<string, []>("op_661"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_5_pad_type_0 = const()[name = tensor<string, []>("current_value_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_5_pad_0 = const()[name = tensor<string, []>("current_value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179332864))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180152128))), name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180152256)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_661, groups = var_611, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_659, weight = layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_668_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_668_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_670_cast_fp16 = mul(x = var_103_cast_fp16_2, y = var_241_cast_fp16)[name = tensor<string, []>("op_670_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_9_cast_fp16 = add(x = var_668_cast_fp16, y = var_670_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_672_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_672_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_674_cast_fp16 = mul(x = var_138_cast_fp16_2, y = var_241_cast_fp16)[name = tensor<string, []>("op_674_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_9_cast_fp16 = add(x = var_672_cast_fp16, y = var_674_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
            tensor<int32, [4]> var_677 = const()[name = tensor<string, []>("op_677"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_678_cast_fp16 = reshape(shape = var_677, x = query_9_cast_fp16)[name = tensor<string, []>("op_678_cast_fp16")];
            tensor<fp16, []> var_679_to_fp16 = const()[name = tensor<string, []>("op_679_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_680_cast_fp16 = mul(x = var_678_cast_fp16, y = var_679_to_fp16)[name = tensor<string, []>("op_680_cast_fp16")];
            tensor<int32, [4]> var_681 = const()[name = tensor<string, []>("op_681"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_682_cast_fp16 = reshape(shape = var_681, x = key_9_cast_fp16)[name = tensor<string, []>("op_682_cast_fp16")];
            tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_680_cast_fp16, y = var_682_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_690_cast_fp16 = softmax(axis = var_604, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_690_cast_fp16")];
            tensor<int32, [4]> var_691 = const()[name = tensor<string, []>("op_691"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_692_cast_fp16 = reshape(shape = var_691, x = value_9_cast_fp16)[name = tensor<string, []>("op_692_cast_fp16")];
            tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_692_cast_fp16, y = var_690_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
            tensor<int32, [4]> var_695 = const()[name = tensor<string, []>("op_695"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_21_cast_fp16 = reshape(shape = var_695, x = attn_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<int32, [2]> var_699 = const()[name = tensor<string, []>("op_699"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_701 = const()[name = tensor<string, []>("op_701"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_31_pad_type_0 = const()[name = tensor<string, []>("obj_31_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_31_pad_0 = const()[name = tensor<string, []>("obj_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180154880))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181793344))), name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181793920)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_31_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_701, groups = var_611, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = var_699, weight = layers_2_self_attn_o_proj_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor<string, []>("obj_31_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_31_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
            tensor<int32, [1]> var_711 = const()[name = tensor<string, []>("op_711"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_15_cast_fp16 = reduce_mean(axes = var_711, keep_dims = var_612, x = inputs_15_cast_fp16)[name = tensor<string, []>("channels_mean_15_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_15_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_sq_15_cast_fp16")];
            tensor<int32, [1]> var_715 = const()[name = tensor<string, []>("op_715"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_716_cast_fp16 = reduce_mean(axes = var_715, keep_dims = var_612, x = zero_mean_sq_15_cast_fp16)[name = tensor<string, []>("op_716_cast_fp16")];
            tensor<fp16, []> var_717_to_fp16 = const()[name = tensor<string, []>("op_717_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_718_cast_fp16 = add(x = var_716_cast_fp16, y = var_717_to_fp16)[name = tensor<string, []>("op_718_cast_fp16")];
            tensor<fp32, []> denom_15_epsilon_0 = const()[name = tensor<string, []>("denom_15_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0, x = var_718_cast_fp16)[name = tensor<string, []>("denom_15_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
            tensor<fp16, [1280]> obj_33_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181796544)))];
            tensor<fp16, [1280]> obj_33_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_33_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181799168)))];
            tensor<fp16, []> obj_33_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_33_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_33_cast_fp16")];
            tensor<int32, [2]> var_733 = const()[name = tensor<string, []>("op_733"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_735 = const()[name = tensor<string, []>("op_735"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_11_pad_type_0 = const()[name = tensor<string, []>("query_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_11_pad_0 = const()[name = tensor<string, []>("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181801792))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182621056))), name = tensor<string, []>("layers_2_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182621184)))];
            tensor<fp16, [1, 1280, 1, 1]> query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_735, groups = var_611, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_733, weight = layers_2_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
            tensor<int32, [2]> var_739 = const()[name = tensor<string, []>("op_739"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_741 = const()[name = tensor<string, []>("op_741"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182623808))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183443072))), name = tensor<string, []>("layers_2_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_11_cast_fp16 = conv(dilations = var_741, groups = var_611, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_739, weight = layers_2_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_11_cast_fp16")];
            tensor<int32, [2]> var_746 = const()[name = tensor<string, []>("op_746"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_748 = const()[name = tensor<string, []>("op_748"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_11_pad_type_0 = const()[name = tensor<string, []>("value_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_11_pad_0 = const()[name = tensor<string, []>("value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183443200))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184262464))), name = tensor<string, []>("layers_2_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184262592)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_748, groups = var_611, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_746, weight = layers_2_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_11_cast_fp16")];
            tensor<int32, [4]> var_752 = const()[name = tensor<string, []>("op_752"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_753_cast_fp16 = reshape(shape = var_752, x = query_11_cast_fp16)[name = tensor<string, []>("op_753_cast_fp16")];
            tensor<fp16, []> var_754_to_fp16 = const()[name = tensor<string, []>("op_754_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_755_cast_fp16 = mul(x = var_753_cast_fp16, y = var_754_to_fp16)[name = tensor<string, []>("op_755_cast_fp16")];
            tensor<int32, [4]> var_756 = const()[name = tensor<string, []>("op_756"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_757_cast_fp16 = reshape(shape = var_756, x = key_11_cast_fp16)[name = tensor<string, []>("op_757_cast_fp16")];
            tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_755_cast_fp16, y = var_757_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_760_cast_fp16 = softmax(axis = var_604, x = mh_w_17_cast_fp16)[name = tensor<string, []>("op_760_cast_fp16")];
            tensor<int32, [4]> var_761 = const()[name = tensor<string, []>("op_761"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_762_cast_fp16 = reshape(shape = var_761, x = value_11_cast_fp16)[name = tensor<string, []>("op_762_cast_fp16")];
            tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_762_cast_fp16, y = var_760_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
            tensor<int32, [4]> var_765 = const()[name = tensor<string, []>("op_765"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_23_cast_fp16 = reshape(shape = var_765, x = attn_11_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<int32, [2]> var_769 = const()[name = tensor<string, []>("op_769"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_771 = const()[name = tensor<string, []>("op_771"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184265216))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185084480))), name = tensor<string, []>("layers_2_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185084608)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_35_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_771, groups = var_611, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_769, weight = layers_2_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
            tensor<int32, [1]> var_777 = const()[name = tensor<string, []>("op_777"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_17_cast_fp16 = reduce_mean(axes = var_777, keep_dims = var_612, x = inputs_17_cast_fp16)[name = tensor<string, []>("channels_mean_17_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor<string, []>("zero_mean_17_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor<string, []>("zero_mean_sq_17_cast_fp16")];
            tensor<int32, [1]> var_781 = const()[name = tensor<string, []>("op_781"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_782_cast_fp16 = reduce_mean(axes = var_781, keep_dims = var_612, x = zero_mean_sq_17_cast_fp16)[name = tensor<string, []>("op_782_cast_fp16")];
            tensor<fp16, []> var_783_to_fp16 = const()[name = tensor<string, []>("op_783_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_784_cast_fp16 = add(x = var_782_cast_fp16, y = var_783_to_fp16)[name = tensor<string, []>("op_784_cast_fp16")];
            tensor<fp32, []> denom_17_epsilon_0 = const()[name = tensor<string, []>("denom_17_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0, x = var_784_cast_fp16)[name = tensor<string, []>("denom_17_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
            tensor<fp16, [1280]> input_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185087232)))];
            tensor<fp16, [1280]> input_25_beta_0_to_fp16 = const()[name = tensor<string, []>("input_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185089856)))];
            tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<int32, [2]> var_795 = const()[name = tensor<string, []>("op_795"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_797 = const()[name = tensor<string, []>("op_797"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_27_pad_type_0 = const()[name = tensor<string, []>("input_27_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_27_pad_0 = const()[name = tensor<string, []>("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185092480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190007744))), name = tensor<string, []>("layers_2_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190007936)))];
            tensor<fp16, [1, 5120, 1, 1]> input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_797, groups = var_611, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_795, weight = layers_2_fc1_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<int32, [2]> var_803 = const()[name = tensor<string, []>("op_803"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_805 = const()[name = tensor<string, []>("op_805"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190018240))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194933504))), name = tensor<string, []>("layers_2_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194933696)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_805, groups = var_611, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_803, weight = layers_2_fc2_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
            tensor<int32, []> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_825 = const()[name = tensor<string, []>("op_825"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_826 = const()[name = tensor<string, []>("op_826"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_838 = const()[name = tensor<string, []>("op_838"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_19_cast_fp16 = reduce_mean(axes = var_838, keep_dims = var_826, x = inputs_19_cast_fp16)[name = tensor<string, []>("channels_mean_19_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor<string, []>("zero_mean_19_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor<string, []>("zero_mean_sq_19_cast_fp16")];
            tensor<int32, [1]> var_842 = const()[name = tensor<string, []>("op_842"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_843_cast_fp16 = reduce_mean(axes = var_842, keep_dims = var_826, x = zero_mean_sq_19_cast_fp16)[name = tensor<string, []>("op_843_cast_fp16")];
            tensor<fp16, []> var_844_to_fp16 = const()[name = tensor<string, []>("op_844_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_845_cast_fp16 = add(x = var_843_cast_fp16, y = var_844_to_fp16)[name = tensor<string, []>("op_845_cast_fp16")];
            tensor<fp32, []> denom_19_epsilon_0 = const()[name = tensor<string, []>("denom_19_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0, x = var_845_cast_fp16)[name = tensor<string, []>("denom_19_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
            tensor<fp16, [1280]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194936320)))];
            tensor<fp16, [1280]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194938944)))];
            tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
            tensor<int32, [2]> var_860 = const()[name = tensor<string, []>("op_860"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_862 = const()[name = tensor<string, []>("op_862"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194941568))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195760832))), name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195760960)))];
            tensor<fp16, [1, 1280, 1, 1]> query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_862, groups = var_825, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_860, weight = layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
            tensor<int32, [2]> var_866 = const()[name = tensor<string, []>("op_866"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_868 = const()[name = tensor<string, []>("op_868"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_7_pad_type_0 = const()[name = tensor<string, []>("current_key_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_7_pad_0 = const()[name = tensor<string, []>("current_key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195763584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196582848))), name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_7_cast_fp16 = conv(dilations = var_868, groups = var_825, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = var_866, weight = layers_3_self_attn_k_proj_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("current_key_7_cast_fp16")];
            tensor<int32, [2]> var_873 = const()[name = tensor<string, []>("op_873"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_875 = const()[name = tensor<string, []>("op_875"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_7_pad_type_0 = const()[name = tensor<string, []>("current_value_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_7_pad_0 = const()[name = tensor<string, []>("current_value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196582976))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197402240))), name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197402368)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_875, groups = var_825, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = var_873, weight = layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("current_value_7_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_882_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_882_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_884_cast_fp16 = mul(x = var_103_cast_fp16_3, y = var_241_cast_fp16)[name = tensor<string, []>("op_884_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_13_cast_fp16 = add(x = var_882_cast_fp16, y = var_884_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_886_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_886_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_888_cast_fp16 = mul(x = var_138_cast_fp16_3, y = var_241_cast_fp16)[name = tensor<string, []>("op_888_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_13_cast_fp16 = add(x = var_886_cast_fp16, y = var_888_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
            tensor<int32, [4]> var_891 = const()[name = tensor<string, []>("op_891"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_892_cast_fp16 = reshape(shape = var_891, x = query_13_cast_fp16)[name = tensor<string, []>("op_892_cast_fp16")];
            tensor<fp16, []> var_893_to_fp16 = const()[name = tensor<string, []>("op_893_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_894_cast_fp16 = mul(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor<string, []>("op_894_cast_fp16")];
            tensor<int32, [4]> var_895 = const()[name = tensor<string, []>("op_895"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_896_cast_fp16 = reshape(shape = var_895, x = key_13_cast_fp16)[name = tensor<string, []>("op_896_cast_fp16")];
            tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_894_cast_fp16, y = var_896_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_904_cast_fp16 = softmax(axis = var_818, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_904_cast_fp16")];
            tensor<int32, [4]> var_905 = const()[name = tensor<string, []>("op_905"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_906_cast_fp16 = reshape(shape = var_905, x = value_13_cast_fp16)[name = tensor<string, []>("op_906_cast_fp16")];
            tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_906_cast_fp16, y = var_904_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
            tensor<int32, [4]> var_909 = const()[name = tensor<string, []>("op_909"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_31_cast_fp16 = reshape(shape = var_909, x = attn_13_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<int32, [2]> var_913 = const()[name = tensor<string, []>("op_913"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_915 = const()[name = tensor<string, []>("op_915"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_43_pad_type_0 = const()[name = tensor<string, []>("obj_43_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_43_pad_0 = const()[name = tensor<string, []>("obj_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197404992))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198633856))), name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198634048)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_43_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_915, groups = var_825, pad = obj_43_pad_0, pad_type = obj_43_pad_type_0, strides = var_913, weight = layers_3_self_attn_o_proj_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_43_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
            tensor<int32, [1]> var_925 = const()[name = tensor<string, []>("op_925"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_21_cast_fp16 = reduce_mean(axes = var_925, keep_dims = var_826, x = inputs_21_cast_fp16)[name = tensor<string, []>("channels_mean_21_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor<string, []>("zero_mean_21_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor<string, []>("zero_mean_sq_21_cast_fp16")];
            tensor<int32, [1]> var_929 = const()[name = tensor<string, []>("op_929"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_930_cast_fp16 = reduce_mean(axes = var_929, keep_dims = var_826, x = zero_mean_sq_21_cast_fp16)[name = tensor<string, []>("op_930_cast_fp16")];
            tensor<fp16, []> var_931_to_fp16 = const()[name = tensor<string, []>("op_931_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_932_cast_fp16 = add(x = var_930_cast_fp16, y = var_931_to_fp16)[name = tensor<string, []>("op_932_cast_fp16")];
            tensor<fp32, []> denom_21_epsilon_0 = const()[name = tensor<string, []>("denom_21_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0, x = var_932_cast_fp16)[name = tensor<string, []>("denom_21_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
            tensor<fp16, [1280]> obj_45_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_45_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198636672)))];
            tensor<fp16, [1280]> obj_45_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_45_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198639296)))];
            tensor<fp16, []> obj_45_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_45_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_45_cast_fp16")];
            tensor<int32, [2]> var_947 = const()[name = tensor<string, []>("op_947"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_949 = const()[name = tensor<string, []>("op_949"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_15_pad_type_0 = const()[name = tensor<string, []>("query_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_15_pad_0 = const()[name = tensor<string, []>("query_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198641920))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199461184))), name = tensor<string, []>("layers_3_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199461312)))];
            tensor<fp16, [1, 1280, 1, 1]> query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_949, groups = var_825, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = var_947, weight = layers_3_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
            tensor<int32, [2]> var_953 = const()[name = tensor<string, []>("op_953"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_955 = const()[name = tensor<string, []>("op_955"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_15_pad_type_0 = const()[name = tensor<string, []>("key_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_15_pad_0 = const()[name = tensor<string, []>("key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199463936))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200692800))), name = tensor<string, []>("layers_3_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_15_cast_fp16 = conv(dilations = var_955, groups = var_825, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = var_953, weight = layers_3_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_15_cast_fp16")];
            tensor<int32, [2]> var_960 = const()[name = tensor<string, []>("op_960"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_962 = const()[name = tensor<string, []>("op_962"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_15_pad_type_0 = const()[name = tensor<string, []>("value_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_15_pad_0 = const()[name = tensor<string, []>("value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200692992))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201512256))), name = tensor<string, []>("layers_3_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201512384)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_962, groups = var_825, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = var_960, weight = layers_3_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_15_cast_fp16")];
            tensor<int32, [4]> var_966 = const()[name = tensor<string, []>("op_966"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_967_cast_fp16 = reshape(shape = var_966, x = query_15_cast_fp16)[name = tensor<string, []>("op_967_cast_fp16")];
            tensor<fp16, []> var_968_to_fp16 = const()[name = tensor<string, []>("op_968_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_969_cast_fp16 = mul(x = var_967_cast_fp16, y = var_968_to_fp16)[name = tensor<string, []>("op_969_cast_fp16")];
            tensor<int32, [4]> var_970 = const()[name = tensor<string, []>("op_970"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_971_cast_fp16 = reshape(shape = var_970, x = key_15_cast_fp16)[name = tensor<string, []>("op_971_cast_fp16")];
            tensor<bool, []> mh_w_23_transpose_x_0 = const()[name = tensor<string, []>("mh_w_23_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_23_transpose_y_0 = const()[name = tensor<string, []>("mh_w_23_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_969_cast_fp16, y = var_971_cast_fp16)[name = tensor<string, []>("mh_w_23_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_974_cast_fp16 = softmax(axis = var_818, x = mh_w_23_cast_fp16)[name = tensor<string, []>("op_974_cast_fp16")];
            tensor<int32, [4]> var_975 = const()[name = tensor<string, []>("op_975"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_976_cast_fp16 = reshape(shape = var_975, x = value_15_cast_fp16)[name = tensor<string, []>("op_976_cast_fp16")];
            tensor<bool, []> attn_15_transpose_x_0 = const()[name = tensor<string, []>("attn_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_15_transpose_y_0 = const()[name = tensor<string, []>("attn_15_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_976_cast_fp16, y = var_974_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
            tensor<int32, [4]> var_979 = const()[name = tensor<string, []>("op_979"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_33_cast_fp16 = reshape(shape = var_979, x = attn_15_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<int32, [2]> var_983 = const()[name = tensor<string, []>("op_983"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_985 = const()[name = tensor<string, []>("op_985"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_47_pad_type_0 = const()[name = tensor<string, []>("obj_47_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_47_pad_0 = const()[name = tensor<string, []>("obj_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201515008))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202743872))), name = tensor<string, []>("layers_3_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202744064)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_47_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_985, groups = var_825, pad = obj_47_pad_0, pad_type = obj_47_pad_type_0, strides = var_983, weight = layers_3_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = tensor<string, []>("obj_47_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_47_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
            tensor<int32, [1]> var_991 = const()[name = tensor<string, []>("op_991"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_23_cast_fp16 = reduce_mean(axes = var_991, keep_dims = var_826, x = inputs_23_cast_fp16)[name = tensor<string, []>("channels_mean_23_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor<string, []>("zero_mean_23_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor<string, []>("zero_mean_sq_23_cast_fp16")];
            tensor<int32, [1]> var_995 = const()[name = tensor<string, []>("op_995"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_996_cast_fp16 = reduce_mean(axes = var_995, keep_dims = var_826, x = zero_mean_sq_23_cast_fp16)[name = tensor<string, []>("op_996_cast_fp16")];
            tensor<fp16, []> var_997_to_fp16 = const()[name = tensor<string, []>("op_997_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_998_cast_fp16 = add(x = var_996_cast_fp16, y = var_997_to_fp16)[name = tensor<string, []>("op_998_cast_fp16")];
            tensor<fp32, []> denom_23_epsilon_0 = const()[name = tensor<string, []>("denom_23_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0, x = var_998_cast_fp16)[name = tensor<string, []>("denom_23_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
            tensor<fp16, [1280]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202746688)))];
            tensor<fp16, [1280]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202749312)))];
            tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<int32, [2]> var_1009 = const()[name = tensor<string, []>("op_1009"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1011 = const()[name = tensor<string, []>("op_1011"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202751936))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207667200))), name = tensor<string, []>("layers_3_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207667392)))];
            tensor<fp16, [1, 5120, 1, 1]> input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_1011, groups = var_825, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_1009, weight = layers_3_fc1_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<string, []> input_39_mode_0 = const()[name = tensor<string, []>("input_39_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<int32, [2]> var_1017 = const()[name = tensor<string, []>("op_1017"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1019 = const()[name = tensor<string, []>("op_1019"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207677696))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214231360))), name = tensor<string, []>("layers_3_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214231936)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_1019, groups = var_825, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_1017, weight = layers_3_fc2_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_25_cast_fp16")];
            tensor<int32, []> var_1032 = const()[name = tensor<string, []>("op_1032"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_1039 = const()[name = tensor<string, []>("op_1039"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_1040 = const()[name = tensor<string, []>("op_1040"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_1052 = const()[name = tensor<string, []>("op_1052"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_25_cast_fp16 = reduce_mean(axes = var_1052, keep_dims = var_1040, x = inputs_25_cast_fp16)[name = tensor<string, []>("channels_mean_25_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor<string, []>("zero_mean_25_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor<string, []>("zero_mean_sq_25_cast_fp16")];
            tensor<int32, [1]> var_1056 = const()[name = tensor<string, []>("op_1056"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1057_cast_fp16 = reduce_mean(axes = var_1056, keep_dims = var_1040, x = zero_mean_sq_25_cast_fp16)[name = tensor<string, []>("op_1057_cast_fp16")];
            tensor<fp16, []> var_1058_to_fp16 = const()[name = tensor<string, []>("op_1058_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1059_cast_fp16 = add(x = var_1057_cast_fp16, y = var_1058_to_fp16)[name = tensor<string, []>("op_1059_cast_fp16")];
            tensor<fp32, []> denom_25_epsilon_0 = const()[name = tensor<string, []>("denom_25_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0, x = var_1059_cast_fp16)[name = tensor<string, []>("denom_25_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
            tensor<fp16, [1280]> obj_49_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_49_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214234560)))];
            tensor<fp16, [1280]> obj_49_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_49_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214237184)))];
            tensor<fp16, []> obj_49_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_49_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
            tensor<int32, [2]> var_1074 = const()[name = tensor<string, []>("op_1074"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1076 = const()[name = tensor<string, []>("op_1076"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_17_pad_type_0 = const()[name = tensor<string, []>("query_17_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_17_pad_0 = const()[name = tensor<string, []>("query_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214239808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215468672))), name = tensor<string, []>("layers_4_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215468864)))];
            tensor<fp16, [1, 1280, 1, 1]> query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = var_1076, groups = var_1039, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = var_1074, weight = layers_4_self_attn_q_proj_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
            tensor<int32, [2]> var_1080 = const()[name = tensor<string, []>("op_1080"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1082 = const()[name = tensor<string, []>("op_1082"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_9_pad_type_0 = const()[name = tensor<string, []>("current_key_9_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_9_pad_0 = const()[name = tensor<string, []>("current_key_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215471488))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216290752))), name = tensor<string, []>("layers_4_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_9_cast_fp16 = conv(dilations = var_1082, groups = var_1039, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = var_1080, weight = layers_4_self_attn_k_proj_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor<string, []>("current_key_9_cast_fp16")];
            tensor<int32, [2]> var_1087 = const()[name = tensor<string, []>("op_1087"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1089 = const()[name = tensor<string, []>("op_1089"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_9_pad_type_0 = const()[name = tensor<string, []>("current_value_9_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_9_pad_0 = const()[name = tensor<string, []>("current_value_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216290880))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217519744))), name = tensor<string, []>("layers_4_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217519936)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = var_1089, groups = var_1039, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = var_1087, weight = layers_4_self_attn_v_proj_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor<string, []>("current_value_9_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1096_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1096_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1098_cast_fp16 = mul(x = var_103_cast_fp16_4, y = var_241_cast_fp16)[name = tensor<string, []>("op_1098_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_17_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1098_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1100_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1100_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1102_cast_fp16 = mul(x = var_138_cast_fp16_4, y = var_241_cast_fp16)[name = tensor<string, []>("op_1102_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_17_cast_fp16 = add(x = var_1100_cast_fp16, y = var_1102_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")];
            tensor<int32, [4]> var_1105 = const()[name = tensor<string, []>("op_1105"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1106_cast_fp16 = reshape(shape = var_1105, x = query_17_cast_fp16)[name = tensor<string, []>("op_1106_cast_fp16")];
            tensor<fp16, []> var_1107_to_fp16 = const()[name = tensor<string, []>("op_1107_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1108_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_1107_to_fp16)[name = tensor<string, []>("op_1108_cast_fp16")];
            tensor<int32, [4]> var_1109 = const()[name = tensor<string, []>("op_1109"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1110_cast_fp16 = reshape(shape = var_1109, x = key_17_cast_fp16)[name = tensor<string, []>("op_1110_cast_fp16")];
            tensor<bool, []> mh_w_25_transpose_x_0 = const()[name = tensor<string, []>("mh_w_25_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_25_transpose_y_0 = const()[name = tensor<string, []>("mh_w_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 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_1108_cast_fp16, y = var_1110_cast_fp16)[name = tensor<string, []>("mh_w_25_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_27_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_1118_cast_fp16 = softmax(axis = var_1032, x = mh_w_27_cast_fp16)[name = tensor<string, []>("op_1118_cast_fp16")];
            tensor<int32, [4]> var_1119 = const()[name = tensor<string, []>("op_1119"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1120_cast_fp16 = reshape(shape = var_1119, x = value_17_cast_fp16)[name = tensor<string, []>("op_1120_cast_fp16")];
            tensor<bool, []> attn_17_transpose_x_0 = const()[name = tensor<string, []>("attn_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_17_transpose_y_0 = const()[name = tensor<string, []>("attn_17_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1120_cast_fp16, y = var_1118_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
            tensor<int32, [4]> var_1123 = const()[name = tensor<string, []>("op_1123"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_41_cast_fp16 = reshape(shape = var_1123, x = attn_17_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<int32, [2]> var_1127 = const()[name = tensor<string, []>("op_1127"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1129 = const()[name = tensor<string, []>("op_1129"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_55_pad_type_0 = const()[name = tensor<string, []>("obj_55_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_55_pad_0 = const()[name = tensor<string, []>("obj_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217522560))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218751424))), name = tensor<string, []>("layers_4_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218751616)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_55_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = var_1129, groups = var_1039, pad = obj_55_pad_0, pad_type = obj_55_pad_type_0, strides = var_1127, weight = layers_4_self_attn_o_proj_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("inputs_27_cast_fp16")];
            tensor<int32, [1]> var_1139 = const()[name = tensor<string, []>("op_1139"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_27_cast_fp16 = reduce_mean(axes = var_1139, keep_dims = var_1040, x = inputs_27_cast_fp16)[name = tensor<string, []>("channels_mean_27_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor<string, []>("zero_mean_27_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor<string, []>("zero_mean_sq_27_cast_fp16")];
            tensor<int32, [1]> var_1143 = const()[name = tensor<string, []>("op_1143"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1144_cast_fp16 = reduce_mean(axes = var_1143, keep_dims = var_1040, x = zero_mean_sq_27_cast_fp16)[name = tensor<string, []>("op_1144_cast_fp16")];
            tensor<fp16, []> var_1145_to_fp16 = const()[name = tensor<string, []>("op_1145_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1146_cast_fp16 = add(x = var_1144_cast_fp16, y = var_1145_to_fp16)[name = tensor<string, []>("op_1146_cast_fp16")];
            tensor<fp32, []> denom_27_epsilon_0 = const()[name = tensor<string, []>("denom_27_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0, x = var_1146_cast_fp16)[name = tensor<string, []>("denom_27_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
            tensor<fp16, [1280]> obj_57_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_57_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218754240)))];
            tensor<fp16, [1280]> obj_57_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_57_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218756864)))];
            tensor<fp16, []> obj_57_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_57_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor<string, []>("obj_57_cast_fp16")];
            tensor<int32, [2]> var_1161 = const()[name = tensor<string, []>("op_1161"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1163 = const()[name = tensor<string, []>("op_1163"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_19_pad_type_0 = const()[name = tensor<string, []>("query_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_19_pad_0 = const()[name = tensor<string, []>("query_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218759488))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219578752))), name = tensor<string, []>("layers_4_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219578880)))];
            tensor<fp16, [1, 1280, 1, 1]> query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = var_1163, groups = var_1039, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = var_1161, weight = layers_4_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
            tensor<int32, [2]> var_1167 = const()[name = tensor<string, []>("op_1167"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1169 = const()[name = tensor<string, []>("op_1169"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_19_pad_type_0 = const()[name = tensor<string, []>("key_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_19_pad_0 = const()[name = tensor<string, []>("key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219581504))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220810368))), name = tensor<string, []>("layers_4_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_19_cast_fp16 = conv(dilations = var_1169, groups = var_1039, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = var_1167, weight = layers_4_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_19_cast_fp16")];
            tensor<int32, [2]> var_1174 = const()[name = tensor<string, []>("op_1174"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1176 = const()[name = tensor<string, []>("op_1176"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_19_pad_type_0 = const()[name = tensor<string, []>("value_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_19_pad_0 = const()[name = tensor<string, []>("value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220810560))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222039424))), name = tensor<string, []>("layers_4_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222039616)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = var_1176, groups = var_1039, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = var_1174, weight = layers_4_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_19_cast_fp16")];
            tensor<int32, [4]> var_1180 = const()[name = tensor<string, []>("op_1180"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1181_cast_fp16 = reshape(shape = var_1180, x = query_19_cast_fp16)[name = tensor<string, []>("op_1181_cast_fp16")];
            tensor<fp16, []> var_1182_to_fp16 = const()[name = tensor<string, []>("op_1182_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1183_cast_fp16 = mul(x = var_1181_cast_fp16, y = var_1182_to_fp16)[name = tensor<string, []>("op_1183_cast_fp16")];
            tensor<int32, [4]> var_1184 = const()[name = tensor<string, []>("op_1184"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1185_cast_fp16 = reshape(shape = var_1184, x = key_19_cast_fp16)[name = tensor<string, []>("op_1185_cast_fp16")];
            tensor<bool, []> mh_w_29_transpose_x_0 = const()[name = tensor<string, []>("mh_w_29_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_29_transpose_y_0 = const()[name = tensor<string, []>("mh_w_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1183_cast_fp16, y = var_1185_cast_fp16)[name = tensor<string, []>("mh_w_29_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_1188_cast_fp16 = softmax(axis = var_1032, x = mh_w_29_cast_fp16)[name = tensor<string, []>("op_1188_cast_fp16")];
            tensor<int32, [4]> var_1189 = const()[name = tensor<string, []>("op_1189"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1190_cast_fp16 = reshape(shape = var_1189, x = value_19_cast_fp16)[name = tensor<string, []>("op_1190_cast_fp16")];
            tensor<bool, []> attn_19_transpose_x_0 = const()[name = tensor<string, []>("attn_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_19_transpose_y_0 = const()[name = tensor<string, []>("attn_19_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1190_cast_fp16, y = var_1188_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
            tensor<int32, [4]> var_1193 = const()[name = tensor<string, []>("op_1193"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_43_cast_fp16 = reshape(shape = var_1193, x = attn_19_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<int32, [2]> var_1197 = const()[name = tensor<string, []>("op_1197"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1199 = const()[name = tensor<string, []>("op_1199"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_59_pad_type_0 = const()[name = tensor<string, []>("obj_59_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_59_pad_0 = const()[name = tensor<string, []>("obj_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_4_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222042240))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223271104))), name = tensor<string, []>("layers_4_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223271296)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_59_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = var_1199, groups = var_1039, pad = obj_59_pad_0, pad_type = obj_59_pad_type_0, strides = var_1197, weight = layers_4_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor<string, []>("obj_59_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_59_cast_fp16)[name = tensor<string, []>("inputs_29_cast_fp16")];
            tensor<int32, [1]> var_1205 = const()[name = tensor<string, []>("op_1205"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_29_cast_fp16 = reduce_mean(axes = var_1205, keep_dims = var_1040, x = inputs_29_cast_fp16)[name = tensor<string, []>("channels_mean_29_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor<string, []>("zero_mean_29_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor<string, []>("zero_mean_sq_29_cast_fp16")];
            tensor<int32, [1]> var_1209 = const()[name = tensor<string, []>("op_1209"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1210_cast_fp16 = reduce_mean(axes = var_1209, keep_dims = var_1040, x = zero_mean_sq_29_cast_fp16)[name = tensor<string, []>("op_1210_cast_fp16")];
            tensor<fp16, []> var_1211_to_fp16 = const()[name = tensor<string, []>("op_1211_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1212_cast_fp16 = add(x = var_1210_cast_fp16, y = var_1211_to_fp16)[name = tensor<string, []>("op_1212_cast_fp16")];
            tensor<fp32, []> denom_29_epsilon_0 = const()[name = tensor<string, []>("denom_29_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0, x = var_1212_cast_fp16)[name = tensor<string, []>("denom_29_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
            tensor<fp16, [1280]> input_45_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_45_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223273920)))];
            tensor<fp16, [1280]> input_45_beta_0_to_fp16 = const()[name = tensor<string, []>("input_45_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223276544)))];
            tensor<fp16, []> input_45_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_45_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<int32, [2]> var_1223 = const()[name = tensor<string, []>("op_1223"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1225 = const()[name = tensor<string, []>("op_1225"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_47_pad_type_0 = const()[name = tensor<string, []>("input_47_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_47_pad_0 = const()[name = tensor<string, []>("input_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_4_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223279168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228194432))), name = tensor<string, []>("layers_4_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_4_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228194624)))];
            tensor<fp16, [1, 5120, 1, 1]> input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = var_1225, groups = var_1039, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = var_1223, weight = layers_4_fc1_weight_to_fp16_palettized, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<string, []> input_49_mode_0 = const()[name = tensor<string, []>("input_49_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<int32, [2]> var_1231 = const()[name = tensor<string, []>("op_1231"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1233 = const()[name = tensor<string, []>("op_1233"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_11_pad_type_0 = const()[name = tensor<string, []>("hidden_states_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = tensor<string, []>("hidden_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_4_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228204928))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233120192))), name = tensor<string, []>("layers_4_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_4_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233120384)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = var_1233, groups = var_1039, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = var_1231, weight = layers_4_fc2_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("hidden_states_11_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor<string, []>("inputs_31_cast_fp16")];
            tensor<int32, []> var_1246 = const()[name = tensor<string, []>("op_1246"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_1253 = const()[name = tensor<string, []>("op_1253"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_1254 = const()[name = tensor<string, []>("op_1254"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_1266 = const()[name = tensor<string, []>("op_1266"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_31_cast_fp16 = reduce_mean(axes = var_1266, keep_dims = var_1254, x = inputs_31_cast_fp16)[name = tensor<string, []>("channels_mean_31_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor<string, []>("zero_mean_31_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor<string, []>("zero_mean_sq_31_cast_fp16")];
            tensor<int32, [1]> var_1270 = const()[name = tensor<string, []>("op_1270"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1271_cast_fp16 = reduce_mean(axes = var_1270, keep_dims = var_1254, x = zero_mean_sq_31_cast_fp16)[name = tensor<string, []>("op_1271_cast_fp16")];
            tensor<fp16, []> var_1272_to_fp16 = const()[name = tensor<string, []>("op_1272_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1273_cast_fp16 = add(x = var_1271_cast_fp16, y = var_1272_to_fp16)[name = tensor<string, []>("op_1273_cast_fp16")];
            tensor<fp32, []> denom_31_epsilon_0 = const()[name = tensor<string, []>("denom_31_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0, x = var_1273_cast_fp16)[name = tensor<string, []>("denom_31_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
            tensor<fp16, [1280]> obj_61_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_61_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233123008)))];
            tensor<fp16, [1280]> obj_61_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_61_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233125632)))];
            tensor<fp16, []> obj_61_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_61_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor<string, []>("obj_61_cast_fp16")];
            tensor<int32, [2]> var_1288 = const()[name = tensor<string, []>("op_1288"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1290 = const()[name = tensor<string, []>("op_1290"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_21_pad_type_0 = const()[name = tensor<string, []>("query_21_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_21_pad_0 = const()[name = tensor<string, []>("query_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233128256))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234357120))), name = tensor<string, []>("layers_5_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234357312)))];
            tensor<fp16, [1, 1280, 1, 1]> query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = var_1290, groups = var_1253, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = var_1288, weight = layers_5_self_attn_q_proj_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
            tensor<int32, [2]> var_1294 = const()[name = tensor<string, []>("op_1294"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1296 = const()[name = tensor<string, []>("op_1296"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_11_pad_type_0 = const()[name = tensor<string, []>("current_key_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_11_pad_0 = const()[name = tensor<string, []>("current_key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234359936))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235588800))), name = tensor<string, []>("layers_5_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_11_cast_fp16 = conv(dilations = var_1296, groups = var_1253, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = var_1294, weight = layers_5_self_attn_k_proj_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor<string, []>("current_key_11_cast_fp16")];
            tensor<int32, [2]> var_1301 = const()[name = tensor<string, []>("op_1301"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1303 = const()[name = tensor<string, []>("op_1303"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_11_pad_type_0 = const()[name = tensor<string, []>("current_value_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_11_pad_0 = const()[name = tensor<string, []>("current_value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235588992))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236408256))), name = tensor<string, []>("layers_5_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236408384)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = var_1303, groups = var_1253, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = var_1301, weight = layers_5_self_attn_v_proj_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor<string, []>("current_value_11_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1310_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1310_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1312_cast_fp16 = mul(x = var_103_cast_fp16_5, y = var_241_cast_fp16)[name = tensor<string, []>("op_1312_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_21_cast_fp16 = add(x = var_1310_cast_fp16, y = var_1312_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1314_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1314_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1316_cast_fp16 = mul(x = var_138_cast_fp16_5, y = var_241_cast_fp16)[name = tensor<string, []>("op_1316_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_21_cast_fp16 = add(x = var_1314_cast_fp16, y = var_1316_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")];
            tensor<int32, [4]> var_1319 = const()[name = tensor<string, []>("op_1319"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1320_cast_fp16 = reshape(shape = var_1319, x = query_21_cast_fp16)[name = tensor<string, []>("op_1320_cast_fp16")];
            tensor<fp16, []> var_1321_to_fp16 = const()[name = tensor<string, []>("op_1321_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1322_cast_fp16 = mul(x = var_1320_cast_fp16, y = var_1321_to_fp16)[name = tensor<string, []>("op_1322_cast_fp16")];
            tensor<int32, [4]> var_1323 = const()[name = tensor<string, []>("op_1323"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1324_cast_fp16 = reshape(shape = var_1323, x = key_21_cast_fp16)[name = tensor<string, []>("op_1324_cast_fp16")];
            tensor<bool, []> mh_w_31_transpose_x_0 = const()[name = tensor<string, []>("mh_w_31_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_31_transpose_y_0 = const()[name = tensor<string, []>("mh_w_31_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_1322_cast_fp16, y = var_1324_cast_fp16)[name = tensor<string, []>("mh_w_31_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_33_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_1332_cast_fp16 = softmax(axis = var_1246, x = mh_w_33_cast_fp16)[name = tensor<string, []>("op_1332_cast_fp16")];
            tensor<int32, [4]> var_1333 = const()[name = tensor<string, []>("op_1333"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1334_cast_fp16 = reshape(shape = var_1333, x = value_21_cast_fp16)[name = tensor<string, []>("op_1334_cast_fp16")];
            tensor<bool, []> attn_21_transpose_x_0 = const()[name = tensor<string, []>("attn_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_21_transpose_y_0 = const()[name = tensor<string, []>("attn_21_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1334_cast_fp16, y = var_1332_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
            tensor<int32, [4]> var_1337 = const()[name = tensor<string, []>("op_1337"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_51_cast_fp16 = reshape(shape = var_1337, x = attn_21_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<int32, [2]> var_1341 = const()[name = tensor<string, []>("op_1341"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1343 = const()[name = tensor<string, []>("op_1343"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_67_pad_type_0 = const()[name = tensor<string, []>("obj_67_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_67_pad_0 = const()[name = tensor<string, []>("obj_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236411008))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237639872))), name = tensor<string, []>("layers_5_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237640064)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_67_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = var_1343, groups = var_1253, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = var_1341, weight = layers_5_self_attn_o_proj_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor<string, []>("obj_67_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_67_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")];
            tensor<int32, [1]> var_1353 = const()[name = tensor<string, []>("op_1353"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_33_cast_fp16 = reduce_mean(axes = var_1353, keep_dims = var_1254, x = inputs_33_cast_fp16)[name = tensor<string, []>("channels_mean_33_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor<string, []>("zero_mean_33_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor<string, []>("zero_mean_sq_33_cast_fp16")];
            tensor<int32, [1]> var_1357 = const()[name = tensor<string, []>("op_1357"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1358_cast_fp16 = reduce_mean(axes = var_1357, keep_dims = var_1254, x = zero_mean_sq_33_cast_fp16)[name = tensor<string, []>("op_1358_cast_fp16")];
            tensor<fp16, []> var_1359_to_fp16 = const()[name = tensor<string, []>("op_1359_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1360_cast_fp16 = add(x = var_1358_cast_fp16, y = var_1359_to_fp16)[name = tensor<string, []>("op_1360_cast_fp16")];
            tensor<fp32, []> denom_33_epsilon_0 = const()[name = tensor<string, []>("denom_33_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0, x = var_1360_cast_fp16)[name = tensor<string, []>("denom_33_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
            tensor<fp16, [1280]> obj_69_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_69_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237642688)))];
            tensor<fp16, [1280]> obj_69_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_69_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237645312)))];
            tensor<fp16, []> obj_69_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_69_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor<string, []>("obj_69_cast_fp16")];
            tensor<int32, [2]> var_1375 = const()[name = tensor<string, []>("op_1375"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1377 = const()[name = tensor<string, []>("op_1377"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_23_pad_type_0 = const()[name = tensor<string, []>("query_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_23_pad_0 = const()[name = tensor<string, []>("query_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237647936))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238467200))), name = tensor<string, []>("layers_5_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238467328)))];
            tensor<fp16, [1, 1280, 1, 1]> query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = var_1377, groups = var_1253, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = var_1375, weight = layers_5_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = tensor<string, []>("query_23_cast_fp16")];
            tensor<int32, [2]> var_1381 = const()[name = tensor<string, []>("op_1381"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1383 = const()[name = tensor<string, []>("op_1383"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_23_pad_type_0 = const()[name = tensor<string, []>("key_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_23_pad_0 = const()[name = tensor<string, []>("key_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238469952))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239289216))), name = tensor<string, []>("layers_5_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_23_cast_fp16 = conv(dilations = var_1383, groups = var_1253, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = var_1381, weight = layers_5_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_23_cast_fp16")];
            tensor<int32, [2]> var_1388 = const()[name = tensor<string, []>("op_1388"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1390 = const()[name = tensor<string, []>("op_1390"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_23_pad_type_0 = const()[name = tensor<string, []>("value_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_23_pad_0 = const()[name = tensor<string, []>("value_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239289344))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240108608))), name = tensor<string, []>("layers_5_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240108736)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = var_1390, groups = var_1253, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = var_1388, weight = layers_5_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_23_cast_fp16")];
            tensor<int32, [4]> var_1394 = const()[name = tensor<string, []>("op_1394"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1395_cast_fp16 = reshape(shape = var_1394, x = query_23_cast_fp16)[name = tensor<string, []>("op_1395_cast_fp16")];
            tensor<fp16, []> var_1396_to_fp16 = const()[name = tensor<string, []>("op_1396_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1397_cast_fp16 = mul(x = var_1395_cast_fp16, y = var_1396_to_fp16)[name = tensor<string, []>("op_1397_cast_fp16")];
            tensor<int32, [4]> var_1398 = const()[name = tensor<string, []>("op_1398"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1399_cast_fp16 = reshape(shape = var_1398, x = key_23_cast_fp16)[name = tensor<string, []>("op_1399_cast_fp16")];
            tensor<bool, []> mh_w_35_transpose_x_0 = const()[name = tensor<string, []>("mh_w_35_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_35_transpose_y_0 = const()[name = tensor<string, []>("mh_w_35_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_1397_cast_fp16, y = var_1399_cast_fp16)[name = tensor<string, []>("mh_w_35_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_1402_cast_fp16 = softmax(axis = var_1246, x = mh_w_35_cast_fp16)[name = tensor<string, []>("op_1402_cast_fp16")];
            tensor<int32, [4]> var_1403 = const()[name = tensor<string, []>("op_1403"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1404_cast_fp16 = reshape(shape = var_1403, x = value_23_cast_fp16)[name = tensor<string, []>("op_1404_cast_fp16")];
            tensor<bool, []> attn_23_transpose_x_0 = const()[name = tensor<string, []>("attn_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_23_transpose_y_0 = const()[name = tensor<string, []>("attn_23_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1404_cast_fp16, y = var_1402_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")];
            tensor<int32, [4]> var_1407 = const()[name = tensor<string, []>("op_1407"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_53_cast_fp16 = reshape(shape = var_1407, x = attn_23_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<int32, [2]> var_1411 = const()[name = tensor<string, []>("op_1411"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1413 = const()[name = tensor<string, []>("op_1413"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_71_pad_type_0 = const()[name = tensor<string, []>("obj_71_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_71_pad_0 = const()[name = tensor<string, []>("obj_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_5_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240111360))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241340224))), name = tensor<string, []>("layers_5_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241340416)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_71_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = var_1413, groups = var_1253, pad = obj_71_pad_0, pad_type = obj_71_pad_type_0, strides = var_1411, weight = layers_5_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_53_cast_fp16)[name = tensor<string, []>("obj_71_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_71_cast_fp16)[name = tensor<string, []>("inputs_35_cast_fp16")];
            tensor<int32, [1]> var_1419 = const()[name = tensor<string, []>("op_1419"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_35_cast_fp16 = reduce_mean(axes = var_1419, keep_dims = var_1254, x = inputs_35_cast_fp16)[name = tensor<string, []>("channels_mean_35_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor<string, []>("zero_mean_35_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor<string, []>("zero_mean_sq_35_cast_fp16")];
            tensor<int32, [1]> var_1423 = const()[name = tensor<string, []>("op_1423"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1424_cast_fp16 = reduce_mean(axes = var_1423, keep_dims = var_1254, x = zero_mean_sq_35_cast_fp16)[name = tensor<string, []>("op_1424_cast_fp16")];
            tensor<fp16, []> var_1425_to_fp16 = const()[name = tensor<string, []>("op_1425_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1426_cast_fp16 = add(x = var_1424_cast_fp16, y = var_1425_to_fp16)[name = tensor<string, []>("op_1426_cast_fp16")];
            tensor<fp32, []> denom_35_epsilon_0 = const()[name = tensor<string, []>("denom_35_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0, x = var_1426_cast_fp16)[name = tensor<string, []>("denom_35_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
            tensor<fp16, [1280]> input_55_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_55_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241343040)))];
            tensor<fp16, [1280]> input_55_beta_0_to_fp16 = const()[name = tensor<string, []>("input_55_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241345664)))];
            tensor<fp16, []> input_55_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_55_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<int32, [2]> var_1437 = const()[name = tensor<string, []>("op_1437"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1439 = const()[name = tensor<string, []>("op_1439"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_57_pad_type_0 = const()[name = tensor<string, []>("input_57_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_57_pad_0 = const()[name = tensor<string, []>("input_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_5_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241348288))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246263552))), name = tensor<string, []>("layers_5_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_5_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246263744)))];
            tensor<fp16, [1, 5120, 1, 1]> input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = var_1439, groups = var_1253, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = var_1437, weight = layers_5_fc1_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<string, []> input_59_mode_0 = const()[name = tensor<string, []>("input_59_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<int32, [2]> var_1445 = const()[name = tensor<string, []>("op_1445"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1447 = const()[name = tensor<string, []>("op_1447"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_13_pad_type_0 = const()[name = tensor<string, []>("hidden_states_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = tensor<string, []>("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_5_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246274048))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251189312))), name = tensor<string, []>("layers_5_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_5_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251189504)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = var_1447, groups = var_1253, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = var_1445, weight = layers_5_fc2_weight_to_fp16_palettized, x = input_59_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor<string, []>("inputs_37_cast_fp16")];
            tensor<int32, []> var_1460 = const()[name = tensor<string, []>("op_1460"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_1467 = const()[name = tensor<string, []>("op_1467"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_1468 = const()[name = tensor<string, []>("op_1468"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_1480 = const()[name = tensor<string, []>("op_1480"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_37_cast_fp16 = reduce_mean(axes = var_1480, keep_dims = var_1468, x = inputs_37_cast_fp16)[name = tensor<string, []>("channels_mean_37_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_37_cast_fp16 = sub(x = inputs_37_cast_fp16, y = channels_mean_37_cast_fp16)[name = tensor<string, []>("zero_mean_37_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = zero_mean_37_cast_fp16)[name = tensor<string, []>("zero_mean_sq_37_cast_fp16")];
            tensor<int32, [1]> var_1484 = const()[name = tensor<string, []>("op_1484"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1485_cast_fp16 = reduce_mean(axes = var_1484, keep_dims = var_1468, x = zero_mean_sq_37_cast_fp16)[name = tensor<string, []>("op_1485_cast_fp16")];
            tensor<fp16, []> var_1486_to_fp16 = const()[name = tensor<string, []>("op_1486_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1487_cast_fp16 = add(x = var_1485_cast_fp16, y = var_1486_to_fp16)[name = tensor<string, []>("op_1487_cast_fp16")];
            tensor<fp32, []> denom_37_epsilon_0 = const()[name = tensor<string, []>("denom_37_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0, x = var_1487_cast_fp16)[name = tensor<string, []>("denom_37_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = denom_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
            tensor<fp16, [1280]> obj_73_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_73_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251192128)))];
            tensor<fp16, [1280]> obj_73_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_73_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251194752)))];
            tensor<fp16, []> obj_73_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_73_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor<string, []>("obj_73_cast_fp16")];
            tensor<int32, [2]> var_1502 = const()[name = tensor<string, []>("op_1502"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1504 = const()[name = tensor<string, []>("op_1504"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_25_pad_type_0 = const()[name = tensor<string, []>("query_25_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_25_pad_0 = const()[name = tensor<string, []>("query_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251197376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252426240))), name = tensor<string, []>("layers_6_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252426432)))];
            tensor<fp16, [1, 1280, 1, 1]> query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = var_1504, groups = var_1467, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = var_1502, weight = layers_6_self_attn_q_proj_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")];
            tensor<int32, [2]> var_1508 = const()[name = tensor<string, []>("op_1508"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1510 = const()[name = tensor<string, []>("op_1510"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_13_pad_type_0 = const()[name = tensor<string, []>("current_key_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_13_pad_0 = const()[name = tensor<string, []>("current_key_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252429056))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253248320))), name = tensor<string, []>("layers_6_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_13_cast_fp16 = conv(dilations = var_1510, groups = var_1467, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = var_1508, weight = layers_6_self_attn_k_proj_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor<string, []>("current_key_13_cast_fp16")];
            tensor<int32, [2]> var_1515 = const()[name = tensor<string, []>("op_1515"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1517 = const()[name = tensor<string, []>("op_1517"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_13_pad_type_0 = const()[name = tensor<string, []>("current_value_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_13_pad_0 = const()[name = tensor<string, []>("current_value_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253248448))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254067712))), name = tensor<string, []>("layers_6_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254067840)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = var_1517, groups = var_1467, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = var_1515, weight = layers_6_self_attn_v_proj_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor<string, []>("current_value_13_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1524_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1524_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1526_cast_fp16 = mul(x = var_103_cast_fp16_6, y = var_241_cast_fp16)[name = tensor<string, []>("op_1526_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_25_cast_fp16 = add(x = var_1524_cast_fp16, y = var_1526_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1528_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1528_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1530_cast_fp16 = mul(x = var_138_cast_fp16_6, y = var_241_cast_fp16)[name = tensor<string, []>("op_1530_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_25_cast_fp16 = add(x = var_1528_cast_fp16, y = var_1530_cast_fp16)[name = tensor<string, []>("value_25_cast_fp16")];
            tensor<int32, [4]> var_1533 = const()[name = tensor<string, []>("op_1533"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1534_cast_fp16 = reshape(shape = var_1533, x = query_25_cast_fp16)[name = tensor<string, []>("op_1534_cast_fp16")];
            tensor<fp16, []> var_1535_to_fp16 = const()[name = tensor<string, []>("op_1535_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1536_cast_fp16 = mul(x = var_1534_cast_fp16, y = var_1535_to_fp16)[name = tensor<string, []>("op_1536_cast_fp16")];
            tensor<int32, [4]> var_1537 = const()[name = tensor<string, []>("op_1537"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1538_cast_fp16 = reshape(shape = var_1537, x = key_25_cast_fp16)[name = tensor<string, []>("op_1538_cast_fp16")];
            tensor<bool, []> mh_w_37_transpose_x_0 = const()[name = tensor<string, []>("mh_w_37_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_37_transpose_y_0 = const()[name = tensor<string, []>("mh_w_37_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1536_cast_fp16, y = var_1538_cast_fp16)[name = tensor<string, []>("mh_w_37_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_39_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_1546_cast_fp16 = softmax(axis = var_1460, x = mh_w_39_cast_fp16)[name = tensor<string, []>("op_1546_cast_fp16")];
            tensor<int32, [4]> var_1547 = const()[name = tensor<string, []>("op_1547"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1548_cast_fp16 = reshape(shape = var_1547, x = value_25_cast_fp16)[name = tensor<string, []>("op_1548_cast_fp16")];
            tensor<bool, []> attn_25_transpose_x_0 = const()[name = tensor<string, []>("attn_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_25_transpose_y_0 = const()[name = tensor<string, []>("attn_25_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1548_cast_fp16, y = var_1546_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
            tensor<int32, [4]> var_1551 = const()[name = tensor<string, []>("op_1551"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_61_cast_fp16 = reshape(shape = var_1551, x = attn_25_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
            tensor<int32, [2]> var_1555 = const()[name = tensor<string, []>("op_1555"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1557 = const()[name = tensor<string, []>("op_1557"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_79_pad_type_0 = const()[name = tensor<string, []>("obj_79_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_79_pad_0 = const()[name = tensor<string, []>("obj_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254070464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255299328))), name = tensor<string, []>("layers_6_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255299520)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_79_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = var_1557, groups = var_1467, pad = obj_79_pad_0, pad_type = obj_79_pad_type_0, strides = var_1555, weight = layers_6_self_attn_o_proj_weight_to_fp16_palettized, x = input_61_cast_fp16)[name = tensor<string, []>("obj_79_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_79_cast_fp16)[name = tensor<string, []>("inputs_39_cast_fp16")];
            tensor<int32, [1]> var_1567 = const()[name = tensor<string, []>("op_1567"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_39_cast_fp16 = reduce_mean(axes = var_1567, keep_dims = var_1468, x = inputs_39_cast_fp16)[name = tensor<string, []>("channels_mean_39_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_39_cast_fp16 = sub(x = inputs_39_cast_fp16, y = channels_mean_39_cast_fp16)[name = tensor<string, []>("zero_mean_39_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = zero_mean_39_cast_fp16)[name = tensor<string, []>("zero_mean_sq_39_cast_fp16")];
            tensor<int32, [1]> var_1571 = const()[name = tensor<string, []>("op_1571"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1572_cast_fp16 = reduce_mean(axes = var_1571, keep_dims = var_1468, x = zero_mean_sq_39_cast_fp16)[name = tensor<string, []>("op_1572_cast_fp16")];
            tensor<fp16, []> var_1573_to_fp16 = const()[name = tensor<string, []>("op_1573_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1574_cast_fp16 = add(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = tensor<string, []>("op_1574_cast_fp16")];
            tensor<fp32, []> denom_39_epsilon_0 = const()[name = tensor<string, []>("denom_39_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0, x = var_1574_cast_fp16)[name = tensor<string, []>("denom_39_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = denom_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
            tensor<fp16, [1280]> obj_81_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_81_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255302144)))];
            tensor<fp16, [1280]> obj_81_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_81_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255304768)))];
            tensor<fp16, []> obj_81_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_81_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor<string, []>("obj_81_cast_fp16")];
            tensor<int32, [2]> var_1589 = const()[name = tensor<string, []>("op_1589"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1591 = const()[name = tensor<string, []>("op_1591"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_27_pad_type_0 = const()[name = tensor<string, []>("query_27_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_27_pad_0 = const()[name = tensor<string, []>("query_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255307392))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256126656))), name = tensor<string, []>("layers_6_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256126784)))];
            tensor<fp16, [1, 1280, 1, 1]> query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = var_1591, groups = var_1467, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = var_1589, weight = layers_6_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = tensor<string, []>("query_27_cast_fp16")];
            tensor<int32, [2]> var_1595 = const()[name = tensor<string, []>("op_1595"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1597 = const()[name = tensor<string, []>("op_1597"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_27_pad_type_0 = const()[name = tensor<string, []>("key_27_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_27_pad_0 = const()[name = tensor<string, []>("key_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256129408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256948672))), name = tensor<string, []>("layers_6_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_27_cast_fp16 = conv(dilations = var_1597, groups = var_1467, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = var_1595, weight = layers_6_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_27_cast_fp16")];
            tensor<int32, [2]> var_1602 = const()[name = tensor<string, []>("op_1602"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1604 = const()[name = tensor<string, []>("op_1604"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_27_pad_type_0 = const()[name = tensor<string, []>("value_27_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_27_pad_0 = const()[name = tensor<string, []>("value_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256948800))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257768064))), name = tensor<string, []>("layers_6_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_6_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257768192)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = var_1604, groups = var_1467, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = var_1602, weight = layers_6_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_27_cast_fp16")];
            tensor<int32, [4]> var_1608 = const()[name = tensor<string, []>("op_1608"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1609_cast_fp16 = reshape(shape = var_1608, x = query_27_cast_fp16)[name = tensor<string, []>("op_1609_cast_fp16")];
            tensor<fp16, []> var_1610_to_fp16 = const()[name = tensor<string, []>("op_1610_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1611_cast_fp16 = mul(x = var_1609_cast_fp16, y = var_1610_to_fp16)[name = tensor<string, []>("op_1611_cast_fp16")];
            tensor<int32, [4]> var_1612 = const()[name = tensor<string, []>("op_1612"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1613_cast_fp16 = reshape(shape = var_1612, x = key_27_cast_fp16)[name = tensor<string, []>("op_1613_cast_fp16")];
            tensor<bool, []> mh_w_41_transpose_x_0 = const()[name = tensor<string, []>("mh_w_41_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_41_transpose_y_0 = const()[name = tensor<string, []>("mh_w_41_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1611_cast_fp16, y = var_1613_cast_fp16)[name = tensor<string, []>("mh_w_41_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_1616_cast_fp16 = softmax(axis = var_1460, x = mh_w_41_cast_fp16)[name = tensor<string, []>("op_1616_cast_fp16")];
            tensor<int32, [4]> var_1617 = const()[name = tensor<string, []>("op_1617"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1618_cast_fp16 = reshape(shape = var_1617, x = value_27_cast_fp16)[name = tensor<string, []>("op_1618_cast_fp16")];
            tensor<bool, []> attn_27_transpose_x_0 = const()[name = tensor<string, []>("attn_27_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_27_transpose_y_0 = const()[name = tensor<string, []>("attn_27_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1618_cast_fp16, y = var_1616_cast_fp16)[name = tensor<string, []>("attn_27_cast_fp16")];
            tensor<int32, [4]> var_1621 = const()[name = tensor<string, []>("op_1621"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_63_cast_fp16 = reshape(shape = var_1621, x = attn_27_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<int32, [2]> var_1625 = const()[name = tensor<string, []>("op_1625"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1627 = const()[name = tensor<string, []>("op_1627"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_83_pad_type_0 = const()[name = tensor<string, []>("obj_83_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_83_pad_0 = const()[name = tensor<string, []>("obj_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_6_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257770816))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259409280))), name = tensor<string, []>("layers_6_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259409856)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_83_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = var_1627, groups = var_1467, pad = obj_83_pad_0, pad_type = obj_83_pad_type_0, strides = var_1625, weight = layers_6_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = tensor<string, []>("obj_83_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_83_cast_fp16)[name = tensor<string, []>("inputs_41_cast_fp16")];
            tensor<int32, [1]> var_1633 = const()[name = tensor<string, []>("op_1633"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_41_cast_fp16 = reduce_mean(axes = var_1633, keep_dims = var_1468, x = inputs_41_cast_fp16)[name = tensor<string, []>("channels_mean_41_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_41_cast_fp16 = sub(x = inputs_41_cast_fp16, y = channels_mean_41_cast_fp16)[name = tensor<string, []>("zero_mean_41_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = zero_mean_41_cast_fp16)[name = tensor<string, []>("zero_mean_sq_41_cast_fp16")];
            tensor<int32, [1]> var_1637 = const()[name = tensor<string, []>("op_1637"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1638_cast_fp16 = reduce_mean(axes = var_1637, keep_dims = var_1468, x = zero_mean_sq_41_cast_fp16)[name = tensor<string, []>("op_1638_cast_fp16")];
            tensor<fp16, []> var_1639_to_fp16 = const()[name = tensor<string, []>("op_1639_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1640_cast_fp16 = add(x = var_1638_cast_fp16, y = var_1639_to_fp16)[name = tensor<string, []>("op_1640_cast_fp16")];
            tensor<fp32, []> denom_41_epsilon_0 = const()[name = tensor<string, []>("denom_41_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0, x = var_1640_cast_fp16)[name = tensor<string, []>("denom_41_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = denom_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
            tensor<fp16, [1280]> input_65_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_65_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259412480)))];
            tensor<fp16, [1280]> input_65_beta_0_to_fp16 = const()[name = tensor<string, []>("input_65_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259415104)))];
            tensor<fp16, []> input_65_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_65_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<int32, [2]> var_1651 = const()[name = tensor<string, []>("op_1651"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1653 = const()[name = tensor<string, []>("op_1653"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_67_pad_type_0 = const()[name = tensor<string, []>("input_67_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_67_pad_0 = const()[name = tensor<string, []>("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_6_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259417728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264332992))), name = tensor<string, []>("layers_6_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_6_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264333184)))];
            tensor<fp16, [1, 5120, 1, 1]> input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = var_1653, groups = var_1467, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = var_1651, weight = layers_6_fc1_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<string, []> input_69_mode_0 = const()[name = tensor<string, []>("input_69_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<int32, [2]> var_1659 = const()[name = tensor<string, []>("op_1659"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1661 = const()[name = tensor<string, []>("op_1661"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_15_pad_type_0 = const()[name = tensor<string, []>("hidden_states_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = tensor<string, []>("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_6_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264343488))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270897152))), name = tensor<string, []>("layers_6_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_6_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270897728)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = var_1661, groups = var_1467, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = var_1659, weight = layers_6_fc2_weight_to_fp16_palettized, x = input_69_cast_fp16)[name = tensor<string, []>("hidden_states_15_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")];
            tensor<int32, []> var_1674 = const()[name = tensor<string, []>("op_1674"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_1681 = const()[name = tensor<string, []>("op_1681"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_1682 = const()[name = tensor<string, []>("op_1682"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_1694 = const()[name = tensor<string, []>("op_1694"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_43_cast_fp16 = reduce_mean(axes = var_1694, keep_dims = var_1682, x = inputs_43_cast_fp16)[name = tensor<string, []>("channels_mean_43_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_43_cast_fp16 = sub(x = inputs_43_cast_fp16, y = channels_mean_43_cast_fp16)[name = tensor<string, []>("zero_mean_43_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = zero_mean_43_cast_fp16)[name = tensor<string, []>("zero_mean_sq_43_cast_fp16")];
            tensor<int32, [1]> var_1698 = const()[name = tensor<string, []>("op_1698"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1699_cast_fp16 = reduce_mean(axes = var_1698, keep_dims = var_1682, x = zero_mean_sq_43_cast_fp16)[name = tensor<string, []>("op_1699_cast_fp16")];
            tensor<fp16, []> var_1700_to_fp16 = const()[name = tensor<string, []>("op_1700_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1701_cast_fp16 = add(x = var_1699_cast_fp16, y = var_1700_to_fp16)[name = tensor<string, []>("op_1701_cast_fp16")];
            tensor<fp32, []> denom_43_epsilon_0 = const()[name = tensor<string, []>("denom_43_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0, x = var_1701_cast_fp16)[name = tensor<string, []>("denom_43_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = denom_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
            tensor<fp16, [1280]> obj_85_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_85_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270900352)))];
            tensor<fp16, [1280]> obj_85_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_85_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270902976)))];
            tensor<fp16, []> obj_85_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_85_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor<string, []>("obj_85_cast_fp16")];
            tensor<int32, [2]> var_1716 = const()[name = tensor<string, []>("op_1716"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1718 = const()[name = tensor<string, []>("op_1718"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_29_pad_type_0 = const()[name = tensor<string, []>("query_29_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_29_pad_0 = const()[name = tensor<string, []>("query_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270905600))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272134464))), name = tensor<string, []>("layers_7_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272134656)))];
            tensor<fp16, [1, 1280, 1, 1]> query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = var_1718, groups = var_1681, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = var_1716, weight = layers_7_self_attn_q_proj_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")];
            tensor<int32, [2]> var_1722 = const()[name = tensor<string, []>("op_1722"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1724 = const()[name = tensor<string, []>("op_1724"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_15_pad_type_0 = const()[name = tensor<string, []>("current_key_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_15_pad_0 = const()[name = tensor<string, []>("current_key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272137280))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272956544))), name = tensor<string, []>("layers_7_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_15_cast_fp16 = conv(dilations = var_1724, groups = var_1681, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = var_1722, weight = layers_7_self_attn_k_proj_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor<string, []>("current_key_15_cast_fp16")];
            tensor<int32, [2]> var_1729 = const()[name = tensor<string, []>("op_1729"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1731 = const()[name = tensor<string, []>("op_1731"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_15_pad_type_0 = const()[name = tensor<string, []>("current_value_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_15_pad_0 = const()[name = tensor<string, []>("current_value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272956672))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273775936))), name = tensor<string, []>("layers_7_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273776064)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = var_1731, groups = var_1681, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = var_1729, weight = layers_7_self_attn_v_proj_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor<string, []>("current_value_15_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1738_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1738_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1740_cast_fp16 = mul(x = var_103_cast_fp16_7, y = var_241_cast_fp16)[name = tensor<string, []>("op_1740_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_29_cast_fp16 = add(x = var_1738_cast_fp16, y = var_1740_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1742_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1742_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1744_cast_fp16 = mul(x = var_138_cast_fp16_7, y = var_241_cast_fp16)[name = tensor<string, []>("op_1744_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_29_cast_fp16 = add(x = var_1742_cast_fp16, y = var_1744_cast_fp16)[name = tensor<string, []>("value_29_cast_fp16")];
            tensor<int32, [4]> var_1747 = const()[name = tensor<string, []>("op_1747"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1748_cast_fp16 = reshape(shape = var_1747, x = query_29_cast_fp16)[name = tensor<string, []>("op_1748_cast_fp16")];
            tensor<fp16, []> var_1749_to_fp16 = const()[name = tensor<string, []>("op_1749_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1750_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_1749_to_fp16)[name = tensor<string, []>("op_1750_cast_fp16")];
            tensor<int32, [4]> var_1751 = const()[name = tensor<string, []>("op_1751"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1752_cast_fp16 = reshape(shape = var_1751, x = key_29_cast_fp16)[name = tensor<string, []>("op_1752_cast_fp16")];
            tensor<bool, []> mh_w_43_transpose_x_0 = const()[name = tensor<string, []>("mh_w_43_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_43_transpose_y_0 = const()[name = tensor<string, []>("mh_w_43_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_1750_cast_fp16, y = var_1752_cast_fp16)[name = tensor<string, []>("mh_w_43_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_45_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_1760_cast_fp16 = softmax(axis = var_1674, x = mh_w_45_cast_fp16)[name = tensor<string, []>("op_1760_cast_fp16")];
            tensor<int32, [4]> var_1761 = const()[name = tensor<string, []>("op_1761"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1762_cast_fp16 = reshape(shape = var_1761, x = value_29_cast_fp16)[name = tensor<string, []>("op_1762_cast_fp16")];
            tensor<bool, []> attn_29_transpose_x_0 = const()[name = tensor<string, []>("attn_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_29_transpose_y_0 = const()[name = tensor<string, []>("attn_29_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1762_cast_fp16, y = var_1760_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
            tensor<int32, [4]> var_1765 = const()[name = tensor<string, []>("op_1765"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_71_cast_fp16 = reshape(shape = var_1765, x = attn_29_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<int32, [2]> var_1769 = const()[name = tensor<string, []>("op_1769"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1771 = const()[name = tensor<string, []>("op_1771"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_91_pad_type_0 = const()[name = tensor<string, []>("obj_91_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_91_pad_0 = const()[name = tensor<string, []>("obj_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273778688))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275007552))), name = tensor<string, []>("layers_7_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275007744)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_91_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = var_1771, groups = var_1681, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = var_1769, weight = layers_7_self_attn_o_proj_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = tensor<string, []>("obj_91_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_91_cast_fp16)[name = tensor<string, []>("inputs_45_cast_fp16")];
            tensor<int32, [1]> var_1781 = const()[name = tensor<string, []>("op_1781"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_45_cast_fp16 = reduce_mean(axes = var_1781, keep_dims = var_1682, x = inputs_45_cast_fp16)[name = tensor<string, []>("channels_mean_45_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_45_cast_fp16 = sub(x = inputs_45_cast_fp16, y = channels_mean_45_cast_fp16)[name = tensor<string, []>("zero_mean_45_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = zero_mean_45_cast_fp16)[name = tensor<string, []>("zero_mean_sq_45_cast_fp16")];
            tensor<int32, [1]> var_1785 = const()[name = tensor<string, []>("op_1785"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1786_cast_fp16 = reduce_mean(axes = var_1785, keep_dims = var_1682, x = zero_mean_sq_45_cast_fp16)[name = tensor<string, []>("op_1786_cast_fp16")];
            tensor<fp16, []> var_1787_to_fp16 = const()[name = tensor<string, []>("op_1787_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1788_cast_fp16 = add(x = var_1786_cast_fp16, y = var_1787_to_fp16)[name = tensor<string, []>("op_1788_cast_fp16")];
            tensor<fp32, []> denom_45_epsilon_0 = const()[name = tensor<string, []>("denom_45_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0, x = var_1788_cast_fp16)[name = tensor<string, []>("denom_45_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = denom_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
            tensor<fp16, [1280]> obj_93_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_93_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275010368)))];
            tensor<fp16, [1280]> obj_93_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_93_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275012992)))];
            tensor<fp16, []> obj_93_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_93_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor<string, []>("obj_93_cast_fp16")];
            tensor<int32, [2]> var_1803 = const()[name = tensor<string, []>("op_1803"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1805 = const()[name = tensor<string, []>("op_1805"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_31_pad_type_0 = const()[name = tensor<string, []>("query_31_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_31_pad_0 = const()[name = tensor<string, []>("query_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275015616))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275834880))), name = tensor<string, []>("layers_7_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275835008)))];
            tensor<fp16, [1, 1280, 1, 1]> query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = var_1805, groups = var_1681, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = var_1803, weight = layers_7_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor<string, []>("query_31_cast_fp16")];
            tensor<int32, [2]> var_1809 = const()[name = tensor<string, []>("op_1809"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1811 = const()[name = tensor<string, []>("op_1811"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_31_pad_type_0 = const()[name = tensor<string, []>("key_31_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_31_pad_0 = const()[name = tensor<string, []>("key_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275837632))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276656896))), name = tensor<string, []>("layers_7_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_31_cast_fp16 = conv(dilations = var_1811, groups = var_1681, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = var_1809, weight = layers_7_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_31_cast_fp16")];
            tensor<int32, [2]> var_1816 = const()[name = tensor<string, []>("op_1816"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1818 = const()[name = tensor<string, []>("op_1818"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_31_pad_type_0 = const()[name = tensor<string, []>("value_31_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_31_pad_0 = const()[name = tensor<string, []>("value_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276657024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277885888))), name = tensor<string, []>("layers_7_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_7_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277886080)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = var_1818, groups = var_1681, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = var_1816, weight = layers_7_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_31_cast_fp16")];
            tensor<int32, [4]> var_1822 = const()[name = tensor<string, []>("op_1822"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1823_cast_fp16 = reshape(shape = var_1822, x = query_31_cast_fp16)[name = tensor<string, []>("op_1823_cast_fp16")];
            tensor<fp16, []> var_1824_to_fp16 = const()[name = tensor<string, []>("op_1824_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1825_cast_fp16 = mul(x = var_1823_cast_fp16, y = var_1824_to_fp16)[name = tensor<string, []>("op_1825_cast_fp16")];
            tensor<int32, [4]> var_1826 = const()[name = tensor<string, []>("op_1826"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1827_cast_fp16 = reshape(shape = var_1826, x = key_31_cast_fp16)[name = tensor<string, []>("op_1827_cast_fp16")];
            tensor<bool, []> mh_w_47_transpose_x_0 = const()[name = tensor<string, []>("mh_w_47_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_47_transpose_y_0 = const()[name = tensor<string, []>("mh_w_47_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_1825_cast_fp16, y = var_1827_cast_fp16)[name = tensor<string, []>("mh_w_47_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_1830_cast_fp16 = softmax(axis = var_1674, x = mh_w_47_cast_fp16)[name = tensor<string, []>("op_1830_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = tensor<string, []>("op_1831"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_1832_cast_fp16 = reshape(shape = var_1831, x = value_31_cast_fp16)[name = tensor<string, []>("op_1832_cast_fp16")];
            tensor<bool, []> attn_31_transpose_x_0 = const()[name = tensor<string, []>("attn_31_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_31_transpose_y_0 = const()[name = tensor<string, []>("attn_31_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_1832_cast_fp16, y = var_1830_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")];
            tensor<int32, [4]> var_1835 = const()[name = tensor<string, []>("op_1835"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_73_cast_fp16 = reshape(shape = var_1835, x = attn_31_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<int32, [2]> var_1839 = const()[name = tensor<string, []>("op_1839"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1841 = const()[name = tensor<string, []>("op_1841"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_95_pad_type_0 = const()[name = tensor<string, []>("obj_95_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_95_pad_0 = const()[name = tensor<string, []>("obj_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_7_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277888704))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279117568))), name = tensor<string, []>("layers_7_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279117760)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_95_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = var_1841, groups = var_1681, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = var_1839, weight = layers_7_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_73_cast_fp16)[name = tensor<string, []>("obj_95_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_95_cast_fp16)[name = tensor<string, []>("inputs_47_cast_fp16")];
            tensor<int32, [1]> var_1847 = const()[name = tensor<string, []>("op_1847"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_47_cast_fp16 = reduce_mean(axes = var_1847, keep_dims = var_1682, x = inputs_47_cast_fp16)[name = tensor<string, []>("channels_mean_47_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_47_cast_fp16 = sub(x = inputs_47_cast_fp16, y = channels_mean_47_cast_fp16)[name = tensor<string, []>("zero_mean_47_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = zero_mean_47_cast_fp16)[name = tensor<string, []>("zero_mean_sq_47_cast_fp16")];
            tensor<int32, [1]> var_1851 = const()[name = tensor<string, []>("op_1851"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1852_cast_fp16 = reduce_mean(axes = var_1851, keep_dims = var_1682, x = zero_mean_sq_47_cast_fp16)[name = tensor<string, []>("op_1852_cast_fp16")];
            tensor<fp16, []> var_1853_to_fp16 = const()[name = tensor<string, []>("op_1853_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1854_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1853_to_fp16)[name = tensor<string, []>("op_1854_cast_fp16")];
            tensor<fp32, []> denom_47_epsilon_0 = const()[name = tensor<string, []>("denom_47_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0, x = var_1854_cast_fp16)[name = tensor<string, []>("denom_47_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = denom_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")];
            tensor<fp16, [1280]> input_75_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_75_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279120384)))];
            tensor<fp16, [1280]> input_75_beta_0_to_fp16 = const()[name = tensor<string, []>("input_75_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279123008)))];
            tensor<fp16, []> input_75_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_75_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<int32, [2]> var_1865 = const()[name = tensor<string, []>("op_1865"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1867 = const()[name = tensor<string, []>("op_1867"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_77_pad_type_0 = const()[name = tensor<string, []>("input_77_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_77_pad_0 = const()[name = tensor<string, []>("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_7_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279125632))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284040896))), name = tensor<string, []>("layers_7_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_7_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284041088)))];
            tensor<fp16, [1, 5120, 1, 1]> input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = var_1867, groups = var_1681, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = var_1865, weight = layers_7_fc1_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<string, []> input_79_mode_0 = const()[name = tensor<string, []>("input_79_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<int32, [2]> var_1873 = const()[name = tensor<string, []>("op_1873"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1875 = const()[name = tensor<string, []>("op_1875"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_17_pad_type_0 = const()[name = tensor<string, []>("hidden_states_17_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = tensor<string, []>("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_7_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284051392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288966656))), name = tensor<string, []>("layers_7_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_7_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288966848)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = var_1875, groups = var_1681, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = var_1873, weight = layers_7_fc2_weight_to_fp16_palettized, x = input_79_cast_fp16)[name = tensor<string, []>("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor<string, []>("inputs_49_cast_fp16")];
            tensor<int32, []> var_1888 = const()[name = tensor<string, []>("op_1888"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_1895 = const()[name = tensor<string, []>("op_1895"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_1896 = const()[name = tensor<string, []>("op_1896"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_1908 = const()[name = tensor<string, []>("op_1908"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_49_cast_fp16 = reduce_mean(axes = var_1908, keep_dims = var_1896, x = inputs_49_cast_fp16)[name = tensor<string, []>("channels_mean_49_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_49_cast_fp16 = sub(x = inputs_49_cast_fp16, y = channels_mean_49_cast_fp16)[name = tensor<string, []>("zero_mean_49_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = zero_mean_49_cast_fp16)[name = tensor<string, []>("zero_mean_sq_49_cast_fp16")];
            tensor<int32, [1]> var_1912 = const()[name = tensor<string, []>("op_1912"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_1913_cast_fp16 = reduce_mean(axes = var_1912, keep_dims = var_1896, x = zero_mean_sq_49_cast_fp16)[name = tensor<string, []>("op_1913_cast_fp16")];
            tensor<fp16, []> var_1914_to_fp16 = const()[name = tensor<string, []>("op_1914_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_1915_cast_fp16 = add(x = var_1913_cast_fp16, y = var_1914_to_fp16)[name = tensor<string, []>("op_1915_cast_fp16")];
            tensor<fp32, []> denom_49_epsilon_0 = const()[name = tensor<string, []>("denom_49_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0, x = var_1915_cast_fp16)[name = tensor<string, []>("denom_49_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = denom_49_cast_fp16)[name = tensor<string, []>("out_49_cast_fp16")];
            tensor<fp16, [1280]> obj_97_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_97_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288969472)))];
            tensor<fp16, [1280]> obj_97_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_97_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288972096)))];
            tensor<fp16, []> obj_97_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_97_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor<string, []>("obj_97_cast_fp16")];
            tensor<int32, [2]> var_1930 = const()[name = tensor<string, []>("op_1930"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1932 = const()[name = tensor<string, []>("op_1932"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_33_pad_type_0 = const()[name = tensor<string, []>("query_33_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_33_pad_0 = const()[name = tensor<string, []>("query_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288974720))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290203584))), name = tensor<string, []>("layers_8_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290203776)))];
            tensor<fp16, [1, 1280, 1, 1]> query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = var_1932, groups = var_1895, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = var_1930, weight = layers_8_self_attn_q_proj_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")];
            tensor<int32, [2]> var_1936 = const()[name = tensor<string, []>("op_1936"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1938 = const()[name = tensor<string, []>("op_1938"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_17_pad_type_0 = const()[name = tensor<string, []>("current_key_17_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_17_pad_0 = const()[name = tensor<string, []>("current_key_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290206400))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291435264))), name = tensor<string, []>("layers_8_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_17_cast_fp16 = conv(dilations = var_1938, groups = var_1895, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = var_1936, weight = layers_8_self_attn_k_proj_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor<string, []>("current_key_17_cast_fp16")];
            tensor<int32, [2]> var_1943 = const()[name = tensor<string, []>("op_1943"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1945 = const()[name = tensor<string, []>("op_1945"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_17_pad_type_0 = const()[name = tensor<string, []>("current_value_17_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_17_pad_0 = const()[name = tensor<string, []>("current_value_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291435456))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292254720))), name = tensor<string, []>("layers_8_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292254848)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = var_1945, groups = var_1895, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = var_1943, weight = layers_8_self_attn_v_proj_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor<string, []>("current_value_17_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1952_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1952_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1954_cast_fp16 = mul(x = var_103_cast_fp16_8, y = var_241_cast_fp16)[name = tensor<string, []>("op_1954_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_33_cast_fp16 = add(x = var_1952_cast_fp16, y = var_1954_cast_fp16)[name = tensor<string, []>("key_33_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1956_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_1956_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_1958_cast_fp16 = mul(x = var_138_cast_fp16_8, y = var_241_cast_fp16)[name = tensor<string, []>("op_1958_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_33_cast_fp16 = add(x = var_1956_cast_fp16, y = var_1958_cast_fp16)[name = tensor<string, []>("value_33_cast_fp16")];
            tensor<int32, [4]> var_1961 = const()[name = tensor<string, []>("op_1961"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_1962_cast_fp16 = reshape(shape = var_1961, x = query_33_cast_fp16)[name = tensor<string, []>("op_1962_cast_fp16")];
            tensor<fp16, []> var_1963_to_fp16 = const()[name = tensor<string, []>("op_1963_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_1964_cast_fp16 = mul(x = var_1962_cast_fp16, y = var_1963_to_fp16)[name = tensor<string, []>("op_1964_cast_fp16")];
            tensor<int32, [4]> var_1965 = const()[name = tensor<string, []>("op_1965"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1966_cast_fp16 = reshape(shape = var_1965, x = key_33_cast_fp16)[name = tensor<string, []>("op_1966_cast_fp16")];
            tensor<bool, []> mh_w_49_transpose_x_0 = const()[name = tensor<string, []>("mh_w_49_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_49_transpose_y_0 = const()[name = tensor<string, []>("mh_w_49_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 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_1964_cast_fp16, y = var_1966_cast_fp16)[name = tensor<string, []>("mh_w_49_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_51_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_1974_cast_fp16 = softmax(axis = var_1888, x = mh_w_51_cast_fp16)[name = tensor<string, []>("op_1974_cast_fp16")];
            tensor<int32, [4]> var_1975 = const()[name = tensor<string, []>("op_1975"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_1976_cast_fp16 = reshape(shape = var_1975, x = value_33_cast_fp16)[name = tensor<string, []>("op_1976_cast_fp16")];
            tensor<bool, []> attn_33_transpose_x_0 = const()[name = tensor<string, []>("attn_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_33_transpose_y_0 = const()[name = tensor<string, []>("attn_33_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_1976_cast_fp16, y = var_1974_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
            tensor<int32, [4]> var_1979 = const()[name = tensor<string, []>("op_1979"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_81_cast_fp16 = reshape(shape = var_1979, x = attn_33_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<int32, [2]> var_1983 = const()[name = tensor<string, []>("op_1983"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1985 = const()[name = tensor<string, []>("op_1985"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_103_pad_type_0 = const()[name = tensor<string, []>("obj_103_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_103_pad_0 = const()[name = tensor<string, []>("obj_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292257472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293486336))), name = tensor<string, []>("layers_8_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293486528)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_103_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = var_1985, groups = var_1895, pad = obj_103_pad_0, pad_type = obj_103_pad_type_0, strides = var_1983, weight = layers_8_self_attn_o_proj_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor<string, []>("obj_103_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_103_cast_fp16)[name = tensor<string, []>("inputs_51_cast_fp16")];
            tensor<int32, [1]> var_1995 = const()[name = tensor<string, []>("op_1995"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_51_cast_fp16 = reduce_mean(axes = var_1995, keep_dims = var_1896, x = inputs_51_cast_fp16)[name = tensor<string, []>("channels_mean_51_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_51_cast_fp16 = sub(x = inputs_51_cast_fp16, y = channels_mean_51_cast_fp16)[name = tensor<string, []>("zero_mean_51_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = zero_mean_51_cast_fp16)[name = tensor<string, []>("zero_mean_sq_51_cast_fp16")];
            tensor<int32, [1]> var_1999 = const()[name = tensor<string, []>("op_1999"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2000_cast_fp16 = reduce_mean(axes = var_1999, keep_dims = var_1896, x = zero_mean_sq_51_cast_fp16)[name = tensor<string, []>("op_2000_cast_fp16")];
            tensor<fp16, []> var_2001_to_fp16 = const()[name = tensor<string, []>("op_2001_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2002_cast_fp16 = add(x = var_2000_cast_fp16, y = var_2001_to_fp16)[name = tensor<string, []>("op_2002_cast_fp16")];
            tensor<fp32, []> denom_51_epsilon_0 = const()[name = tensor<string, []>("denom_51_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0, x = var_2002_cast_fp16)[name = tensor<string, []>("denom_51_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = denom_51_cast_fp16)[name = tensor<string, []>("out_51_cast_fp16")];
            tensor<fp16, [1280]> obj_105_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_105_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293489152)))];
            tensor<fp16, [1280]> obj_105_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_105_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293491776)))];
            tensor<fp16, []> obj_105_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_105_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor<string, []>("obj_105_cast_fp16")];
            tensor<int32, [2]> var_2017 = const()[name = tensor<string, []>("op_2017"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2019 = const()[name = tensor<string, []>("op_2019"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_35_pad_type_0 = const()[name = tensor<string, []>("query_35_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_35_pad_0 = const()[name = tensor<string, []>("query_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293494400))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294723264))), name = tensor<string, []>("layers_8_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294723456)))];
            tensor<fp16, [1, 1280, 1, 1]> query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = var_2019, groups = var_1895, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = var_2017, weight = layers_8_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = tensor<string, []>("query_35_cast_fp16")];
            tensor<int32, [2]> var_2023 = const()[name = tensor<string, []>("op_2023"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2025 = const()[name = tensor<string, []>("op_2025"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_35_pad_type_0 = const()[name = tensor<string, []>("key_35_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_35_pad_0 = const()[name = tensor<string, []>("key_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294726080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295954944))), name = tensor<string, []>("layers_8_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_35_cast_fp16 = conv(dilations = var_2025, groups = var_1895, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = var_2023, weight = layers_8_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_35_cast_fp16")];
            tensor<int32, [2]> var_2030 = const()[name = tensor<string, []>("op_2030"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2032 = const()[name = tensor<string, []>("op_2032"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_35_pad_type_0 = const()[name = tensor<string, []>("value_35_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_35_pad_0 = const()[name = tensor<string, []>("value_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295955136))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297593600))), name = tensor<string, []>("layers_8_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_8_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297594176)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = var_2032, groups = var_1895, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = var_2030, weight = layers_8_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_35_cast_fp16")];
            tensor<int32, [4]> var_2036 = const()[name = tensor<string, []>("op_2036"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2037_cast_fp16 = reshape(shape = var_2036, x = query_35_cast_fp16)[name = tensor<string, []>("op_2037_cast_fp16")];
            tensor<fp16, []> var_2038_to_fp16 = const()[name = tensor<string, []>("op_2038_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2039_cast_fp16 = mul(x = var_2037_cast_fp16, y = var_2038_to_fp16)[name = tensor<string, []>("op_2039_cast_fp16")];
            tensor<int32, [4]> var_2040 = const()[name = tensor<string, []>("op_2040"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2041_cast_fp16 = reshape(shape = var_2040, x = key_35_cast_fp16)[name = tensor<string, []>("op_2041_cast_fp16")];
            tensor<bool, []> mh_w_53_transpose_x_0 = const()[name = tensor<string, []>("mh_w_53_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_53_transpose_y_0 = const()[name = tensor<string, []>("mh_w_53_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_2039_cast_fp16, y = var_2041_cast_fp16)[name = tensor<string, []>("mh_w_53_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_2044_cast_fp16 = softmax(axis = var_1888, x = mh_w_53_cast_fp16)[name = tensor<string, []>("op_2044_cast_fp16")];
            tensor<int32, [4]> var_2045 = const()[name = tensor<string, []>("op_2045"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2046_cast_fp16 = reshape(shape = var_2045, x = value_35_cast_fp16)[name = tensor<string, []>("op_2046_cast_fp16")];
            tensor<bool, []> attn_35_transpose_x_0 = const()[name = tensor<string, []>("attn_35_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_35_transpose_y_0 = const()[name = tensor<string, []>("attn_35_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2046_cast_fp16, y = var_2044_cast_fp16)[name = tensor<string, []>("attn_35_cast_fp16")];
            tensor<int32, [4]> var_2049 = const()[name = tensor<string, []>("op_2049"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_83_cast_fp16 = reshape(shape = var_2049, x = attn_35_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<int32, [2]> var_2053 = const()[name = tensor<string, []>("op_2053"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2055 = const()[name = tensor<string, []>("op_2055"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_107_pad_type_0 = const()[name = tensor<string, []>("obj_107_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_107_pad_0 = const()[name = tensor<string, []>("obj_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_8_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297596800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(298825664))), name = tensor<string, []>("layers_8_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(298825856)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_107_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = var_2055, groups = var_1895, pad = obj_107_pad_0, pad_type = obj_107_pad_type_0, strides = var_2053, weight = layers_8_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_83_cast_fp16)[name = tensor<string, []>("obj_107_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_107_cast_fp16)[name = tensor<string, []>("inputs_53_cast_fp16")];
            tensor<int32, [1]> var_2061 = const()[name = tensor<string, []>("op_2061"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_53_cast_fp16 = reduce_mean(axes = var_2061, keep_dims = var_1896, x = inputs_53_cast_fp16)[name = tensor<string, []>("channels_mean_53_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_53_cast_fp16 = sub(x = inputs_53_cast_fp16, y = channels_mean_53_cast_fp16)[name = tensor<string, []>("zero_mean_53_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = zero_mean_53_cast_fp16)[name = tensor<string, []>("zero_mean_sq_53_cast_fp16")];
            tensor<int32, [1]> var_2065 = const()[name = tensor<string, []>("op_2065"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2066_cast_fp16 = reduce_mean(axes = var_2065, keep_dims = var_1896, x = zero_mean_sq_53_cast_fp16)[name = tensor<string, []>("op_2066_cast_fp16")];
            tensor<fp16, []> var_2067_to_fp16 = const()[name = tensor<string, []>("op_2067_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2068_cast_fp16 = add(x = var_2066_cast_fp16, y = var_2067_to_fp16)[name = tensor<string, []>("op_2068_cast_fp16")];
            tensor<fp32, []> denom_53_epsilon_0 = const()[name = tensor<string, []>("denom_53_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0, x = var_2068_cast_fp16)[name = tensor<string, []>("denom_53_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = denom_53_cast_fp16)[name = tensor<string, []>("out_53_cast_fp16")];
            tensor<fp16, [1280]> input_85_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_85_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(298828480)))];
            tensor<fp16, [1280]> input_85_beta_0_to_fp16 = const()[name = tensor<string, []>("input_85_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(298831104)))];
            tensor<fp16, []> input_85_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_85_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<int32, [2]> var_2079 = const()[name = tensor<string, []>("op_2079"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2081 = const()[name = tensor<string, []>("op_2081"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_87_pad_type_0 = const()[name = tensor<string, []>("input_87_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_87_pad_0 = const()[name = tensor<string, []>("input_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_8_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(298833728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303748992))), name = tensor<string, []>("layers_8_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_8_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303749184)))];
            tensor<fp16, [1, 5120, 1, 1]> input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = var_2081, groups = var_1895, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = var_2079, weight = layers_8_fc1_weight_to_fp16_palettized, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<string, []> input_89_mode_0 = const()[name = tensor<string, []>("input_89_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<int32, [2]> var_2087 = const()[name = tensor<string, []>("op_2087"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2089 = const()[name = tensor<string, []>("op_2089"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_19_pad_type_0 = const()[name = tensor<string, []>("hidden_states_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_19_pad_0 = const()[name = tensor<string, []>("hidden_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_8_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303759488)))];
            tensor<fp16, [1280]> layers_8_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316866752)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = var_2089, groups = var_1895, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = var_2087, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("hidden_states_19_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor<string, []>("inputs_55_cast_fp16")];
            tensor<int32, []> var_2102 = const()[name = tensor<string, []>("op_2102"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_2109 = const()[name = tensor<string, []>("op_2109"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_2110 = const()[name = tensor<string, []>("op_2110"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_2122 = const()[name = tensor<string, []>("op_2122"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_55_cast_fp16 = reduce_mean(axes = var_2122, keep_dims = var_2110, x = inputs_55_cast_fp16)[name = tensor<string, []>("channels_mean_55_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_55_cast_fp16 = sub(x = inputs_55_cast_fp16, y = channels_mean_55_cast_fp16)[name = tensor<string, []>("zero_mean_55_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = zero_mean_55_cast_fp16)[name = tensor<string, []>("zero_mean_sq_55_cast_fp16")];
            tensor<int32, [1]> var_2126 = const()[name = tensor<string, []>("op_2126"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2127_cast_fp16 = reduce_mean(axes = var_2126, keep_dims = var_2110, x = zero_mean_sq_55_cast_fp16)[name = tensor<string, []>("op_2127_cast_fp16")];
            tensor<fp16, []> var_2128_to_fp16 = const()[name = tensor<string, []>("op_2128_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2129_cast_fp16 = add(x = var_2127_cast_fp16, y = var_2128_to_fp16)[name = tensor<string, []>("op_2129_cast_fp16")];
            tensor<fp32, []> denom_55_epsilon_0 = const()[name = tensor<string, []>("denom_55_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0, x = var_2129_cast_fp16)[name = tensor<string, []>("denom_55_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = denom_55_cast_fp16)[name = tensor<string, []>("out_55_cast_fp16")];
            tensor<fp16, [1280]> obj_109_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_109_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316869376)))];
            tensor<fp16, [1280]> obj_109_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_109_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316872000)))];
            tensor<fp16, []> obj_109_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_109_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor<string, []>("obj_109_cast_fp16")];
            tensor<int32, [2]> var_2144 = const()[name = tensor<string, []>("op_2144"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2146 = const()[name = tensor<string, []>("op_2146"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_37_pad_type_0 = const()[name = tensor<string, []>("query_37_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_37_pad_0 = const()[name = tensor<string, []>("query_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316874624))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318103488))), name = tensor<string, []>("layers_9_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318103680)))];
            tensor<fp16, [1, 1280, 1, 1]> query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = var_2146, groups = var_2109, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = var_2144, weight = layers_9_self_attn_q_proj_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")];
            tensor<int32, [2]> var_2150 = const()[name = tensor<string, []>("op_2150"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2152 = const()[name = tensor<string, []>("op_2152"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_19_pad_type_0 = const()[name = tensor<string, []>("current_key_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_19_pad_0 = const()[name = tensor<string, []>("current_key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318106304))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318925568))), name = tensor<string, []>("layers_9_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_19_cast_fp16 = conv(dilations = var_2152, groups = var_2109, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = var_2150, weight = layers_9_self_attn_k_proj_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor<string, []>("current_key_19_cast_fp16")];
            tensor<int32, [2]> var_2157 = const()[name = tensor<string, []>("op_2157"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2159 = const()[name = tensor<string, []>("op_2159"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_19_pad_type_0 = const()[name = tensor<string, []>("current_value_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_19_pad_0 = const()[name = tensor<string, []>("current_value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318925696))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319744960))), name = tensor<string, []>("layers_9_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319745088)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = var_2159, groups = var_2109, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = var_2157, weight = layers_9_self_attn_v_proj_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor<string, []>("current_value_19_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2166_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_2166_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2168_cast_fp16 = mul(x = var_103_cast_fp16_9, y = var_241_cast_fp16)[name = tensor<string, []>("op_2168_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_37_cast_fp16 = add(x = var_2166_cast_fp16, y = var_2168_cast_fp16)[name = tensor<string, []>("key_37_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2170_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_2170_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2172_cast_fp16 = mul(x = var_138_cast_fp16_9, y = var_241_cast_fp16)[name = tensor<string, []>("op_2172_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_37_cast_fp16 = add(x = var_2170_cast_fp16, y = var_2172_cast_fp16)[name = tensor<string, []>("value_37_cast_fp16")];
            tensor<int32, [4]> var_2175 = const()[name = tensor<string, []>("op_2175"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2176_cast_fp16 = reshape(shape = var_2175, x = query_37_cast_fp16)[name = tensor<string, []>("op_2176_cast_fp16")];
            tensor<fp16, []> var_2177_to_fp16 = const()[name = tensor<string, []>("op_2177_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2178_cast_fp16 = mul(x = var_2176_cast_fp16, y = var_2177_to_fp16)[name = tensor<string, []>("op_2178_cast_fp16")];
            tensor<int32, [4]> var_2179 = const()[name = tensor<string, []>("op_2179"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_2180_cast_fp16 = reshape(shape = var_2179, x = key_37_cast_fp16)[name = tensor<string, []>("op_2180_cast_fp16")];
            tensor<bool, []> mh_w_55_transpose_x_0 = const()[name = tensor<string, []>("mh_w_55_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_55_transpose_y_0 = const()[name = tensor<string, []>("mh_w_55_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_2178_cast_fp16, y = var_2180_cast_fp16)[name = tensor<string, []>("mh_w_55_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_57_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_2188_cast_fp16 = softmax(axis = var_2102, x = mh_w_57_cast_fp16)[name = tensor<string, []>("op_2188_cast_fp16")];
            tensor<int32, [4]> var_2189 = const()[name = tensor<string, []>("op_2189"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_2190_cast_fp16 = reshape(shape = var_2189, x = value_37_cast_fp16)[name = tensor<string, []>("op_2190_cast_fp16")];
            tensor<bool, []> attn_37_transpose_x_0 = const()[name = tensor<string, []>("attn_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_37_transpose_y_0 = const()[name = tensor<string, []>("attn_37_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2190_cast_fp16, y = var_2188_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
            tensor<int32, [4]> var_2193 = const()[name = tensor<string, []>("op_2193"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_91_cast_fp16 = reshape(shape = var_2193, x = attn_37_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<int32, [2]> var_2197 = const()[name = tensor<string, []>("op_2197"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2199 = const()[name = tensor<string, []>("op_2199"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_115_pad_type_0 = const()[name = tensor<string, []>("obj_115_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_115_pad_0 = const()[name = tensor<string, []>("obj_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319747712))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320976576))), name = tensor<string, []>("layers_9_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320976768)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_115_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = var_2199, groups = var_2109, pad = obj_115_pad_0, pad_type = obj_115_pad_type_0, strides = var_2197, weight = layers_9_self_attn_o_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = tensor<string, []>("obj_115_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_115_cast_fp16)[name = tensor<string, []>("inputs_57_cast_fp16")];
            tensor<int32, [1]> var_2209 = const()[name = tensor<string, []>("op_2209"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_57_cast_fp16 = reduce_mean(axes = var_2209, keep_dims = var_2110, x = inputs_57_cast_fp16)[name = tensor<string, []>("channels_mean_57_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_57_cast_fp16 = sub(x = inputs_57_cast_fp16, y = channels_mean_57_cast_fp16)[name = tensor<string, []>("zero_mean_57_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = zero_mean_57_cast_fp16)[name = tensor<string, []>("zero_mean_sq_57_cast_fp16")];
            tensor<int32, [1]> var_2213 = const()[name = tensor<string, []>("op_2213"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2214_cast_fp16 = reduce_mean(axes = var_2213, keep_dims = var_2110, x = zero_mean_sq_57_cast_fp16)[name = tensor<string, []>("op_2214_cast_fp16")];
            tensor<fp16, []> var_2215_to_fp16 = const()[name = tensor<string, []>("op_2215_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2216_cast_fp16 = add(x = var_2214_cast_fp16, y = var_2215_to_fp16)[name = tensor<string, []>("op_2216_cast_fp16")];
            tensor<fp32, []> denom_57_epsilon_0 = const()[name = tensor<string, []>("denom_57_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0, x = var_2216_cast_fp16)[name = tensor<string, []>("denom_57_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = denom_57_cast_fp16)[name = tensor<string, []>("out_57_cast_fp16")];
            tensor<fp16, [1280]> obj_117_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_117_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320979392)))];
            tensor<fp16, [1280]> obj_117_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_117_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320982016)))];
            tensor<fp16, []> obj_117_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_117_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor<string, []>("obj_117_cast_fp16")];
            tensor<int32, [2]> var_2231 = const()[name = tensor<string, []>("op_2231"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2233 = const()[name = tensor<string, []>("op_2233"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_39_pad_type_0 = const()[name = tensor<string, []>("query_39_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_39_pad_0 = const()[name = tensor<string, []>("query_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320984640))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321803904))), name = tensor<string, []>("layers_9_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321804032)))];
            tensor<fp16, [1, 1280, 1, 1]> query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = var_2233, groups = var_2109, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = var_2231, weight = layers_9_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = tensor<string, []>("query_39_cast_fp16")];
            tensor<int32, [2]> var_2237 = const()[name = tensor<string, []>("op_2237"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2239 = const()[name = tensor<string, []>("op_2239"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_39_pad_type_0 = const()[name = tensor<string, []>("key_39_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_39_pad_0 = const()[name = tensor<string, []>("key_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321806656))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322625920))), name = tensor<string, []>("layers_9_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_39_cast_fp16 = conv(dilations = var_2239, groups = var_2109, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = var_2237, weight = layers_9_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_39_cast_fp16")];
            tensor<int32, [2]> var_2244 = const()[name = tensor<string, []>("op_2244"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2246 = const()[name = tensor<string, []>("op_2246"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_39_pad_type_0 = const()[name = tensor<string, []>("value_39_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_39_pad_0 = const()[name = tensor<string, []>("value_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322626048))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323445312))), name = tensor<string, []>("layers_9_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_9_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323445440)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = var_2246, groups = var_2109, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = var_2244, weight = layers_9_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_39_cast_fp16")];
            tensor<int32, [4]> var_2250 = const()[name = tensor<string, []>("op_2250"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2251_cast_fp16 = reshape(shape = var_2250, x = query_39_cast_fp16)[name = tensor<string, []>("op_2251_cast_fp16")];
            tensor<fp16, []> var_2252_to_fp16 = const()[name = tensor<string, []>("op_2252_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2253_cast_fp16 = mul(x = var_2251_cast_fp16, y = var_2252_to_fp16)[name = tensor<string, []>("op_2253_cast_fp16")];
            tensor<int32, [4]> var_2254 = const()[name = tensor<string, []>("op_2254"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2255_cast_fp16 = reshape(shape = var_2254, x = key_39_cast_fp16)[name = tensor<string, []>("op_2255_cast_fp16")];
            tensor<bool, []> mh_w_59_transpose_x_0 = const()[name = tensor<string, []>("mh_w_59_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_59_transpose_y_0 = const()[name = tensor<string, []>("mh_w_59_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_2253_cast_fp16, y = var_2255_cast_fp16)[name = tensor<string, []>("mh_w_59_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_2258_cast_fp16 = softmax(axis = var_2102, x = mh_w_59_cast_fp16)[name = tensor<string, []>("op_2258_cast_fp16")];
            tensor<int32, [4]> var_2259 = const()[name = tensor<string, []>("op_2259"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2260_cast_fp16 = reshape(shape = var_2259, x = value_39_cast_fp16)[name = tensor<string, []>("op_2260_cast_fp16")];
            tensor<bool, []> attn_39_transpose_x_0 = const()[name = tensor<string, []>("attn_39_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_39_transpose_y_0 = const()[name = tensor<string, []>("attn_39_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2260_cast_fp16, y = var_2258_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")];
            tensor<int32, [4]> var_2263 = const()[name = tensor<string, []>("op_2263"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_93_cast_fp16 = reshape(shape = var_2263, x = attn_39_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<int32, [2]> var_2267 = const()[name = tensor<string, []>("op_2267"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2269 = const()[name = tensor<string, []>("op_2269"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_119_pad_type_0 = const()[name = tensor<string, []>("obj_119_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_119_pad_0 = const()[name = tensor<string, []>("obj_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_9_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323448064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324676928))), name = tensor<string, []>("layers_9_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324677120)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_119_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = var_2269, groups = var_2109, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = var_2267, weight = layers_9_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_93_cast_fp16)[name = tensor<string, []>("obj_119_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_119_cast_fp16)[name = tensor<string, []>("inputs_59_cast_fp16")];
            tensor<int32, [1]> var_2275 = const()[name = tensor<string, []>("op_2275"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_59_cast_fp16 = reduce_mean(axes = var_2275, keep_dims = var_2110, x = inputs_59_cast_fp16)[name = tensor<string, []>("channels_mean_59_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_59_cast_fp16 = sub(x = inputs_59_cast_fp16, y = channels_mean_59_cast_fp16)[name = tensor<string, []>("zero_mean_59_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = zero_mean_59_cast_fp16)[name = tensor<string, []>("zero_mean_sq_59_cast_fp16")];
            tensor<int32, [1]> var_2279 = const()[name = tensor<string, []>("op_2279"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2280_cast_fp16 = reduce_mean(axes = var_2279, keep_dims = var_2110, x = zero_mean_sq_59_cast_fp16)[name = tensor<string, []>("op_2280_cast_fp16")];
            tensor<fp16, []> var_2281_to_fp16 = const()[name = tensor<string, []>("op_2281_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2282_cast_fp16 = add(x = var_2280_cast_fp16, y = var_2281_to_fp16)[name = tensor<string, []>("op_2282_cast_fp16")];
            tensor<fp32, []> denom_59_epsilon_0 = const()[name = tensor<string, []>("denom_59_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0, x = var_2282_cast_fp16)[name = tensor<string, []>("denom_59_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = denom_59_cast_fp16)[name = tensor<string, []>("out_59_cast_fp16")];
            tensor<fp16, [1280]> input_95_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_95_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324679744)))];
            tensor<fp16, [1280]> input_95_beta_0_to_fp16 = const()[name = tensor<string, []>("input_95_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324682368)))];
            tensor<fp16, []> input_95_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_95_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<int32, [2]> var_2293 = const()[name = tensor<string, []>("op_2293"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2295 = const()[name = tensor<string, []>("op_2295"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_97_pad_type_0 = const()[name = tensor<string, []>("input_97_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_97_pad_0 = const()[name = tensor<string, []>("input_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_9_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324684992))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329600256))), name = tensor<string, []>("layers_9_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_9_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329600448)))];
            tensor<fp16, [1, 5120, 1, 1]> input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = var_2295, groups = var_2109, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = var_2293, weight = layers_9_fc1_weight_to_fp16_palettized, x = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
            tensor<string, []> input_99_mode_0 = const()[name = tensor<string, []>("input_99_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<int32, [2]> var_2301 = const()[name = tensor<string, []>("op_2301"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2303 = const()[name = tensor<string, []>("op_2303"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_21_pad_type_0 = const()[name = tensor<string, []>("hidden_states_21_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_21_pad_0 = const()[name = tensor<string, []>("hidden_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_9_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329610752))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336164416))), name = tensor<string, []>("layers_9_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_9_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336164992)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = var_2303, groups = var_2109, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = var_2301, weight = layers_9_fc2_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = tensor<string, []>("hidden_states_21_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor<string, []>("inputs_61_cast_fp16")];
            tensor<int32, []> var_2316 = const()[name = tensor<string, []>("op_2316"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_2323 = const()[name = tensor<string, []>("op_2323"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_2324 = const()[name = tensor<string, []>("op_2324"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_2336 = const()[name = tensor<string, []>("op_2336"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_61_cast_fp16 = reduce_mean(axes = var_2336, keep_dims = var_2324, x = inputs_61_cast_fp16)[name = tensor<string, []>("channels_mean_61_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_61_cast_fp16 = sub(x = inputs_61_cast_fp16, y = channels_mean_61_cast_fp16)[name = tensor<string, []>("zero_mean_61_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = zero_mean_61_cast_fp16)[name = tensor<string, []>("zero_mean_sq_61_cast_fp16")];
            tensor<int32, [1]> var_2340 = const()[name = tensor<string, []>("op_2340"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2341_cast_fp16 = reduce_mean(axes = var_2340, keep_dims = var_2324, x = zero_mean_sq_61_cast_fp16)[name = tensor<string, []>("op_2341_cast_fp16")];
            tensor<fp16, []> var_2342_to_fp16 = const()[name = tensor<string, []>("op_2342_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2343_cast_fp16 = add(x = var_2341_cast_fp16, y = var_2342_to_fp16)[name = tensor<string, []>("op_2343_cast_fp16")];
            tensor<fp32, []> denom_61_epsilon_0 = const()[name = tensor<string, []>("denom_61_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0, x = var_2343_cast_fp16)[name = tensor<string, []>("denom_61_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = denom_61_cast_fp16)[name = tensor<string, []>("out_61_cast_fp16")];
            tensor<fp16, [1280]> obj_121_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_121_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336167616)))];
            tensor<fp16, [1280]> obj_121_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_121_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336170240)))];
            tensor<fp16, []> obj_121_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_121_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor<string, []>("obj_121_cast_fp16")];
            tensor<int32, [2]> var_2358 = const()[name = tensor<string, []>("op_2358"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2360 = const()[name = tensor<string, []>("op_2360"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_41_pad_type_0 = const()[name = tensor<string, []>("query_41_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_41_pad_0 = const()[name = tensor<string, []>("query_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336172864))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337401728))), name = tensor<string, []>("layers_10_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337401920)))];
            tensor<fp16, [1, 1280, 1, 1]> query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = var_2360, groups = var_2323, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = var_2358, weight = layers_10_self_attn_q_proj_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")];
            tensor<int32, [2]> var_2364 = const()[name = tensor<string, []>("op_2364"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2366 = const()[name = tensor<string, []>("op_2366"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_21_pad_type_0 = const()[name = tensor<string, []>("current_key_21_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_21_pad_0 = const()[name = tensor<string, []>("current_key_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337404544))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339043008))), name = tensor<string, []>("layers_10_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_21_cast_fp16 = conv(dilations = var_2366, groups = var_2323, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = var_2364, weight = layers_10_self_attn_k_proj_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor<string, []>("current_key_21_cast_fp16")];
            tensor<int32, [2]> var_2371 = const()[name = tensor<string, []>("op_2371"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2373 = const()[name = tensor<string, []>("op_2373"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_21_pad_type_0 = const()[name = tensor<string, []>("current_value_21_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_21_pad_0 = const()[name = tensor<string, []>("current_value_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339043584))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340272448))), name = tensor<string, []>("layers_10_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340272640)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = var_2373, groups = var_2323, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = var_2371, weight = layers_10_self_attn_v_proj_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor<string, []>("current_value_21_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2380_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_2380_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2382_cast_fp16 = mul(x = var_103_cast_fp16_10, y = var_241_cast_fp16)[name = tensor<string, []>("op_2382_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_41_cast_fp16 = add(x = var_2380_cast_fp16, y = var_2382_cast_fp16)[name = tensor<string, []>("key_41_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2384_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_2384_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2386_cast_fp16 = mul(x = var_138_cast_fp16_10, y = var_241_cast_fp16)[name = tensor<string, []>("op_2386_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_41_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2386_cast_fp16)[name = tensor<string, []>("value_41_cast_fp16")];
            tensor<int32, [4]> var_2389 = const()[name = tensor<string, []>("op_2389"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2390_cast_fp16 = reshape(shape = var_2389, x = query_41_cast_fp16)[name = tensor<string, []>("op_2390_cast_fp16")];
            tensor<fp16, []> var_2391_to_fp16 = const()[name = tensor<string, []>("op_2391_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2392_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_2391_to_fp16)[name = tensor<string, []>("op_2392_cast_fp16")];
            tensor<int32, [4]> var_2393 = const()[name = tensor<string, []>("op_2393"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_2394_cast_fp16 = reshape(shape = var_2393, x = key_41_cast_fp16)[name = tensor<string, []>("op_2394_cast_fp16")];
            tensor<bool, []> mh_w_61_transpose_x_0 = const()[name = tensor<string, []>("mh_w_61_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_61_transpose_y_0 = const()[name = tensor<string, []>("mh_w_61_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_2392_cast_fp16, y = var_2394_cast_fp16)[name = tensor<string, []>("mh_w_61_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_63_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_2402_cast_fp16 = softmax(axis = var_2316, x = mh_w_63_cast_fp16)[name = tensor<string, []>("op_2402_cast_fp16")];
            tensor<int32, [4]> var_2403 = const()[name = tensor<string, []>("op_2403"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_2404_cast_fp16 = reshape(shape = var_2403, x = value_41_cast_fp16)[name = tensor<string, []>("op_2404_cast_fp16")];
            tensor<bool, []> attn_41_transpose_x_0 = const()[name = tensor<string, []>("attn_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_41_transpose_y_0 = const()[name = tensor<string, []>("attn_41_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2404_cast_fp16, y = var_2402_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
            tensor<int32, [4]> var_2407 = const()[name = tensor<string, []>("op_2407"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_101_cast_fp16 = reshape(shape = var_2407, x = attn_41_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<int32, [2]> var_2411 = const()[name = tensor<string, []>("op_2411"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2413 = const()[name = tensor<string, []>("op_2413"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_127_pad_type_0 = const()[name = tensor<string, []>("obj_127_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_127_pad_0 = const()[name = tensor<string, []>("obj_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340275264))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341504128))), name = tensor<string, []>("layers_10_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341504320)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_127_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = var_2413, groups = var_2323, pad = obj_127_pad_0, pad_type = obj_127_pad_type_0, strides = var_2411, weight = layers_10_self_attn_o_proj_weight_to_fp16_palettized, x = input_101_cast_fp16)[name = tensor<string, []>("obj_127_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_127_cast_fp16)[name = tensor<string, []>("inputs_63_cast_fp16")];
            tensor<int32, [1]> var_2423 = const()[name = tensor<string, []>("op_2423"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_63_cast_fp16 = reduce_mean(axes = var_2423, keep_dims = var_2324, x = inputs_63_cast_fp16)[name = tensor<string, []>("channels_mean_63_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_63_cast_fp16 = sub(x = inputs_63_cast_fp16, y = channels_mean_63_cast_fp16)[name = tensor<string, []>("zero_mean_63_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = zero_mean_63_cast_fp16)[name = tensor<string, []>("zero_mean_sq_63_cast_fp16")];
            tensor<int32, [1]> var_2427 = const()[name = tensor<string, []>("op_2427"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2428_cast_fp16 = reduce_mean(axes = var_2427, keep_dims = var_2324, x = zero_mean_sq_63_cast_fp16)[name = tensor<string, []>("op_2428_cast_fp16")];
            tensor<fp16, []> var_2429_to_fp16 = const()[name = tensor<string, []>("op_2429_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2430_cast_fp16 = add(x = var_2428_cast_fp16, y = var_2429_to_fp16)[name = tensor<string, []>("op_2430_cast_fp16")];
            tensor<fp32, []> denom_63_epsilon_0 = const()[name = tensor<string, []>("denom_63_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0, x = var_2430_cast_fp16)[name = tensor<string, []>("denom_63_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = denom_63_cast_fp16)[name = tensor<string, []>("out_63_cast_fp16")];
            tensor<fp16, [1280]> obj_129_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_129_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341506944)))];
            tensor<fp16, [1280]> obj_129_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_129_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341509568)))];
            tensor<fp16, []> obj_129_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_129_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_129_cast_fp16 = batch_norm(beta = obj_129_beta_0_to_fp16, epsilon = obj_129_epsilon_0_to_fp16, gamma = obj_129_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor<string, []>("obj_129_cast_fp16")];
            tensor<int32, [2]> var_2445 = const()[name = tensor<string, []>("op_2445"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2447 = const()[name = tensor<string, []>("op_2447"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_43_pad_type_0 = const()[name = tensor<string, []>("query_43_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_43_pad_0 = const()[name = tensor<string, []>("query_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341512192))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342331456))), name = tensor<string, []>("layers_10_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342331584)))];
            tensor<fp16, [1, 1280, 1, 1]> query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = var_2447, groups = var_2323, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = var_2445, weight = layers_10_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_129_cast_fp16)[name = tensor<string, []>("query_43_cast_fp16")];
            tensor<int32, [2]> var_2451 = const()[name = tensor<string, []>("op_2451"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2453 = const()[name = tensor<string, []>("op_2453"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_43_pad_type_0 = const()[name = tensor<string, []>("key_43_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_43_pad_0 = const()[name = tensor<string, []>("key_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342334208))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343153472))), name = tensor<string, []>("layers_10_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_43_cast_fp16 = conv(dilations = var_2453, groups = var_2323, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = var_2451, weight = layers_10_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_43_cast_fp16")];
            tensor<int32, [2]> var_2458 = const()[name = tensor<string, []>("op_2458"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2460 = const()[name = tensor<string, []>("op_2460"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_43_pad_type_0 = const()[name = tensor<string, []>("value_43_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_43_pad_0 = const()[name = tensor<string, []>("value_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343153600))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343972864))), name = tensor<string, []>("layers_10_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_10_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343972992)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = var_2460, groups = var_2323, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = var_2458, weight = layers_10_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_43_cast_fp16")];
            tensor<int32, [4]> var_2464 = const()[name = tensor<string, []>("op_2464"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2465_cast_fp16 = reshape(shape = var_2464, x = query_43_cast_fp16)[name = tensor<string, []>("op_2465_cast_fp16")];
            tensor<fp16, []> var_2466_to_fp16 = const()[name = tensor<string, []>("op_2466_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2467_cast_fp16 = mul(x = var_2465_cast_fp16, y = var_2466_to_fp16)[name = tensor<string, []>("op_2467_cast_fp16")];
            tensor<int32, [4]> var_2468 = const()[name = tensor<string, []>("op_2468"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2469_cast_fp16 = reshape(shape = var_2468, x = key_43_cast_fp16)[name = tensor<string, []>("op_2469_cast_fp16")];
            tensor<bool, []> mh_w_65_transpose_x_0 = const()[name = tensor<string, []>("mh_w_65_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_65_transpose_y_0 = const()[name = tensor<string, []>("mh_w_65_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_2467_cast_fp16, y = var_2469_cast_fp16)[name = tensor<string, []>("mh_w_65_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_2472_cast_fp16 = softmax(axis = var_2316, x = mh_w_65_cast_fp16)[name = tensor<string, []>("op_2472_cast_fp16")];
            tensor<int32, [4]> var_2473 = const()[name = tensor<string, []>("op_2473"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2474_cast_fp16 = reshape(shape = var_2473, x = value_43_cast_fp16)[name = tensor<string, []>("op_2474_cast_fp16")];
            tensor<bool, []> attn_43_transpose_x_0 = const()[name = tensor<string, []>("attn_43_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_43_transpose_y_0 = const()[name = tensor<string, []>("attn_43_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2474_cast_fp16, y = var_2472_cast_fp16)[name = tensor<string, []>("attn_43_cast_fp16")];
            tensor<int32, [4]> var_2477 = const()[name = tensor<string, []>("op_2477"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_103_cast_fp16 = reshape(shape = var_2477, x = attn_43_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<int32, [2]> var_2481 = const()[name = tensor<string, []>("op_2481"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2483 = const()[name = tensor<string, []>("op_2483"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_131_pad_type_0 = const()[name = tensor<string, []>("obj_131_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_131_pad_0 = const()[name = tensor<string, []>("obj_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_10_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343975616))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345204480))), name = tensor<string, []>("layers_10_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345204672)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_131_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = var_2483, groups = var_2323, pad = obj_131_pad_0, pad_type = obj_131_pad_type_0, strides = var_2481, weight = layers_10_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_103_cast_fp16)[name = tensor<string, []>("obj_131_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_131_cast_fp16)[name = tensor<string, []>("inputs_65_cast_fp16")];
            tensor<int32, [1]> var_2489 = const()[name = tensor<string, []>("op_2489"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_65_cast_fp16 = reduce_mean(axes = var_2489, keep_dims = var_2324, x = inputs_65_cast_fp16)[name = tensor<string, []>("channels_mean_65_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_65_cast_fp16 = sub(x = inputs_65_cast_fp16, y = channels_mean_65_cast_fp16)[name = tensor<string, []>("zero_mean_65_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = zero_mean_65_cast_fp16)[name = tensor<string, []>("zero_mean_sq_65_cast_fp16")];
            tensor<int32, [1]> var_2493 = const()[name = tensor<string, []>("op_2493"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2494_cast_fp16 = reduce_mean(axes = var_2493, keep_dims = var_2324, x = zero_mean_sq_65_cast_fp16)[name = tensor<string, []>("op_2494_cast_fp16")];
            tensor<fp16, []> var_2495_to_fp16 = const()[name = tensor<string, []>("op_2495_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2496_cast_fp16 = add(x = var_2494_cast_fp16, y = var_2495_to_fp16)[name = tensor<string, []>("op_2496_cast_fp16")];
            tensor<fp32, []> denom_65_epsilon_0 = const()[name = tensor<string, []>("denom_65_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0, x = var_2496_cast_fp16)[name = tensor<string, []>("denom_65_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = denom_65_cast_fp16)[name = tensor<string, []>("out_65_cast_fp16")];
            tensor<fp16, [1280]> input_105_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_105_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345207296)))];
            tensor<fp16, [1280]> input_105_beta_0_to_fp16 = const()[name = tensor<string, []>("input_105_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345209920)))];
            tensor<fp16, []> input_105_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_105_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<int32, [2]> var_2507 = const()[name = tensor<string, []>("op_2507"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2509 = const()[name = tensor<string, []>("op_2509"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_107_pad_type_0 = const()[name = tensor<string, []>("input_107_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_107_pad_0 = const()[name = tensor<string, []>("input_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_10_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345212544))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350127808))), name = tensor<string, []>("layers_10_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_10_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350128000)))];
            tensor<fp16, [1, 5120, 1, 1]> input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = var_2509, groups = var_2323, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = var_2507, weight = layers_10_fc1_weight_to_fp16_palettized, x = input_105_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<string, []> input_109_mode_0 = const()[name = tensor<string, []>("input_109_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<int32, [2]> var_2515 = const()[name = tensor<string, []>("op_2515"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2517 = const()[name = tensor<string, []>("op_2517"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_23_pad_type_0 = const()[name = tensor<string, []>("hidden_states_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = tensor<string, []>("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_10_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350138304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355053568))), name = tensor<string, []>("layers_10_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_10_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355053760)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = var_2517, groups = var_2323, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = var_2515, weight = layers_10_fc2_weight_to_fp16_palettized, x = input_109_cast_fp16)[name = tensor<string, []>("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor<string, []>("inputs_67_cast_fp16")];
            tensor<int32, []> var_2530 = const()[name = tensor<string, []>("op_2530"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_2537 = const()[name = tensor<string, []>("op_2537"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_2538 = const()[name = tensor<string, []>("op_2538"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_2550 = const()[name = tensor<string, []>("op_2550"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_67_cast_fp16 = reduce_mean(axes = var_2550, keep_dims = var_2538, x = inputs_67_cast_fp16)[name = tensor<string, []>("channels_mean_67_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_67_cast_fp16 = sub(x = inputs_67_cast_fp16, y = channels_mean_67_cast_fp16)[name = tensor<string, []>("zero_mean_67_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = zero_mean_67_cast_fp16)[name = tensor<string, []>("zero_mean_sq_67_cast_fp16")];
            tensor<int32, [1]> var_2554 = const()[name = tensor<string, []>("op_2554"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2555_cast_fp16 = reduce_mean(axes = var_2554, keep_dims = var_2538, x = zero_mean_sq_67_cast_fp16)[name = tensor<string, []>("op_2555_cast_fp16")];
            tensor<fp16, []> var_2556_to_fp16 = const()[name = tensor<string, []>("op_2556_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2557_cast_fp16 = add(x = var_2555_cast_fp16, y = var_2556_to_fp16)[name = tensor<string, []>("op_2557_cast_fp16")];
            tensor<fp32, []> denom_67_epsilon_0 = const()[name = tensor<string, []>("denom_67_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0, x = var_2557_cast_fp16)[name = tensor<string, []>("denom_67_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = denom_67_cast_fp16)[name = tensor<string, []>("out_67_cast_fp16")];
            tensor<fp16, [1280]> obj_133_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_133_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355056384)))];
            tensor<fp16, [1280]> obj_133_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_133_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355059008)))];
            tensor<fp16, []> obj_133_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_133_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor<string, []>("obj_133_cast_fp16")];
            tensor<int32, [2]> var_2572 = const()[name = tensor<string, []>("op_2572"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2574 = const()[name = tensor<string, []>("op_2574"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_45_pad_type_0 = const()[name = tensor<string, []>("query_45_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_45_pad_0 = const()[name = tensor<string, []>("query_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355061632))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356290496))), name = tensor<string, []>("layers_11_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356290688)))];
            tensor<fp16, [1, 1280, 1, 1]> query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = var_2574, groups = var_2537, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = var_2572, weight = layers_11_self_attn_q_proj_weight_to_fp16_palettized, x = obj_133_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")];
            tensor<int32, [2]> var_2578 = const()[name = tensor<string, []>("op_2578"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2580 = const()[name = tensor<string, []>("op_2580"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_23_pad_type_0 = const()[name = tensor<string, []>("current_key_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_23_pad_0 = const()[name = tensor<string, []>("current_key_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356293312))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357522176))), name = tensor<string, []>("layers_11_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_23_cast_fp16 = conv(dilations = var_2580, groups = var_2537, pad = current_key_23_pad_0, pad_type = current_key_23_pad_type_0, strides = var_2578, weight = layers_11_self_attn_k_proj_weight_to_fp16_palettized, x = obj_133_cast_fp16)[name = tensor<string, []>("current_key_23_cast_fp16")];
            tensor<int32, [2]> var_2585 = const()[name = tensor<string, []>("op_2585"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2587 = const()[name = tensor<string, []>("op_2587"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_23_pad_type_0 = const()[name = tensor<string, []>("current_value_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_23_pad_0 = const()[name = tensor<string, []>("current_value_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357522368))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358341632))), name = tensor<string, []>("layers_11_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358341760)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = var_2587, groups = var_2537, pad = current_value_23_pad_0, pad_type = current_value_23_pad_type_0, strides = var_2585, weight = layers_11_self_attn_v_proj_weight_to_fp16_palettized, x = obj_133_cast_fp16)[name = tensor<string, []>("current_value_23_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2594_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_2594_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2596_cast_fp16 = mul(x = var_103_cast_fp16_11, y = var_241_cast_fp16)[name = tensor<string, []>("op_2596_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_45_cast_fp16 = add(x = var_2594_cast_fp16, y = var_2596_cast_fp16)[name = tensor<string, []>("key_45_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2598_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_2598_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2600_cast_fp16 = mul(x = var_138_cast_fp16_11, y = var_241_cast_fp16)[name = tensor<string, []>("op_2600_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_45_cast_fp16 = add(x = var_2598_cast_fp16, y = var_2600_cast_fp16)[name = tensor<string, []>("value_45_cast_fp16")];
            tensor<int32, [4]> var_2603 = const()[name = tensor<string, []>("op_2603"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2604_cast_fp16 = reshape(shape = var_2603, x = query_45_cast_fp16)[name = tensor<string, []>("op_2604_cast_fp16")];
            tensor<fp16, []> var_2605_to_fp16 = const()[name = tensor<string, []>("op_2605_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2606_cast_fp16 = mul(x = var_2604_cast_fp16, y = var_2605_to_fp16)[name = tensor<string, []>("op_2606_cast_fp16")];
            tensor<int32, [4]> var_2607 = const()[name = tensor<string, []>("op_2607"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_2608_cast_fp16 = reshape(shape = var_2607, x = key_45_cast_fp16)[name = tensor<string, []>("op_2608_cast_fp16")];
            tensor<bool, []> mh_w_67_transpose_x_0 = const()[name = tensor<string, []>("mh_w_67_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_67_transpose_y_0 = const()[name = tensor<string, []>("mh_w_67_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_2606_cast_fp16, y = var_2608_cast_fp16)[name = tensor<string, []>("mh_w_67_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_69_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_2616_cast_fp16 = softmax(axis = var_2530, x = mh_w_69_cast_fp16)[name = tensor<string, []>("op_2616_cast_fp16")];
            tensor<int32, [4]> var_2617 = const()[name = tensor<string, []>("op_2617"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_2618_cast_fp16 = reshape(shape = var_2617, x = value_45_cast_fp16)[name = tensor<string, []>("op_2618_cast_fp16")];
            tensor<bool, []> attn_45_transpose_x_0 = const()[name = tensor<string, []>("attn_45_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_45_transpose_y_0 = const()[name = tensor<string, []>("attn_45_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2618_cast_fp16, y = var_2616_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
            tensor<int32, [4]> var_2621 = const()[name = tensor<string, []>("op_2621"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_111_cast_fp16 = reshape(shape = var_2621, x = attn_45_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<int32, [2]> var_2625 = const()[name = tensor<string, []>("op_2625"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2627 = const()[name = tensor<string, []>("op_2627"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_139_pad_type_0 = const()[name = tensor<string, []>("obj_139_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_139_pad_0 = const()[name = tensor<string, []>("obj_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358344384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359573248))), name = tensor<string, []>("layers_11_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359573440)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_139_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = var_2627, groups = var_2537, pad = obj_139_pad_0, pad_type = obj_139_pad_type_0, strides = var_2625, weight = layers_11_self_attn_o_proj_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor<string, []>("obj_139_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_139_cast_fp16)[name = tensor<string, []>("inputs_69_cast_fp16")];
            tensor<int32, [1]> var_2637 = const()[name = tensor<string, []>("op_2637"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_69_cast_fp16 = reduce_mean(axes = var_2637, keep_dims = var_2538, x = inputs_69_cast_fp16)[name = tensor<string, []>("channels_mean_69_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_69_cast_fp16 = sub(x = inputs_69_cast_fp16, y = channels_mean_69_cast_fp16)[name = tensor<string, []>("zero_mean_69_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = zero_mean_69_cast_fp16)[name = tensor<string, []>("zero_mean_sq_69_cast_fp16")];
            tensor<int32, [1]> var_2641 = const()[name = tensor<string, []>("op_2641"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2642_cast_fp16 = reduce_mean(axes = var_2641, keep_dims = var_2538, x = zero_mean_sq_69_cast_fp16)[name = tensor<string, []>("op_2642_cast_fp16")];
            tensor<fp16, []> var_2643_to_fp16 = const()[name = tensor<string, []>("op_2643_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2644_cast_fp16 = add(x = var_2642_cast_fp16, y = var_2643_to_fp16)[name = tensor<string, []>("op_2644_cast_fp16")];
            tensor<fp32, []> denom_69_epsilon_0 = const()[name = tensor<string, []>("denom_69_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0, x = var_2644_cast_fp16)[name = tensor<string, []>("denom_69_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = denom_69_cast_fp16)[name = tensor<string, []>("out_69_cast_fp16")];
            tensor<fp16, [1280]> obj_141_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_141_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359576064)))];
            tensor<fp16, [1280]> obj_141_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_141_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359578688)))];
            tensor<fp16, []> obj_141_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_141_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor<string, []>("obj_141_cast_fp16")];
            tensor<int32, [2]> var_2659 = const()[name = tensor<string, []>("op_2659"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2661 = const()[name = tensor<string, []>("op_2661"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_47_pad_type_0 = const()[name = tensor<string, []>("query_47_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_47_pad_0 = const()[name = tensor<string, []>("query_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359581312))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360400576))), name = tensor<string, []>("layers_11_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360400704)))];
            tensor<fp16, [1, 1280, 1, 1]> query_47_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = var_2661, groups = var_2537, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = var_2659, weight = layers_11_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_141_cast_fp16)[name = tensor<string, []>("query_47_cast_fp16")];
            tensor<int32, [2]> var_2665 = const()[name = tensor<string, []>("op_2665"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2667 = const()[name = tensor<string, []>("op_2667"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_47_pad_type_0 = const()[name = tensor<string, []>("key_47_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_47_pad_0 = const()[name = tensor<string, []>("key_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360403328))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361222592))), name = tensor<string, []>("layers_11_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_47_cast_fp16 = conv(dilations = var_2667, groups = var_2537, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = var_2665, weight = layers_11_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_47_cast_fp16")];
            tensor<int32, [2]> var_2672 = const()[name = tensor<string, []>("op_2672"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2674 = const()[name = tensor<string, []>("op_2674"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_47_pad_type_0 = const()[name = tensor<string, []>("value_47_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_47_pad_0 = const()[name = tensor<string, []>("value_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361222720))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362041984))), name = tensor<string, []>("layers_11_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_11_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362042112)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_47_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = var_2674, groups = var_2537, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = var_2672, weight = layers_11_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_47_cast_fp16")];
            tensor<int32, [4]> var_2678 = const()[name = tensor<string, []>("op_2678"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2679_cast_fp16 = reshape(shape = var_2678, x = query_47_cast_fp16)[name = tensor<string, []>("op_2679_cast_fp16")];
            tensor<fp16, []> var_2680_to_fp16 = const()[name = tensor<string, []>("op_2680_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2681_cast_fp16 = mul(x = var_2679_cast_fp16, y = var_2680_to_fp16)[name = tensor<string, []>("op_2681_cast_fp16")];
            tensor<int32, [4]> var_2682 = const()[name = tensor<string, []>("op_2682"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2683_cast_fp16 = reshape(shape = var_2682, x = key_47_cast_fp16)[name = tensor<string, []>("op_2683_cast_fp16")];
            tensor<bool, []> mh_w_71_transpose_x_0 = const()[name = tensor<string, []>("mh_w_71_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_71_transpose_y_0 = const()[name = tensor<string, []>("mh_w_71_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_71_cast_fp16 = matmul(transpose_x = mh_w_71_transpose_x_0, transpose_y = mh_w_71_transpose_y_0, x = var_2681_cast_fp16, y = var_2683_cast_fp16)[name = tensor<string, []>("mh_w_71_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_2686_cast_fp16 = softmax(axis = var_2530, x = mh_w_71_cast_fp16)[name = tensor<string, []>("op_2686_cast_fp16")];
            tensor<int32, [4]> var_2687 = const()[name = tensor<string, []>("op_2687"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2688_cast_fp16 = reshape(shape = var_2687, x = value_47_cast_fp16)[name = tensor<string, []>("op_2688_cast_fp16")];
            tensor<bool, []> attn_47_transpose_x_0 = const()[name = tensor<string, []>("attn_47_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_47_transpose_y_0 = const()[name = tensor<string, []>("attn_47_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2688_cast_fp16, y = var_2686_cast_fp16)[name = tensor<string, []>("attn_47_cast_fp16")];
            tensor<int32, [4]> var_2691 = const()[name = tensor<string, []>("op_2691"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_113_cast_fp16 = reshape(shape = var_2691, x = attn_47_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<int32, [2]> var_2695 = const()[name = tensor<string, []>("op_2695"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2697 = const()[name = tensor<string, []>("op_2697"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_143_pad_type_0 = const()[name = tensor<string, []>("obj_143_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_143_pad_0 = const()[name = tensor<string, []>("obj_143_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_11_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362044736))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362864000))), name = tensor<string, []>("layers_11_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362864128)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_143_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = var_2697, groups = var_2537, pad = obj_143_pad_0, pad_type = obj_143_pad_type_0, strides = var_2695, weight = layers_11_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = tensor<string, []>("obj_143_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_143_cast_fp16)[name = tensor<string, []>("inputs_71_cast_fp16")];
            tensor<int32, [1]> var_2703 = const()[name = tensor<string, []>("op_2703"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_71_cast_fp16 = reduce_mean(axes = var_2703, keep_dims = var_2538, x = inputs_71_cast_fp16)[name = tensor<string, []>("channels_mean_71_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_71_cast_fp16 = sub(x = inputs_71_cast_fp16, y = channels_mean_71_cast_fp16)[name = tensor<string, []>("zero_mean_71_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = zero_mean_71_cast_fp16)[name = tensor<string, []>("zero_mean_sq_71_cast_fp16")];
            tensor<int32, [1]> var_2707 = const()[name = tensor<string, []>("op_2707"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2708_cast_fp16 = reduce_mean(axes = var_2707, keep_dims = var_2538, x = zero_mean_sq_71_cast_fp16)[name = tensor<string, []>("op_2708_cast_fp16")];
            tensor<fp16, []> var_2709_to_fp16 = const()[name = tensor<string, []>("op_2709_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2710_cast_fp16 = add(x = var_2708_cast_fp16, y = var_2709_to_fp16)[name = tensor<string, []>("op_2710_cast_fp16")];
            tensor<fp32, []> denom_71_epsilon_0 = const()[name = tensor<string, []>("denom_71_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0, x = var_2710_cast_fp16)[name = tensor<string, []>("denom_71_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = denom_71_cast_fp16)[name = tensor<string, []>("out_71_cast_fp16")];
            tensor<fp16, [1280]> input_115_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_115_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362866752)))];
            tensor<fp16, [1280]> input_115_beta_0_to_fp16 = const()[name = tensor<string, []>("input_115_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362869376)))];
            tensor<fp16, []> input_115_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_115_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<int32, [2]> var_2721 = const()[name = tensor<string, []>("op_2721"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2723 = const()[name = tensor<string, []>("op_2723"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_117_pad_type_0 = const()[name = tensor<string, []>("input_117_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_117_pad_0 = const()[name = tensor<string, []>("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_11_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362872000))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369425664))), name = tensor<string, []>("layers_11_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_11_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369426240)))];
            tensor<fp16, [1, 5120, 1, 1]> input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = var_2723, groups = var_2537, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = var_2721, weight = layers_11_fc1_weight_to_fp16_palettized, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<string, []> input_119_mode_0 = const()[name = tensor<string, []>("input_119_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<int32, [2]> var_2729 = const()[name = tensor<string, []>("op_2729"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2731 = const()[name = tensor<string, []>("op_2731"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_25_pad_type_0 = const()[name = tensor<string, []>("hidden_states_25_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_25_pad_0 = const()[name = tensor<string, []>("hidden_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_11_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369436544)))];
            tensor<fp16, [1280]> layers_11_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382543808)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = var_2731, groups = var_2537, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = var_2729, weight = layers_11_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("hidden_states_25_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor<string, []>("inputs_73_cast_fp16")];
            tensor<int32, []> var_2744 = const()[name = tensor<string, []>("op_2744"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_2751 = const()[name = tensor<string, []>("op_2751"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_2752 = const()[name = tensor<string, []>("op_2752"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_2764 = const()[name = tensor<string, []>("op_2764"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_73_cast_fp16 = reduce_mean(axes = var_2764, keep_dims = var_2752, x = inputs_73_cast_fp16)[name = tensor<string, []>("channels_mean_73_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_73_cast_fp16 = sub(x = inputs_73_cast_fp16, y = channels_mean_73_cast_fp16)[name = tensor<string, []>("zero_mean_73_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = zero_mean_73_cast_fp16)[name = tensor<string, []>("zero_mean_sq_73_cast_fp16")];
            tensor<int32, [1]> var_2768 = const()[name = tensor<string, []>("op_2768"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2769_cast_fp16 = reduce_mean(axes = var_2768, keep_dims = var_2752, x = zero_mean_sq_73_cast_fp16)[name = tensor<string, []>("op_2769_cast_fp16")];
            tensor<fp16, []> var_2770_to_fp16 = const()[name = tensor<string, []>("op_2770_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2771_cast_fp16 = add(x = var_2769_cast_fp16, y = var_2770_to_fp16)[name = tensor<string, []>("op_2771_cast_fp16")];
            tensor<fp32, []> denom_73_epsilon_0 = const()[name = tensor<string, []>("denom_73_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_73_cast_fp16 = rsqrt(epsilon = denom_73_epsilon_0, x = var_2771_cast_fp16)[name = tensor<string, []>("denom_73_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = denom_73_cast_fp16)[name = tensor<string, []>("out_73_cast_fp16")];
            tensor<fp16, [1280]> obj_145_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_145_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382546432)))];
            tensor<fp16, [1280]> obj_145_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_145_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382549056)))];
            tensor<fp16, []> obj_145_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_145_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_145_cast_fp16 = batch_norm(beta = obj_145_beta_0_to_fp16, epsilon = obj_145_epsilon_0_to_fp16, gamma = obj_145_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor<string, []>("obj_145_cast_fp16")];
            tensor<int32, [2]> var_2786 = const()[name = tensor<string, []>("op_2786"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2788 = const()[name = tensor<string, []>("op_2788"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_49_pad_type_0 = const()[name = tensor<string, []>("query_49_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_49_pad_0 = const()[name = tensor<string, []>("query_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382551680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383780544))), name = tensor<string, []>("layers_12_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383780736)))];
            tensor<fp16, [1, 1280, 1, 1]> query_49_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = var_2788, groups = var_2751, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = var_2786, weight = layers_12_self_attn_q_proj_weight_to_fp16_palettized, x = obj_145_cast_fp16)[name = tensor<string, []>("query_49_cast_fp16")];
            tensor<int32, [2]> var_2792 = const()[name = tensor<string, []>("op_2792"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2794 = const()[name = tensor<string, []>("op_2794"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_25_pad_type_0 = const()[name = tensor<string, []>("current_key_25_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_25_pad_0 = const()[name = tensor<string, []>("current_key_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383783360))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384602624))), name = tensor<string, []>("layers_12_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_25_cast_fp16 = conv(dilations = var_2794, groups = var_2751, pad = current_key_25_pad_0, pad_type = current_key_25_pad_type_0, strides = var_2792, weight = layers_12_self_attn_k_proj_weight_to_fp16_palettized, x = obj_145_cast_fp16)[name = tensor<string, []>("current_key_25_cast_fp16")];
            tensor<int32, [2]> var_2799 = const()[name = tensor<string, []>("op_2799"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2801 = const()[name = tensor<string, []>("op_2801"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_25_pad_type_0 = const()[name = tensor<string, []>("current_value_25_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_25_pad_0 = const()[name = tensor<string, []>("current_value_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384602752))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385422016))), name = tensor<string, []>("layers_12_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385422144)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = var_2801, groups = var_2751, pad = current_value_25_pad_0, pad_type = current_value_25_pad_type_0, strides = var_2799, weight = layers_12_self_attn_v_proj_weight_to_fp16_palettized, x = obj_145_cast_fp16)[name = tensor<string, []>("current_value_25_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2808_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_2808_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2810_cast_fp16 = mul(x = var_103_cast_fp16_12, y = var_241_cast_fp16)[name = tensor<string, []>("op_2810_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_49_cast_fp16 = add(x = var_2808_cast_fp16, y = var_2810_cast_fp16)[name = tensor<string, []>("key_49_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2812_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_2812_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_2814_cast_fp16 = mul(x = var_138_cast_fp16_12, y = var_241_cast_fp16)[name = tensor<string, []>("op_2814_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_49_cast_fp16 = add(x = var_2812_cast_fp16, y = var_2814_cast_fp16)[name = tensor<string, []>("value_49_cast_fp16")];
            tensor<int32, [4]> var_2817 = const()[name = tensor<string, []>("op_2817"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2818_cast_fp16 = reshape(shape = var_2817, x = query_49_cast_fp16)[name = tensor<string, []>("op_2818_cast_fp16")];
            tensor<fp16, []> var_2819_to_fp16 = const()[name = tensor<string, []>("op_2819_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2820_cast_fp16 = mul(x = var_2818_cast_fp16, y = var_2819_to_fp16)[name = tensor<string, []>("op_2820_cast_fp16")];
            tensor<int32, [4]> var_2821 = const()[name = tensor<string, []>("op_2821"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_2822_cast_fp16 = reshape(shape = var_2821, x = key_49_cast_fp16)[name = tensor<string, []>("op_2822_cast_fp16")];
            tensor<bool, []> mh_w_73_transpose_x_0 = const()[name = tensor<string, []>("mh_w_73_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_73_transpose_y_0 = const()[name = tensor<string, []>("mh_w_73_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 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_2820_cast_fp16, y = var_2822_cast_fp16)[name = tensor<string, []>("mh_w_73_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_75_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_2830_cast_fp16 = softmax(axis = var_2744, x = mh_w_75_cast_fp16)[name = tensor<string, []>("op_2830_cast_fp16")];
            tensor<int32, [4]> var_2831 = const()[name = tensor<string, []>("op_2831"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_2832_cast_fp16 = reshape(shape = var_2831, x = value_49_cast_fp16)[name = tensor<string, []>("op_2832_cast_fp16")];
            tensor<bool, []> attn_49_transpose_x_0 = const()[name = tensor<string, []>("attn_49_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_49_transpose_y_0 = const()[name = tensor<string, []>("attn_49_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_2832_cast_fp16, y = var_2830_cast_fp16)[name = tensor<string, []>("attn_49_cast_fp16")];
            tensor<int32, [4]> var_2835 = const()[name = tensor<string, []>("op_2835"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_121_cast_fp16 = reshape(shape = var_2835, x = attn_49_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<int32, [2]> var_2839 = const()[name = tensor<string, []>("op_2839"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2841 = const()[name = tensor<string, []>("op_2841"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_151_pad_type_0 = const()[name = tensor<string, []>("obj_151_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_151_pad_0 = const()[name = tensor<string, []>("obj_151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385424768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386653632))), name = tensor<string, []>("layers_12_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386653824)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_151_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = var_2841, groups = var_2751, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = var_2839, weight = layers_12_self_attn_o_proj_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = tensor<string, []>("obj_151_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_151_cast_fp16)[name = tensor<string, []>("inputs_75_cast_fp16")];
            tensor<int32, [1]> var_2851 = const()[name = tensor<string, []>("op_2851"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_75_cast_fp16 = reduce_mean(axes = var_2851, keep_dims = var_2752, x = inputs_75_cast_fp16)[name = tensor<string, []>("channels_mean_75_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_75_cast_fp16 = sub(x = inputs_75_cast_fp16, y = channels_mean_75_cast_fp16)[name = tensor<string, []>("zero_mean_75_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = zero_mean_75_cast_fp16)[name = tensor<string, []>("zero_mean_sq_75_cast_fp16")];
            tensor<int32, [1]> var_2855 = const()[name = tensor<string, []>("op_2855"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2856_cast_fp16 = reduce_mean(axes = var_2855, keep_dims = var_2752, x = zero_mean_sq_75_cast_fp16)[name = tensor<string, []>("op_2856_cast_fp16")];
            tensor<fp16, []> var_2857_to_fp16 = const()[name = tensor<string, []>("op_2857_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2858_cast_fp16 = add(x = var_2856_cast_fp16, y = var_2857_to_fp16)[name = tensor<string, []>("op_2858_cast_fp16")];
            tensor<fp32, []> denom_75_epsilon_0 = const()[name = tensor<string, []>("denom_75_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_75_cast_fp16 = rsqrt(epsilon = denom_75_epsilon_0, x = var_2858_cast_fp16)[name = tensor<string, []>("denom_75_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = denom_75_cast_fp16)[name = tensor<string, []>("out_75_cast_fp16")];
            tensor<fp16, [1280]> obj_153_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_153_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386656448)))];
            tensor<fp16, [1280]> obj_153_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_153_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386659072)))];
            tensor<fp16, []> obj_153_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_153_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_153_cast_fp16 = batch_norm(beta = obj_153_beta_0_to_fp16, epsilon = obj_153_epsilon_0_to_fp16, gamma = obj_153_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor<string, []>("obj_153_cast_fp16")];
            tensor<int32, [2]> var_2873 = const()[name = tensor<string, []>("op_2873"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2875 = const()[name = tensor<string, []>("op_2875"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_51_pad_type_0 = const()[name = tensor<string, []>("query_51_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_51_pad_0 = const()[name = tensor<string, []>("query_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386661696))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387480960))), name = tensor<string, []>("layers_12_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_12_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387481088)))];
            tensor<fp16, [1, 1280, 1, 1]> query_51_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_bias_to_fp16, dilations = var_2875, groups = var_2751, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = var_2873, weight = layers_12_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_153_cast_fp16)[name = tensor<string, []>("query_51_cast_fp16")];
            tensor<int32, [2]> var_2879 = const()[name = tensor<string, []>("op_2879"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2881 = const()[name = tensor<string, []>("op_2881"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_51_pad_type_0 = const()[name = tensor<string, []>("key_51_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_51_pad_0 = const()[name = tensor<string, []>("key_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387483712))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388302976))), name = tensor<string, []>("layers_12_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_51_cast_fp16 = conv(dilations = var_2881, groups = var_2751, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = var_2879, weight = layers_12_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_51_cast_fp16")];
            tensor<int32, [2]> var_2886 = const()[name = tensor<string, []>("op_2886"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2888 = const()[name = tensor<string, []>("op_2888"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_51_pad_type_0 = const()[name = tensor<string, []>("value_51_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_51_pad_0 = const()[name = tensor<string, []>("value_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388303104))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389122368))), name = tensor<string, []>("layers_12_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_12_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389122496)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_51_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_bias_to_fp16, dilations = var_2888, groups = var_2751, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = var_2886, weight = layers_12_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_51_cast_fp16")];
            tensor<int32, [4]> var_2892 = const()[name = tensor<string, []>("op_2892"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_2893_cast_fp16 = reshape(shape = var_2892, x = query_51_cast_fp16)[name = tensor<string, []>("op_2893_cast_fp16")];
            tensor<fp16, []> var_2894_to_fp16 = const()[name = tensor<string, []>("op_2894_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_2895_cast_fp16 = mul(x = var_2893_cast_fp16, y = var_2894_to_fp16)[name = tensor<string, []>("op_2895_cast_fp16")];
            tensor<int32, [4]> var_2896 = const()[name = tensor<string, []>("op_2896"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2897_cast_fp16 = reshape(shape = var_2896, x = key_51_cast_fp16)[name = tensor<string, []>("op_2897_cast_fp16")];
            tensor<bool, []> mh_w_77_transpose_x_0 = const()[name = tensor<string, []>("mh_w_77_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_77_transpose_y_0 = const()[name = tensor<string, []>("mh_w_77_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_2895_cast_fp16, y = var_2897_cast_fp16)[name = tensor<string, []>("mh_w_77_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_2900_cast_fp16 = softmax(axis = var_2744, x = mh_w_77_cast_fp16)[name = tensor<string, []>("op_2900_cast_fp16")];
            tensor<int32, [4]> var_2901 = const()[name = tensor<string, []>("op_2901"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_2902_cast_fp16 = reshape(shape = var_2901, x = value_51_cast_fp16)[name = tensor<string, []>("op_2902_cast_fp16")];
            tensor<bool, []> attn_51_transpose_x_0 = const()[name = tensor<string, []>("attn_51_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_51_transpose_y_0 = const()[name = tensor<string, []>("attn_51_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_2902_cast_fp16, y = var_2900_cast_fp16)[name = tensor<string, []>("attn_51_cast_fp16")];
            tensor<int32, [4]> var_2905 = const()[name = tensor<string, []>("op_2905"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_123_cast_fp16 = reshape(shape = var_2905, x = attn_51_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<int32, [2]> var_2909 = const()[name = tensor<string, []>("op_2909"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2911 = const()[name = tensor<string, []>("op_2911"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_155_pad_type_0 = const()[name = tensor<string, []>("obj_155_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_155_pad_0 = const()[name = tensor<string, []>("obj_155_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_12_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389125120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390353984))), name = tensor<string, []>("layers_12_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_12_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390354176)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_155_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_bias_to_fp16, dilations = var_2911, groups = var_2751, pad = obj_155_pad_0, pad_type = obj_155_pad_type_0, strides = var_2909, weight = layers_12_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_123_cast_fp16)[name = tensor<string, []>("obj_155_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_155_cast_fp16)[name = tensor<string, []>("inputs_77_cast_fp16")];
            tensor<int32, [1]> var_2917 = const()[name = tensor<string, []>("op_2917"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_77_cast_fp16 = reduce_mean(axes = var_2917, keep_dims = var_2752, x = inputs_77_cast_fp16)[name = tensor<string, []>("channels_mean_77_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_77_cast_fp16 = sub(x = inputs_77_cast_fp16, y = channels_mean_77_cast_fp16)[name = tensor<string, []>("zero_mean_77_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = zero_mean_77_cast_fp16)[name = tensor<string, []>("zero_mean_sq_77_cast_fp16")];
            tensor<int32, [1]> var_2921 = const()[name = tensor<string, []>("op_2921"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2922_cast_fp16 = reduce_mean(axes = var_2921, keep_dims = var_2752, x = zero_mean_sq_77_cast_fp16)[name = tensor<string, []>("op_2922_cast_fp16")];
            tensor<fp16, []> var_2923_to_fp16 = const()[name = tensor<string, []>("op_2923_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2924_cast_fp16 = add(x = var_2922_cast_fp16, y = var_2923_to_fp16)[name = tensor<string, []>("op_2924_cast_fp16")];
            tensor<fp32, []> denom_77_epsilon_0 = const()[name = tensor<string, []>("denom_77_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_77_cast_fp16 = rsqrt(epsilon = denom_77_epsilon_0, x = var_2924_cast_fp16)[name = tensor<string, []>("denom_77_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = denom_77_cast_fp16)[name = tensor<string, []>("out_77_cast_fp16")];
            tensor<fp16, [1280]> input_125_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_125_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390356800)))];
            tensor<fp16, [1280]> input_125_beta_0_to_fp16 = const()[name = tensor<string, []>("input_125_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390359424)))];
            tensor<fp16, []> input_125_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_125_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_125_cast_fp16 = batch_norm(beta = input_125_beta_0_to_fp16, epsilon = input_125_epsilon_0_to_fp16, gamma = input_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
            tensor<int32, [2]> var_2935 = const()[name = tensor<string, []>("op_2935"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2937 = const()[name = tensor<string, []>("op_2937"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_127_pad_type_0 = const()[name = tensor<string, []>("input_127_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_127_pad_0 = const()[name = tensor<string, []>("input_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_12_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390362048))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396915712))), name = tensor<string, []>("layers_12_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_12_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396916288)))];
            tensor<fp16, [1, 5120, 1, 1]> input_127_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = var_2937, groups = var_2751, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = var_2935, weight = layers_12_fc1_weight_to_fp16_palettized, x = input_125_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<string, []> input_129_mode_0 = const()[name = tensor<string, []>("input_129_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_129_cast_fp16 = gelu(mode = input_129_mode_0, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<int32, [2]> var_2943 = const()[name = tensor<string, []>("op_2943"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2945 = const()[name = tensor<string, []>("op_2945"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_27_pad_type_0 = const()[name = tensor<string, []>("hidden_states_27_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = tensor<string, []>("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_12_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396926592))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401841856))), name = tensor<string, []>("layers_12_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_12_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401842048)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_27_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = var_2945, groups = var_2751, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = var_2943, weight = layers_12_fc2_weight_to_fp16_palettized, x = input_129_cast_fp16)[name = tensor<string, []>("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor<string, []>("inputs_79_cast_fp16")];
            tensor<int32, []> var_2958 = const()[name = tensor<string, []>("op_2958"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_2965 = const()[name = tensor<string, []>("op_2965"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_2966 = const()[name = tensor<string, []>("op_2966"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_2978 = const()[name = tensor<string, []>("op_2978"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_79_cast_fp16 = reduce_mean(axes = var_2978, keep_dims = var_2966, x = inputs_79_cast_fp16)[name = tensor<string, []>("channels_mean_79_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_79_cast_fp16 = sub(x = inputs_79_cast_fp16, y = channels_mean_79_cast_fp16)[name = tensor<string, []>("zero_mean_79_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = zero_mean_79_cast_fp16)[name = tensor<string, []>("zero_mean_sq_79_cast_fp16")];
            tensor<int32, [1]> var_2982 = const()[name = tensor<string, []>("op_2982"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_2983_cast_fp16 = reduce_mean(axes = var_2982, keep_dims = var_2966, x = zero_mean_sq_79_cast_fp16)[name = tensor<string, []>("op_2983_cast_fp16")];
            tensor<fp16, []> var_2984_to_fp16 = const()[name = tensor<string, []>("op_2984_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_2985_cast_fp16 = add(x = var_2983_cast_fp16, y = var_2984_to_fp16)[name = tensor<string, []>("op_2985_cast_fp16")];
            tensor<fp32, []> denom_79_epsilon_0 = const()[name = tensor<string, []>("denom_79_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_79_cast_fp16 = rsqrt(epsilon = denom_79_epsilon_0, x = var_2985_cast_fp16)[name = tensor<string, []>("denom_79_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = denom_79_cast_fp16)[name = tensor<string, []>("out_79_cast_fp16")];
            tensor<fp16, [1280]> obj_157_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_157_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401844672)))];
            tensor<fp16, [1280]> obj_157_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_157_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401847296)))];
            tensor<fp16, []> obj_157_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_157_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_157_cast_fp16 = batch_norm(beta = obj_157_beta_0_to_fp16, epsilon = obj_157_epsilon_0_to_fp16, gamma = obj_157_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor<string, []>("obj_157_cast_fp16")];
            tensor<int32, [2]> var_3000 = const()[name = tensor<string, []>("op_3000"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3002 = const()[name = tensor<string, []>("op_3002"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_53_pad_type_0 = const()[name = tensor<string, []>("query_53_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_53_pad_0 = const()[name = tensor<string, []>("query_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401849920))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403078784))), name = tensor<string, []>("layers_13_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403078976)))];
            tensor<fp16, [1, 1280, 1, 1]> query_53_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = var_3002, groups = var_2965, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = var_3000, weight = layers_13_self_attn_q_proj_weight_to_fp16_palettized, x = obj_157_cast_fp16)[name = tensor<string, []>("query_53_cast_fp16")];
            tensor<int32, [2]> var_3006 = const()[name = tensor<string, []>("op_3006"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3008 = const()[name = tensor<string, []>("op_3008"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_27_pad_type_0 = const()[name = tensor<string, []>("current_key_27_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_27_pad_0 = const()[name = tensor<string, []>("current_key_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403081600))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404310464))), name = tensor<string, []>("layers_13_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_27_cast_fp16 = conv(dilations = var_3008, groups = var_2965, pad = current_key_27_pad_0, pad_type = current_key_27_pad_type_0, strides = var_3006, weight = layers_13_self_attn_k_proj_weight_to_fp16_palettized, x = obj_157_cast_fp16)[name = tensor<string, []>("current_key_27_cast_fp16")];
            tensor<int32, [2]> var_3013 = const()[name = tensor<string, []>("op_3013"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3015 = const()[name = tensor<string, []>("op_3015"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_27_pad_type_0 = const()[name = tensor<string, []>("current_value_27_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_27_pad_0 = const()[name = tensor<string, []>("current_value_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404310656))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405129920))), name = tensor<string, []>("layers_13_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405130048)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = var_3015, groups = var_2965, pad = current_value_27_pad_0, pad_type = current_value_27_pad_type_0, strides = var_3013, weight = layers_13_self_attn_v_proj_weight_to_fp16_palettized, x = obj_157_cast_fp16)[name = tensor<string, []>("current_value_27_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3022_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3022_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3024_cast_fp16 = mul(x = var_103_cast_fp16_13, y = var_241_cast_fp16)[name = tensor<string, []>("op_3024_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_53_cast_fp16 = add(x = var_3022_cast_fp16, y = var_3024_cast_fp16)[name = tensor<string, []>("key_53_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3026_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3026_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3028_cast_fp16 = mul(x = var_138_cast_fp16_13, y = var_241_cast_fp16)[name = tensor<string, []>("op_3028_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_53_cast_fp16 = add(x = var_3026_cast_fp16, y = var_3028_cast_fp16)[name = tensor<string, []>("value_53_cast_fp16")];
            tensor<int32, [4]> var_3031 = const()[name = tensor<string, []>("op_3031"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3032_cast_fp16 = reshape(shape = var_3031, x = query_53_cast_fp16)[name = tensor<string, []>("op_3032_cast_fp16")];
            tensor<fp16, []> var_3033_to_fp16 = const()[name = tensor<string, []>("op_3033_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3034_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_3033_to_fp16)[name = tensor<string, []>("op_3034_cast_fp16")];
            tensor<int32, [4]> var_3035 = const()[name = tensor<string, []>("op_3035"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3036_cast_fp16 = reshape(shape = var_3035, x = key_53_cast_fp16)[name = tensor<string, []>("op_3036_cast_fp16")];
            tensor<bool, []> mh_w_79_transpose_x_0 = const()[name = tensor<string, []>("mh_w_79_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_79_transpose_y_0 = const()[name = tensor<string, []>("mh_w_79_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_79_cast_fp16 = matmul(transpose_x = mh_w_79_transpose_x_0, transpose_y = mh_w_79_transpose_y_0, x = var_3034_cast_fp16, y = var_3036_cast_fp16)[name = tensor<string, []>("mh_w_79_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_81_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_3044_cast_fp16 = softmax(axis = var_2958, x = mh_w_81_cast_fp16)[name = tensor<string, []>("op_3044_cast_fp16")];
            tensor<int32, [4]> var_3045 = const()[name = tensor<string, []>("op_3045"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3046_cast_fp16 = reshape(shape = var_3045, x = value_53_cast_fp16)[name = tensor<string, []>("op_3046_cast_fp16")];
            tensor<bool, []> attn_53_transpose_x_0 = const()[name = tensor<string, []>("attn_53_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_53_transpose_y_0 = const()[name = tensor<string, []>("attn_53_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3046_cast_fp16, y = var_3044_cast_fp16)[name = tensor<string, []>("attn_53_cast_fp16")];
            tensor<int32, [4]> var_3049 = const()[name = tensor<string, []>("op_3049"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_131_cast_fp16 = reshape(shape = var_3049, x = attn_53_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<int32, [2]> var_3053 = const()[name = tensor<string, []>("op_3053"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3055 = const()[name = tensor<string, []>("op_3055"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_163_pad_type_0 = const()[name = tensor<string, []>("obj_163_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_163_pad_0 = const()[name = tensor<string, []>("obj_163_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405132672))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406361536))), name = tensor<string, []>("layers_13_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406361728)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_163_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = var_3055, groups = var_2965, pad = obj_163_pad_0, pad_type = obj_163_pad_type_0, strides = var_3053, weight = layers_13_self_attn_o_proj_weight_to_fp16_palettized, x = input_131_cast_fp16)[name = tensor<string, []>("obj_163_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_163_cast_fp16)[name = tensor<string, []>("inputs_81_cast_fp16")];
            tensor<int32, [1]> var_3065 = const()[name = tensor<string, []>("op_3065"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_81_cast_fp16 = reduce_mean(axes = var_3065, keep_dims = var_2966, x = inputs_81_cast_fp16)[name = tensor<string, []>("channels_mean_81_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_81_cast_fp16 = sub(x = inputs_81_cast_fp16, y = channels_mean_81_cast_fp16)[name = tensor<string, []>("zero_mean_81_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = zero_mean_81_cast_fp16)[name = tensor<string, []>("zero_mean_sq_81_cast_fp16")];
            tensor<int32, [1]> var_3069 = const()[name = tensor<string, []>("op_3069"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3070_cast_fp16 = reduce_mean(axes = var_3069, keep_dims = var_2966, x = zero_mean_sq_81_cast_fp16)[name = tensor<string, []>("op_3070_cast_fp16")];
            tensor<fp16, []> var_3071_to_fp16 = const()[name = tensor<string, []>("op_3071_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3072_cast_fp16 = add(x = var_3070_cast_fp16, y = var_3071_to_fp16)[name = tensor<string, []>("op_3072_cast_fp16")];
            tensor<fp32, []> denom_81_epsilon_0 = const()[name = tensor<string, []>("denom_81_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_81_cast_fp16 = rsqrt(epsilon = denom_81_epsilon_0, x = var_3072_cast_fp16)[name = tensor<string, []>("denom_81_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = denom_81_cast_fp16)[name = tensor<string, []>("out_81_cast_fp16")];
            tensor<fp16, [1280]> obj_165_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_165_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406364352)))];
            tensor<fp16, [1280]> obj_165_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_165_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406366976)))];
            tensor<fp16, []> obj_165_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_165_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_165_cast_fp16 = batch_norm(beta = obj_165_beta_0_to_fp16, epsilon = obj_165_epsilon_0_to_fp16, gamma = obj_165_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor<string, []>("obj_165_cast_fp16")];
            tensor<int32, [2]> var_3087 = const()[name = tensor<string, []>("op_3087"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3089 = const()[name = tensor<string, []>("op_3089"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_55_pad_type_0 = const()[name = tensor<string, []>("query_55_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_55_pad_0 = const()[name = tensor<string, []>("query_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406369600))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407188864))), name = tensor<string, []>("layers_13_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_13_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407188992)))];
            tensor<fp16, [1, 1280, 1, 1]> query_55_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_bias_to_fp16, dilations = var_3089, groups = var_2965, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = var_3087, weight = layers_13_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_165_cast_fp16)[name = tensor<string, []>("query_55_cast_fp16")];
            tensor<int32, [2]> var_3093 = const()[name = tensor<string, []>("op_3093"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3095 = const()[name = tensor<string, []>("op_3095"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_55_pad_type_0 = const()[name = tensor<string, []>("key_55_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_55_pad_0 = const()[name = tensor<string, []>("key_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407191616))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408010880))), name = tensor<string, []>("layers_13_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_55_cast_fp16 = conv(dilations = var_3095, groups = var_2965, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = var_3093, weight = layers_13_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_55_cast_fp16")];
            tensor<int32, [2]> var_3100 = const()[name = tensor<string, []>("op_3100"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3102 = const()[name = tensor<string, []>("op_3102"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_55_pad_type_0 = const()[name = tensor<string, []>("value_55_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_55_pad_0 = const()[name = tensor<string, []>("value_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408011008))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408830272))), name = tensor<string, []>("layers_13_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_13_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408830400)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_55_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_bias_to_fp16, dilations = var_3102, groups = var_2965, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = var_3100, weight = layers_13_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_55_cast_fp16")];
            tensor<int32, [4]> var_3106 = const()[name = tensor<string, []>("op_3106"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3107_cast_fp16 = reshape(shape = var_3106, x = query_55_cast_fp16)[name = tensor<string, []>("op_3107_cast_fp16")];
            tensor<fp16, []> var_3108_to_fp16 = const()[name = tensor<string, []>("op_3108_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3109_cast_fp16 = mul(x = var_3107_cast_fp16, y = var_3108_to_fp16)[name = tensor<string, []>("op_3109_cast_fp16")];
            tensor<int32, [4]> var_3110 = const()[name = tensor<string, []>("op_3110"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3111_cast_fp16 = reshape(shape = var_3110, x = key_55_cast_fp16)[name = tensor<string, []>("op_3111_cast_fp16")];
            tensor<bool, []> mh_w_83_transpose_x_0 = const()[name = tensor<string, []>("mh_w_83_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_83_transpose_y_0 = const()[name = tensor<string, []>("mh_w_83_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_83_cast_fp16 = matmul(transpose_x = mh_w_83_transpose_x_0, transpose_y = mh_w_83_transpose_y_0, x = var_3109_cast_fp16, y = var_3111_cast_fp16)[name = tensor<string, []>("mh_w_83_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_3114_cast_fp16 = softmax(axis = var_2958, x = mh_w_83_cast_fp16)[name = tensor<string, []>("op_3114_cast_fp16")];
            tensor<int32, [4]> var_3115 = const()[name = tensor<string, []>("op_3115"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3116_cast_fp16 = reshape(shape = var_3115, x = value_55_cast_fp16)[name = tensor<string, []>("op_3116_cast_fp16")];
            tensor<bool, []> attn_55_transpose_x_0 = const()[name = tensor<string, []>("attn_55_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_55_transpose_y_0 = const()[name = tensor<string, []>("attn_55_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3116_cast_fp16, y = var_3114_cast_fp16)[name = tensor<string, []>("attn_55_cast_fp16")];
            tensor<int32, [4]> var_3119 = const()[name = tensor<string, []>("op_3119"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_133_cast_fp16 = reshape(shape = var_3119, x = attn_55_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<int32, [2]> var_3123 = const()[name = tensor<string, []>("op_3123"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3125 = const()[name = tensor<string, []>("op_3125"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_167_pad_type_0 = const()[name = tensor<string, []>("obj_167_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_167_pad_0 = const()[name = tensor<string, []>("obj_167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_13_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408833024))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409652288))), name = tensor<string, []>("layers_13_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_13_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409652416)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_167_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_bias_to_fp16, dilations = var_3125, groups = var_2965, pad = obj_167_pad_0, pad_type = obj_167_pad_type_0, strides = var_3123, weight = layers_13_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = tensor<string, []>("obj_167_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_167_cast_fp16)[name = tensor<string, []>("inputs_83_cast_fp16")];
            tensor<int32, [1]> var_3131 = const()[name = tensor<string, []>("op_3131"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_83_cast_fp16 = reduce_mean(axes = var_3131, keep_dims = var_2966, x = inputs_83_cast_fp16)[name = tensor<string, []>("channels_mean_83_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_83_cast_fp16 = sub(x = inputs_83_cast_fp16, y = channels_mean_83_cast_fp16)[name = tensor<string, []>("zero_mean_83_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = zero_mean_83_cast_fp16)[name = tensor<string, []>("zero_mean_sq_83_cast_fp16")];
            tensor<int32, [1]> var_3135 = const()[name = tensor<string, []>("op_3135"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3136_cast_fp16 = reduce_mean(axes = var_3135, keep_dims = var_2966, x = zero_mean_sq_83_cast_fp16)[name = tensor<string, []>("op_3136_cast_fp16")];
            tensor<fp16, []> var_3137_to_fp16 = const()[name = tensor<string, []>("op_3137_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3138_cast_fp16 = add(x = var_3136_cast_fp16, y = var_3137_to_fp16)[name = tensor<string, []>("op_3138_cast_fp16")];
            tensor<fp32, []> denom_83_epsilon_0 = const()[name = tensor<string, []>("denom_83_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_83_cast_fp16 = rsqrt(epsilon = denom_83_epsilon_0, x = var_3138_cast_fp16)[name = tensor<string, []>("denom_83_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = denom_83_cast_fp16)[name = tensor<string, []>("out_83_cast_fp16")];
            tensor<fp16, [1280]> input_135_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_135_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409655040)))];
            tensor<fp16, [1280]> input_135_beta_0_to_fp16 = const()[name = tensor<string, []>("input_135_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409657664)))];
            tensor<fp16, []> input_135_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_135_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_135_cast_fp16 = batch_norm(beta = input_135_beta_0_to_fp16, epsilon = input_135_epsilon_0_to_fp16, gamma = input_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<int32, [2]> var_3149 = const()[name = tensor<string, []>("op_3149"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3151 = const()[name = tensor<string, []>("op_3151"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_137_pad_type_0 = const()[name = tensor<string, []>("input_137_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_137_pad_0 = const()[name = tensor<string, []>("input_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_13_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409660288))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414575552))), name = tensor<string, []>("layers_13_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_13_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414575744)))];
            tensor<fp16, [1, 5120, 1, 1]> input_137_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = var_3151, groups = var_2965, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = var_3149, weight = layers_13_fc1_weight_to_fp16_palettized, x = input_135_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
            tensor<string, []> input_139_mode_0 = const()[name = tensor<string, []>("input_139_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = input_137_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
            tensor<int32, [2]> var_3157 = const()[name = tensor<string, []>("op_3157"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3159 = const()[name = tensor<string, []>("op_3159"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_29_pad_type_0 = const()[name = tensor<string, []>("hidden_states_29_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_29_pad_0 = const()[name = tensor<string, []>("hidden_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_13_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414586048))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419501312))), name = tensor<string, []>("layers_13_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_13_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419501504)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_29_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = var_3159, groups = var_2965, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = var_3157, weight = layers_13_fc2_weight_to_fp16_palettized, x = input_139_cast_fp16)[name = tensor<string, []>("hidden_states_29_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor<string, []>("inputs_85_cast_fp16")];
            tensor<int32, []> var_3172 = const()[name = tensor<string, []>("op_3172"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_3179 = const()[name = tensor<string, []>("op_3179"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_3180 = const()[name = tensor<string, []>("op_3180"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_3192 = const()[name = tensor<string, []>("op_3192"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_85_cast_fp16 = reduce_mean(axes = var_3192, keep_dims = var_3180, x = inputs_85_cast_fp16)[name = tensor<string, []>("channels_mean_85_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_85_cast_fp16 = sub(x = inputs_85_cast_fp16, y = channels_mean_85_cast_fp16)[name = tensor<string, []>("zero_mean_85_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = zero_mean_85_cast_fp16)[name = tensor<string, []>("zero_mean_sq_85_cast_fp16")];
            tensor<int32, [1]> var_3196 = const()[name = tensor<string, []>("op_3196"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3197_cast_fp16 = reduce_mean(axes = var_3196, keep_dims = var_3180, x = zero_mean_sq_85_cast_fp16)[name = tensor<string, []>("op_3197_cast_fp16")];
            tensor<fp16, []> var_3198_to_fp16 = const()[name = tensor<string, []>("op_3198_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3199_cast_fp16 = add(x = var_3197_cast_fp16, y = var_3198_to_fp16)[name = tensor<string, []>("op_3199_cast_fp16")];
            tensor<fp32, []> denom_85_epsilon_0 = const()[name = tensor<string, []>("denom_85_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_85_cast_fp16 = rsqrt(epsilon = denom_85_epsilon_0, x = var_3199_cast_fp16)[name = tensor<string, []>("denom_85_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = denom_85_cast_fp16)[name = tensor<string, []>("out_85_cast_fp16")];
            tensor<fp16, [1280]> obj_169_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_169_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419504128)))];
            tensor<fp16, [1280]> obj_169_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_169_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419506752)))];
            tensor<fp16, []> obj_169_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_169_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor<string, []>("obj_169_cast_fp16")];
            tensor<int32, [2]> var_3214 = const()[name = tensor<string, []>("op_3214"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3216 = const()[name = tensor<string, []>("op_3216"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_57_pad_type_0 = const()[name = tensor<string, []>("query_57_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_57_pad_0 = const()[name = tensor<string, []>("query_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419509376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420738240))), name = tensor<string, []>("layers_14_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420738432)))];
            tensor<fp16, [1, 1280, 1, 1]> query_57_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = var_3216, groups = var_3179, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = var_3214, weight = layers_14_self_attn_q_proj_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor<string, []>("query_57_cast_fp16")];
            tensor<int32, [2]> var_3220 = const()[name = tensor<string, []>("op_3220"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3222 = const()[name = tensor<string, []>("op_3222"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_29_pad_type_0 = const()[name = tensor<string, []>("current_key_29_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_29_pad_0 = const()[name = tensor<string, []>("current_key_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420741056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421969920))), name = tensor<string, []>("layers_14_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_29_cast_fp16 = conv(dilations = var_3222, groups = var_3179, pad = current_key_29_pad_0, pad_type = current_key_29_pad_type_0, strides = var_3220, weight = layers_14_self_attn_k_proj_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor<string, []>("current_key_29_cast_fp16")];
            tensor<int32, [2]> var_3227 = const()[name = tensor<string, []>("op_3227"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3229 = const()[name = tensor<string, []>("op_3229"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_29_pad_type_0 = const()[name = tensor<string, []>("current_value_29_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_29_pad_0 = const()[name = tensor<string, []>("current_value_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421970112))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422789376))), name = tensor<string, []>("layers_14_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422789504)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = var_3229, groups = var_3179, pad = current_value_29_pad_0, pad_type = current_value_29_pad_type_0, strides = var_3227, weight = layers_14_self_attn_v_proj_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor<string, []>("current_value_29_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3236_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3236_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3238_cast_fp16 = mul(x = var_103_cast_fp16_14, y = var_241_cast_fp16)[name = tensor<string, []>("op_3238_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_57_cast_fp16 = add(x = var_3236_cast_fp16, y = var_3238_cast_fp16)[name = tensor<string, []>("key_57_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3240_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3240_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3242_cast_fp16 = mul(x = var_138_cast_fp16_14, y = var_241_cast_fp16)[name = tensor<string, []>("op_3242_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_57_cast_fp16 = add(x = var_3240_cast_fp16, y = var_3242_cast_fp16)[name = tensor<string, []>("value_57_cast_fp16")];
            tensor<int32, [4]> var_3245 = const()[name = tensor<string, []>("op_3245"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3246_cast_fp16 = reshape(shape = var_3245, x = query_57_cast_fp16)[name = tensor<string, []>("op_3246_cast_fp16")];
            tensor<fp16, []> var_3247_to_fp16 = const()[name = tensor<string, []>("op_3247_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3248_cast_fp16 = mul(x = var_3246_cast_fp16, y = var_3247_to_fp16)[name = tensor<string, []>("op_3248_cast_fp16")];
            tensor<int32, [4]> var_3249 = const()[name = tensor<string, []>("op_3249"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3250_cast_fp16 = reshape(shape = var_3249, x = key_57_cast_fp16)[name = tensor<string, []>("op_3250_cast_fp16")];
            tensor<bool, []> mh_w_85_transpose_x_0 = const()[name = tensor<string, []>("mh_w_85_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_85_transpose_y_0 = const()[name = tensor<string, []>("mh_w_85_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_3248_cast_fp16, y = var_3250_cast_fp16)[name = tensor<string, []>("mh_w_85_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_87_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_3258_cast_fp16 = softmax(axis = var_3172, x = mh_w_87_cast_fp16)[name = tensor<string, []>("op_3258_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = tensor<string, []>("op_3259"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3260_cast_fp16 = reshape(shape = var_3259, x = value_57_cast_fp16)[name = tensor<string, []>("op_3260_cast_fp16")];
            tensor<bool, []> attn_57_transpose_x_0 = const()[name = tensor<string, []>("attn_57_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_57_transpose_y_0 = const()[name = tensor<string, []>("attn_57_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3260_cast_fp16, y = var_3258_cast_fp16)[name = tensor<string, []>("attn_57_cast_fp16")];
            tensor<int32, [4]> var_3263 = const()[name = tensor<string, []>("op_3263"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_141_cast_fp16 = reshape(shape = var_3263, x = attn_57_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<int32, [2]> var_3267 = const()[name = tensor<string, []>("op_3267"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3269 = const()[name = tensor<string, []>("op_3269"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_175_pad_type_0 = const()[name = tensor<string, []>("obj_175_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_175_pad_0 = const()[name = tensor<string, []>("obj_175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422792128))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423611392))), name = tensor<string, []>("layers_14_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423611520)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_175_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = var_3269, groups = var_3179, pad = obj_175_pad_0, pad_type = obj_175_pad_type_0, strides = var_3267, weight = layers_14_self_attn_o_proj_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor<string, []>("obj_175_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_175_cast_fp16)[name = tensor<string, []>("inputs_87_cast_fp16")];
            tensor<int32, [1]> var_3279 = const()[name = tensor<string, []>("op_3279"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_87_cast_fp16 = reduce_mean(axes = var_3279, keep_dims = var_3180, x = inputs_87_cast_fp16)[name = tensor<string, []>("channels_mean_87_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_87_cast_fp16 = sub(x = inputs_87_cast_fp16, y = channels_mean_87_cast_fp16)[name = tensor<string, []>("zero_mean_87_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = zero_mean_87_cast_fp16)[name = tensor<string, []>("zero_mean_sq_87_cast_fp16")];
            tensor<int32, [1]> var_3283 = const()[name = tensor<string, []>("op_3283"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3284_cast_fp16 = reduce_mean(axes = var_3283, keep_dims = var_3180, x = zero_mean_sq_87_cast_fp16)[name = tensor<string, []>("op_3284_cast_fp16")];
            tensor<fp16, []> var_3285_to_fp16 = const()[name = tensor<string, []>("op_3285_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3286_cast_fp16 = add(x = var_3284_cast_fp16, y = var_3285_to_fp16)[name = tensor<string, []>("op_3286_cast_fp16")];
            tensor<fp32, []> denom_87_epsilon_0 = const()[name = tensor<string, []>("denom_87_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_87_cast_fp16 = rsqrt(epsilon = denom_87_epsilon_0, x = var_3286_cast_fp16)[name = tensor<string, []>("denom_87_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = denom_87_cast_fp16)[name = tensor<string, []>("out_87_cast_fp16")];
            tensor<fp16, [1280]> obj_177_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_177_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423614144)))];
            tensor<fp16, [1280]> obj_177_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_177_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423616768)))];
            tensor<fp16, []> obj_177_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_177_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor<string, []>("obj_177_cast_fp16")];
            tensor<int32, [2]> var_3301 = const()[name = tensor<string, []>("op_3301"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3303 = const()[name = tensor<string, []>("op_3303"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_59_pad_type_0 = const()[name = tensor<string, []>("query_59_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_59_pad_0 = const()[name = tensor<string, []>("query_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423619392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424848256))), name = tensor<string, []>("layers_14_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_14_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424848448)))];
            tensor<fp16, [1, 1280, 1, 1]> query_59_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_bias_to_fp16, dilations = var_3303, groups = var_3179, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = var_3301, weight = layers_14_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_177_cast_fp16)[name = tensor<string, []>("query_59_cast_fp16")];
            tensor<int32, [2]> var_3307 = const()[name = tensor<string, []>("op_3307"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3309 = const()[name = tensor<string, []>("op_3309"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_59_pad_type_0 = const()[name = tensor<string, []>("key_59_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_59_pad_0 = const()[name = tensor<string, []>("key_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424851072))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425670336))), name = tensor<string, []>("layers_14_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_59_cast_fp16 = conv(dilations = var_3309, groups = var_3179, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = var_3307, weight = layers_14_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_59_cast_fp16")];
            tensor<int32, [2]> var_3314 = const()[name = tensor<string, []>("op_3314"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3316 = const()[name = tensor<string, []>("op_3316"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_59_pad_type_0 = const()[name = tensor<string, []>("value_59_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_59_pad_0 = const()[name = tensor<string, []>("value_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425670464))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426489728))), name = tensor<string, []>("layers_14_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_14_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426489856)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_59_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_bias_to_fp16, dilations = var_3316, groups = var_3179, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = var_3314, weight = layers_14_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_59_cast_fp16")];
            tensor<int32, [4]> var_3320 = const()[name = tensor<string, []>("op_3320"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3321_cast_fp16 = reshape(shape = var_3320, x = query_59_cast_fp16)[name = tensor<string, []>("op_3321_cast_fp16")];
            tensor<fp16, []> var_3322_to_fp16 = const()[name = tensor<string, []>("op_3322_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3323_cast_fp16 = mul(x = var_3321_cast_fp16, y = var_3322_to_fp16)[name = tensor<string, []>("op_3323_cast_fp16")];
            tensor<int32, [4]> var_3324 = const()[name = tensor<string, []>("op_3324"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3325_cast_fp16 = reshape(shape = var_3324, x = key_59_cast_fp16)[name = tensor<string, []>("op_3325_cast_fp16")];
            tensor<bool, []> mh_w_89_transpose_x_0 = const()[name = tensor<string, []>("mh_w_89_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_89_transpose_y_0 = const()[name = tensor<string, []>("mh_w_89_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_3323_cast_fp16, y = var_3325_cast_fp16)[name = tensor<string, []>("mh_w_89_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_3328_cast_fp16 = softmax(axis = var_3172, x = mh_w_89_cast_fp16)[name = tensor<string, []>("op_3328_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = tensor<string, []>("op_3329"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3330_cast_fp16 = reshape(shape = var_3329, x = value_59_cast_fp16)[name = tensor<string, []>("op_3330_cast_fp16")];
            tensor<bool, []> attn_59_transpose_x_0 = const()[name = tensor<string, []>("attn_59_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_59_transpose_y_0 = const()[name = tensor<string, []>("attn_59_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3330_cast_fp16, y = var_3328_cast_fp16)[name = tensor<string, []>("attn_59_cast_fp16")];
            tensor<int32, [4]> var_3333 = const()[name = tensor<string, []>("op_3333"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_143_cast_fp16 = reshape(shape = var_3333, x = attn_59_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<int32, [2]> var_3337 = const()[name = tensor<string, []>("op_3337"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3339 = const()[name = tensor<string, []>("op_3339"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_179_pad_type_0 = const()[name = tensor<string, []>("obj_179_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_179_pad_0 = const()[name = tensor<string, []>("obj_179_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_14_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426492480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427721344))), name = tensor<string, []>("layers_14_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_14_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427721536)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_179_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_bias_to_fp16, dilations = var_3339, groups = var_3179, pad = obj_179_pad_0, pad_type = obj_179_pad_type_0, strides = var_3337, weight = layers_14_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_143_cast_fp16)[name = tensor<string, []>("obj_179_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_179_cast_fp16)[name = tensor<string, []>("inputs_89_cast_fp16")];
            tensor<int32, [1]> var_3345 = const()[name = tensor<string, []>("op_3345"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_89_cast_fp16 = reduce_mean(axes = var_3345, keep_dims = var_3180, x = inputs_89_cast_fp16)[name = tensor<string, []>("channels_mean_89_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_89_cast_fp16 = sub(x = inputs_89_cast_fp16, y = channels_mean_89_cast_fp16)[name = tensor<string, []>("zero_mean_89_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = zero_mean_89_cast_fp16)[name = tensor<string, []>("zero_mean_sq_89_cast_fp16")];
            tensor<int32, [1]> var_3349 = const()[name = tensor<string, []>("op_3349"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3350_cast_fp16 = reduce_mean(axes = var_3349, keep_dims = var_3180, x = zero_mean_sq_89_cast_fp16)[name = tensor<string, []>("op_3350_cast_fp16")];
            tensor<fp16, []> var_3351_to_fp16 = const()[name = tensor<string, []>("op_3351_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3352_cast_fp16 = add(x = var_3350_cast_fp16, y = var_3351_to_fp16)[name = tensor<string, []>("op_3352_cast_fp16")];
            tensor<fp32, []> denom_89_epsilon_0 = const()[name = tensor<string, []>("denom_89_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_89_cast_fp16 = rsqrt(epsilon = denom_89_epsilon_0, x = var_3352_cast_fp16)[name = tensor<string, []>("denom_89_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = denom_89_cast_fp16)[name = tensor<string, []>("out_89_cast_fp16")];
            tensor<fp16, [1280]> input_145_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_145_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427724160)))];
            tensor<fp16, [1280]> input_145_beta_0_to_fp16 = const()[name = tensor<string, []>("input_145_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427726784)))];
            tensor<fp16, []> input_145_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_145_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_145_cast_fp16 = batch_norm(beta = input_145_beta_0_to_fp16, epsilon = input_145_epsilon_0_to_fp16, gamma = input_145_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<int32, [2]> var_3363 = const()[name = tensor<string, []>("op_3363"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3365 = const()[name = tensor<string, []>("op_3365"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_147_pad_type_0 = const()[name = tensor<string, []>("input_147_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_147_pad_0 = const()[name = tensor<string, []>("input_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_14_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427729408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432644672))), name = tensor<string, []>("layers_14_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_14_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432644864)))];
            tensor<fp16, [1, 5120, 1, 1]> input_147_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = var_3365, groups = var_3179, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = var_3363, weight = layers_14_fc1_weight_to_fp16_palettized, x = input_145_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<string, []> input_149_mode_0 = const()[name = tensor<string, []>("input_149_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = input_147_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<int32, [2]> var_3371 = const()[name = tensor<string, []>("op_3371"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3373 = const()[name = tensor<string, []>("op_3373"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_31_pad_type_0 = const()[name = tensor<string, []>("hidden_states_31_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_31_pad_0 = const()[name = tensor<string, []>("hidden_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_14_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432655168)))];
            tensor<fp16, [1280]> layers_14_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445762432)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_31_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = var_3373, groups = var_3179, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = var_3371, weight = layers_14_fc2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("hidden_states_31_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor<string, []>("inputs_91_cast_fp16")];
            tensor<int32, []> var_3386 = const()[name = tensor<string, []>("op_3386"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_3393 = const()[name = tensor<string, []>("op_3393"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_3394 = const()[name = tensor<string, []>("op_3394"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_3406 = const()[name = tensor<string, []>("op_3406"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_91_cast_fp16 = reduce_mean(axes = var_3406, keep_dims = var_3394, x = inputs_91_cast_fp16)[name = tensor<string, []>("channels_mean_91_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_91_cast_fp16 = sub(x = inputs_91_cast_fp16, y = channels_mean_91_cast_fp16)[name = tensor<string, []>("zero_mean_91_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = zero_mean_91_cast_fp16)[name = tensor<string, []>("zero_mean_sq_91_cast_fp16")];
            tensor<int32, [1]> var_3410 = const()[name = tensor<string, []>("op_3410"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3411_cast_fp16 = reduce_mean(axes = var_3410, keep_dims = var_3394, x = zero_mean_sq_91_cast_fp16)[name = tensor<string, []>("op_3411_cast_fp16")];
            tensor<fp16, []> var_3412_to_fp16 = const()[name = tensor<string, []>("op_3412_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3413_cast_fp16 = add(x = var_3411_cast_fp16, y = var_3412_to_fp16)[name = tensor<string, []>("op_3413_cast_fp16")];
            tensor<fp32, []> denom_91_epsilon_0 = const()[name = tensor<string, []>("denom_91_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_91_cast_fp16 = rsqrt(epsilon = denom_91_epsilon_0, x = var_3413_cast_fp16)[name = tensor<string, []>("denom_91_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = denom_91_cast_fp16)[name = tensor<string, []>("out_91_cast_fp16")];
            tensor<fp16, [1280]> obj_181_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_181_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445765056)))];
            tensor<fp16, [1280]> obj_181_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_181_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445767680)))];
            tensor<fp16, []> obj_181_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_181_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_181_cast_fp16 = batch_norm(beta = obj_181_beta_0_to_fp16, epsilon = obj_181_epsilon_0_to_fp16, gamma = obj_181_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor<string, []>("obj_181_cast_fp16")];
            tensor<int32, [2]> var_3428 = const()[name = tensor<string, []>("op_3428"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3430 = const()[name = tensor<string, []>("op_3430"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_61_pad_type_0 = const()[name = tensor<string, []>("query_61_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_61_pad_0 = const()[name = tensor<string, []>("query_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445770304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446999168))), name = tensor<string, []>("layers_15_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446999360)))];
            tensor<fp16, [1, 1280, 1, 1]> query_61_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = var_3430, groups = var_3393, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = var_3428, weight = layers_15_self_attn_q_proj_weight_to_fp16_palettized, x = obj_181_cast_fp16)[name = tensor<string, []>("query_61_cast_fp16")];
            tensor<int32, [2]> var_3434 = const()[name = tensor<string, []>("op_3434"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3436 = const()[name = tensor<string, []>("op_3436"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_31_pad_type_0 = const()[name = tensor<string, []>("current_key_31_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_31_pad_0 = const()[name = tensor<string, []>("current_key_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447001984))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447821248))), name = tensor<string, []>("layers_15_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_31_cast_fp16 = conv(dilations = var_3436, groups = var_3393, pad = current_key_31_pad_0, pad_type = current_key_31_pad_type_0, strides = var_3434, weight = layers_15_self_attn_k_proj_weight_to_fp16_palettized, x = obj_181_cast_fp16)[name = tensor<string, []>("current_key_31_cast_fp16")];
            tensor<int32, [2]> var_3441 = const()[name = tensor<string, []>("op_3441"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3443 = const()[name = tensor<string, []>("op_3443"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_31_pad_type_0 = const()[name = tensor<string, []>("current_value_31_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_31_pad_0 = const()[name = tensor<string, []>("current_value_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447821376))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448640640))), name = tensor<string, []>("layers_15_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448640768)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = var_3443, groups = var_3393, pad = current_value_31_pad_0, pad_type = current_value_31_pad_type_0, strides = var_3441, weight = layers_15_self_attn_v_proj_weight_to_fp16_palettized, x = obj_181_cast_fp16)[name = tensor<string, []>("current_value_31_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3450_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3450_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3452_cast_fp16 = mul(x = var_103_cast_fp16_15, y = var_241_cast_fp16)[name = tensor<string, []>("op_3452_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_61_cast_fp16 = add(x = var_3450_cast_fp16, y = var_3452_cast_fp16)[name = tensor<string, []>("key_61_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3454_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3454_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3456_cast_fp16 = mul(x = var_138_cast_fp16_15, y = var_241_cast_fp16)[name = tensor<string, []>("op_3456_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_61_cast_fp16 = add(x = var_3454_cast_fp16, y = var_3456_cast_fp16)[name = tensor<string, []>("value_61_cast_fp16")];
            tensor<int32, [4]> var_3459 = const()[name = tensor<string, []>("op_3459"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3460_cast_fp16 = reshape(shape = var_3459, x = query_61_cast_fp16)[name = tensor<string, []>("op_3460_cast_fp16")];
            tensor<fp16, []> var_3461_to_fp16 = const()[name = tensor<string, []>("op_3461_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3462_cast_fp16 = mul(x = var_3460_cast_fp16, y = var_3461_to_fp16)[name = tensor<string, []>("op_3462_cast_fp16")];
            tensor<int32, [4]> var_3463 = const()[name = tensor<string, []>("op_3463"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3464_cast_fp16 = reshape(shape = var_3463, x = key_61_cast_fp16)[name = tensor<string, []>("op_3464_cast_fp16")];
            tensor<bool, []> mh_w_91_transpose_x_0 = const()[name = tensor<string, []>("mh_w_91_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_91_transpose_y_0 = const()[name = tensor<string, []>("mh_w_91_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_91_cast_fp16 = matmul(transpose_x = mh_w_91_transpose_x_0, transpose_y = mh_w_91_transpose_y_0, x = var_3462_cast_fp16, y = var_3464_cast_fp16)[name = tensor<string, []>("mh_w_91_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_93_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_3472_cast_fp16 = softmax(axis = var_3386, x = mh_w_93_cast_fp16)[name = tensor<string, []>("op_3472_cast_fp16")];
            tensor<int32, [4]> var_3473 = const()[name = tensor<string, []>("op_3473"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3474_cast_fp16 = reshape(shape = var_3473, x = value_61_cast_fp16)[name = tensor<string, []>("op_3474_cast_fp16")];
            tensor<bool, []> attn_61_transpose_x_0 = const()[name = tensor<string, []>("attn_61_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_61_transpose_y_0 = const()[name = tensor<string, []>("attn_61_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3474_cast_fp16, y = var_3472_cast_fp16)[name = tensor<string, []>("attn_61_cast_fp16")];
            tensor<int32, [4]> var_3477 = const()[name = tensor<string, []>("op_3477"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_151_cast_fp16 = reshape(shape = var_3477, x = attn_61_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<int32, [2]> var_3481 = const()[name = tensor<string, []>("op_3481"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3483 = const()[name = tensor<string, []>("op_3483"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_187_pad_type_0 = const()[name = tensor<string, []>("obj_187_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_187_pad_0 = const()[name = tensor<string, []>("obj_187_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448643392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449872256))), name = tensor<string, []>("layers_15_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449872448)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_187_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = var_3483, groups = var_3393, pad = obj_187_pad_0, pad_type = obj_187_pad_type_0, strides = var_3481, weight = layers_15_self_attn_o_proj_weight_to_fp16_palettized, x = input_151_cast_fp16)[name = tensor<string, []>("obj_187_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_187_cast_fp16)[name = tensor<string, []>("inputs_93_cast_fp16")];
            tensor<int32, [1]> var_3493 = const()[name = tensor<string, []>("op_3493"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_93_cast_fp16 = reduce_mean(axes = var_3493, keep_dims = var_3394, x = inputs_93_cast_fp16)[name = tensor<string, []>("channels_mean_93_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_93_cast_fp16 = sub(x = inputs_93_cast_fp16, y = channels_mean_93_cast_fp16)[name = tensor<string, []>("zero_mean_93_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = zero_mean_93_cast_fp16)[name = tensor<string, []>("zero_mean_sq_93_cast_fp16")];
            tensor<int32, [1]> var_3497 = const()[name = tensor<string, []>("op_3497"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3498_cast_fp16 = reduce_mean(axes = var_3497, keep_dims = var_3394, x = zero_mean_sq_93_cast_fp16)[name = tensor<string, []>("op_3498_cast_fp16")];
            tensor<fp16, []> var_3499_to_fp16 = const()[name = tensor<string, []>("op_3499_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3500_cast_fp16 = add(x = var_3498_cast_fp16, y = var_3499_to_fp16)[name = tensor<string, []>("op_3500_cast_fp16")];
            tensor<fp32, []> denom_93_epsilon_0 = const()[name = tensor<string, []>("denom_93_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_93_cast_fp16 = rsqrt(epsilon = denom_93_epsilon_0, x = var_3500_cast_fp16)[name = tensor<string, []>("denom_93_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = denom_93_cast_fp16)[name = tensor<string, []>("out_93_cast_fp16")];
            tensor<fp16, [1280]> obj_189_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_189_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449875072)))];
            tensor<fp16, [1280]> obj_189_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_189_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449877696)))];
            tensor<fp16, []> obj_189_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_189_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_189_cast_fp16 = batch_norm(beta = obj_189_beta_0_to_fp16, epsilon = obj_189_epsilon_0_to_fp16, gamma = obj_189_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor<string, []>("obj_189_cast_fp16")];
            tensor<int32, [2]> var_3515 = const()[name = tensor<string, []>("op_3515"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3517 = const()[name = tensor<string, []>("op_3517"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_63_pad_type_0 = const()[name = tensor<string, []>("query_63_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_63_pad_0 = const()[name = tensor<string, []>("query_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449880320))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451109184))), name = tensor<string, []>("layers_15_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_15_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451109376)))];
            tensor<fp16, [1, 1280, 1, 1]> query_63_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_bias_to_fp16, dilations = var_3517, groups = var_3393, pad = query_63_pad_0, pad_type = query_63_pad_type_0, strides = var_3515, weight = layers_15_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_189_cast_fp16)[name = tensor<string, []>("query_63_cast_fp16")];
            tensor<int32, [2]> var_3521 = const()[name = tensor<string, []>("op_3521"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3523 = const()[name = tensor<string, []>("op_3523"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_63_pad_type_0 = const()[name = tensor<string, []>("key_63_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_63_pad_0 = const()[name = tensor<string, []>("key_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451112000))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452340864))), name = tensor<string, []>("layers_15_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_63_cast_fp16 = conv(dilations = var_3523, groups = var_3393, pad = key_63_pad_0, pad_type = key_63_pad_type_0, strides = var_3521, weight = layers_15_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_63_cast_fp16")];
            tensor<int32, [2]> var_3528 = const()[name = tensor<string, []>("op_3528"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3530 = const()[name = tensor<string, []>("op_3530"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_63_pad_type_0 = const()[name = tensor<string, []>("value_63_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_63_pad_0 = const()[name = tensor<string, []>("value_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452341056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453569920))), name = tensor<string, []>("layers_15_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_15_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453570112)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_63_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_bias_to_fp16, dilations = var_3530, groups = var_3393, pad = value_63_pad_0, pad_type = value_63_pad_type_0, strides = var_3528, weight = layers_15_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_63_cast_fp16")];
            tensor<int32, [4]> var_3534 = const()[name = tensor<string, []>("op_3534"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3535_cast_fp16 = reshape(shape = var_3534, x = query_63_cast_fp16)[name = tensor<string, []>("op_3535_cast_fp16")];
            tensor<fp16, []> var_3536_to_fp16 = const()[name = tensor<string, []>("op_3536_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3537_cast_fp16 = mul(x = var_3535_cast_fp16, y = var_3536_to_fp16)[name = tensor<string, []>("op_3537_cast_fp16")];
            tensor<int32, [4]> var_3538 = const()[name = tensor<string, []>("op_3538"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3539_cast_fp16 = reshape(shape = var_3538, x = key_63_cast_fp16)[name = tensor<string, []>("op_3539_cast_fp16")];
            tensor<bool, []> mh_w_95_transpose_x_0 = const()[name = tensor<string, []>("mh_w_95_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_95_transpose_y_0 = const()[name = tensor<string, []>("mh_w_95_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_95_cast_fp16 = matmul(transpose_x = mh_w_95_transpose_x_0, transpose_y = mh_w_95_transpose_y_0, x = var_3537_cast_fp16, y = var_3539_cast_fp16)[name = tensor<string, []>("mh_w_95_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_3542_cast_fp16 = softmax(axis = var_3386, x = mh_w_95_cast_fp16)[name = tensor<string, []>("op_3542_cast_fp16")];
            tensor<int32, [4]> var_3543 = const()[name = tensor<string, []>("op_3543"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3544_cast_fp16 = reshape(shape = var_3543, x = value_63_cast_fp16)[name = tensor<string, []>("op_3544_cast_fp16")];
            tensor<bool, []> attn_63_transpose_x_0 = const()[name = tensor<string, []>("attn_63_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_63_transpose_y_0 = const()[name = tensor<string, []>("attn_63_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_63_cast_fp16 = matmul(transpose_x = attn_63_transpose_x_0, transpose_y = attn_63_transpose_y_0, x = var_3544_cast_fp16, y = var_3542_cast_fp16)[name = tensor<string, []>("attn_63_cast_fp16")];
            tensor<int32, [4]> var_3547 = const()[name = tensor<string, []>("op_3547"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_153_cast_fp16 = reshape(shape = var_3547, x = attn_63_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<int32, [2]> var_3551 = const()[name = tensor<string, []>("op_3551"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3553 = const()[name = tensor<string, []>("op_3553"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_191_pad_type_0 = const()[name = tensor<string, []>("obj_191_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_191_pad_0 = const()[name = tensor<string, []>("obj_191_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_15_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453572736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454801600))), name = tensor<string, []>("layers_15_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_15_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454801792)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_191_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_bias_to_fp16, dilations = var_3553, groups = var_3393, pad = obj_191_pad_0, pad_type = obj_191_pad_type_0, strides = var_3551, weight = layers_15_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_153_cast_fp16)[name = tensor<string, []>("obj_191_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_191_cast_fp16)[name = tensor<string, []>("inputs_95_cast_fp16")];
            tensor<int32, [1]> var_3559 = const()[name = tensor<string, []>("op_3559"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_95_cast_fp16 = reduce_mean(axes = var_3559, keep_dims = var_3394, x = inputs_95_cast_fp16)[name = tensor<string, []>("channels_mean_95_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_95_cast_fp16 = sub(x = inputs_95_cast_fp16, y = channels_mean_95_cast_fp16)[name = tensor<string, []>("zero_mean_95_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = zero_mean_95_cast_fp16)[name = tensor<string, []>("zero_mean_sq_95_cast_fp16")];
            tensor<int32, [1]> var_3563 = const()[name = tensor<string, []>("op_3563"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3564_cast_fp16 = reduce_mean(axes = var_3563, keep_dims = var_3394, x = zero_mean_sq_95_cast_fp16)[name = tensor<string, []>("op_3564_cast_fp16")];
            tensor<fp16, []> var_3565_to_fp16 = const()[name = tensor<string, []>("op_3565_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3566_cast_fp16 = add(x = var_3564_cast_fp16, y = var_3565_to_fp16)[name = tensor<string, []>("op_3566_cast_fp16")];
            tensor<fp32, []> denom_95_epsilon_0 = const()[name = tensor<string, []>("denom_95_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_95_cast_fp16 = rsqrt(epsilon = denom_95_epsilon_0, x = var_3566_cast_fp16)[name = tensor<string, []>("denom_95_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = denom_95_cast_fp16)[name = tensor<string, []>("out_95_cast_fp16")];
            tensor<fp16, [1280]> input_155_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_155_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454804416)))];
            tensor<fp16, [1280]> input_155_beta_0_to_fp16 = const()[name = tensor<string, []>("input_155_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454807040)))];
            tensor<fp16, []> input_155_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_155_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<int32, [2]> var_3577 = const()[name = tensor<string, []>("op_3577"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3579 = const()[name = tensor<string, []>("op_3579"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_157_pad_type_0 = const()[name = tensor<string, []>("input_157_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_157_pad_0 = const()[name = tensor<string, []>("input_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_15_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454809664))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461363328))), name = tensor<string, []>("layers_15_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_15_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461363904)))];
            tensor<fp16, [1, 5120, 1, 1]> input_157_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = var_3579, groups = var_3393, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = var_3577, weight = layers_15_fc1_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<string, []> input_159_mode_0 = const()[name = tensor<string, []>("input_159_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<int32, [2]> var_3585 = const()[name = tensor<string, []>("op_3585"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3587 = const()[name = tensor<string, []>("op_3587"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_33_pad_type_0 = const()[name = tensor<string, []>("hidden_states_33_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = tensor<string, []>("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_15_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461374208)))];
            tensor<fp16, [1280]> layers_15_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474481472)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_33_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = var_3587, groups = var_3393, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = var_3585, weight = layers_15_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor<string, []>("inputs_97_cast_fp16")];
            tensor<int32, []> var_3600 = const()[name = tensor<string, []>("op_3600"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_3607 = const()[name = tensor<string, []>("op_3607"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_3608 = const()[name = tensor<string, []>("op_3608"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_3620 = const()[name = tensor<string, []>("op_3620"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_97_cast_fp16 = reduce_mean(axes = var_3620, keep_dims = var_3608, x = inputs_97_cast_fp16)[name = tensor<string, []>("channels_mean_97_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_97_cast_fp16 = sub(x = inputs_97_cast_fp16, y = channels_mean_97_cast_fp16)[name = tensor<string, []>("zero_mean_97_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = zero_mean_97_cast_fp16)[name = tensor<string, []>("zero_mean_sq_97_cast_fp16")];
            tensor<int32, [1]> var_3624 = const()[name = tensor<string, []>("op_3624"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3625_cast_fp16 = reduce_mean(axes = var_3624, keep_dims = var_3608, x = zero_mean_sq_97_cast_fp16)[name = tensor<string, []>("op_3625_cast_fp16")];
            tensor<fp16, []> var_3626_to_fp16 = const()[name = tensor<string, []>("op_3626_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3627_cast_fp16 = add(x = var_3625_cast_fp16, y = var_3626_to_fp16)[name = tensor<string, []>("op_3627_cast_fp16")];
            tensor<fp32, []> denom_97_epsilon_0 = const()[name = tensor<string, []>("denom_97_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_97_cast_fp16 = rsqrt(epsilon = denom_97_epsilon_0, x = var_3627_cast_fp16)[name = tensor<string, []>("denom_97_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = denom_97_cast_fp16)[name = tensor<string, []>("out_97_cast_fp16")];
            tensor<fp16, [1280]> obj_193_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_193_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474484096)))];
            tensor<fp16, [1280]> obj_193_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_193_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474486720)))];
            tensor<fp16, []> obj_193_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_193_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_193_cast_fp16 = batch_norm(beta = obj_193_beta_0_to_fp16, epsilon = obj_193_epsilon_0_to_fp16, gamma = obj_193_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor<string, []>("obj_193_cast_fp16")];
            tensor<int32, [2]> var_3642 = const()[name = tensor<string, []>("op_3642"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3644 = const()[name = tensor<string, []>("op_3644"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_65_pad_type_0 = const()[name = tensor<string, []>("query_65_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_65_pad_0 = const()[name = tensor<string, []>("query_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474489344))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475308608))), name = tensor<string, []>("layers_16_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475308736)))];
            tensor<fp16, [1, 1280, 1, 1]> query_65_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = var_3644, groups = var_3607, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = var_3642, weight = layers_16_self_attn_q_proj_weight_to_fp16_palettized, x = obj_193_cast_fp16)[name = tensor<string, []>("query_65_cast_fp16")];
            tensor<int32, [2]> var_3648 = const()[name = tensor<string, []>("op_3648"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3650 = const()[name = tensor<string, []>("op_3650"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_33_pad_type_0 = const()[name = tensor<string, []>("current_key_33_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_33_pad_0 = const()[name = tensor<string, []>("current_key_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475311360))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476540224))), name = tensor<string, []>("layers_16_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_33_cast_fp16 = conv(dilations = var_3650, groups = var_3607, pad = current_key_33_pad_0, pad_type = current_key_33_pad_type_0, strides = var_3648, weight = layers_16_self_attn_k_proj_weight_to_fp16_palettized, x = obj_193_cast_fp16)[name = tensor<string, []>("current_key_33_cast_fp16")];
            tensor<int32, [2]> var_3655 = const()[name = tensor<string, []>("op_3655"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3657 = const()[name = tensor<string, []>("op_3657"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_33_pad_type_0 = const()[name = tensor<string, []>("current_value_33_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_33_pad_0 = const()[name = tensor<string, []>("current_value_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476540416))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477359680))), name = tensor<string, []>("layers_16_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477359808)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = var_3657, groups = var_3607, pad = current_value_33_pad_0, pad_type = current_value_33_pad_type_0, strides = var_3655, weight = layers_16_self_attn_v_proj_weight_to_fp16_palettized, x = obj_193_cast_fp16)[name = tensor<string, []>("current_value_33_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3664_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3664_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3666_cast_fp16 = mul(x = var_103_cast_fp16_16, y = var_241_cast_fp16)[name = tensor<string, []>("op_3666_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_65_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3666_cast_fp16)[name = tensor<string, []>("key_65_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3668_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3668_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3670_cast_fp16 = mul(x = var_138_cast_fp16_16, y = var_241_cast_fp16)[name = tensor<string, []>("op_3670_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_65_cast_fp16 = add(x = var_3668_cast_fp16, y = var_3670_cast_fp16)[name = tensor<string, []>("value_65_cast_fp16")];
            tensor<int32, [4]> var_3673 = const()[name = tensor<string, []>("op_3673"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3674_cast_fp16 = reshape(shape = var_3673, x = query_65_cast_fp16)[name = tensor<string, []>("op_3674_cast_fp16")];
            tensor<fp16, []> var_3675_to_fp16 = const()[name = tensor<string, []>("op_3675_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3676_cast_fp16 = mul(x = var_3674_cast_fp16, y = var_3675_to_fp16)[name = tensor<string, []>("op_3676_cast_fp16")];
            tensor<int32, [4]> var_3677 = const()[name = tensor<string, []>("op_3677"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3678_cast_fp16 = reshape(shape = var_3677, x = key_65_cast_fp16)[name = tensor<string, []>("op_3678_cast_fp16")];
            tensor<bool, []> mh_w_97_transpose_x_0 = const()[name = tensor<string, []>("mh_w_97_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_97_transpose_y_0 = const()[name = tensor<string, []>("mh_w_97_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_97_cast_fp16 = matmul(transpose_x = mh_w_97_transpose_x_0, transpose_y = mh_w_97_transpose_y_0, x = var_3676_cast_fp16, y = var_3678_cast_fp16)[name = tensor<string, []>("mh_w_97_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_99_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_3686_cast_fp16 = softmax(axis = var_3600, x = mh_w_99_cast_fp16)[name = tensor<string, []>("op_3686_cast_fp16")];
            tensor<int32, [4]> var_3687 = const()[name = tensor<string, []>("op_3687"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3688_cast_fp16 = reshape(shape = var_3687, x = value_65_cast_fp16)[name = tensor<string, []>("op_3688_cast_fp16")];
            tensor<bool, []> attn_65_transpose_x_0 = const()[name = tensor<string, []>("attn_65_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_65_transpose_y_0 = const()[name = tensor<string, []>("attn_65_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_0, transpose_y = attn_65_transpose_y_0, x = var_3688_cast_fp16, y = var_3686_cast_fp16)[name = tensor<string, []>("attn_65_cast_fp16")];
            tensor<int32, [4]> var_3691 = const()[name = tensor<string, []>("op_3691"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_161_cast_fp16 = reshape(shape = var_3691, x = attn_65_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<int32, [2]> var_3695 = const()[name = tensor<string, []>("op_3695"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3697 = const()[name = tensor<string, []>("op_3697"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_199_pad_type_0 = const()[name = tensor<string, []>("obj_199_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_199_pad_0 = const()[name = tensor<string, []>("obj_199_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477362432))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478591296))), name = tensor<string, []>("layers_16_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478591488)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_199_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = var_3697, groups = var_3607, pad = obj_199_pad_0, pad_type = obj_199_pad_type_0, strides = var_3695, weight = layers_16_self_attn_o_proj_weight_to_fp16_palettized, x = input_161_cast_fp16)[name = tensor<string, []>("obj_199_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_199_cast_fp16)[name = tensor<string, []>("inputs_99_cast_fp16")];
            tensor<int32, [1]> var_3707 = const()[name = tensor<string, []>("op_3707"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_99_cast_fp16 = reduce_mean(axes = var_3707, keep_dims = var_3608, x = inputs_99_cast_fp16)[name = tensor<string, []>("channels_mean_99_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_99_cast_fp16 = sub(x = inputs_99_cast_fp16, y = channels_mean_99_cast_fp16)[name = tensor<string, []>("zero_mean_99_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = zero_mean_99_cast_fp16)[name = tensor<string, []>("zero_mean_sq_99_cast_fp16")];
            tensor<int32, [1]> var_3711 = const()[name = tensor<string, []>("op_3711"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3712_cast_fp16 = reduce_mean(axes = var_3711, keep_dims = var_3608, x = zero_mean_sq_99_cast_fp16)[name = tensor<string, []>("op_3712_cast_fp16")];
            tensor<fp16, []> var_3713_to_fp16 = const()[name = tensor<string, []>("op_3713_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3714_cast_fp16 = add(x = var_3712_cast_fp16, y = var_3713_to_fp16)[name = tensor<string, []>("op_3714_cast_fp16")];
            tensor<fp32, []> denom_99_epsilon_0 = const()[name = tensor<string, []>("denom_99_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_99_cast_fp16 = rsqrt(epsilon = denom_99_epsilon_0, x = var_3714_cast_fp16)[name = tensor<string, []>("denom_99_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = denom_99_cast_fp16)[name = tensor<string, []>("out_99_cast_fp16")];
            tensor<fp16, [1280]> obj_201_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_201_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478594112)))];
            tensor<fp16, [1280]> obj_201_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_201_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478596736)))];
            tensor<fp16, []> obj_201_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_201_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_201_cast_fp16 = batch_norm(beta = obj_201_beta_0_to_fp16, epsilon = obj_201_epsilon_0_to_fp16, gamma = obj_201_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor<string, []>("obj_201_cast_fp16")];
            tensor<int32, [2]> var_3729 = const()[name = tensor<string, []>("op_3729"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3731 = const()[name = tensor<string, []>("op_3731"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_67_pad_type_0 = const()[name = tensor<string, []>("query_67_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_67_pad_0 = const()[name = tensor<string, []>("query_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478599360)))];
            tensor<fp16, [1280]> layers_16_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481876224)))];
            tensor<fp16, [1, 1280, 1, 1]> query_67_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_bias_to_fp16, dilations = var_3731, groups = var_3607, pad = query_67_pad_0, pad_type = query_67_pad_type_0, strides = var_3729, weight = layers_16_encoder_attn_q_proj_weight_to_fp16, x = obj_201_cast_fp16)[name = tensor<string, []>("query_67_cast_fp16")];
            tensor<int32, [2]> var_3735 = const()[name = tensor<string, []>("op_3735"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3737 = const()[name = tensor<string, []>("op_3737"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_67_pad_type_0 = const()[name = tensor<string, []>("key_67_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_67_pad_0 = const()[name = tensor<string, []>("key_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481878848))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483517312))), name = tensor<string, []>("layers_16_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_67_cast_fp16 = conv(dilations = var_3737, groups = var_3607, pad = key_67_pad_0, pad_type = key_67_pad_type_0, strides = var_3735, weight = layers_16_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_67_cast_fp16")];
            tensor<int32, [2]> var_3742 = const()[name = tensor<string, []>("op_3742"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3744 = const()[name = tensor<string, []>("op_3744"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_67_pad_type_0 = const()[name = tensor<string, []>("value_67_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_67_pad_0 = const()[name = tensor<string, []>("value_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483517888))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484746752))), name = tensor<string, []>("layers_16_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_16_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484746944)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_67_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_bias_to_fp16, dilations = var_3744, groups = var_3607, pad = value_67_pad_0, pad_type = value_67_pad_type_0, strides = var_3742, weight = layers_16_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_67_cast_fp16")];
            tensor<int32, [4]> var_3748 = const()[name = tensor<string, []>("op_3748"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3749_cast_fp16 = reshape(shape = var_3748, x = query_67_cast_fp16)[name = tensor<string, []>("op_3749_cast_fp16")];
            tensor<fp16, []> var_3750_to_fp16 = const()[name = tensor<string, []>("op_3750_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3751_cast_fp16 = mul(x = var_3749_cast_fp16, y = var_3750_to_fp16)[name = tensor<string, []>("op_3751_cast_fp16")];
            tensor<int32, [4]> var_3752 = const()[name = tensor<string, []>("op_3752"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3753_cast_fp16 = reshape(shape = var_3752, x = key_67_cast_fp16)[name = tensor<string, []>("op_3753_cast_fp16")];
            tensor<bool, []> mh_w_101_transpose_x_0 = const()[name = tensor<string, []>("mh_w_101_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_101_transpose_y_0 = const()[name = tensor<string, []>("mh_w_101_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_101_cast_fp16 = matmul(transpose_x = mh_w_101_transpose_x_0, transpose_y = mh_w_101_transpose_y_0, x = var_3751_cast_fp16, y = var_3753_cast_fp16)[name = tensor<string, []>("mh_w_101_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_3756_cast_fp16 = softmax(axis = var_3600, x = mh_w_101_cast_fp16)[name = tensor<string, []>("op_3756_cast_fp16")];
            tensor<int32, [4]> var_3757 = const()[name = tensor<string, []>("op_3757"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3758_cast_fp16 = reshape(shape = var_3757, x = value_67_cast_fp16)[name = tensor<string, []>("op_3758_cast_fp16")];
            tensor<bool, []> attn_67_transpose_x_0 = const()[name = tensor<string, []>("attn_67_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_67_transpose_y_0 = const()[name = tensor<string, []>("attn_67_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_67_cast_fp16 = matmul(transpose_x = attn_67_transpose_x_0, transpose_y = attn_67_transpose_y_0, x = var_3758_cast_fp16, y = var_3756_cast_fp16)[name = tensor<string, []>("attn_67_cast_fp16")];
            tensor<int32, [4]> var_3761 = const()[name = tensor<string, []>("op_3761"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_163_cast_fp16 = reshape(shape = var_3761, x = attn_67_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<int32, [2]> var_3765 = const()[name = tensor<string, []>("op_3765"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3767 = const()[name = tensor<string, []>("op_3767"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_203_pad_type_0 = const()[name = tensor<string, []>("obj_203_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_203_pad_0 = const()[name = tensor<string, []>("obj_203_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_16_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484749568)))];
            tensor<fp16, [1280]> layers_16_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488026432)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_203_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_bias_to_fp16, dilations = var_3767, groups = var_3607, pad = obj_203_pad_0, pad_type = obj_203_pad_type_0, strides = var_3765, weight = layers_16_encoder_attn_o_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("obj_203_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_203_cast_fp16)[name = tensor<string, []>("inputs_101_cast_fp16")];
            tensor<int32, [1]> var_3773 = const()[name = tensor<string, []>("op_3773"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_101_cast_fp16 = reduce_mean(axes = var_3773, keep_dims = var_3608, x = inputs_101_cast_fp16)[name = tensor<string, []>("channels_mean_101_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_101_cast_fp16 = sub(x = inputs_101_cast_fp16, y = channels_mean_101_cast_fp16)[name = tensor<string, []>("zero_mean_101_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = zero_mean_101_cast_fp16)[name = tensor<string, []>("zero_mean_sq_101_cast_fp16")];
            tensor<int32, [1]> var_3777 = const()[name = tensor<string, []>("op_3777"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3778_cast_fp16 = reduce_mean(axes = var_3777, keep_dims = var_3608, x = zero_mean_sq_101_cast_fp16)[name = tensor<string, []>("op_3778_cast_fp16")];
            tensor<fp16, []> var_3779_to_fp16 = const()[name = tensor<string, []>("op_3779_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3780_cast_fp16 = add(x = var_3778_cast_fp16, y = var_3779_to_fp16)[name = tensor<string, []>("op_3780_cast_fp16")];
            tensor<fp32, []> denom_101_epsilon_0 = const()[name = tensor<string, []>("denom_101_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_101_cast_fp16 = rsqrt(epsilon = denom_101_epsilon_0, x = var_3780_cast_fp16)[name = tensor<string, []>("denom_101_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = denom_101_cast_fp16)[name = tensor<string, []>("out_101_cast_fp16")];
            tensor<fp16, [1280]> input_165_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_165_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488029056)))];
            tensor<fp16, [1280]> input_165_beta_0_to_fp16 = const()[name = tensor<string, []>("input_165_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488031680)))];
            tensor<fp16, []> input_165_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_165_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_165_cast_fp16 = batch_norm(beta = input_165_beta_0_to_fp16, epsilon = input_165_epsilon_0_to_fp16, gamma = input_165_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<int32, [2]> var_3791 = const()[name = tensor<string, []>("op_3791"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3793 = const()[name = tensor<string, []>("op_3793"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_167_pad_type_0 = const()[name = tensor<string, []>("input_167_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_167_pad_0 = const()[name = tensor<string, []>("input_167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_16_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488034304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492949568))), name = tensor<string, []>("layers_16_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_16_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492949760)))];
            tensor<fp16, [1, 5120, 1, 1]> input_167_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = var_3793, groups = var_3607, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = var_3791, weight = layers_16_fc1_weight_to_fp16_palettized, x = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<string, []> input_169_mode_0 = const()[name = tensor<string, []>("input_169_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<int32, [2]> var_3799 = const()[name = tensor<string, []>("op_3799"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3801 = const()[name = tensor<string, []>("op_3801"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_35_pad_type_0 = const()[name = tensor<string, []>("hidden_states_35_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_35_pad_0 = const()[name = tensor<string, []>("hidden_states_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_16_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492960064)))];
            tensor<fp16, [1280]> layers_16_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506067328)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_35_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = var_3801, groups = var_3607, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = var_3799, weight = layers_16_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("hidden_states_35_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor<string, []>("inputs_103_cast_fp16")];
            tensor<int32, []> var_3814 = const()[name = tensor<string, []>("op_3814"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_3821 = const()[name = tensor<string, []>("op_3821"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_3822 = const()[name = tensor<string, []>("op_3822"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_3834 = const()[name = tensor<string, []>("op_3834"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_103_cast_fp16 = reduce_mean(axes = var_3834, keep_dims = var_3822, x = inputs_103_cast_fp16)[name = tensor<string, []>("channels_mean_103_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_103_cast_fp16 = sub(x = inputs_103_cast_fp16, y = channels_mean_103_cast_fp16)[name = tensor<string, []>("zero_mean_103_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = zero_mean_103_cast_fp16)[name = tensor<string, []>("zero_mean_sq_103_cast_fp16")];
            tensor<int32, [1]> var_3838 = const()[name = tensor<string, []>("op_3838"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3839_cast_fp16 = reduce_mean(axes = var_3838, keep_dims = var_3822, x = zero_mean_sq_103_cast_fp16)[name = tensor<string, []>("op_3839_cast_fp16")];
            tensor<fp16, []> var_3840_to_fp16 = const()[name = tensor<string, []>("op_3840_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3841_cast_fp16 = add(x = var_3839_cast_fp16, y = var_3840_to_fp16)[name = tensor<string, []>("op_3841_cast_fp16")];
            tensor<fp32, []> denom_103_epsilon_0 = const()[name = tensor<string, []>("denom_103_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_103_cast_fp16 = rsqrt(epsilon = denom_103_epsilon_0, x = var_3841_cast_fp16)[name = tensor<string, []>("denom_103_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = denom_103_cast_fp16)[name = tensor<string, []>("out_103_cast_fp16")];
            tensor<fp16, [1280]> obj_205_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_205_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506069952)))];
            tensor<fp16, [1280]> obj_205_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_205_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506072576)))];
            tensor<fp16, []> obj_205_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_205_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor<string, []>("obj_205_cast_fp16")];
            tensor<int32, [2]> var_3856 = const()[name = tensor<string, []>("op_3856"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3858 = const()[name = tensor<string, []>("op_3858"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_69_pad_type_0 = const()[name = tensor<string, []>("query_69_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_69_pad_0 = const()[name = tensor<string, []>("query_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506075200))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(507304064))), name = tensor<string, []>("layers_17_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(507304256)))];
            tensor<fp16, [1, 1280, 1, 1]> query_69_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = var_3858, groups = var_3821, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = var_3856, weight = layers_17_self_attn_q_proj_weight_to_fp16_palettized, x = obj_205_cast_fp16)[name = tensor<string, []>("query_69_cast_fp16")];
            tensor<int32, [2]> var_3862 = const()[name = tensor<string, []>("op_3862"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3864 = const()[name = tensor<string, []>("op_3864"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_35_pad_type_0 = const()[name = tensor<string, []>("current_key_35_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_35_pad_0 = const()[name = tensor<string, []>("current_key_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(507306880))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508126144))), name = tensor<string, []>("layers_17_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_35_cast_fp16 = conv(dilations = var_3864, groups = var_3821, pad = current_key_35_pad_0, pad_type = current_key_35_pad_type_0, strides = var_3862, weight = layers_17_self_attn_k_proj_weight_to_fp16_palettized, x = obj_205_cast_fp16)[name = tensor<string, []>("current_key_35_cast_fp16")];
            tensor<int32, [2]> var_3869 = const()[name = tensor<string, []>("op_3869"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3871 = const()[name = tensor<string, []>("op_3871"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_35_pad_type_0 = const()[name = tensor<string, []>("current_value_35_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_35_pad_0 = const()[name = tensor<string, []>("current_value_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508126272))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508945536))), name = tensor<string, []>("layers_17_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508945664)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = var_3871, groups = var_3821, pad = current_value_35_pad_0, pad_type = current_value_35_pad_type_0, strides = var_3869, weight = layers_17_self_attn_v_proj_weight_to_fp16_palettized, x = obj_205_cast_fp16)[name = tensor<string, []>("current_value_35_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3878_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3878_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3880_cast_fp16 = mul(x = var_103_cast_fp16_17, y = var_241_cast_fp16)[name = tensor<string, []>("op_3880_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_69_cast_fp16 = add(x = var_3878_cast_fp16, y = var_3880_cast_fp16)[name = tensor<string, []>("key_69_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3882_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_3882_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_3884_cast_fp16 = mul(x = var_138_cast_fp16_17, y = var_241_cast_fp16)[name = tensor<string, []>("op_3884_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_69_cast_fp16 = add(x = var_3882_cast_fp16, y = var_3884_cast_fp16)[name = tensor<string, []>("value_69_cast_fp16")];
            tensor<int32, [4]> var_3887 = const()[name = tensor<string, []>("op_3887"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3888_cast_fp16 = reshape(shape = var_3887, x = query_69_cast_fp16)[name = tensor<string, []>("op_3888_cast_fp16")];
            tensor<fp16, []> var_3889_to_fp16 = const()[name = tensor<string, []>("op_3889_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3890_cast_fp16 = mul(x = var_3888_cast_fp16, y = var_3889_to_fp16)[name = tensor<string, []>("op_3890_cast_fp16")];
            tensor<int32, [4]> var_3891 = const()[name = tensor<string, []>("op_3891"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3892_cast_fp16 = reshape(shape = var_3891, x = key_69_cast_fp16)[name = tensor<string, []>("op_3892_cast_fp16")];
            tensor<bool, []> mh_w_103_transpose_x_0 = const()[name = tensor<string, []>("mh_w_103_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_103_transpose_y_0 = const()[name = tensor<string, []>("mh_w_103_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_103_cast_fp16 = matmul(transpose_x = mh_w_103_transpose_x_0, transpose_y = mh_w_103_transpose_y_0, x = var_3890_cast_fp16, y = var_3892_cast_fp16)[name = tensor<string, []>("mh_w_103_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_105_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_3900_cast_fp16 = softmax(axis = var_3814, x = mh_w_105_cast_fp16)[name = tensor<string, []>("op_3900_cast_fp16")];
            tensor<int32, [4]> var_3901 = const()[name = tensor<string, []>("op_3901"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_3902_cast_fp16 = reshape(shape = var_3901, x = value_69_cast_fp16)[name = tensor<string, []>("op_3902_cast_fp16")];
            tensor<bool, []> attn_69_transpose_x_0 = const()[name = tensor<string, []>("attn_69_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_69_transpose_y_0 = const()[name = tensor<string, []>("attn_69_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_0, transpose_y = attn_69_transpose_y_0, x = var_3902_cast_fp16, y = var_3900_cast_fp16)[name = tensor<string, []>("attn_69_cast_fp16")];
            tensor<int32, [4]> var_3905 = const()[name = tensor<string, []>("op_3905"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_171_cast_fp16 = reshape(shape = var_3905, x = attn_69_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
            tensor<int32, [2]> var_3909 = const()[name = tensor<string, []>("op_3909"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3911 = const()[name = tensor<string, []>("op_3911"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_211_pad_type_0 = const()[name = tensor<string, []>("obj_211_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_211_pad_0 = const()[name = tensor<string, []>("obj_211_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508948288))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509767552))), name = tensor<string, []>("layers_17_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509767680)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_211_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = var_3911, groups = var_3821, pad = obj_211_pad_0, pad_type = obj_211_pad_type_0, strides = var_3909, weight = layers_17_self_attn_o_proj_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = tensor<string, []>("obj_211_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_211_cast_fp16)[name = tensor<string, []>("inputs_105_cast_fp16")];
            tensor<int32, [1]> var_3921 = const()[name = tensor<string, []>("op_3921"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_105_cast_fp16 = reduce_mean(axes = var_3921, keep_dims = var_3822, x = inputs_105_cast_fp16)[name = tensor<string, []>("channels_mean_105_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_105_cast_fp16 = sub(x = inputs_105_cast_fp16, y = channels_mean_105_cast_fp16)[name = tensor<string, []>("zero_mean_105_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = zero_mean_105_cast_fp16)[name = tensor<string, []>("zero_mean_sq_105_cast_fp16")];
            tensor<int32, [1]> var_3925 = const()[name = tensor<string, []>("op_3925"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3926_cast_fp16 = reduce_mean(axes = var_3925, keep_dims = var_3822, x = zero_mean_sq_105_cast_fp16)[name = tensor<string, []>("op_3926_cast_fp16")];
            tensor<fp16, []> var_3927_to_fp16 = const()[name = tensor<string, []>("op_3927_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3928_cast_fp16 = add(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = tensor<string, []>("op_3928_cast_fp16")];
            tensor<fp32, []> denom_105_epsilon_0 = const()[name = tensor<string, []>("denom_105_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_105_cast_fp16 = rsqrt(epsilon = denom_105_epsilon_0, x = var_3928_cast_fp16)[name = tensor<string, []>("denom_105_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = denom_105_cast_fp16)[name = tensor<string, []>("out_105_cast_fp16")];
            tensor<fp16, [1280]> obj_213_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_213_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509770304)))];
            tensor<fp16, [1280]> obj_213_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_213_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509772928)))];
            tensor<fp16, []> obj_213_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_213_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor<string, []>("obj_213_cast_fp16")];
            tensor<int32, [2]> var_3943 = const()[name = tensor<string, []>("op_3943"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3945 = const()[name = tensor<string, []>("op_3945"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_71_pad_type_0 = const()[name = tensor<string, []>("query_71_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_71_pad_0 = const()[name = tensor<string, []>("query_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509775552))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(511004416))), name = tensor<string, []>("layers_17_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_17_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(511004608)))];
            tensor<fp16, [1, 1280, 1, 1]> query_71_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_bias_to_fp16, dilations = var_3945, groups = var_3821, pad = query_71_pad_0, pad_type = query_71_pad_type_0, strides = var_3943, weight = layers_17_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_213_cast_fp16)[name = tensor<string, []>("query_71_cast_fp16")];
            tensor<int32, [2]> var_3949 = const()[name = tensor<string, []>("op_3949"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3951 = const()[name = tensor<string, []>("op_3951"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_71_pad_type_0 = const()[name = tensor<string, []>("key_71_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_71_pad_0 = const()[name = tensor<string, []>("key_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(511007232))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(511826496))), name = tensor<string, []>("layers_17_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_71_cast_fp16 = conv(dilations = var_3951, groups = var_3821, pad = key_71_pad_0, pad_type = key_71_pad_type_0, strides = var_3949, weight = layers_17_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_71_cast_fp16")];
            tensor<int32, [2]> var_3956 = const()[name = tensor<string, []>("op_3956"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3958 = const()[name = tensor<string, []>("op_3958"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_71_pad_type_0 = const()[name = tensor<string, []>("value_71_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_71_pad_0 = const()[name = tensor<string, []>("value_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(511826624))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513465088))), name = tensor<string, []>("layers_17_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_17_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513465664)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_71_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_bias_to_fp16, dilations = var_3958, groups = var_3821, pad = value_71_pad_0, pad_type = value_71_pad_type_0, strides = var_3956, weight = layers_17_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_71_cast_fp16")];
            tensor<int32, [4]> var_3962 = const()[name = tensor<string, []>("op_3962"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_3963_cast_fp16 = reshape(shape = var_3962, x = query_71_cast_fp16)[name = tensor<string, []>("op_3963_cast_fp16")];
            tensor<fp16, []> var_3964_to_fp16 = const()[name = tensor<string, []>("op_3964_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_3965_cast_fp16 = mul(x = var_3963_cast_fp16, y = var_3964_to_fp16)[name = tensor<string, []>("op_3965_cast_fp16")];
            tensor<int32, [4]> var_3966 = const()[name = tensor<string, []>("op_3966"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3967_cast_fp16 = reshape(shape = var_3966, x = key_71_cast_fp16)[name = tensor<string, []>("op_3967_cast_fp16")];
            tensor<bool, []> mh_w_107_transpose_x_0 = const()[name = tensor<string, []>("mh_w_107_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_107_transpose_y_0 = const()[name = tensor<string, []>("mh_w_107_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_107_cast_fp16 = matmul(transpose_x = mh_w_107_transpose_x_0, transpose_y = mh_w_107_transpose_y_0, x = var_3965_cast_fp16, y = var_3967_cast_fp16)[name = tensor<string, []>("mh_w_107_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_3970_cast_fp16 = softmax(axis = var_3814, x = mh_w_107_cast_fp16)[name = tensor<string, []>("op_3970_cast_fp16")];
            tensor<int32, [4]> var_3971 = const()[name = tensor<string, []>("op_3971"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_3972_cast_fp16 = reshape(shape = var_3971, x = value_71_cast_fp16)[name = tensor<string, []>("op_3972_cast_fp16")];
            tensor<bool, []> attn_71_transpose_x_0 = const()[name = tensor<string, []>("attn_71_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_71_transpose_y_0 = const()[name = tensor<string, []>("attn_71_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_71_cast_fp16 = matmul(transpose_x = attn_71_transpose_x_0, transpose_y = attn_71_transpose_y_0, x = var_3972_cast_fp16, y = var_3970_cast_fp16)[name = tensor<string, []>("attn_71_cast_fp16")];
            tensor<int32, [4]> var_3975 = const()[name = tensor<string, []>("op_3975"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_173_cast_fp16 = reshape(shape = var_3975, x = attn_71_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
            tensor<int32, [2]> var_3979 = const()[name = tensor<string, []>("op_3979"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_3981 = const()[name = tensor<string, []>("op_3981"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_215_pad_type_0 = const()[name = tensor<string, []>("obj_215_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_215_pad_0 = const()[name = tensor<string, []>("obj_215_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_17_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513468288))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(515106752))), name = tensor<string, []>("layers_17_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_17_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(515107328)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_215_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_bias_to_fp16, dilations = var_3981, groups = var_3821, pad = obj_215_pad_0, pad_type = obj_215_pad_type_0, strides = var_3979, weight = layers_17_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = tensor<string, []>("obj_215_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_215_cast_fp16)[name = tensor<string, []>("inputs_107_cast_fp16")];
            tensor<int32, [1]> var_3987 = const()[name = tensor<string, []>("op_3987"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_107_cast_fp16 = reduce_mean(axes = var_3987, keep_dims = var_3822, x = inputs_107_cast_fp16)[name = tensor<string, []>("channels_mean_107_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_107_cast_fp16 = sub(x = inputs_107_cast_fp16, y = channels_mean_107_cast_fp16)[name = tensor<string, []>("zero_mean_107_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = zero_mean_107_cast_fp16)[name = tensor<string, []>("zero_mean_sq_107_cast_fp16")];
            tensor<int32, [1]> var_3991 = const()[name = tensor<string, []>("op_3991"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_3992_cast_fp16 = reduce_mean(axes = var_3991, keep_dims = var_3822, x = zero_mean_sq_107_cast_fp16)[name = tensor<string, []>("op_3992_cast_fp16")];
            tensor<fp16, []> var_3993_to_fp16 = const()[name = tensor<string, []>("op_3993_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_3994_cast_fp16 = add(x = var_3992_cast_fp16, y = var_3993_to_fp16)[name = tensor<string, []>("op_3994_cast_fp16")];
            tensor<fp32, []> denom_107_epsilon_0 = const()[name = tensor<string, []>("denom_107_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_107_cast_fp16 = rsqrt(epsilon = denom_107_epsilon_0, x = var_3994_cast_fp16)[name = tensor<string, []>("denom_107_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = denom_107_cast_fp16)[name = tensor<string, []>("out_107_cast_fp16")];
            tensor<fp16, [1280]> input_175_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_175_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(515109952)))];
            tensor<fp16, [1280]> input_175_beta_0_to_fp16 = const()[name = tensor<string, []>("input_175_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(515112576)))];
            tensor<fp16, []> input_175_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_175_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_175_cast_fp16 = batch_norm(beta = input_175_beta_0_to_fp16, epsilon = input_175_epsilon_0_to_fp16, gamma = input_175_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<int32, [2]> var_4005 = const()[name = tensor<string, []>("op_4005"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4007 = const()[name = tensor<string, []>("op_4007"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_177_pad_type_0 = const()[name = tensor<string, []>("input_177_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_177_pad_0 = const()[name = tensor<string, []>("input_177_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_17_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(515115200))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520030464))), name = tensor<string, []>("layers_17_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_17_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520030656)))];
            tensor<fp16, [1, 5120, 1, 1]> input_177_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = var_4007, groups = var_3821, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = var_4005, weight = layers_17_fc1_weight_to_fp16_palettized, x = input_175_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
            tensor<string, []> input_179_mode_0 = const()[name = tensor<string, []>("input_179_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = input_177_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<int32, [2]> var_4013 = const()[name = tensor<string, []>("op_4013"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4015 = const()[name = tensor<string, []>("op_4015"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_37_pad_type_0 = const()[name = tensor<string, []>("hidden_states_37_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = tensor<string, []>("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_17_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520040960))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526594624))), name = tensor<string, []>("layers_17_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_17_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526595200)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_37_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = var_4015, groups = var_3821, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = var_4013, weight = layers_17_fc2_weight_to_fp16_palettized, x = input_179_cast_fp16)[name = tensor<string, []>("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor<string, []>("inputs_109_cast_fp16")];
            tensor<int32, []> var_4028 = const()[name = tensor<string, []>("op_4028"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_4035 = const()[name = tensor<string, []>("op_4035"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_4036 = const()[name = tensor<string, []>("op_4036"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_4048 = const()[name = tensor<string, []>("op_4048"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_109_cast_fp16 = reduce_mean(axes = var_4048, keep_dims = var_4036, x = inputs_109_cast_fp16)[name = tensor<string, []>("channels_mean_109_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_109_cast_fp16 = sub(x = inputs_109_cast_fp16, y = channels_mean_109_cast_fp16)[name = tensor<string, []>("zero_mean_109_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = zero_mean_109_cast_fp16)[name = tensor<string, []>("zero_mean_sq_109_cast_fp16")];
            tensor<int32, [1]> var_4052 = const()[name = tensor<string, []>("op_4052"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4053_cast_fp16 = reduce_mean(axes = var_4052, keep_dims = var_4036, x = zero_mean_sq_109_cast_fp16)[name = tensor<string, []>("op_4053_cast_fp16")];
            tensor<fp16, []> var_4054_to_fp16 = const()[name = tensor<string, []>("op_4054_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4055_cast_fp16 = add(x = var_4053_cast_fp16, y = var_4054_to_fp16)[name = tensor<string, []>("op_4055_cast_fp16")];
            tensor<fp32, []> denom_109_epsilon_0 = const()[name = tensor<string, []>("denom_109_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_109_cast_fp16 = rsqrt(epsilon = denom_109_epsilon_0, x = var_4055_cast_fp16)[name = tensor<string, []>("denom_109_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = denom_109_cast_fp16)[name = tensor<string, []>("out_109_cast_fp16")];
            tensor<fp16, [1280]> obj_217_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_217_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526597824)))];
            tensor<fp16, [1280]> obj_217_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_217_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526600448)))];
            tensor<fp16, []> obj_217_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_217_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_217_cast_fp16 = batch_norm(beta = obj_217_beta_0_to_fp16, epsilon = obj_217_epsilon_0_to_fp16, gamma = obj_217_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor<string, []>("obj_217_cast_fp16")];
            tensor<int32, [2]> var_4070 = const()[name = tensor<string, []>("op_4070"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4072 = const()[name = tensor<string, []>("op_4072"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_73_pad_type_0 = const()[name = tensor<string, []>("query_73_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_73_pad_0 = const()[name = tensor<string, []>("query_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526603072))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527831936))), name = tensor<string, []>("layers_18_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527832128)))];
            tensor<fp16, [1, 1280, 1, 1]> query_73_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = var_4072, groups = var_4035, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = var_4070, weight = layers_18_self_attn_q_proj_weight_to_fp16_palettized, x = obj_217_cast_fp16)[name = tensor<string, []>("query_73_cast_fp16")];
            tensor<int32, [2]> var_4076 = const()[name = tensor<string, []>("op_4076"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4078 = const()[name = tensor<string, []>("op_4078"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_37_pad_type_0 = const()[name = tensor<string, []>("current_key_37_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_37_pad_0 = const()[name = tensor<string, []>("current_key_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527834752))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528654016))), name = tensor<string, []>("layers_18_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_37_cast_fp16 = conv(dilations = var_4078, groups = var_4035, pad = current_key_37_pad_0, pad_type = current_key_37_pad_type_0, strides = var_4076, weight = layers_18_self_attn_k_proj_weight_to_fp16_palettized, x = obj_217_cast_fp16)[name = tensor<string, []>("current_key_37_cast_fp16")];
            tensor<int32, [2]> var_4083 = const()[name = tensor<string, []>("op_4083"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4085 = const()[name = tensor<string, []>("op_4085"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_37_pad_type_0 = const()[name = tensor<string, []>("current_value_37_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_37_pad_0 = const()[name = tensor<string, []>("current_value_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528654144))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529473408))), name = tensor<string, []>("layers_18_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529473536)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = var_4085, groups = var_4035, pad = current_value_37_pad_0, pad_type = current_value_37_pad_type_0, strides = var_4083, weight = layers_18_self_attn_v_proj_weight_to_fp16_palettized, x = obj_217_cast_fp16)[name = tensor<string, []>("current_value_37_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4092_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4092_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4094_cast_fp16 = mul(x = var_103_cast_fp16_18, y = var_241_cast_fp16)[name = tensor<string, []>("op_4094_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_73_cast_fp16 = add(x = var_4092_cast_fp16, y = var_4094_cast_fp16)[name = tensor<string, []>("key_73_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4096_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4096_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4098_cast_fp16 = mul(x = var_138_cast_fp16_18, y = var_241_cast_fp16)[name = tensor<string, []>("op_4098_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_73_cast_fp16 = add(x = var_4096_cast_fp16, y = var_4098_cast_fp16)[name = tensor<string, []>("value_73_cast_fp16")];
            tensor<int32, [4]> var_4101 = const()[name = tensor<string, []>("op_4101"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4102_cast_fp16 = reshape(shape = var_4101, x = query_73_cast_fp16)[name = tensor<string, []>("op_4102_cast_fp16")];
            tensor<fp16, []> var_4103_to_fp16 = const()[name = tensor<string, []>("op_4103_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4104_cast_fp16 = mul(x = var_4102_cast_fp16, y = var_4103_to_fp16)[name = tensor<string, []>("op_4104_cast_fp16")];
            tensor<int32, [4]> var_4105 = const()[name = tensor<string, []>("op_4105"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4106_cast_fp16 = reshape(shape = var_4105, x = key_73_cast_fp16)[name = tensor<string, []>("op_4106_cast_fp16")];
            tensor<bool, []> mh_w_109_transpose_x_0 = const()[name = tensor<string, []>("mh_w_109_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_109_transpose_y_0 = const()[name = tensor<string, []>("mh_w_109_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_109_cast_fp16 = matmul(transpose_x = mh_w_109_transpose_x_0, transpose_y = mh_w_109_transpose_y_0, x = var_4104_cast_fp16, y = var_4106_cast_fp16)[name = tensor<string, []>("mh_w_109_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_111_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_4114_cast_fp16 = softmax(axis = var_4028, x = mh_w_111_cast_fp16)[name = tensor<string, []>("op_4114_cast_fp16")];
            tensor<int32, [4]> var_4115 = const()[name = tensor<string, []>("op_4115"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4116_cast_fp16 = reshape(shape = var_4115, x = value_73_cast_fp16)[name = tensor<string, []>("op_4116_cast_fp16")];
            tensor<bool, []> attn_73_transpose_x_0 = const()[name = tensor<string, []>("attn_73_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_73_transpose_y_0 = const()[name = tensor<string, []>("attn_73_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_0, transpose_y = attn_73_transpose_y_0, x = var_4116_cast_fp16, y = var_4114_cast_fp16)[name = tensor<string, []>("attn_73_cast_fp16")];
            tensor<int32, [4]> var_4119 = const()[name = tensor<string, []>("op_4119"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_181_cast_fp16 = reshape(shape = var_4119, x = attn_73_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
            tensor<int32, [2]> var_4123 = const()[name = tensor<string, []>("op_4123"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4125 = const()[name = tensor<string, []>("op_4125"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_223_pad_type_0 = const()[name = tensor<string, []>("obj_223_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_223_pad_0 = const()[name = tensor<string, []>("obj_223_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529476160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530705024))), name = tensor<string, []>("layers_18_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530705216)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_223_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = var_4125, groups = var_4035, pad = obj_223_pad_0, pad_type = obj_223_pad_type_0, strides = var_4123, weight = layers_18_self_attn_o_proj_weight_to_fp16_palettized, x = input_181_cast_fp16)[name = tensor<string, []>("obj_223_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_223_cast_fp16)[name = tensor<string, []>("inputs_111_cast_fp16")];
            tensor<int32, [1]> var_4135 = const()[name = tensor<string, []>("op_4135"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_111_cast_fp16 = reduce_mean(axes = var_4135, keep_dims = var_4036, x = inputs_111_cast_fp16)[name = tensor<string, []>("channels_mean_111_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_111_cast_fp16 = sub(x = inputs_111_cast_fp16, y = channels_mean_111_cast_fp16)[name = tensor<string, []>("zero_mean_111_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = zero_mean_111_cast_fp16)[name = tensor<string, []>("zero_mean_sq_111_cast_fp16")];
            tensor<int32, [1]> var_4139 = const()[name = tensor<string, []>("op_4139"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4140_cast_fp16 = reduce_mean(axes = var_4139, keep_dims = var_4036, x = zero_mean_sq_111_cast_fp16)[name = tensor<string, []>("op_4140_cast_fp16")];
            tensor<fp16, []> var_4141_to_fp16 = const()[name = tensor<string, []>("op_4141_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4142_cast_fp16 = add(x = var_4140_cast_fp16, y = var_4141_to_fp16)[name = tensor<string, []>("op_4142_cast_fp16")];
            tensor<fp32, []> denom_111_epsilon_0 = const()[name = tensor<string, []>("denom_111_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_111_cast_fp16 = rsqrt(epsilon = denom_111_epsilon_0, x = var_4142_cast_fp16)[name = tensor<string, []>("denom_111_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = denom_111_cast_fp16)[name = tensor<string, []>("out_111_cast_fp16")];
            tensor<fp16, [1280]> obj_225_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_225_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530707840)))];
            tensor<fp16, [1280]> obj_225_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_225_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530710464)))];
            tensor<fp16, []> obj_225_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_225_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_225_cast_fp16 = batch_norm(beta = obj_225_beta_0_to_fp16, epsilon = obj_225_epsilon_0_to_fp16, gamma = obj_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor<string, []>("obj_225_cast_fp16")];
            tensor<int32, [2]> var_4157 = const()[name = tensor<string, []>("op_4157"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4159 = const()[name = tensor<string, []>("op_4159"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_75_pad_type_0 = const()[name = tensor<string, []>("query_75_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_75_pad_0 = const()[name = tensor<string, []>("query_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530713088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(531941952))), name = tensor<string, []>("layers_18_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_18_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(531942144)))];
            tensor<fp16, [1, 1280, 1, 1]> query_75_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_bias_to_fp16, dilations = var_4159, groups = var_4035, pad = query_75_pad_0, pad_type = query_75_pad_type_0, strides = var_4157, weight = layers_18_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_225_cast_fp16)[name = tensor<string, []>("query_75_cast_fp16")];
            tensor<int32, [2]> var_4163 = const()[name = tensor<string, []>("op_4163"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4165 = const()[name = tensor<string, []>("op_4165"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_75_pad_type_0 = const()[name = tensor<string, []>("key_75_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_75_pad_0 = const()[name = tensor<string, []>("key_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(531944768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533173632))), name = tensor<string, []>("layers_18_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_75_cast_fp16 = conv(dilations = var_4165, groups = var_4035, pad = key_75_pad_0, pad_type = key_75_pad_type_0, strides = var_4163, weight = layers_18_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_75_cast_fp16")];
            tensor<int32, [2]> var_4170 = const()[name = tensor<string, []>("op_4170"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4172 = const()[name = tensor<string, []>("op_4172"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_75_pad_type_0 = const()[name = tensor<string, []>("value_75_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_75_pad_0 = const()[name = tensor<string, []>("value_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533173824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(534402688))), name = tensor<string, []>("layers_18_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_18_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(534402880)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_75_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_bias_to_fp16, dilations = var_4172, groups = var_4035, pad = value_75_pad_0, pad_type = value_75_pad_type_0, strides = var_4170, weight = layers_18_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_75_cast_fp16")];
            tensor<int32, [4]> var_4176 = const()[name = tensor<string, []>("op_4176"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4177_cast_fp16 = reshape(shape = var_4176, x = query_75_cast_fp16)[name = tensor<string, []>("op_4177_cast_fp16")];
            tensor<fp16, []> var_4178_to_fp16 = const()[name = tensor<string, []>("op_4178_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4179_cast_fp16 = mul(x = var_4177_cast_fp16, y = var_4178_to_fp16)[name = tensor<string, []>("op_4179_cast_fp16")];
            tensor<int32, [4]> var_4180 = const()[name = tensor<string, []>("op_4180"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4181_cast_fp16 = reshape(shape = var_4180, x = key_75_cast_fp16)[name = tensor<string, []>("op_4181_cast_fp16")];
            tensor<bool, []> mh_w_113_transpose_x_0 = const()[name = tensor<string, []>("mh_w_113_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_113_transpose_y_0 = const()[name = tensor<string, []>("mh_w_113_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_113_cast_fp16 = matmul(transpose_x = mh_w_113_transpose_x_0, transpose_y = mh_w_113_transpose_y_0, x = var_4179_cast_fp16, y = var_4181_cast_fp16)[name = tensor<string, []>("mh_w_113_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_4184_cast_fp16 = softmax(axis = var_4028, x = mh_w_113_cast_fp16)[name = tensor<string, []>("op_4184_cast_fp16")];
            tensor<int32, [4]> var_4185 = const()[name = tensor<string, []>("op_4185"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4186_cast_fp16 = reshape(shape = var_4185, x = value_75_cast_fp16)[name = tensor<string, []>("op_4186_cast_fp16")];
            tensor<bool, []> attn_75_transpose_x_0 = const()[name = tensor<string, []>("attn_75_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_75_transpose_y_0 = const()[name = tensor<string, []>("attn_75_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_75_cast_fp16 = matmul(transpose_x = attn_75_transpose_x_0, transpose_y = attn_75_transpose_y_0, x = var_4186_cast_fp16, y = var_4184_cast_fp16)[name = tensor<string, []>("attn_75_cast_fp16")];
            tensor<int32, [4]> var_4189 = const()[name = tensor<string, []>("op_4189"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_183_cast_fp16 = reshape(shape = var_4189, x = attn_75_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
            tensor<int32, [2]> var_4193 = const()[name = tensor<string, []>("op_4193"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4195 = const()[name = tensor<string, []>("op_4195"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_227_pad_type_0 = const()[name = tensor<string, []>("obj_227_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_227_pad_0 = const()[name = tensor<string, []>("obj_227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_18_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(534405504))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535634368))), name = tensor<string, []>("layers_18_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_18_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535634560)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_227_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_bias_to_fp16, dilations = var_4195, groups = var_4035, pad = obj_227_pad_0, pad_type = obj_227_pad_type_0, strides = var_4193, weight = layers_18_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = tensor<string, []>("obj_227_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_227_cast_fp16)[name = tensor<string, []>("inputs_113_cast_fp16")];
            tensor<int32, [1]> var_4201 = const()[name = tensor<string, []>("op_4201"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_113_cast_fp16 = reduce_mean(axes = var_4201, keep_dims = var_4036, x = inputs_113_cast_fp16)[name = tensor<string, []>("channels_mean_113_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_113_cast_fp16 = sub(x = inputs_113_cast_fp16, y = channels_mean_113_cast_fp16)[name = tensor<string, []>("zero_mean_113_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = zero_mean_113_cast_fp16)[name = tensor<string, []>("zero_mean_sq_113_cast_fp16")];
            tensor<int32, [1]> var_4205 = const()[name = tensor<string, []>("op_4205"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4206_cast_fp16 = reduce_mean(axes = var_4205, keep_dims = var_4036, x = zero_mean_sq_113_cast_fp16)[name = tensor<string, []>("op_4206_cast_fp16")];
            tensor<fp16, []> var_4207_to_fp16 = const()[name = tensor<string, []>("op_4207_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4208_cast_fp16 = add(x = var_4206_cast_fp16, y = var_4207_to_fp16)[name = tensor<string, []>("op_4208_cast_fp16")];
            tensor<fp32, []> denom_113_epsilon_0 = const()[name = tensor<string, []>("denom_113_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_113_cast_fp16 = rsqrt(epsilon = denom_113_epsilon_0, x = var_4208_cast_fp16)[name = tensor<string, []>("denom_113_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = denom_113_cast_fp16)[name = tensor<string, []>("out_113_cast_fp16")];
            tensor<fp16, [1280]> input_185_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_185_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535637184)))];
            tensor<fp16, [1280]> input_185_beta_0_to_fp16 = const()[name = tensor<string, []>("input_185_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535639808)))];
            tensor<fp16, []> input_185_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_185_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_185_cast_fp16 = batch_norm(beta = input_185_beta_0_to_fp16, epsilon = input_185_epsilon_0_to_fp16, gamma = input_185_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
            tensor<int32, [2]> var_4219 = const()[name = tensor<string, []>("op_4219"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4221 = const()[name = tensor<string, []>("op_4221"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_187_pad_type_0 = const()[name = tensor<string, []>("input_187_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_187_pad_0 = const()[name = tensor<string, []>("input_187_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_18_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535642432))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540557696))), name = tensor<string, []>("layers_18_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_18_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540557888)))];
            tensor<fp16, [1, 5120, 1, 1]> input_187_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = var_4221, groups = var_4035, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = var_4219, weight = layers_18_fc1_weight_to_fp16_palettized, x = input_185_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<string, []> input_189_mode_0 = const()[name = tensor<string, []>("input_189_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = input_187_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<int32, [2]> var_4227 = const()[name = tensor<string, []>("op_4227"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4229 = const()[name = tensor<string, []>("op_4229"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_39_pad_type_0 = const()[name = tensor<string, []>("hidden_states_39_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_39_pad_0 = const()[name = tensor<string, []>("hidden_states_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_18_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540568192))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545483456))), name = tensor<string, []>("layers_18_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_18_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545483648)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_39_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = var_4229, groups = var_4035, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = var_4227, weight = layers_18_fc2_weight_to_fp16_palettized, x = input_189_cast_fp16)[name = tensor<string, []>("hidden_states_39_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor<string, []>("inputs_115_cast_fp16")];
            tensor<int32, []> var_4242 = const()[name = tensor<string, []>("op_4242"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_4249 = const()[name = tensor<string, []>("op_4249"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_4250 = const()[name = tensor<string, []>("op_4250"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_4262 = const()[name = tensor<string, []>("op_4262"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_115_cast_fp16 = reduce_mean(axes = var_4262, keep_dims = var_4250, x = inputs_115_cast_fp16)[name = tensor<string, []>("channels_mean_115_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_115_cast_fp16 = sub(x = inputs_115_cast_fp16, y = channels_mean_115_cast_fp16)[name = tensor<string, []>("zero_mean_115_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = zero_mean_115_cast_fp16)[name = tensor<string, []>("zero_mean_sq_115_cast_fp16")];
            tensor<int32, [1]> var_4266 = const()[name = tensor<string, []>("op_4266"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4267_cast_fp16 = reduce_mean(axes = var_4266, keep_dims = var_4250, x = zero_mean_sq_115_cast_fp16)[name = tensor<string, []>("op_4267_cast_fp16")];
            tensor<fp16, []> var_4268_to_fp16 = const()[name = tensor<string, []>("op_4268_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4269_cast_fp16 = add(x = var_4267_cast_fp16, y = var_4268_to_fp16)[name = tensor<string, []>("op_4269_cast_fp16")];
            tensor<fp32, []> denom_115_epsilon_0 = const()[name = tensor<string, []>("denom_115_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_115_cast_fp16 = rsqrt(epsilon = denom_115_epsilon_0, x = var_4269_cast_fp16)[name = tensor<string, []>("denom_115_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = denom_115_cast_fp16)[name = tensor<string, []>("out_115_cast_fp16")];
            tensor<fp16, [1280]> obj_229_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_229_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545486272)))];
            tensor<fp16, [1280]> obj_229_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_229_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545488896)))];
            tensor<fp16, []> obj_229_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_229_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_229_cast_fp16 = batch_norm(beta = obj_229_beta_0_to_fp16, epsilon = obj_229_epsilon_0_to_fp16, gamma = obj_229_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor<string, []>("obj_229_cast_fp16")];
            tensor<int32, [2]> var_4284 = const()[name = tensor<string, []>("op_4284"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4286 = const()[name = tensor<string, []>("op_4286"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_77_pad_type_0 = const()[name = tensor<string, []>("query_77_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_77_pad_0 = const()[name = tensor<string, []>("query_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545491520))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546310784))), name = tensor<string, []>("layers_19_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546310912)))];
            tensor<fp16, [1, 1280, 1, 1]> query_77_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = var_4286, groups = var_4249, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = var_4284, weight = layers_19_self_attn_q_proj_weight_to_fp16_palettized, x = obj_229_cast_fp16)[name = tensor<string, []>("query_77_cast_fp16")];
            tensor<int32, [2]> var_4290 = const()[name = tensor<string, []>("op_4290"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4292 = const()[name = tensor<string, []>("op_4292"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_39_pad_type_0 = const()[name = tensor<string, []>("current_key_39_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_39_pad_0 = const()[name = tensor<string, []>("current_key_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546313536))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547132800))), name = tensor<string, []>("layers_19_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_39_cast_fp16 = conv(dilations = var_4292, groups = var_4249, pad = current_key_39_pad_0, pad_type = current_key_39_pad_type_0, strides = var_4290, weight = layers_19_self_attn_k_proj_weight_to_fp16_palettized, x = obj_229_cast_fp16)[name = tensor<string, []>("current_key_39_cast_fp16")];
            tensor<int32, [2]> var_4297 = const()[name = tensor<string, []>("op_4297"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4299 = const()[name = tensor<string, []>("op_4299"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_39_pad_type_0 = const()[name = tensor<string, []>("current_value_39_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_39_pad_0 = const()[name = tensor<string, []>("current_value_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547132928))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547952192))), name = tensor<string, []>("layers_19_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547952320)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = var_4299, groups = var_4249, pad = current_value_39_pad_0, pad_type = current_value_39_pad_type_0, strides = var_4297, weight = layers_19_self_attn_v_proj_weight_to_fp16_palettized, x = obj_229_cast_fp16)[name = tensor<string, []>("current_value_39_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4306_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4306_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4308_cast_fp16 = mul(x = var_103_cast_fp16_19, y = var_241_cast_fp16)[name = tensor<string, []>("op_4308_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_77_cast_fp16 = add(x = var_4306_cast_fp16, y = var_4308_cast_fp16)[name = tensor<string, []>("key_77_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4310_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4310_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4312_cast_fp16 = mul(x = var_138_cast_fp16_19, y = var_241_cast_fp16)[name = tensor<string, []>("op_4312_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_77_cast_fp16 = add(x = var_4310_cast_fp16, y = var_4312_cast_fp16)[name = tensor<string, []>("value_77_cast_fp16")];
            tensor<int32, [4]> var_4315 = const()[name = tensor<string, []>("op_4315"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4316_cast_fp16 = reshape(shape = var_4315, x = query_77_cast_fp16)[name = tensor<string, []>("op_4316_cast_fp16")];
            tensor<fp16, []> var_4317_to_fp16 = const()[name = tensor<string, []>("op_4317_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4318_cast_fp16 = mul(x = var_4316_cast_fp16, y = var_4317_to_fp16)[name = tensor<string, []>("op_4318_cast_fp16")];
            tensor<int32, [4]> var_4319 = const()[name = tensor<string, []>("op_4319"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4320_cast_fp16 = reshape(shape = var_4319, x = key_77_cast_fp16)[name = tensor<string, []>("op_4320_cast_fp16")];
            tensor<bool, []> mh_w_115_transpose_x_0 = const()[name = tensor<string, []>("mh_w_115_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_115_transpose_y_0 = const()[name = tensor<string, []>("mh_w_115_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_115_cast_fp16 = matmul(transpose_x = mh_w_115_transpose_x_0, transpose_y = mh_w_115_transpose_y_0, x = var_4318_cast_fp16, y = var_4320_cast_fp16)[name = tensor<string, []>("mh_w_115_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_117_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_4328_cast_fp16 = softmax(axis = var_4242, x = mh_w_117_cast_fp16)[name = tensor<string, []>("op_4328_cast_fp16")];
            tensor<int32, [4]> var_4329 = const()[name = tensor<string, []>("op_4329"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4330_cast_fp16 = reshape(shape = var_4329, x = value_77_cast_fp16)[name = tensor<string, []>("op_4330_cast_fp16")];
            tensor<bool, []> attn_77_transpose_x_0 = const()[name = tensor<string, []>("attn_77_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_77_transpose_y_0 = const()[name = tensor<string, []>("attn_77_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_0, transpose_y = attn_77_transpose_y_0, x = var_4330_cast_fp16, y = var_4328_cast_fp16)[name = tensor<string, []>("attn_77_cast_fp16")];
            tensor<int32, [4]> var_4333 = const()[name = tensor<string, []>("op_4333"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_191_cast_fp16 = reshape(shape = var_4333, x = attn_77_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
            tensor<int32, [2]> var_4337 = const()[name = tensor<string, []>("op_4337"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4339 = const()[name = tensor<string, []>("op_4339"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_235_pad_type_0 = const()[name = tensor<string, []>("obj_235_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_235_pad_0 = const()[name = tensor<string, []>("obj_235_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547954944))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548774208))), name = tensor<string, []>("layers_19_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548774336)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_235_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = var_4339, groups = var_4249, pad = obj_235_pad_0, pad_type = obj_235_pad_type_0, strides = var_4337, weight = layers_19_self_attn_o_proj_weight_to_fp16_palettized, x = input_191_cast_fp16)[name = tensor<string, []>("obj_235_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_235_cast_fp16)[name = tensor<string, []>("inputs_117_cast_fp16")];
            tensor<int32, [1]> var_4349 = const()[name = tensor<string, []>("op_4349"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_117_cast_fp16 = reduce_mean(axes = var_4349, keep_dims = var_4250, x = inputs_117_cast_fp16)[name = tensor<string, []>("channels_mean_117_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_117_cast_fp16 = sub(x = inputs_117_cast_fp16, y = channels_mean_117_cast_fp16)[name = tensor<string, []>("zero_mean_117_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = zero_mean_117_cast_fp16)[name = tensor<string, []>("zero_mean_sq_117_cast_fp16")];
            tensor<int32, [1]> var_4353 = const()[name = tensor<string, []>("op_4353"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4354_cast_fp16 = reduce_mean(axes = var_4353, keep_dims = var_4250, x = zero_mean_sq_117_cast_fp16)[name = tensor<string, []>("op_4354_cast_fp16")];
            tensor<fp16, []> var_4355_to_fp16 = const()[name = tensor<string, []>("op_4355_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4356_cast_fp16 = add(x = var_4354_cast_fp16, y = var_4355_to_fp16)[name = tensor<string, []>("op_4356_cast_fp16")];
            tensor<fp32, []> denom_117_epsilon_0 = const()[name = tensor<string, []>("denom_117_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_117_cast_fp16 = rsqrt(epsilon = denom_117_epsilon_0, x = var_4356_cast_fp16)[name = tensor<string, []>("denom_117_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = denom_117_cast_fp16)[name = tensor<string, []>("out_117_cast_fp16")];
            tensor<fp16, [1280]> obj_237_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_237_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548776960)))];
            tensor<fp16, [1280]> obj_237_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_237_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548779584)))];
            tensor<fp16, []> obj_237_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_237_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_237_cast_fp16 = batch_norm(beta = obj_237_beta_0_to_fp16, epsilon = obj_237_epsilon_0_to_fp16, gamma = obj_237_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor<string, []>("obj_237_cast_fp16")];
            tensor<int32, [2]> var_4371 = const()[name = tensor<string, []>("op_4371"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4373 = const()[name = tensor<string, []>("op_4373"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_79_pad_type_0 = const()[name = tensor<string, []>("query_79_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_79_pad_0 = const()[name = tensor<string, []>("query_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548782208)))];
            tensor<fp16, [1280]> layers_19_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552059072)))];
            tensor<fp16, [1, 1280, 1, 1]> query_79_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_bias_to_fp16, dilations = var_4373, groups = var_4249, pad = query_79_pad_0, pad_type = query_79_pad_type_0, strides = var_4371, weight = layers_19_encoder_attn_q_proj_weight_to_fp16, x = obj_237_cast_fp16)[name = tensor<string, []>("query_79_cast_fp16")];
            tensor<int32, [2]> var_4377 = const()[name = tensor<string, []>("op_4377"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4379 = const()[name = tensor<string, []>("op_4379"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_79_pad_type_0 = const()[name = tensor<string, []>("key_79_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_79_pad_0 = const()[name = tensor<string, []>("key_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552061696))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553290560))), name = tensor<string, []>("layers_19_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_79_cast_fp16 = conv(dilations = var_4379, groups = var_4249, pad = key_79_pad_0, pad_type = key_79_pad_type_0, strides = var_4377, weight = layers_19_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_79_cast_fp16")];
            tensor<int32, [2]> var_4384 = const()[name = tensor<string, []>("op_4384"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4386 = const()[name = tensor<string, []>("op_4386"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_79_pad_type_0 = const()[name = tensor<string, []>("value_79_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_79_pad_0 = const()[name = tensor<string, []>("value_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553290752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554519616))), name = tensor<string, []>("layers_19_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_19_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554519808)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_79_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_bias_to_fp16, dilations = var_4386, groups = var_4249, pad = value_79_pad_0, pad_type = value_79_pad_type_0, strides = var_4384, weight = layers_19_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_79_cast_fp16")];
            tensor<int32, [4]> var_4390 = const()[name = tensor<string, []>("op_4390"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4391_cast_fp16 = reshape(shape = var_4390, x = query_79_cast_fp16)[name = tensor<string, []>("op_4391_cast_fp16")];
            tensor<fp16, []> var_4392_to_fp16 = const()[name = tensor<string, []>("op_4392_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = tensor<string, []>("op_4393_cast_fp16")];
            tensor<int32, [4]> var_4394 = const()[name = tensor<string, []>("op_4394"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4395_cast_fp16 = reshape(shape = var_4394, x = key_79_cast_fp16)[name = tensor<string, []>("op_4395_cast_fp16")];
            tensor<bool, []> mh_w_119_transpose_x_0 = const()[name = tensor<string, []>("mh_w_119_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_119_transpose_y_0 = const()[name = tensor<string, []>("mh_w_119_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_119_cast_fp16 = matmul(transpose_x = mh_w_119_transpose_x_0, transpose_y = mh_w_119_transpose_y_0, x = var_4393_cast_fp16, y = var_4395_cast_fp16)[name = tensor<string, []>("mh_w_119_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_4398_cast_fp16 = softmax(axis = var_4242, x = mh_w_119_cast_fp16)[name = tensor<string, []>("op_4398_cast_fp16")];
            tensor<int32, [4]> var_4399 = const()[name = tensor<string, []>("op_4399"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4400_cast_fp16 = reshape(shape = var_4399, x = value_79_cast_fp16)[name = tensor<string, []>("op_4400_cast_fp16")];
            tensor<bool, []> attn_79_transpose_x_0 = const()[name = tensor<string, []>("attn_79_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_79_transpose_y_0 = const()[name = tensor<string, []>("attn_79_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_79_cast_fp16 = matmul(transpose_x = attn_79_transpose_x_0, transpose_y = attn_79_transpose_y_0, x = var_4400_cast_fp16, y = var_4398_cast_fp16)[name = tensor<string, []>("attn_79_cast_fp16")];
            tensor<int32, [4]> var_4403 = const()[name = tensor<string, []>("op_4403"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_193_cast_fp16 = reshape(shape = var_4403, x = attn_79_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<int32, [2]> var_4407 = const()[name = tensor<string, []>("op_4407"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4409 = const()[name = tensor<string, []>("op_4409"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_239_pad_type_0 = const()[name = tensor<string, []>("obj_239_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_239_pad_0 = const()[name = tensor<string, []>("obj_239_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_19_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554522432))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(556160896))), name = tensor<string, []>("layers_19_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_19_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(556161472)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_239_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_bias_to_fp16, dilations = var_4409, groups = var_4249, pad = obj_239_pad_0, pad_type = obj_239_pad_type_0, strides = var_4407, weight = layers_19_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_193_cast_fp16)[name = tensor<string, []>("obj_239_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_239_cast_fp16)[name = tensor<string, []>("inputs_119_cast_fp16")];
            tensor<int32, [1]> var_4415 = const()[name = tensor<string, []>("op_4415"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_119_cast_fp16 = reduce_mean(axes = var_4415, keep_dims = var_4250, x = inputs_119_cast_fp16)[name = tensor<string, []>("channels_mean_119_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_119_cast_fp16 = sub(x = inputs_119_cast_fp16, y = channels_mean_119_cast_fp16)[name = tensor<string, []>("zero_mean_119_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = zero_mean_119_cast_fp16)[name = tensor<string, []>("zero_mean_sq_119_cast_fp16")];
            tensor<int32, [1]> var_4419 = const()[name = tensor<string, []>("op_4419"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4420_cast_fp16 = reduce_mean(axes = var_4419, keep_dims = var_4250, x = zero_mean_sq_119_cast_fp16)[name = tensor<string, []>("op_4420_cast_fp16")];
            tensor<fp16, []> var_4421_to_fp16 = const()[name = tensor<string, []>("op_4421_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4422_cast_fp16 = add(x = var_4420_cast_fp16, y = var_4421_to_fp16)[name = tensor<string, []>("op_4422_cast_fp16")];
            tensor<fp32, []> denom_119_epsilon_0 = const()[name = tensor<string, []>("denom_119_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_119_cast_fp16 = rsqrt(epsilon = denom_119_epsilon_0, x = var_4422_cast_fp16)[name = tensor<string, []>("denom_119_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = denom_119_cast_fp16)[name = tensor<string, []>("out_119_cast_fp16")];
            tensor<fp16, [1280]> input_195_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_195_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(556164096)))];
            tensor<fp16, [1280]> input_195_beta_0_to_fp16 = const()[name = tensor<string, []>("input_195_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(556166720)))];
            tensor<fp16, []> input_195_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_195_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<int32, [2]> var_4433 = const()[name = tensor<string, []>("op_4433"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4435 = const()[name = tensor<string, []>("op_4435"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_197_pad_type_0 = const()[name = tensor<string, []>("input_197_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_197_pad_0 = const()[name = tensor<string, []>("input_197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_19_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(556169344)))];
            tensor<fp16, [5120]> layers_19_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(569276608)))];
            tensor<fp16, [1, 5120, 1, 1]> input_197_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = var_4435, groups = var_4249, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = var_4433, weight = layers_19_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
            tensor<string, []> input_199_mode_0 = const()[name = tensor<string, []>("input_199_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<int32, [2]> var_4441 = const()[name = tensor<string, []>("op_4441"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4443 = const()[name = tensor<string, []>("op_4443"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_41_pad_type_0 = const()[name = tensor<string, []>("hidden_states_41_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_41_pad_0 = const()[name = tensor<string, []>("hidden_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_19_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(569286912))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575840576))), name = tensor<string, []>("layers_19_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_19_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575841152)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_41_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = var_4443, groups = var_4249, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = var_4441, weight = layers_19_fc2_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = tensor<string, []>("hidden_states_41_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor<string, []>("inputs_121_cast_fp16")];
            tensor<int32, []> var_4456 = const()[name = tensor<string, []>("op_4456"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_4463 = const()[name = tensor<string, []>("op_4463"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_4464 = const()[name = tensor<string, []>("op_4464"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_4476 = const()[name = tensor<string, []>("op_4476"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_121_cast_fp16 = reduce_mean(axes = var_4476, keep_dims = var_4464, x = inputs_121_cast_fp16)[name = tensor<string, []>("channels_mean_121_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_121_cast_fp16 = sub(x = inputs_121_cast_fp16, y = channels_mean_121_cast_fp16)[name = tensor<string, []>("zero_mean_121_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = zero_mean_121_cast_fp16)[name = tensor<string, []>("zero_mean_sq_121_cast_fp16")];
            tensor<int32, [1]> var_4480 = const()[name = tensor<string, []>("op_4480"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4481_cast_fp16 = reduce_mean(axes = var_4480, keep_dims = var_4464, x = zero_mean_sq_121_cast_fp16)[name = tensor<string, []>("op_4481_cast_fp16")];
            tensor<fp16, []> var_4482_to_fp16 = const()[name = tensor<string, []>("op_4482_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4483_cast_fp16 = add(x = var_4481_cast_fp16, y = var_4482_to_fp16)[name = tensor<string, []>("op_4483_cast_fp16")];
            tensor<fp32, []> denom_121_epsilon_0 = const()[name = tensor<string, []>("denom_121_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_121_cast_fp16 = rsqrt(epsilon = denom_121_epsilon_0, x = var_4483_cast_fp16)[name = tensor<string, []>("denom_121_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = denom_121_cast_fp16)[name = tensor<string, []>("out_121_cast_fp16")];
            tensor<fp16, [1280]> obj_241_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_241_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575843776)))];
            tensor<fp16, [1280]> obj_241_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_241_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575846400)))];
            tensor<fp16, []> obj_241_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_241_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_241_cast_fp16 = batch_norm(beta = obj_241_beta_0_to_fp16, epsilon = obj_241_epsilon_0_to_fp16, gamma = obj_241_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor<string, []>("obj_241_cast_fp16")];
            tensor<int32, [2]> var_4498 = const()[name = tensor<string, []>("op_4498"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4500 = const()[name = tensor<string, []>("op_4500"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_81_pad_type_0 = const()[name = tensor<string, []>("query_81_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_81_pad_0 = const()[name = tensor<string, []>("query_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575849024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577077888))), name = tensor<string, []>("layers_20_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577078080)))];
            tensor<fp16, [1, 1280, 1, 1]> query_81_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = var_4500, groups = var_4463, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = var_4498, weight = layers_20_self_attn_q_proj_weight_to_fp16_palettized, x = obj_241_cast_fp16)[name = tensor<string, []>("query_81_cast_fp16")];
            tensor<int32, [2]> var_4504 = const()[name = tensor<string, []>("op_4504"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4506 = const()[name = tensor<string, []>("op_4506"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_41_pad_type_0 = const()[name = tensor<string, []>("current_key_41_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_41_pad_0 = const()[name = tensor<string, []>("current_key_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577080704))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577899968))), name = tensor<string, []>("layers_20_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_41_cast_fp16 = conv(dilations = var_4506, groups = var_4463, pad = current_key_41_pad_0, pad_type = current_key_41_pad_type_0, strides = var_4504, weight = layers_20_self_attn_k_proj_weight_to_fp16_palettized, x = obj_241_cast_fp16)[name = tensor<string, []>("current_key_41_cast_fp16")];
            tensor<int32, [2]> var_4511 = const()[name = tensor<string, []>("op_4511"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4513 = const()[name = tensor<string, []>("op_4513"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_41_pad_type_0 = const()[name = tensor<string, []>("current_value_41_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_41_pad_0 = const()[name = tensor<string, []>("current_value_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577900096))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578719360))), name = tensor<string, []>("layers_20_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578719488)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = var_4513, groups = var_4463, pad = current_value_41_pad_0, pad_type = current_value_41_pad_type_0, strides = var_4511, weight = layers_20_self_attn_v_proj_weight_to_fp16_palettized, x = obj_241_cast_fp16)[name = tensor<string, []>("current_value_41_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4520_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4520_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4522_cast_fp16 = mul(x = var_103_cast_fp16_20, y = var_241_cast_fp16)[name = tensor<string, []>("op_4522_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_81_cast_fp16 = add(x = var_4520_cast_fp16, y = var_4522_cast_fp16)[name = tensor<string, []>("key_81_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4524_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4524_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4526_cast_fp16 = mul(x = var_138_cast_fp16_20, y = var_241_cast_fp16)[name = tensor<string, []>("op_4526_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_81_cast_fp16 = add(x = var_4524_cast_fp16, y = var_4526_cast_fp16)[name = tensor<string, []>("value_81_cast_fp16")];
            tensor<int32, [4]> var_4529 = const()[name = tensor<string, []>("op_4529"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4530_cast_fp16 = reshape(shape = var_4529, x = query_81_cast_fp16)[name = tensor<string, []>("op_4530_cast_fp16")];
            tensor<fp16, []> var_4531_to_fp16 = const()[name = tensor<string, []>("op_4531_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4532_cast_fp16 = mul(x = var_4530_cast_fp16, y = var_4531_to_fp16)[name = tensor<string, []>("op_4532_cast_fp16")];
            tensor<int32, [4]> var_4533 = const()[name = tensor<string, []>("op_4533"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4534_cast_fp16 = reshape(shape = var_4533, x = key_81_cast_fp16)[name = tensor<string, []>("op_4534_cast_fp16")];
            tensor<bool, []> mh_w_121_transpose_x_0 = const()[name = tensor<string, []>("mh_w_121_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_121_transpose_y_0 = const()[name = tensor<string, []>("mh_w_121_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_121_cast_fp16 = matmul(transpose_x = mh_w_121_transpose_x_0, transpose_y = mh_w_121_transpose_y_0, x = var_4532_cast_fp16, y = var_4534_cast_fp16)[name = tensor<string, []>("mh_w_121_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_123_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_4542_cast_fp16 = softmax(axis = var_4456, x = mh_w_123_cast_fp16)[name = tensor<string, []>("op_4542_cast_fp16")];
            tensor<int32, [4]> var_4543 = const()[name = tensor<string, []>("op_4543"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4544_cast_fp16 = reshape(shape = var_4543, x = value_81_cast_fp16)[name = tensor<string, []>("op_4544_cast_fp16")];
            tensor<bool, []> attn_81_transpose_x_0 = const()[name = tensor<string, []>("attn_81_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_81_transpose_y_0 = const()[name = tensor<string, []>("attn_81_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_0, transpose_y = attn_81_transpose_y_0, x = var_4544_cast_fp16, y = var_4542_cast_fp16)[name = tensor<string, []>("attn_81_cast_fp16")];
            tensor<int32, [4]> var_4547 = const()[name = tensor<string, []>("op_4547"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_201_cast_fp16 = reshape(shape = var_4547, x = attn_81_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")];
            tensor<int32, [2]> var_4551 = const()[name = tensor<string, []>("op_4551"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4553 = const()[name = tensor<string, []>("op_4553"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_247_pad_type_0 = const()[name = tensor<string, []>("obj_247_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_247_pad_0 = const()[name = tensor<string, []>("obj_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578722112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579950976))), name = tensor<string, []>("layers_20_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579951168)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_247_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = var_4553, groups = var_4463, pad = obj_247_pad_0, pad_type = obj_247_pad_type_0, strides = var_4551, weight = layers_20_self_attn_o_proj_weight_to_fp16_palettized, x = input_201_cast_fp16)[name = tensor<string, []>("obj_247_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_247_cast_fp16)[name = tensor<string, []>("inputs_123_cast_fp16")];
            tensor<int32, [1]> var_4563 = const()[name = tensor<string, []>("op_4563"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_123_cast_fp16 = reduce_mean(axes = var_4563, keep_dims = var_4464, x = inputs_123_cast_fp16)[name = tensor<string, []>("channels_mean_123_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_123_cast_fp16 = sub(x = inputs_123_cast_fp16, y = channels_mean_123_cast_fp16)[name = tensor<string, []>("zero_mean_123_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = zero_mean_123_cast_fp16)[name = tensor<string, []>("zero_mean_sq_123_cast_fp16")];
            tensor<int32, [1]> var_4567 = const()[name = tensor<string, []>("op_4567"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4568_cast_fp16 = reduce_mean(axes = var_4567, keep_dims = var_4464, x = zero_mean_sq_123_cast_fp16)[name = tensor<string, []>("op_4568_cast_fp16")];
            tensor<fp16, []> var_4569_to_fp16 = const()[name = tensor<string, []>("op_4569_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4570_cast_fp16 = add(x = var_4568_cast_fp16, y = var_4569_to_fp16)[name = tensor<string, []>("op_4570_cast_fp16")];
            tensor<fp32, []> denom_123_epsilon_0 = const()[name = tensor<string, []>("denom_123_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_123_cast_fp16 = rsqrt(epsilon = denom_123_epsilon_0, x = var_4570_cast_fp16)[name = tensor<string, []>("denom_123_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = denom_123_cast_fp16)[name = tensor<string, []>("out_123_cast_fp16")];
            tensor<fp16, [1280]> obj_249_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_249_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579953792)))];
            tensor<fp16, [1280]> obj_249_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_249_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579956416)))];
            tensor<fp16, []> obj_249_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_249_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_249_cast_fp16 = batch_norm(beta = obj_249_beta_0_to_fp16, epsilon = obj_249_epsilon_0_to_fp16, gamma = obj_249_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor<string, []>("obj_249_cast_fp16")];
            tensor<int32, [2]> var_4585 = const()[name = tensor<string, []>("op_4585"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4587 = const()[name = tensor<string, []>("op_4587"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_83_pad_type_0 = const()[name = tensor<string, []>("query_83_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_83_pad_0 = const()[name = tensor<string, []>("query_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579959040))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581597504))), name = tensor<string, []>("layers_20_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_20_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581598080)))];
            tensor<fp16, [1, 1280, 1, 1]> query_83_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_bias_to_fp16, dilations = var_4587, groups = var_4463, pad = query_83_pad_0, pad_type = query_83_pad_type_0, strides = var_4585, weight = layers_20_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_249_cast_fp16)[name = tensor<string, []>("query_83_cast_fp16")];
            tensor<int32, [2]> var_4591 = const()[name = tensor<string, []>("op_4591"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4593 = const()[name = tensor<string, []>("op_4593"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_83_pad_type_0 = const()[name = tensor<string, []>("key_83_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_83_pad_0 = const()[name = tensor<string, []>("key_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581600704)))];
            tensor<fp16, [1, 1280, 1, 1500]> key_83_cast_fp16 = conv(dilations = var_4593, groups = var_4463, pad = key_83_pad_0, pad_type = key_83_pad_type_0, strides = var_4591, weight = layers_20_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_83_cast_fp16")];
            tensor<int32, [2]> var_4598 = const()[name = tensor<string, []>("op_4598"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4600 = const()[name = tensor<string, []>("op_4600"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_83_pad_type_0 = const()[name = tensor<string, []>("value_83_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_83_pad_0 = const()[name = tensor<string, []>("value_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(584877568))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586106432))), name = tensor<string, []>("layers_20_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_20_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586106624)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_83_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_bias_to_fp16, dilations = var_4600, groups = var_4463, pad = value_83_pad_0, pad_type = value_83_pad_type_0, strides = var_4598, weight = layers_20_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_83_cast_fp16")];
            tensor<int32, [4]> var_4604 = const()[name = tensor<string, []>("op_4604"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4605_cast_fp16 = reshape(shape = var_4604, x = query_83_cast_fp16)[name = tensor<string, []>("op_4605_cast_fp16")];
            tensor<fp16, []> var_4606_to_fp16 = const()[name = tensor<string, []>("op_4606_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4607_cast_fp16 = mul(x = var_4605_cast_fp16, y = var_4606_to_fp16)[name = tensor<string, []>("op_4607_cast_fp16")];
            tensor<int32, [4]> var_4608 = const()[name = tensor<string, []>("op_4608"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4609_cast_fp16 = reshape(shape = var_4608, x = key_83_cast_fp16)[name = tensor<string, []>("op_4609_cast_fp16")];
            tensor<bool, []> mh_w_125_transpose_x_0 = const()[name = tensor<string, []>("mh_w_125_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_125_transpose_y_0 = const()[name = tensor<string, []>("mh_w_125_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_125_cast_fp16 = matmul(transpose_x = mh_w_125_transpose_x_0, transpose_y = mh_w_125_transpose_y_0, x = var_4607_cast_fp16, y = var_4609_cast_fp16)[name = tensor<string, []>("mh_w_125_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_4612_cast_fp16 = softmax(axis = var_4456, x = mh_w_125_cast_fp16)[name = tensor<string, []>("op_4612_cast_fp16")];
            tensor<int32, [4]> var_4613 = const()[name = tensor<string, []>("op_4613"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4614_cast_fp16 = reshape(shape = var_4613, x = value_83_cast_fp16)[name = tensor<string, []>("op_4614_cast_fp16")];
            tensor<bool, []> attn_83_transpose_x_0 = const()[name = tensor<string, []>("attn_83_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_83_transpose_y_0 = const()[name = tensor<string, []>("attn_83_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_83_cast_fp16 = matmul(transpose_x = attn_83_transpose_x_0, transpose_y = attn_83_transpose_y_0, x = var_4614_cast_fp16, y = var_4612_cast_fp16)[name = tensor<string, []>("attn_83_cast_fp16")];
            tensor<int32, [4]> var_4617 = const()[name = tensor<string, []>("op_4617"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_203_cast_fp16 = reshape(shape = var_4617, x = attn_83_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<int32, [2]> var_4621 = const()[name = tensor<string, []>("op_4621"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4623 = const()[name = tensor<string, []>("op_4623"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_251_pad_type_0 = const()[name = tensor<string, []>("obj_251_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_251_pad_0 = const()[name = tensor<string, []>("obj_251_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_20_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586109248)))];
            tensor<fp16, [1280]> layers_20_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589386112)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_251_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_bias_to_fp16, dilations = var_4623, groups = var_4463, pad = obj_251_pad_0, pad_type = obj_251_pad_type_0, strides = var_4621, weight = layers_20_encoder_attn_o_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("obj_251_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_251_cast_fp16)[name = tensor<string, []>("inputs_125_cast_fp16")];
            tensor<int32, [1]> var_4629 = const()[name = tensor<string, []>("op_4629"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_125_cast_fp16 = reduce_mean(axes = var_4629, keep_dims = var_4464, x = inputs_125_cast_fp16)[name = tensor<string, []>("channels_mean_125_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_125_cast_fp16 = sub(x = inputs_125_cast_fp16, y = channels_mean_125_cast_fp16)[name = tensor<string, []>("zero_mean_125_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = zero_mean_125_cast_fp16)[name = tensor<string, []>("zero_mean_sq_125_cast_fp16")];
            tensor<int32, [1]> var_4633 = const()[name = tensor<string, []>("op_4633"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4634_cast_fp16 = reduce_mean(axes = var_4633, keep_dims = var_4464, x = zero_mean_sq_125_cast_fp16)[name = tensor<string, []>("op_4634_cast_fp16")];
            tensor<fp16, []> var_4635_to_fp16 = const()[name = tensor<string, []>("op_4635_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4636_cast_fp16 = add(x = var_4634_cast_fp16, y = var_4635_to_fp16)[name = tensor<string, []>("op_4636_cast_fp16")];
            tensor<fp32, []> denom_125_epsilon_0 = const()[name = tensor<string, []>("denom_125_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_125_cast_fp16 = rsqrt(epsilon = denom_125_epsilon_0, x = var_4636_cast_fp16)[name = tensor<string, []>("denom_125_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = denom_125_cast_fp16)[name = tensor<string, []>("out_125_cast_fp16")];
            tensor<fp16, [1280]> input_205_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_205_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589388736)))];
            tensor<fp16, [1280]> input_205_beta_0_to_fp16 = const()[name = tensor<string, []>("input_205_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589391360)))];
            tensor<fp16, []> input_205_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_205_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_205_cast_fp16 = batch_norm(beta = input_205_beta_0_to_fp16, epsilon = input_205_epsilon_0_to_fp16, gamma = input_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
            tensor<int32, [2]> var_4647 = const()[name = tensor<string, []>("op_4647"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4649 = const()[name = tensor<string, []>("op_4649"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_20_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589393984))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594309248))), name = tensor<string, []>("layers_20_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_20_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594309440)))];
            tensor<fp16, [1, 5120, 1, 1]> input_207_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = var_4649, groups = var_4463, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = var_4647, weight = layers_20_fc1_weight_to_fp16_palettized, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<string, []> input_209_mode_0 = const()[name = tensor<string, []>("input_209_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_209_cast_fp16 = gelu(mode = input_209_mode_0, x = input_207_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
            tensor<int32, [2]> var_4655 = const()[name = tensor<string, []>("op_4655"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4657 = const()[name = tensor<string, []>("op_4657"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_43_pad_type_0 = const()[name = tensor<string, []>("hidden_states_43_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = tensor<string, []>("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_20_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594319744))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(600873408))), name = tensor<string, []>("layers_20_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_20_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(600873984)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_43_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = var_4657, groups = var_4463, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = var_4655, weight = layers_20_fc2_weight_to_fp16_palettized, x = input_209_cast_fp16)[name = tensor<string, []>("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor<string, []>("inputs_127_cast_fp16")];
            tensor<int32, []> var_4670 = const()[name = tensor<string, []>("op_4670"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_4677 = const()[name = tensor<string, []>("op_4677"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_4678 = const()[name = tensor<string, []>("op_4678"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_4690 = const()[name = tensor<string, []>("op_4690"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_127_cast_fp16 = reduce_mean(axes = var_4690, keep_dims = var_4678, x = inputs_127_cast_fp16)[name = tensor<string, []>("channels_mean_127_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_127_cast_fp16 = sub(x = inputs_127_cast_fp16, y = channels_mean_127_cast_fp16)[name = tensor<string, []>("zero_mean_127_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = zero_mean_127_cast_fp16)[name = tensor<string, []>("zero_mean_sq_127_cast_fp16")];
            tensor<int32, [1]> var_4694 = const()[name = tensor<string, []>("op_4694"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4695_cast_fp16 = reduce_mean(axes = var_4694, keep_dims = var_4678, x = zero_mean_sq_127_cast_fp16)[name = tensor<string, []>("op_4695_cast_fp16")];
            tensor<fp16, []> var_4696_to_fp16 = const()[name = tensor<string, []>("op_4696_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4697_cast_fp16 = add(x = var_4695_cast_fp16, y = var_4696_to_fp16)[name = tensor<string, []>("op_4697_cast_fp16")];
            tensor<fp32, []> denom_127_epsilon_0 = const()[name = tensor<string, []>("denom_127_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_127_cast_fp16 = rsqrt(epsilon = denom_127_epsilon_0, x = var_4697_cast_fp16)[name = tensor<string, []>("denom_127_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = denom_127_cast_fp16)[name = tensor<string, []>("out_127_cast_fp16")];
            tensor<fp16, [1280]> obj_253_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_253_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(600876608)))];
            tensor<fp16, [1280]> obj_253_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_253_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(600879232)))];
            tensor<fp16, []> obj_253_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_253_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_253_cast_fp16 = batch_norm(beta = obj_253_beta_0_to_fp16, epsilon = obj_253_epsilon_0_to_fp16, gamma = obj_253_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor<string, []>("obj_253_cast_fp16")];
            tensor<int32, [2]> var_4712 = const()[name = tensor<string, []>("op_4712"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4714 = const()[name = tensor<string, []>("op_4714"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_85_pad_type_0 = const()[name = tensor<string, []>("query_85_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_85_pad_0 = const()[name = tensor<string, []>("query_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(600881856))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(601701120))), name = tensor<string, []>("layers_21_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(601701248)))];
            tensor<fp16, [1, 1280, 1, 1]> query_85_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = var_4714, groups = var_4677, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = var_4712, weight = layers_21_self_attn_q_proj_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor<string, []>("query_85_cast_fp16")];
            tensor<int32, [2]> var_4718 = const()[name = tensor<string, []>("op_4718"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4720 = const()[name = tensor<string, []>("op_4720"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_43_pad_type_0 = const()[name = tensor<string, []>("current_key_43_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_43_pad_0 = const()[name = tensor<string, []>("current_key_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(601703872))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602523136))), name = tensor<string, []>("layers_21_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_43_cast_fp16 = conv(dilations = var_4720, groups = var_4677, pad = current_key_43_pad_0, pad_type = current_key_43_pad_type_0, strides = var_4718, weight = layers_21_self_attn_k_proj_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor<string, []>("current_key_43_cast_fp16")];
            tensor<int32, [2]> var_4725 = const()[name = tensor<string, []>("op_4725"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4727 = const()[name = tensor<string, []>("op_4727"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_43_pad_type_0 = const()[name = tensor<string, []>("current_value_43_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_43_pad_0 = const()[name = tensor<string, []>("current_value_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602523264))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(603342528))), name = tensor<string, []>("layers_21_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(603342656)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = var_4727, groups = var_4677, pad = current_value_43_pad_0, pad_type = current_value_43_pad_type_0, strides = var_4725, weight = layers_21_self_attn_v_proj_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor<string, []>("current_value_43_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4734_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4734_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4736_cast_fp16 = mul(x = var_103_cast_fp16_21, y = var_241_cast_fp16)[name = tensor<string, []>("op_4736_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_85_cast_fp16 = add(x = var_4734_cast_fp16, y = var_4736_cast_fp16)[name = tensor<string, []>("key_85_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4738_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4738_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4740_cast_fp16 = mul(x = var_138_cast_fp16_21, y = var_241_cast_fp16)[name = tensor<string, []>("op_4740_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_85_cast_fp16 = add(x = var_4738_cast_fp16, y = var_4740_cast_fp16)[name = tensor<string, []>("value_85_cast_fp16")];
            tensor<int32, [4]> var_4743 = const()[name = tensor<string, []>("op_4743"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4744_cast_fp16 = reshape(shape = var_4743, x = query_85_cast_fp16)[name = tensor<string, []>("op_4744_cast_fp16")];
            tensor<fp16, []> var_4745_to_fp16 = const()[name = tensor<string, []>("op_4745_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4746_cast_fp16 = mul(x = var_4744_cast_fp16, y = var_4745_to_fp16)[name = tensor<string, []>("op_4746_cast_fp16")];
            tensor<int32, [4]> var_4747 = const()[name = tensor<string, []>("op_4747"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4748_cast_fp16 = reshape(shape = var_4747, x = key_85_cast_fp16)[name = tensor<string, []>("op_4748_cast_fp16")];
            tensor<bool, []> mh_w_127_transpose_x_0 = const()[name = tensor<string, []>("mh_w_127_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_127_transpose_y_0 = const()[name = tensor<string, []>("mh_w_127_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_127_cast_fp16 = matmul(transpose_x = mh_w_127_transpose_x_0, transpose_y = mh_w_127_transpose_y_0, x = var_4746_cast_fp16, y = var_4748_cast_fp16)[name = tensor<string, []>("mh_w_127_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_129_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_4756_cast_fp16 = softmax(axis = var_4670, x = mh_w_129_cast_fp16)[name = tensor<string, []>("op_4756_cast_fp16")];
            tensor<int32, [4]> var_4757 = const()[name = tensor<string, []>("op_4757"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4758_cast_fp16 = reshape(shape = var_4757, x = value_85_cast_fp16)[name = tensor<string, []>("op_4758_cast_fp16")];
            tensor<bool, []> attn_85_transpose_x_0 = const()[name = tensor<string, []>("attn_85_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_85_transpose_y_0 = const()[name = tensor<string, []>("attn_85_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_0, transpose_y = attn_85_transpose_y_0, x = var_4758_cast_fp16, y = var_4756_cast_fp16)[name = tensor<string, []>("attn_85_cast_fp16")];
            tensor<int32, [4]> var_4761 = const()[name = tensor<string, []>("op_4761"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_211_cast_fp16 = reshape(shape = var_4761, x = attn_85_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
            tensor<int32, [2]> var_4765 = const()[name = tensor<string, []>("op_4765"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4767 = const()[name = tensor<string, []>("op_4767"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_259_pad_type_0 = const()[name = tensor<string, []>("obj_259_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_259_pad_0 = const()[name = tensor<string, []>("obj_259_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(603345280))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604574144))), name = tensor<string, []>("layers_21_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604574336)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_259_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = var_4767, groups = var_4677, pad = obj_259_pad_0, pad_type = obj_259_pad_type_0, strides = var_4765, weight = layers_21_self_attn_o_proj_weight_to_fp16_palettized, x = input_211_cast_fp16)[name = tensor<string, []>("obj_259_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_259_cast_fp16)[name = tensor<string, []>("inputs_129_cast_fp16")];
            tensor<int32, [1]> var_4777 = const()[name = tensor<string, []>("op_4777"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_129_cast_fp16 = reduce_mean(axes = var_4777, keep_dims = var_4678, x = inputs_129_cast_fp16)[name = tensor<string, []>("channels_mean_129_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_129_cast_fp16 = sub(x = inputs_129_cast_fp16, y = channels_mean_129_cast_fp16)[name = tensor<string, []>("zero_mean_129_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = zero_mean_129_cast_fp16)[name = tensor<string, []>("zero_mean_sq_129_cast_fp16")];
            tensor<int32, [1]> var_4781 = const()[name = tensor<string, []>("op_4781"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4782_cast_fp16 = reduce_mean(axes = var_4781, keep_dims = var_4678, x = zero_mean_sq_129_cast_fp16)[name = tensor<string, []>("op_4782_cast_fp16")];
            tensor<fp16, []> var_4783_to_fp16 = const()[name = tensor<string, []>("op_4783_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4784_cast_fp16 = add(x = var_4782_cast_fp16, y = var_4783_to_fp16)[name = tensor<string, []>("op_4784_cast_fp16")];
            tensor<fp32, []> denom_129_epsilon_0 = const()[name = tensor<string, []>("denom_129_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_129_cast_fp16 = rsqrt(epsilon = denom_129_epsilon_0, x = var_4784_cast_fp16)[name = tensor<string, []>("denom_129_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = denom_129_cast_fp16)[name = tensor<string, []>("out_129_cast_fp16")];
            tensor<fp16, [1280]> obj_261_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_261_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604576960)))];
            tensor<fp16, [1280]> obj_261_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_261_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604579584)))];
            tensor<fp16, []> obj_261_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_261_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_261_cast_fp16 = batch_norm(beta = obj_261_beta_0_to_fp16, epsilon = obj_261_epsilon_0_to_fp16, gamma = obj_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor<string, []>("obj_261_cast_fp16")];
            tensor<int32, [2]> var_4799 = const()[name = tensor<string, []>("op_4799"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4801 = const()[name = tensor<string, []>("op_4801"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_87_pad_type_0 = const()[name = tensor<string, []>("query_87_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_87_pad_0 = const()[name = tensor<string, []>("query_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604582208))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(605401472))), name = tensor<string, []>("layers_21_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_21_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(605401600)))];
            tensor<fp16, [1, 1280, 1, 1]> query_87_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_bias_to_fp16, dilations = var_4801, groups = var_4677, pad = query_87_pad_0, pad_type = query_87_pad_type_0, strides = var_4799, weight = layers_21_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_261_cast_fp16)[name = tensor<string, []>("query_87_cast_fp16")];
            tensor<int32, [2]> var_4805 = const()[name = tensor<string, []>("op_4805"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4807 = const()[name = tensor<string, []>("op_4807"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_87_pad_type_0 = const()[name = tensor<string, []>("key_87_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_87_pad_0 = const()[name = tensor<string, []>("key_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(605404224))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(606633088))), name = tensor<string, []>("layers_21_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_87_cast_fp16 = conv(dilations = var_4807, groups = var_4677, pad = key_87_pad_0, pad_type = key_87_pad_type_0, strides = var_4805, weight = layers_21_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_87_cast_fp16")];
            tensor<int32, [2]> var_4812 = const()[name = tensor<string, []>("op_4812"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4814 = const()[name = tensor<string, []>("op_4814"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_87_pad_type_0 = const()[name = tensor<string, []>("value_87_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_87_pad_0 = const()[name = tensor<string, []>("value_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(606633280))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(607452544))), name = tensor<string, []>("layers_21_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_21_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(607452672)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_87_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_bias_to_fp16, dilations = var_4814, groups = var_4677, pad = value_87_pad_0, pad_type = value_87_pad_type_0, strides = var_4812, weight = layers_21_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_87_cast_fp16")];
            tensor<int32, [4]> var_4818 = const()[name = tensor<string, []>("op_4818"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4819_cast_fp16 = reshape(shape = var_4818, x = query_87_cast_fp16)[name = tensor<string, []>("op_4819_cast_fp16")];
            tensor<fp16, []> var_4820_to_fp16 = const()[name = tensor<string, []>("op_4820_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4821_cast_fp16 = mul(x = var_4819_cast_fp16, y = var_4820_to_fp16)[name = tensor<string, []>("op_4821_cast_fp16")];
            tensor<int32, [4]> var_4822 = const()[name = tensor<string, []>("op_4822"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4823_cast_fp16 = reshape(shape = var_4822, x = key_87_cast_fp16)[name = tensor<string, []>("op_4823_cast_fp16")];
            tensor<bool, []> mh_w_131_transpose_x_0 = const()[name = tensor<string, []>("mh_w_131_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_131_transpose_y_0 = const()[name = tensor<string, []>("mh_w_131_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_131_cast_fp16 = matmul(transpose_x = mh_w_131_transpose_x_0, transpose_y = mh_w_131_transpose_y_0, x = var_4821_cast_fp16, y = var_4823_cast_fp16)[name = tensor<string, []>("mh_w_131_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_4826_cast_fp16 = softmax(axis = var_4670, x = mh_w_131_cast_fp16)[name = tensor<string, []>("op_4826_cast_fp16")];
            tensor<int32, [4]> var_4827 = const()[name = tensor<string, []>("op_4827"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_4828_cast_fp16 = reshape(shape = var_4827, x = value_87_cast_fp16)[name = tensor<string, []>("op_4828_cast_fp16")];
            tensor<bool, []> attn_87_transpose_x_0 = const()[name = tensor<string, []>("attn_87_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_87_transpose_y_0 = const()[name = tensor<string, []>("attn_87_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_87_cast_fp16 = matmul(transpose_x = attn_87_transpose_x_0, transpose_y = attn_87_transpose_y_0, x = var_4828_cast_fp16, y = var_4826_cast_fp16)[name = tensor<string, []>("attn_87_cast_fp16")];
            tensor<int32, [4]> var_4831 = const()[name = tensor<string, []>("op_4831"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_213_cast_fp16 = reshape(shape = var_4831, x = attn_87_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
            tensor<int32, [2]> var_4835 = const()[name = tensor<string, []>("op_4835"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4837 = const()[name = tensor<string, []>("op_4837"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_263_pad_type_0 = const()[name = tensor<string, []>("obj_263_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_263_pad_0 = const()[name = tensor<string, []>("obj_263_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_21_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(607455296))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(609093760))), name = tensor<string, []>("layers_21_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_21_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(609094336)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_263_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_bias_to_fp16, dilations = var_4837, groups = var_4677, pad = obj_263_pad_0, pad_type = obj_263_pad_type_0, strides = var_4835, weight = layers_21_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = tensor<string, []>("obj_263_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_263_cast_fp16)[name = tensor<string, []>("inputs_131_cast_fp16")];
            tensor<int32, [1]> var_4843 = const()[name = tensor<string, []>("op_4843"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_131_cast_fp16 = reduce_mean(axes = var_4843, keep_dims = var_4678, x = inputs_131_cast_fp16)[name = tensor<string, []>("channels_mean_131_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_131_cast_fp16 = sub(x = inputs_131_cast_fp16, y = channels_mean_131_cast_fp16)[name = tensor<string, []>("zero_mean_131_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = zero_mean_131_cast_fp16)[name = tensor<string, []>("zero_mean_sq_131_cast_fp16")];
            tensor<int32, [1]> var_4847 = const()[name = tensor<string, []>("op_4847"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4848_cast_fp16 = reduce_mean(axes = var_4847, keep_dims = var_4678, x = zero_mean_sq_131_cast_fp16)[name = tensor<string, []>("op_4848_cast_fp16")];
            tensor<fp16, []> var_4849_to_fp16 = const()[name = tensor<string, []>("op_4849_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4850_cast_fp16 = add(x = var_4848_cast_fp16, y = var_4849_to_fp16)[name = tensor<string, []>("op_4850_cast_fp16")];
            tensor<fp32, []> denom_131_epsilon_0 = const()[name = tensor<string, []>("denom_131_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_131_cast_fp16 = rsqrt(epsilon = denom_131_epsilon_0, x = var_4850_cast_fp16)[name = tensor<string, []>("denom_131_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = denom_131_cast_fp16)[name = tensor<string, []>("out_131_cast_fp16")];
            tensor<fp16, [1280]> input_215_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_215_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(609096960)))];
            tensor<fp16, [1280]> input_215_beta_0_to_fp16 = const()[name = tensor<string, []>("input_215_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(609099584)))];
            tensor<fp16, []> input_215_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_215_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_215_cast_fp16 = batch_norm(beta = input_215_beta_0_to_fp16, epsilon = input_215_epsilon_0_to_fp16, gamma = input_215_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
            tensor<int32, [2]> var_4861 = const()[name = tensor<string, []>("op_4861"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4863 = const()[name = tensor<string, []>("op_4863"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_217_pad_type_0 = const()[name = tensor<string, []>("input_217_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_217_pad_0 = const()[name = tensor<string, []>("input_217_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_21_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(609102208))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(614017472))), name = tensor<string, []>("layers_21_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_21_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(614017664)))];
            tensor<fp16, [1, 5120, 1, 1]> input_217_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = var_4863, groups = var_4677, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = var_4861, weight = layers_21_fc1_weight_to_fp16_palettized, x = input_215_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
            tensor<string, []> input_219_mode_0 = const()[name = tensor<string, []>("input_219_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<int32, [2]> var_4869 = const()[name = tensor<string, []>("op_4869"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4871 = const()[name = tensor<string, []>("op_4871"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_45_pad_type_0 = const()[name = tensor<string, []>("hidden_states_45_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_45_pad_0 = const()[name = tensor<string, []>("hidden_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_21_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(614027968))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(620581632))), name = tensor<string, []>("layers_21_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_21_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(620582208)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_45_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = var_4871, groups = var_4677, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = var_4869, weight = layers_21_fc2_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = tensor<string, []>("hidden_states_45_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor<string, []>("inputs_133_cast_fp16")];
            tensor<int32, []> var_4884 = const()[name = tensor<string, []>("op_4884"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_4891 = const()[name = tensor<string, []>("op_4891"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_4892 = const()[name = tensor<string, []>("op_4892"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_4904 = const()[name = tensor<string, []>("op_4904"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_133_cast_fp16 = reduce_mean(axes = var_4904, keep_dims = var_4892, x = inputs_133_cast_fp16)[name = tensor<string, []>("channels_mean_133_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_133_cast_fp16 = sub(x = inputs_133_cast_fp16, y = channels_mean_133_cast_fp16)[name = tensor<string, []>("zero_mean_133_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = zero_mean_133_cast_fp16)[name = tensor<string, []>("zero_mean_sq_133_cast_fp16")];
            tensor<int32, [1]> var_4908 = const()[name = tensor<string, []>("op_4908"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4909_cast_fp16 = reduce_mean(axes = var_4908, keep_dims = var_4892, x = zero_mean_sq_133_cast_fp16)[name = tensor<string, []>("op_4909_cast_fp16")];
            tensor<fp16, []> var_4910_to_fp16 = const()[name = tensor<string, []>("op_4910_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4911_cast_fp16 = add(x = var_4909_cast_fp16, y = var_4910_to_fp16)[name = tensor<string, []>("op_4911_cast_fp16")];
            tensor<fp32, []> denom_133_epsilon_0 = const()[name = tensor<string, []>("denom_133_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_133_cast_fp16 = rsqrt(epsilon = denom_133_epsilon_0, x = var_4911_cast_fp16)[name = tensor<string, []>("denom_133_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = denom_133_cast_fp16)[name = tensor<string, []>("out_133_cast_fp16")];
            tensor<fp16, [1280]> obj_265_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_265_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(620584832)))];
            tensor<fp16, [1280]> obj_265_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_265_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(620587456)))];
            tensor<fp16, []> obj_265_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_265_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_265_cast_fp16 = batch_norm(beta = obj_265_beta_0_to_fp16, epsilon = obj_265_epsilon_0_to_fp16, gamma = obj_265_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor<string, []>("obj_265_cast_fp16")];
            tensor<int32, [2]> var_4926 = const()[name = tensor<string, []>("op_4926"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4928 = const()[name = tensor<string, []>("op_4928"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_89_pad_type_0 = const()[name = tensor<string, []>("query_89_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_89_pad_0 = const()[name = tensor<string, []>("query_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(620590080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(621818944))), name = tensor<string, []>("layers_22_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(621819136)))];
            tensor<fp16, [1, 1280, 1, 1]> query_89_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = var_4928, groups = var_4891, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = var_4926, weight = layers_22_self_attn_q_proj_weight_to_fp16_palettized, x = obj_265_cast_fp16)[name = tensor<string, []>("query_89_cast_fp16")];
            tensor<int32, [2]> var_4932 = const()[name = tensor<string, []>("op_4932"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4934 = const()[name = tensor<string, []>("op_4934"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_45_pad_type_0 = const()[name = tensor<string, []>("current_key_45_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_45_pad_0 = const()[name = tensor<string, []>("current_key_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(621821760))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622641024))), name = tensor<string, []>("layers_22_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_45_cast_fp16 = conv(dilations = var_4934, groups = var_4891, pad = current_key_45_pad_0, pad_type = current_key_45_pad_type_0, strides = var_4932, weight = layers_22_self_attn_k_proj_weight_to_fp16_palettized, x = obj_265_cast_fp16)[name = tensor<string, []>("current_key_45_cast_fp16")];
            tensor<int32, [2]> var_4939 = const()[name = tensor<string, []>("op_4939"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4941 = const()[name = tensor<string, []>("op_4941"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_45_pad_type_0 = const()[name = tensor<string, []>("current_value_45_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_45_pad_0 = const()[name = tensor<string, []>("current_value_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622641152))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623460416))), name = tensor<string, []>("layers_22_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623460544)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = var_4941, groups = var_4891, pad = current_value_45_pad_0, pad_type = current_value_45_pad_type_0, strides = var_4939, weight = layers_22_self_attn_v_proj_weight_to_fp16_palettized, x = obj_265_cast_fp16)[name = tensor<string, []>("current_value_45_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4948_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4948_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4950_cast_fp16 = mul(x = var_103_cast_fp16_22, y = var_241_cast_fp16)[name = tensor<string, []>("op_4950_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_89_cast_fp16 = add(x = var_4948_cast_fp16, y = var_4950_cast_fp16)[name = tensor<string, []>("key_89_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4952_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_4952_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_4954_cast_fp16 = mul(x = var_138_cast_fp16_22, y = var_241_cast_fp16)[name = tensor<string, []>("op_4954_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_89_cast_fp16 = add(x = var_4952_cast_fp16, y = var_4954_cast_fp16)[name = tensor<string, []>("value_89_cast_fp16")];
            tensor<int32, [4]> var_4957 = const()[name = tensor<string, []>("op_4957"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_4958_cast_fp16 = reshape(shape = var_4957, x = query_89_cast_fp16)[name = tensor<string, []>("op_4958_cast_fp16")];
            tensor<fp16, []> var_4959_to_fp16 = const()[name = tensor<string, []>("op_4959_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_4960_cast_fp16 = mul(x = var_4958_cast_fp16, y = var_4959_to_fp16)[name = tensor<string, []>("op_4960_cast_fp16")];
            tensor<int32, [4]> var_4961 = const()[name = tensor<string, []>("op_4961"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4962_cast_fp16 = reshape(shape = var_4961, x = key_89_cast_fp16)[name = tensor<string, []>("op_4962_cast_fp16")];
            tensor<bool, []> mh_w_133_transpose_x_0 = const()[name = tensor<string, []>("mh_w_133_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_133_transpose_y_0 = const()[name = tensor<string, []>("mh_w_133_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_133_cast_fp16 = matmul(transpose_x = mh_w_133_transpose_x_0, transpose_y = mh_w_133_transpose_y_0, x = var_4960_cast_fp16, y = var_4962_cast_fp16)[name = tensor<string, []>("mh_w_133_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_135_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_4970_cast_fp16 = softmax(axis = var_4884, x = mh_w_135_cast_fp16)[name = tensor<string, []>("op_4970_cast_fp16")];
            tensor<int32, [4]> var_4971 = const()[name = tensor<string, []>("op_4971"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_4972_cast_fp16 = reshape(shape = var_4971, x = value_89_cast_fp16)[name = tensor<string, []>("op_4972_cast_fp16")];
            tensor<bool, []> attn_89_transpose_x_0 = const()[name = tensor<string, []>("attn_89_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_89_transpose_y_0 = const()[name = tensor<string, []>("attn_89_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_0, transpose_y = attn_89_transpose_y_0, x = var_4972_cast_fp16, y = var_4970_cast_fp16)[name = tensor<string, []>("attn_89_cast_fp16")];
            tensor<int32, [4]> var_4975 = const()[name = tensor<string, []>("op_4975"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_221_cast_fp16 = reshape(shape = var_4975, x = attn_89_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
            tensor<int32, [2]> var_4979 = const()[name = tensor<string, []>("op_4979"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4981 = const()[name = tensor<string, []>("op_4981"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_271_pad_type_0 = const()[name = tensor<string, []>("obj_271_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_271_pad_0 = const()[name = tensor<string, []>("obj_271_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623463168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(624692032))), name = tensor<string, []>("layers_22_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(624692224)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_271_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = var_4981, groups = var_4891, pad = obj_271_pad_0, pad_type = obj_271_pad_type_0, strides = var_4979, weight = layers_22_self_attn_o_proj_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = tensor<string, []>("obj_271_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_271_cast_fp16)[name = tensor<string, []>("inputs_135_cast_fp16")];
            tensor<int32, [1]> var_4991 = const()[name = tensor<string, []>("op_4991"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_135_cast_fp16 = reduce_mean(axes = var_4991, keep_dims = var_4892, x = inputs_135_cast_fp16)[name = tensor<string, []>("channels_mean_135_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_135_cast_fp16 = sub(x = inputs_135_cast_fp16, y = channels_mean_135_cast_fp16)[name = tensor<string, []>("zero_mean_135_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = zero_mean_135_cast_fp16)[name = tensor<string, []>("zero_mean_sq_135_cast_fp16")];
            tensor<int32, [1]> var_4995 = const()[name = tensor<string, []>("op_4995"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_4996_cast_fp16 = reduce_mean(axes = var_4995, keep_dims = var_4892, x = zero_mean_sq_135_cast_fp16)[name = tensor<string, []>("op_4996_cast_fp16")];
            tensor<fp16, []> var_4997_to_fp16 = const()[name = tensor<string, []>("op_4997_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_4998_cast_fp16 = add(x = var_4996_cast_fp16, y = var_4997_to_fp16)[name = tensor<string, []>("op_4998_cast_fp16")];
            tensor<fp32, []> denom_135_epsilon_0 = const()[name = tensor<string, []>("denom_135_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_135_cast_fp16 = rsqrt(epsilon = denom_135_epsilon_0, x = var_4998_cast_fp16)[name = tensor<string, []>("denom_135_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = denom_135_cast_fp16)[name = tensor<string, []>("out_135_cast_fp16")];
            tensor<fp16, [1280]> obj_273_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_273_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(624694848)))];
            tensor<fp16, [1280]> obj_273_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_273_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(624697472)))];
            tensor<fp16, []> obj_273_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_273_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_273_cast_fp16 = batch_norm(beta = obj_273_beta_0_to_fp16, epsilon = obj_273_epsilon_0_to_fp16, gamma = obj_273_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor<string, []>("obj_273_cast_fp16")];
            tensor<int32, [2]> var_5013 = const()[name = tensor<string, []>("op_5013"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5015 = const()[name = tensor<string, []>("op_5015"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_91_pad_type_0 = const()[name = tensor<string, []>("query_91_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_91_pad_0 = const()[name = tensor<string, []>("query_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(624700096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(625928960))), name = tensor<string, []>("layers_22_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_22_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(625929152)))];
            tensor<fp16, [1, 1280, 1, 1]> query_91_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_bias_to_fp16, dilations = var_5015, groups = var_4891, pad = query_91_pad_0, pad_type = query_91_pad_type_0, strides = var_5013, weight = layers_22_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_273_cast_fp16)[name = tensor<string, []>("query_91_cast_fp16")];
            tensor<int32, [2]> var_5019 = const()[name = tensor<string, []>("op_5019"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5021 = const()[name = tensor<string, []>("op_5021"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_91_pad_type_0 = const()[name = tensor<string, []>("key_91_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_91_pad_0 = const()[name = tensor<string, []>("key_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(625931776))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(626751040))), name = tensor<string, []>("layers_22_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_91_cast_fp16 = conv(dilations = var_5021, groups = var_4891, pad = key_91_pad_0, pad_type = key_91_pad_type_0, strides = var_5019, weight = layers_22_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_91_cast_fp16")];
            tensor<int32, [2]> var_5026 = const()[name = tensor<string, []>("op_5026"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5028 = const()[name = tensor<string, []>("op_5028"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_91_pad_type_0 = const()[name = tensor<string, []>("value_91_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_91_pad_0 = const()[name = tensor<string, []>("value_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(626751168))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627570432))), name = tensor<string, []>("layers_22_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_22_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627570560)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_91_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_bias_to_fp16, dilations = var_5028, groups = var_4891, pad = value_91_pad_0, pad_type = value_91_pad_type_0, strides = var_5026, weight = layers_22_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_91_cast_fp16")];
            tensor<int32, [4]> var_5032 = const()[name = tensor<string, []>("op_5032"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5033_cast_fp16 = reshape(shape = var_5032, x = query_91_cast_fp16)[name = tensor<string, []>("op_5033_cast_fp16")];
            tensor<fp16, []> var_5034_to_fp16 = const()[name = tensor<string, []>("op_5034_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5035_cast_fp16 = mul(x = var_5033_cast_fp16, y = var_5034_to_fp16)[name = tensor<string, []>("op_5035_cast_fp16")];
            tensor<int32, [4]> var_5036 = const()[name = tensor<string, []>("op_5036"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5037_cast_fp16 = reshape(shape = var_5036, x = key_91_cast_fp16)[name = tensor<string, []>("op_5037_cast_fp16")];
            tensor<bool, []> mh_w_137_transpose_x_0 = const()[name = tensor<string, []>("mh_w_137_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_137_transpose_y_0 = const()[name = tensor<string, []>("mh_w_137_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_137_cast_fp16 = matmul(transpose_x = mh_w_137_transpose_x_0, transpose_y = mh_w_137_transpose_y_0, x = var_5035_cast_fp16, y = var_5037_cast_fp16)[name = tensor<string, []>("mh_w_137_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_5040_cast_fp16 = softmax(axis = var_4884, x = mh_w_137_cast_fp16)[name = tensor<string, []>("op_5040_cast_fp16")];
            tensor<int32, [4]> var_5041 = const()[name = tensor<string, []>("op_5041"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5042_cast_fp16 = reshape(shape = var_5041, x = value_91_cast_fp16)[name = tensor<string, []>("op_5042_cast_fp16")];
            tensor<bool, []> attn_91_transpose_x_0 = const()[name = tensor<string, []>("attn_91_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_91_transpose_y_0 = const()[name = tensor<string, []>("attn_91_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_91_cast_fp16 = matmul(transpose_x = attn_91_transpose_x_0, transpose_y = attn_91_transpose_y_0, x = var_5042_cast_fp16, y = var_5040_cast_fp16)[name = tensor<string, []>("attn_91_cast_fp16")];
            tensor<int32, [4]> var_5045 = const()[name = tensor<string, []>("op_5045"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_223_cast_fp16 = reshape(shape = var_5045, x = attn_91_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
            tensor<int32, [2]> var_5049 = const()[name = tensor<string, []>("op_5049"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5051 = const()[name = tensor<string, []>("op_5051"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_275_pad_type_0 = const()[name = tensor<string, []>("obj_275_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_275_pad_0 = const()[name = tensor<string, []>("obj_275_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_22_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627573184))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(628802048))), name = tensor<string, []>("layers_22_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_22_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(628802240)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_275_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_bias_to_fp16, dilations = var_5051, groups = var_4891, pad = obj_275_pad_0, pad_type = obj_275_pad_type_0, strides = var_5049, weight = layers_22_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_223_cast_fp16)[name = tensor<string, []>("obj_275_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_275_cast_fp16)[name = tensor<string, []>("inputs_137_cast_fp16")];
            tensor<int32, [1]> var_5057 = const()[name = tensor<string, []>("op_5057"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_137_cast_fp16 = reduce_mean(axes = var_5057, keep_dims = var_4892, x = inputs_137_cast_fp16)[name = tensor<string, []>("channels_mean_137_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_137_cast_fp16 = sub(x = inputs_137_cast_fp16, y = channels_mean_137_cast_fp16)[name = tensor<string, []>("zero_mean_137_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = zero_mean_137_cast_fp16)[name = tensor<string, []>("zero_mean_sq_137_cast_fp16")];
            tensor<int32, [1]> var_5061 = const()[name = tensor<string, []>("op_5061"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5062_cast_fp16 = reduce_mean(axes = var_5061, keep_dims = var_4892, x = zero_mean_sq_137_cast_fp16)[name = tensor<string, []>("op_5062_cast_fp16")];
            tensor<fp16, []> var_5063_to_fp16 = const()[name = tensor<string, []>("op_5063_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5064_cast_fp16 = add(x = var_5062_cast_fp16, y = var_5063_to_fp16)[name = tensor<string, []>("op_5064_cast_fp16")];
            tensor<fp32, []> denom_137_epsilon_0 = const()[name = tensor<string, []>("denom_137_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_137_cast_fp16 = rsqrt(epsilon = denom_137_epsilon_0, x = var_5064_cast_fp16)[name = tensor<string, []>("denom_137_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = denom_137_cast_fp16)[name = tensor<string, []>("out_137_cast_fp16")];
            tensor<fp16, [1280]> input_225_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_225_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(628804864)))];
            tensor<fp16, [1280]> input_225_beta_0_to_fp16 = const()[name = tensor<string, []>("input_225_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(628807488)))];
            tensor<fp16, []> input_225_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_225_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_225_cast_fp16 = batch_norm(beta = input_225_beta_0_to_fp16, epsilon = input_225_epsilon_0_to_fp16, gamma = input_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
            tensor<int32, [2]> var_5075 = const()[name = tensor<string, []>("op_5075"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5077 = const()[name = tensor<string, []>("op_5077"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_227_pad_type_0 = const()[name = tensor<string, []>("input_227_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_227_pad_0 = const()[name = tensor<string, []>("input_227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_22_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(628810112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(633725376))), name = tensor<string, []>("layers_22_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_22_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(633725568)))];
            tensor<fp16, [1, 5120, 1, 1]> input_227_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = var_5077, groups = var_4891, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = var_5075, weight = layers_22_fc1_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
            tensor<string, []> input_229_mode_0 = const()[name = tensor<string, []>("input_229_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = input_227_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
            tensor<int32, [2]> var_5083 = const()[name = tensor<string, []>("op_5083"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5085 = const()[name = tensor<string, []>("op_5085"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_47_pad_type_0 = const()[name = tensor<string, []>("hidden_states_47_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = tensor<string, []>("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_22_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(633735872))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(640289536))), name = tensor<string, []>("layers_22_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_22_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(640290112)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_47_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = var_5085, groups = var_4891, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = var_5083, weight = layers_22_fc2_weight_to_fp16_palettized, x = input_229_cast_fp16)[name = tensor<string, []>("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor<string, []>("inputs_139_cast_fp16")];
            tensor<int32, []> var_5098 = const()[name = tensor<string, []>("op_5098"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_5105 = const()[name = tensor<string, []>("op_5105"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_5106 = const()[name = tensor<string, []>("op_5106"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_5118 = const()[name = tensor<string, []>("op_5118"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_139_cast_fp16 = reduce_mean(axes = var_5118, keep_dims = var_5106, x = inputs_139_cast_fp16)[name = tensor<string, []>("channels_mean_139_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_139_cast_fp16 = sub(x = inputs_139_cast_fp16, y = channels_mean_139_cast_fp16)[name = tensor<string, []>("zero_mean_139_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = zero_mean_139_cast_fp16)[name = tensor<string, []>("zero_mean_sq_139_cast_fp16")];
            tensor<int32, [1]> var_5122 = const()[name = tensor<string, []>("op_5122"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5123_cast_fp16 = reduce_mean(axes = var_5122, keep_dims = var_5106, x = zero_mean_sq_139_cast_fp16)[name = tensor<string, []>("op_5123_cast_fp16")];
            tensor<fp16, []> var_5124_to_fp16 = const()[name = tensor<string, []>("op_5124_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5125_cast_fp16 = add(x = var_5123_cast_fp16, y = var_5124_to_fp16)[name = tensor<string, []>("op_5125_cast_fp16")];
            tensor<fp32, []> denom_139_epsilon_0 = const()[name = tensor<string, []>("denom_139_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_139_cast_fp16 = rsqrt(epsilon = denom_139_epsilon_0, x = var_5125_cast_fp16)[name = tensor<string, []>("denom_139_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = denom_139_cast_fp16)[name = tensor<string, []>("out_139_cast_fp16")];
            tensor<fp16, [1280]> obj_277_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_277_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(640292736)))];
            tensor<fp16, [1280]> obj_277_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_277_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(640295360)))];
            tensor<fp16, []> obj_277_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_277_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_277_cast_fp16 = batch_norm(beta = obj_277_beta_0_to_fp16, epsilon = obj_277_epsilon_0_to_fp16, gamma = obj_277_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor<string, []>("obj_277_cast_fp16")];
            tensor<int32, [2]> var_5140 = const()[name = tensor<string, []>("op_5140"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5142 = const()[name = tensor<string, []>("op_5142"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_93_pad_type_0 = const()[name = tensor<string, []>("query_93_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_93_pad_0 = const()[name = tensor<string, []>("query_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(640297984))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641117248))), name = tensor<string, []>("layers_23_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641117376)))];
            tensor<fp16, [1, 1280, 1, 1]> query_93_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = var_5142, groups = var_5105, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = var_5140, weight = layers_23_self_attn_q_proj_weight_to_fp16_palettized, x = obj_277_cast_fp16)[name = tensor<string, []>("query_93_cast_fp16")];
            tensor<int32, [2]> var_5146 = const()[name = tensor<string, []>("op_5146"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5148 = const()[name = tensor<string, []>("op_5148"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_47_pad_type_0 = const()[name = tensor<string, []>("current_key_47_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_47_pad_0 = const()[name = tensor<string, []>("current_key_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641120000))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641939264))), name = tensor<string, []>("layers_23_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_47_cast_fp16 = conv(dilations = var_5148, groups = var_5105, pad = current_key_47_pad_0, pad_type = current_key_47_pad_type_0, strides = var_5146, weight = layers_23_self_attn_k_proj_weight_to_fp16_palettized, x = obj_277_cast_fp16)[name = tensor<string, []>("current_key_47_cast_fp16")];
            tensor<int32, [2]> var_5153 = const()[name = tensor<string, []>("op_5153"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5155 = const()[name = tensor<string, []>("op_5155"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_47_pad_type_0 = const()[name = tensor<string, []>("current_value_47_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_47_pad_0 = const()[name = tensor<string, []>("current_value_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641939392))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(642758656))), name = tensor<string, []>("layers_23_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(642758784)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_47_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = var_5155, groups = var_5105, pad = current_value_47_pad_0, pad_type = current_value_47_pad_type_0, strides = var_5153, weight = layers_23_self_attn_v_proj_weight_to_fp16_palettized, x = obj_277_cast_fp16)[name = tensor<string, []>("current_value_47_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5162_cast_fp16 = mul(x = current_key_47_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_5162_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5164_cast_fp16 = mul(x = var_103_cast_fp16_23, y = var_241_cast_fp16)[name = tensor<string, []>("op_5164_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_93_cast_fp16 = add(x = var_5162_cast_fp16, y = var_5164_cast_fp16)[name = tensor<string, []>("key_93_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5166_cast_fp16 = mul(x = current_value_47_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_5166_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5168_cast_fp16 = mul(x = var_138_cast_fp16_23, y = var_241_cast_fp16)[name = tensor<string, []>("op_5168_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_93_cast_fp16 = add(x = var_5166_cast_fp16, y = var_5168_cast_fp16)[name = tensor<string, []>("value_93_cast_fp16")];
            tensor<int32, [4]> var_5171 = const()[name = tensor<string, []>("op_5171"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5172_cast_fp16 = reshape(shape = var_5171, x = query_93_cast_fp16)[name = tensor<string, []>("op_5172_cast_fp16")];
            tensor<fp16, []> var_5173_to_fp16 = const()[name = tensor<string, []>("op_5173_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5174_cast_fp16 = mul(x = var_5172_cast_fp16, y = var_5173_to_fp16)[name = tensor<string, []>("op_5174_cast_fp16")];
            tensor<int32, [4]> var_5175 = const()[name = tensor<string, []>("op_5175"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_5176_cast_fp16 = reshape(shape = var_5175, x = key_93_cast_fp16)[name = tensor<string, []>("op_5176_cast_fp16")];
            tensor<bool, []> mh_w_139_transpose_x_0 = const()[name = tensor<string, []>("mh_w_139_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_139_transpose_y_0 = const()[name = tensor<string, []>("mh_w_139_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_139_cast_fp16 = matmul(transpose_x = mh_w_139_transpose_x_0, transpose_y = mh_w_139_transpose_y_0, x = var_5174_cast_fp16, y = var_5176_cast_fp16)[name = tensor<string, []>("mh_w_139_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_141_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_5184_cast_fp16 = softmax(axis = var_5098, x = mh_w_141_cast_fp16)[name = tensor<string, []>("op_5184_cast_fp16")];
            tensor<int32, [4]> var_5185 = const()[name = tensor<string, []>("op_5185"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_5186_cast_fp16 = reshape(shape = var_5185, x = value_93_cast_fp16)[name = tensor<string, []>("op_5186_cast_fp16")];
            tensor<bool, []> attn_93_transpose_x_0 = const()[name = tensor<string, []>("attn_93_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_93_transpose_y_0 = const()[name = tensor<string, []>("attn_93_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_0, transpose_y = attn_93_transpose_y_0, x = var_5186_cast_fp16, y = var_5184_cast_fp16)[name = tensor<string, []>("attn_93_cast_fp16")];
            tensor<int32, [4]> var_5189 = const()[name = tensor<string, []>("op_5189"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_231_cast_fp16 = reshape(shape = var_5189, x = attn_93_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
            tensor<int32, [2]> var_5193 = const()[name = tensor<string, []>("op_5193"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5195 = const()[name = tensor<string, []>("op_5195"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_283_pad_type_0 = const()[name = tensor<string, []>("obj_283_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_283_pad_0 = const()[name = tensor<string, []>("obj_283_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(642761408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(643580672))), name = tensor<string, []>("layers_23_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(643580800)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_283_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = var_5195, groups = var_5105, pad = obj_283_pad_0, pad_type = obj_283_pad_type_0, strides = var_5193, weight = layers_23_self_attn_o_proj_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = tensor<string, []>("obj_283_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_283_cast_fp16)[name = tensor<string, []>("inputs_141_cast_fp16")];
            tensor<int32, [1]> var_5205 = const()[name = tensor<string, []>("op_5205"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_141_cast_fp16 = reduce_mean(axes = var_5205, keep_dims = var_5106, x = inputs_141_cast_fp16)[name = tensor<string, []>("channels_mean_141_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_141_cast_fp16 = sub(x = inputs_141_cast_fp16, y = channels_mean_141_cast_fp16)[name = tensor<string, []>("zero_mean_141_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = zero_mean_141_cast_fp16)[name = tensor<string, []>("zero_mean_sq_141_cast_fp16")];
            tensor<int32, [1]> var_5209 = const()[name = tensor<string, []>("op_5209"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5210_cast_fp16 = reduce_mean(axes = var_5209, keep_dims = var_5106, x = zero_mean_sq_141_cast_fp16)[name = tensor<string, []>("op_5210_cast_fp16")];
            tensor<fp16, []> var_5211_to_fp16 = const()[name = tensor<string, []>("op_5211_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5212_cast_fp16 = add(x = var_5210_cast_fp16, y = var_5211_to_fp16)[name = tensor<string, []>("op_5212_cast_fp16")];
            tensor<fp32, []> denom_141_epsilon_0 = const()[name = tensor<string, []>("denom_141_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_141_cast_fp16 = rsqrt(epsilon = denom_141_epsilon_0, x = var_5212_cast_fp16)[name = tensor<string, []>("denom_141_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = denom_141_cast_fp16)[name = tensor<string, []>("out_141_cast_fp16")];
            tensor<fp16, [1280]> obj_285_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_285_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(643583424)))];
            tensor<fp16, [1280]> obj_285_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_285_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(643586048)))];
            tensor<fp16, []> obj_285_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_285_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_285_cast_fp16 = batch_norm(beta = obj_285_beta_0_to_fp16, epsilon = obj_285_epsilon_0_to_fp16, gamma = obj_285_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor<string, []>("obj_285_cast_fp16")];
            tensor<int32, [2]> var_5227 = const()[name = tensor<string, []>("op_5227"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5229 = const()[name = tensor<string, []>("op_5229"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_95_pad_type_0 = const()[name = tensor<string, []>("query_95_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_95_pad_0 = const()[name = tensor<string, []>("query_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(643588672))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644817536))), name = tensor<string, []>("layers_23_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_23_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644817728)))];
            tensor<fp16, [1, 1280, 1, 1]> query_95_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_bias_to_fp16, dilations = var_5229, groups = var_5105, pad = query_95_pad_0, pad_type = query_95_pad_type_0, strides = var_5227, weight = layers_23_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_285_cast_fp16)[name = tensor<string, []>("query_95_cast_fp16")];
            tensor<int32, [2]> var_5233 = const()[name = tensor<string, []>("op_5233"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5235 = const()[name = tensor<string, []>("op_5235"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_95_pad_type_0 = const()[name = tensor<string, []>("key_95_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_95_pad_0 = const()[name = tensor<string, []>("key_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644820352))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(645639616))), name = tensor<string, []>("layers_23_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_95_cast_fp16 = conv(dilations = var_5235, groups = var_5105, pad = key_95_pad_0, pad_type = key_95_pad_type_0, strides = var_5233, weight = layers_23_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_95_cast_fp16")];
            tensor<int32, [2]> var_5240 = const()[name = tensor<string, []>("op_5240"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5242 = const()[name = tensor<string, []>("op_5242"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_95_pad_type_0 = const()[name = tensor<string, []>("value_95_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_95_pad_0 = const()[name = tensor<string, []>("value_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(645639744))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647278208))), name = tensor<string, []>("layers_23_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_23_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647278784)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_95_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_bias_to_fp16, dilations = var_5242, groups = var_5105, pad = value_95_pad_0, pad_type = value_95_pad_type_0, strides = var_5240, weight = layers_23_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_95_cast_fp16")];
            tensor<int32, [4]> var_5246 = const()[name = tensor<string, []>("op_5246"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5247_cast_fp16 = reshape(shape = var_5246, x = query_95_cast_fp16)[name = tensor<string, []>("op_5247_cast_fp16")];
            tensor<fp16, []> var_5248_to_fp16 = const()[name = tensor<string, []>("op_5248_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5249_cast_fp16 = mul(x = var_5247_cast_fp16, y = var_5248_to_fp16)[name = tensor<string, []>("op_5249_cast_fp16")];
            tensor<int32, [4]> var_5250 = const()[name = tensor<string, []>("op_5250"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5251_cast_fp16 = reshape(shape = var_5250, x = key_95_cast_fp16)[name = tensor<string, []>("op_5251_cast_fp16")];
            tensor<bool, []> mh_w_143_transpose_x_0 = const()[name = tensor<string, []>("mh_w_143_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_143_transpose_y_0 = const()[name = tensor<string, []>("mh_w_143_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_143_cast_fp16 = matmul(transpose_x = mh_w_143_transpose_x_0, transpose_y = mh_w_143_transpose_y_0, x = var_5249_cast_fp16, y = var_5251_cast_fp16)[name = tensor<string, []>("mh_w_143_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_5254_cast_fp16 = softmax(axis = var_5098, x = mh_w_143_cast_fp16)[name = tensor<string, []>("op_5254_cast_fp16")];
            tensor<int32, [4]> var_5255 = const()[name = tensor<string, []>("op_5255"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5256_cast_fp16 = reshape(shape = var_5255, x = value_95_cast_fp16)[name = tensor<string, []>("op_5256_cast_fp16")];
            tensor<bool, []> attn_95_transpose_x_0 = const()[name = tensor<string, []>("attn_95_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_95_transpose_y_0 = const()[name = tensor<string, []>("attn_95_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_95_cast_fp16 = matmul(transpose_x = attn_95_transpose_x_0, transpose_y = attn_95_transpose_y_0, x = var_5256_cast_fp16, y = var_5254_cast_fp16)[name = tensor<string, []>("attn_95_cast_fp16")];
            tensor<int32, [4]> var_5259 = const()[name = tensor<string, []>("op_5259"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_233_cast_fp16 = reshape(shape = var_5259, x = attn_95_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
            tensor<int32, [2]> var_5263 = const()[name = tensor<string, []>("op_5263"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5265 = const()[name = tensor<string, []>("op_5265"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_287_pad_type_0 = const()[name = tensor<string, []>("obj_287_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_287_pad_0 = const()[name = tensor<string, []>("obj_287_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_23_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647281408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648510272))), name = tensor<string, []>("layers_23_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_23_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648510464)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_287_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_bias_to_fp16, dilations = var_5265, groups = var_5105, pad = obj_287_pad_0, pad_type = obj_287_pad_type_0, strides = var_5263, weight = layers_23_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_233_cast_fp16)[name = tensor<string, []>("obj_287_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_287_cast_fp16)[name = tensor<string, []>("inputs_143_cast_fp16")];
            tensor<int32, [1]> var_5271 = const()[name = tensor<string, []>("op_5271"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_143_cast_fp16 = reduce_mean(axes = var_5271, keep_dims = var_5106, x = inputs_143_cast_fp16)[name = tensor<string, []>("channels_mean_143_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_143_cast_fp16 = sub(x = inputs_143_cast_fp16, y = channels_mean_143_cast_fp16)[name = tensor<string, []>("zero_mean_143_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = zero_mean_143_cast_fp16)[name = tensor<string, []>("zero_mean_sq_143_cast_fp16")];
            tensor<int32, [1]> var_5275 = const()[name = tensor<string, []>("op_5275"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5276_cast_fp16 = reduce_mean(axes = var_5275, keep_dims = var_5106, x = zero_mean_sq_143_cast_fp16)[name = tensor<string, []>("op_5276_cast_fp16")];
            tensor<fp16, []> var_5277_to_fp16 = const()[name = tensor<string, []>("op_5277_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5278_cast_fp16 = add(x = var_5276_cast_fp16, y = var_5277_to_fp16)[name = tensor<string, []>("op_5278_cast_fp16")];
            tensor<fp32, []> denom_143_epsilon_0 = const()[name = tensor<string, []>("denom_143_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_143_cast_fp16 = rsqrt(epsilon = denom_143_epsilon_0, x = var_5278_cast_fp16)[name = tensor<string, []>("denom_143_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = denom_143_cast_fp16)[name = tensor<string, []>("out_143_cast_fp16")];
            tensor<fp16, [1280]> input_235_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_235_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648513088)))];
            tensor<fp16, [1280]> input_235_beta_0_to_fp16 = const()[name = tensor<string, []>("input_235_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648515712)))];
            tensor<fp16, []> input_235_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_235_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
            tensor<int32, [2]> var_5289 = const()[name = tensor<string, []>("op_5289"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5291 = const()[name = tensor<string, []>("op_5291"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_237_pad_type_0 = const()[name = tensor<string, []>("input_237_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_237_pad_0 = const()[name = tensor<string, []>("input_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_23_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648518336))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653433600))), name = tensor<string, []>("layers_23_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_23_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653433792)))];
            tensor<fp16, [1, 5120, 1, 1]> input_237_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = var_5291, groups = var_5105, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = var_5289, weight = layers_23_fc1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")];
            tensor<string, []> input_239_mode_0 = const()[name = tensor<string, []>("input_239_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = input_237_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
            tensor<int32, [2]> var_5297 = const()[name = tensor<string, []>("op_5297"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5299 = const()[name = tensor<string, []>("op_5299"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_49_pad_type_0 = const()[name = tensor<string, []>("hidden_states_49_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_49_pad_0 = const()[name = tensor<string, []>("hidden_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_23_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653444096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658359360))), name = tensor<string, []>("layers_23_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_23_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658359552)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_49_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = var_5299, groups = var_5105, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = var_5297, weight = layers_23_fc2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = tensor<string, []>("hidden_states_49_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_145_cast_fp16 = add(x = inputs_143_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor<string, []>("inputs_145_cast_fp16")];
            tensor<int32, []> var_5312 = const()[name = tensor<string, []>("op_5312"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_5319 = const()[name = tensor<string, []>("op_5319"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_5320 = const()[name = tensor<string, []>("op_5320"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_5332 = const()[name = tensor<string, []>("op_5332"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_145_cast_fp16 = reduce_mean(axes = var_5332, keep_dims = var_5320, x = inputs_145_cast_fp16)[name = tensor<string, []>("channels_mean_145_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_145_cast_fp16 = sub(x = inputs_145_cast_fp16, y = channels_mean_145_cast_fp16)[name = tensor<string, []>("zero_mean_145_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = zero_mean_145_cast_fp16)[name = tensor<string, []>("zero_mean_sq_145_cast_fp16")];
            tensor<int32, [1]> var_5336 = const()[name = tensor<string, []>("op_5336"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5337_cast_fp16 = reduce_mean(axes = var_5336, keep_dims = var_5320, x = zero_mean_sq_145_cast_fp16)[name = tensor<string, []>("op_5337_cast_fp16")];
            tensor<fp16, []> var_5338_to_fp16 = const()[name = tensor<string, []>("op_5338_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5339_cast_fp16 = add(x = var_5337_cast_fp16, y = var_5338_to_fp16)[name = tensor<string, []>("op_5339_cast_fp16")];
            tensor<fp32, []> denom_145_epsilon_0 = const()[name = tensor<string, []>("denom_145_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_145_cast_fp16 = rsqrt(epsilon = denom_145_epsilon_0, x = var_5339_cast_fp16)[name = tensor<string, []>("denom_145_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = denom_145_cast_fp16)[name = tensor<string, []>("out_145_cast_fp16")];
            tensor<fp16, [1280]> obj_289_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_289_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658362176)))];
            tensor<fp16, [1280]> obj_289_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_289_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658364800)))];
            tensor<fp16, []> obj_289_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_289_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_289_cast_fp16 = batch_norm(beta = obj_289_beta_0_to_fp16, epsilon = obj_289_epsilon_0_to_fp16, gamma = obj_289_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor<string, []>("obj_289_cast_fp16")];
            tensor<int32, [2]> var_5354 = const()[name = tensor<string, []>("op_5354"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5356 = const()[name = tensor<string, []>("op_5356"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_97_pad_type_0 = const()[name = tensor<string, []>("query_97_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_97_pad_0 = const()[name = tensor<string, []>("query_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658367424))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659186688))), name = tensor<string, []>("layers_24_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659186816)))];
            tensor<fp16, [1, 1280, 1, 1]> query_97_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_bias_to_fp16, dilations = var_5356, groups = var_5319, pad = query_97_pad_0, pad_type = query_97_pad_type_0, strides = var_5354, weight = layers_24_self_attn_q_proj_weight_to_fp16_palettized, x = obj_289_cast_fp16)[name = tensor<string, []>("query_97_cast_fp16")];
            tensor<int32, [2]> var_5360 = const()[name = tensor<string, []>("op_5360"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5362 = const()[name = tensor<string, []>("op_5362"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_49_pad_type_0 = const()[name = tensor<string, []>("current_key_49_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_49_pad_0 = const()[name = tensor<string, []>("current_key_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659189440))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(660008704))), name = tensor<string, []>("layers_24_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_49_cast_fp16 = conv(dilations = var_5362, groups = var_5319, pad = current_key_49_pad_0, pad_type = current_key_49_pad_type_0, strides = var_5360, weight = layers_24_self_attn_k_proj_weight_to_fp16_palettized, x = obj_289_cast_fp16)[name = tensor<string, []>("current_key_49_cast_fp16")];
            tensor<int32, [2]> var_5367 = const()[name = tensor<string, []>("op_5367"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5369 = const()[name = tensor<string, []>("op_5369"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_49_pad_type_0 = const()[name = tensor<string, []>("current_value_49_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_49_pad_0 = const()[name = tensor<string, []>("current_value_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(660008832))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(660828096))), name = tensor<string, []>("layers_24_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(660828224)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_49_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_bias_to_fp16, dilations = var_5369, groups = var_5319, pad = current_value_49_pad_0, pad_type = current_value_49_pad_type_0, strides = var_5367, weight = layers_24_self_attn_v_proj_weight_to_fp16_palettized, x = obj_289_cast_fp16)[name = tensor<string, []>("current_value_49_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5376_cast_fp16 = mul(x = current_key_49_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_5376_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5378_cast_fp16 = mul(x = var_103_cast_fp16_24, y = var_241_cast_fp16)[name = tensor<string, []>("op_5378_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_97_cast_fp16 = add(x = var_5376_cast_fp16, y = var_5378_cast_fp16)[name = tensor<string, []>("key_97_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5380_cast_fp16 = mul(x = current_value_49_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_5380_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5382_cast_fp16 = mul(x = var_138_cast_fp16_24, y = var_241_cast_fp16)[name = tensor<string, []>("op_5382_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_97_cast_fp16 = add(x = var_5380_cast_fp16, y = var_5382_cast_fp16)[name = tensor<string, []>("value_97_cast_fp16")];
            tensor<int32, [4]> var_5385 = const()[name = tensor<string, []>("op_5385"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5386_cast_fp16 = reshape(shape = var_5385, x = query_97_cast_fp16)[name = tensor<string, []>("op_5386_cast_fp16")];
            tensor<fp16, []> var_5387_to_fp16 = const()[name = tensor<string, []>("op_5387_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5388_cast_fp16 = mul(x = var_5386_cast_fp16, y = var_5387_to_fp16)[name = tensor<string, []>("op_5388_cast_fp16")];
            tensor<int32, [4]> var_5389 = const()[name = tensor<string, []>("op_5389"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_5390_cast_fp16 = reshape(shape = var_5389, x = key_97_cast_fp16)[name = tensor<string, []>("op_5390_cast_fp16")];
            tensor<bool, []> mh_w_145_transpose_x_0 = const()[name = tensor<string, []>("mh_w_145_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_145_transpose_y_0 = const()[name = tensor<string, []>("mh_w_145_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_145_cast_fp16 = matmul(transpose_x = mh_w_145_transpose_x_0, transpose_y = mh_w_145_transpose_y_0, x = var_5388_cast_fp16, y = var_5390_cast_fp16)[name = tensor<string, []>("mh_w_145_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_147_cast_fp16 = add(x = mh_w_145_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_147_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_5398_cast_fp16 = softmax(axis = var_5312, x = mh_w_147_cast_fp16)[name = tensor<string, []>("op_5398_cast_fp16")];
            tensor<int32, [4]> var_5399 = const()[name = tensor<string, []>("op_5399"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_5400_cast_fp16 = reshape(shape = var_5399, x = value_97_cast_fp16)[name = tensor<string, []>("op_5400_cast_fp16")];
            tensor<bool, []> attn_97_transpose_x_0 = const()[name = tensor<string, []>("attn_97_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_97_transpose_y_0 = const()[name = tensor<string, []>("attn_97_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_97_cast_fp16 = matmul(transpose_x = attn_97_transpose_x_0, transpose_y = attn_97_transpose_y_0, x = var_5400_cast_fp16, y = var_5398_cast_fp16)[name = tensor<string, []>("attn_97_cast_fp16")];
            tensor<int32, [4]> var_5403 = const()[name = tensor<string, []>("op_5403"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_241_cast_fp16 = reshape(shape = var_5403, x = attn_97_cast_fp16)[name = tensor<string, []>("input_241_cast_fp16")];
            tensor<int32, [2]> var_5407 = const()[name = tensor<string, []>("op_5407"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5409 = const()[name = tensor<string, []>("op_5409"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_295_pad_type_0 = const()[name = tensor<string, []>("obj_295_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_295_pad_0 = const()[name = tensor<string, []>("obj_295_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(660830848))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(662059712))), name = tensor<string, []>("layers_24_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_24_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(662059904)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_295_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_bias_to_fp16, dilations = var_5409, groups = var_5319, pad = obj_295_pad_0, pad_type = obj_295_pad_type_0, strides = var_5407, weight = layers_24_self_attn_o_proj_weight_to_fp16_palettized, x = input_241_cast_fp16)[name = tensor<string, []>("obj_295_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_147_cast_fp16 = add(x = inputs_145_cast_fp16, y = obj_295_cast_fp16)[name = tensor<string, []>("inputs_147_cast_fp16")];
            tensor<int32, [1]> var_5419 = const()[name = tensor<string, []>("op_5419"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_147_cast_fp16 = reduce_mean(axes = var_5419, keep_dims = var_5320, x = inputs_147_cast_fp16)[name = tensor<string, []>("channels_mean_147_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_147_cast_fp16 = sub(x = inputs_147_cast_fp16, y = channels_mean_147_cast_fp16)[name = tensor<string, []>("zero_mean_147_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = zero_mean_147_cast_fp16)[name = tensor<string, []>("zero_mean_sq_147_cast_fp16")];
            tensor<int32, [1]> var_5423 = const()[name = tensor<string, []>("op_5423"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5424_cast_fp16 = reduce_mean(axes = var_5423, keep_dims = var_5320, x = zero_mean_sq_147_cast_fp16)[name = tensor<string, []>("op_5424_cast_fp16")];
            tensor<fp16, []> var_5425_to_fp16 = const()[name = tensor<string, []>("op_5425_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5426_cast_fp16 = add(x = var_5424_cast_fp16, y = var_5425_to_fp16)[name = tensor<string, []>("op_5426_cast_fp16")];
            tensor<fp32, []> denom_147_epsilon_0 = const()[name = tensor<string, []>("denom_147_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_147_cast_fp16 = rsqrt(epsilon = denom_147_epsilon_0, x = var_5426_cast_fp16)[name = tensor<string, []>("denom_147_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = denom_147_cast_fp16)[name = tensor<string, []>("out_147_cast_fp16")];
            tensor<fp16, [1280]> obj_297_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_297_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(662062528)))];
            tensor<fp16, [1280]> obj_297_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_297_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(662065152)))];
            tensor<fp16, []> obj_297_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_297_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_297_cast_fp16 = batch_norm(beta = obj_297_beta_0_to_fp16, epsilon = obj_297_epsilon_0_to_fp16, gamma = obj_297_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor<string, []>("obj_297_cast_fp16")];
            tensor<int32, [2]> var_5441 = const()[name = tensor<string, []>("op_5441"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5443 = const()[name = tensor<string, []>("op_5443"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_99_pad_type_0 = const()[name = tensor<string, []>("query_99_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_99_pad_0 = const()[name = tensor<string, []>("query_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(662067776))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(663296640))), name = tensor<string, []>("layers_24_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_24_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(663296832)))];
            tensor<fp16, [1, 1280, 1, 1]> query_99_cast_fp16 = conv(bias = layers_24_encoder_attn_q_proj_bias_to_fp16, dilations = var_5443, groups = var_5319, pad = query_99_pad_0, pad_type = query_99_pad_type_0, strides = var_5441, weight = layers_24_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_297_cast_fp16)[name = tensor<string, []>("query_99_cast_fp16")];
            tensor<int32, [2]> var_5447 = const()[name = tensor<string, []>("op_5447"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5449 = const()[name = tensor<string, []>("op_5449"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_99_pad_type_0 = const()[name = tensor<string, []>("key_99_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_99_pad_0 = const()[name = tensor<string, []>("key_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(663299456))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664528320))), name = tensor<string, []>("layers_24_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_99_cast_fp16 = conv(dilations = var_5449, groups = var_5319, pad = key_99_pad_0, pad_type = key_99_pad_type_0, strides = var_5447, weight = layers_24_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_99_cast_fp16")];
            tensor<int32, [2]> var_5454 = const()[name = tensor<string, []>("op_5454"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5456 = const()[name = tensor<string, []>("op_5456"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_99_pad_type_0 = const()[name = tensor<string, []>("value_99_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_99_pad_0 = const()[name = tensor<string, []>("value_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664528512))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(665757376))), name = tensor<string, []>("layers_24_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_24_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(665757568)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_99_cast_fp16 = conv(bias = layers_24_encoder_attn_v_proj_bias_to_fp16, dilations = var_5456, groups = var_5319, pad = value_99_pad_0, pad_type = value_99_pad_type_0, strides = var_5454, weight = layers_24_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_99_cast_fp16")];
            tensor<int32, [4]> var_5460 = const()[name = tensor<string, []>("op_5460"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5461_cast_fp16 = reshape(shape = var_5460, x = query_99_cast_fp16)[name = tensor<string, []>("op_5461_cast_fp16")];
            tensor<fp16, []> var_5462_to_fp16 = const()[name = tensor<string, []>("op_5462_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5463_cast_fp16 = mul(x = var_5461_cast_fp16, y = var_5462_to_fp16)[name = tensor<string, []>("op_5463_cast_fp16")];
            tensor<int32, [4]> var_5464 = const()[name = tensor<string, []>("op_5464"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5465_cast_fp16 = reshape(shape = var_5464, x = key_99_cast_fp16)[name = tensor<string, []>("op_5465_cast_fp16")];
            tensor<bool, []> mh_w_149_transpose_x_0 = const()[name = tensor<string, []>("mh_w_149_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_149_transpose_y_0 = const()[name = tensor<string, []>("mh_w_149_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_149_cast_fp16 = matmul(transpose_x = mh_w_149_transpose_x_0, transpose_y = mh_w_149_transpose_y_0, x = var_5463_cast_fp16, y = var_5465_cast_fp16)[name = tensor<string, []>("mh_w_149_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_5468_cast_fp16 = softmax(axis = var_5312, x = mh_w_149_cast_fp16)[name = tensor<string, []>("op_5468_cast_fp16")];
            tensor<int32, [4]> var_5469 = const()[name = tensor<string, []>("op_5469"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5470_cast_fp16 = reshape(shape = var_5469, x = value_99_cast_fp16)[name = tensor<string, []>("op_5470_cast_fp16")];
            tensor<bool, []> attn_99_transpose_x_0 = const()[name = tensor<string, []>("attn_99_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_99_transpose_y_0 = const()[name = tensor<string, []>("attn_99_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_99_cast_fp16 = matmul(transpose_x = attn_99_transpose_x_0, transpose_y = attn_99_transpose_y_0, x = var_5470_cast_fp16, y = var_5468_cast_fp16)[name = tensor<string, []>("attn_99_cast_fp16")];
            tensor<int32, [4]> var_5473 = const()[name = tensor<string, []>("op_5473"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_243_cast_fp16 = reshape(shape = var_5473, x = attn_99_cast_fp16)[name = tensor<string, []>("input_243_cast_fp16")];
            tensor<int32, [2]> var_5477 = const()[name = tensor<string, []>("op_5477"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5479 = const()[name = tensor<string, []>("op_5479"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_299_pad_type_0 = const()[name = tensor<string, []>("obj_299_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_299_pad_0 = const()[name = tensor<string, []>("obj_299_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_24_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(665760192))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(666989056))), name = tensor<string, []>("layers_24_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_24_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(666989248)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_299_cast_fp16 = conv(bias = layers_24_encoder_attn_o_proj_bias_to_fp16, dilations = var_5479, groups = var_5319, pad = obj_299_pad_0, pad_type = obj_299_pad_type_0, strides = var_5477, weight = layers_24_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_243_cast_fp16)[name = tensor<string, []>("obj_299_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = obj_299_cast_fp16)[name = tensor<string, []>("inputs_149_cast_fp16")];
            tensor<int32, [1]> var_5485 = const()[name = tensor<string, []>("op_5485"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_149_cast_fp16 = reduce_mean(axes = var_5485, keep_dims = var_5320, x = inputs_149_cast_fp16)[name = tensor<string, []>("channels_mean_149_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_149_cast_fp16 = sub(x = inputs_149_cast_fp16, y = channels_mean_149_cast_fp16)[name = tensor<string, []>("zero_mean_149_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = zero_mean_149_cast_fp16)[name = tensor<string, []>("zero_mean_sq_149_cast_fp16")];
            tensor<int32, [1]> var_5489 = const()[name = tensor<string, []>("op_5489"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5490_cast_fp16 = reduce_mean(axes = var_5489, keep_dims = var_5320, x = zero_mean_sq_149_cast_fp16)[name = tensor<string, []>("op_5490_cast_fp16")];
            tensor<fp16, []> var_5491_to_fp16 = const()[name = tensor<string, []>("op_5491_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5492_cast_fp16 = add(x = var_5490_cast_fp16, y = var_5491_to_fp16)[name = tensor<string, []>("op_5492_cast_fp16")];
            tensor<fp32, []> denom_149_epsilon_0 = const()[name = tensor<string, []>("denom_149_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_149_cast_fp16 = rsqrt(epsilon = denom_149_epsilon_0, x = var_5492_cast_fp16)[name = tensor<string, []>("denom_149_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = denom_149_cast_fp16)[name = tensor<string, []>("out_149_cast_fp16")];
            tensor<fp16, [1280]> input_245_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_245_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(666991872)))];
            tensor<fp16, [1280]> input_245_beta_0_to_fp16 = const()[name = tensor<string, []>("input_245_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(666994496)))];
            tensor<fp16, []> input_245_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_245_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_245_cast_fp16 = batch_norm(beta = input_245_beta_0_to_fp16, epsilon = input_245_epsilon_0_to_fp16, gamma = input_245_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
            tensor<int32, [2]> var_5503 = const()[name = tensor<string, []>("op_5503"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5505 = const()[name = tensor<string, []>("op_5505"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_247_pad_type_0 = const()[name = tensor<string, []>("input_247_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_247_pad_0 = const()[name = tensor<string, []>("input_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_24_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(666997120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(671912384))), name = tensor<string, []>("layers_24_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_24_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(671912576)))];
            tensor<fp16, [1, 5120, 1, 1]> input_247_cast_fp16 = conv(bias = layers_24_fc1_bias_to_fp16, dilations = var_5505, groups = var_5319, pad = input_247_pad_0, pad_type = input_247_pad_type_0, strides = var_5503, weight = layers_24_fc1_weight_to_fp16_palettized, x = input_245_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
            tensor<string, []> input_249_mode_0 = const()[name = tensor<string, []>("input_249_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_249_cast_fp16 = gelu(mode = input_249_mode_0, x = input_247_cast_fp16)[name = tensor<string, []>("input_249_cast_fp16")];
            tensor<int32, [2]> var_5511 = const()[name = tensor<string, []>("op_5511"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5513 = const()[name = tensor<string, []>("op_5513"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_51_pad_type_0 = const()[name = tensor<string, []>("hidden_states_51_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_51_pad_0 = const()[name = tensor<string, []>("hidden_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_24_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(671922880))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(676838144))), name = tensor<string, []>("layers_24_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_24_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(676838336)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_51_cast_fp16 = conv(bias = layers_24_fc2_bias_to_fp16, dilations = var_5513, groups = var_5319, pad = hidden_states_51_pad_0, pad_type = hidden_states_51_pad_type_0, strides = var_5511, weight = layers_24_fc2_weight_to_fp16_palettized, x = input_249_cast_fp16)[name = tensor<string, []>("hidden_states_51_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_151_cast_fp16 = add(x = inputs_149_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor<string, []>("inputs_151_cast_fp16")];
            tensor<int32, []> var_5526 = const()[name = tensor<string, []>("op_5526"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_5533 = const()[name = tensor<string, []>("op_5533"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_5534 = const()[name = tensor<string, []>("op_5534"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_5546 = const()[name = tensor<string, []>("op_5546"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_151_cast_fp16 = reduce_mean(axes = var_5546, keep_dims = var_5534, x = inputs_151_cast_fp16)[name = tensor<string, []>("channels_mean_151_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_151_cast_fp16 = sub(x = inputs_151_cast_fp16, y = channels_mean_151_cast_fp16)[name = tensor<string, []>("zero_mean_151_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = zero_mean_151_cast_fp16)[name = tensor<string, []>("zero_mean_sq_151_cast_fp16")];
            tensor<int32, [1]> var_5550 = const()[name = tensor<string, []>("op_5550"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5551_cast_fp16 = reduce_mean(axes = var_5550, keep_dims = var_5534, x = zero_mean_sq_151_cast_fp16)[name = tensor<string, []>("op_5551_cast_fp16")];
            tensor<fp16, []> var_5552_to_fp16 = const()[name = tensor<string, []>("op_5552_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5553_cast_fp16 = add(x = var_5551_cast_fp16, y = var_5552_to_fp16)[name = tensor<string, []>("op_5553_cast_fp16")];
            tensor<fp32, []> denom_151_epsilon_0 = const()[name = tensor<string, []>("denom_151_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_151_cast_fp16 = rsqrt(epsilon = denom_151_epsilon_0, x = var_5553_cast_fp16)[name = tensor<string, []>("denom_151_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = denom_151_cast_fp16)[name = tensor<string, []>("out_151_cast_fp16")];
            tensor<fp16, [1280]> obj_301_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_301_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(676840960)))];
            tensor<fp16, [1280]> obj_301_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_301_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(676843584)))];
            tensor<fp16, []> obj_301_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_301_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_301_cast_fp16 = batch_norm(beta = obj_301_beta_0_to_fp16, epsilon = obj_301_epsilon_0_to_fp16, gamma = obj_301_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor<string, []>("obj_301_cast_fp16")];
            tensor<int32, [2]> var_5568 = const()[name = tensor<string, []>("op_5568"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5570 = const()[name = tensor<string, []>("op_5570"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_101_pad_type_0 = const()[name = tensor<string, []>("query_101_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_101_pad_0 = const()[name = tensor<string, []>("query_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(676846208))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(677665472))), name = tensor<string, []>("layers_25_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(677665600)))];
            tensor<fp16, [1, 1280, 1, 1]> query_101_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_bias_to_fp16, dilations = var_5570, groups = var_5533, pad = query_101_pad_0, pad_type = query_101_pad_type_0, strides = var_5568, weight = layers_25_self_attn_q_proj_weight_to_fp16_palettized, x = obj_301_cast_fp16)[name = tensor<string, []>("query_101_cast_fp16")];
            tensor<int32, [2]> var_5574 = const()[name = tensor<string, []>("op_5574"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5576 = const()[name = tensor<string, []>("op_5576"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_51_pad_type_0 = const()[name = tensor<string, []>("current_key_51_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_51_pad_0 = const()[name = tensor<string, []>("current_key_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(677668224))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678897088))), name = tensor<string, []>("layers_25_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_51_cast_fp16 = conv(dilations = var_5576, groups = var_5533, pad = current_key_51_pad_0, pad_type = current_key_51_pad_type_0, strides = var_5574, weight = layers_25_self_attn_k_proj_weight_to_fp16_palettized, x = obj_301_cast_fp16)[name = tensor<string, []>("current_key_51_cast_fp16")];
            tensor<int32, [2]> var_5581 = const()[name = tensor<string, []>("op_5581"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5583 = const()[name = tensor<string, []>("op_5583"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_51_pad_type_0 = const()[name = tensor<string, []>("current_value_51_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_51_pad_0 = const()[name = tensor<string, []>("current_value_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678897280))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(679716544))), name = tensor<string, []>("layers_25_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(679716672)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_51_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_bias_to_fp16, dilations = var_5583, groups = var_5533, pad = current_value_51_pad_0, pad_type = current_value_51_pad_type_0, strides = var_5581, weight = layers_25_self_attn_v_proj_weight_to_fp16_palettized, x = obj_301_cast_fp16)[name = tensor<string, []>("current_value_51_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5590_cast_fp16 = mul(x = current_key_51_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_5590_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5592_cast_fp16 = mul(x = var_103_cast_fp16_25, y = var_241_cast_fp16)[name = tensor<string, []>("op_5592_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_101_cast_fp16 = add(x = var_5590_cast_fp16, y = var_5592_cast_fp16)[name = tensor<string, []>("key_101_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5594_cast_fp16 = mul(x = current_value_51_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_5594_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5596_cast_fp16 = mul(x = var_138_cast_fp16_25, y = var_241_cast_fp16)[name = tensor<string, []>("op_5596_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_101_cast_fp16 = add(x = var_5594_cast_fp16, y = var_5596_cast_fp16)[name = tensor<string, []>("value_101_cast_fp16")];
            tensor<int32, [4]> var_5599 = const()[name = tensor<string, []>("op_5599"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5600_cast_fp16 = reshape(shape = var_5599, x = query_101_cast_fp16)[name = tensor<string, []>("op_5600_cast_fp16")];
            tensor<fp16, []> var_5601_to_fp16 = const()[name = tensor<string, []>("op_5601_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5602_cast_fp16 = mul(x = var_5600_cast_fp16, y = var_5601_to_fp16)[name = tensor<string, []>("op_5602_cast_fp16")];
            tensor<int32, [4]> var_5603 = const()[name = tensor<string, []>("op_5603"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_5604_cast_fp16 = reshape(shape = var_5603, x = key_101_cast_fp16)[name = tensor<string, []>("op_5604_cast_fp16")];
            tensor<bool, []> mh_w_151_transpose_x_0 = const()[name = tensor<string, []>("mh_w_151_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_151_transpose_y_0 = const()[name = tensor<string, []>("mh_w_151_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_151_cast_fp16 = matmul(transpose_x = mh_w_151_transpose_x_0, transpose_y = mh_w_151_transpose_y_0, x = var_5602_cast_fp16, y = var_5604_cast_fp16)[name = tensor<string, []>("mh_w_151_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_153_cast_fp16 = add(x = mh_w_151_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_153_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_5612_cast_fp16 = softmax(axis = var_5526, x = mh_w_153_cast_fp16)[name = tensor<string, []>("op_5612_cast_fp16")];
            tensor<int32, [4]> var_5613 = const()[name = tensor<string, []>("op_5613"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_5614_cast_fp16 = reshape(shape = var_5613, x = value_101_cast_fp16)[name = tensor<string, []>("op_5614_cast_fp16")];
            tensor<bool, []> attn_101_transpose_x_0 = const()[name = tensor<string, []>("attn_101_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_101_transpose_y_0 = const()[name = tensor<string, []>("attn_101_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_101_cast_fp16 = matmul(transpose_x = attn_101_transpose_x_0, transpose_y = attn_101_transpose_y_0, x = var_5614_cast_fp16, y = var_5612_cast_fp16)[name = tensor<string, []>("attn_101_cast_fp16")];
            tensor<int32, [4]> var_5617 = const()[name = tensor<string, []>("op_5617"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_251_cast_fp16 = reshape(shape = var_5617, x = attn_101_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
            tensor<int32, [2]> var_5621 = const()[name = tensor<string, []>("op_5621"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5623 = const()[name = tensor<string, []>("op_5623"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_307_pad_type_0 = const()[name = tensor<string, []>("obj_307_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_307_pad_0 = const()[name = tensor<string, []>("obj_307_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(679719296))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680948160))), name = tensor<string, []>("layers_25_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_25_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680948352)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_307_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_bias_to_fp16, dilations = var_5623, groups = var_5533, pad = obj_307_pad_0, pad_type = obj_307_pad_type_0, strides = var_5621, weight = layers_25_self_attn_o_proj_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = tensor<string, []>("obj_307_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = obj_307_cast_fp16)[name = tensor<string, []>("inputs_153_cast_fp16")];
            tensor<int32, [1]> var_5633 = const()[name = tensor<string, []>("op_5633"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_153_cast_fp16 = reduce_mean(axes = var_5633, keep_dims = var_5534, x = inputs_153_cast_fp16)[name = tensor<string, []>("channels_mean_153_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_153_cast_fp16 = sub(x = inputs_153_cast_fp16, y = channels_mean_153_cast_fp16)[name = tensor<string, []>("zero_mean_153_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = zero_mean_153_cast_fp16)[name = tensor<string, []>("zero_mean_sq_153_cast_fp16")];
            tensor<int32, [1]> var_5637 = const()[name = tensor<string, []>("op_5637"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5638_cast_fp16 = reduce_mean(axes = var_5637, keep_dims = var_5534, x = zero_mean_sq_153_cast_fp16)[name = tensor<string, []>("op_5638_cast_fp16")];
            tensor<fp16, []> var_5639_to_fp16 = const()[name = tensor<string, []>("op_5639_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5640_cast_fp16 = add(x = var_5638_cast_fp16, y = var_5639_to_fp16)[name = tensor<string, []>("op_5640_cast_fp16")];
            tensor<fp32, []> denom_153_epsilon_0 = const()[name = tensor<string, []>("denom_153_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_153_cast_fp16 = rsqrt(epsilon = denom_153_epsilon_0, x = var_5640_cast_fp16)[name = tensor<string, []>("denom_153_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = denom_153_cast_fp16)[name = tensor<string, []>("out_153_cast_fp16")];
            tensor<fp16, [1280]> obj_309_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_309_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680950976)))];
            tensor<fp16, [1280]> obj_309_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_309_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680953600)))];
            tensor<fp16, []> obj_309_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_309_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_309_cast_fp16 = batch_norm(beta = obj_309_beta_0_to_fp16, epsilon = obj_309_epsilon_0_to_fp16, gamma = obj_309_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_153_cast_fp16)[name = tensor<string, []>("obj_309_cast_fp16")];
            tensor<int32, [2]> var_5655 = const()[name = tensor<string, []>("op_5655"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5657 = const()[name = tensor<string, []>("op_5657"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_103_pad_type_0 = const()[name = tensor<string, []>("query_103_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_103_pad_0 = const()[name = tensor<string, []>("query_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680956224))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(682185088))), name = tensor<string, []>("layers_25_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_25_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(682185280)))];
            tensor<fp16, [1, 1280, 1, 1]> query_103_cast_fp16 = conv(bias = layers_25_encoder_attn_q_proj_bias_to_fp16, dilations = var_5657, groups = var_5533, pad = query_103_pad_0, pad_type = query_103_pad_type_0, strides = var_5655, weight = layers_25_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_309_cast_fp16)[name = tensor<string, []>("query_103_cast_fp16")];
            tensor<int32, [2]> var_5661 = const()[name = tensor<string, []>("op_5661"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5663 = const()[name = tensor<string, []>("op_5663"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_103_pad_type_0 = const()[name = tensor<string, []>("key_103_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_103_pad_0 = const()[name = tensor<string, []>("key_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(682187904))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683007168))), name = tensor<string, []>("layers_25_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_103_cast_fp16 = conv(dilations = var_5663, groups = var_5533, pad = key_103_pad_0, pad_type = key_103_pad_type_0, strides = var_5661, weight = layers_25_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_103_cast_fp16")];
            tensor<int32, [2]> var_5668 = const()[name = tensor<string, []>("op_5668"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5670 = const()[name = tensor<string, []>("op_5670"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_103_pad_type_0 = const()[name = tensor<string, []>("value_103_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_103_pad_0 = const()[name = tensor<string, []>("value_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683007296))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(684236160))), name = tensor<string, []>("layers_25_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_25_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(684236352)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_103_cast_fp16 = conv(bias = layers_25_encoder_attn_v_proj_bias_to_fp16, dilations = var_5670, groups = var_5533, pad = value_103_pad_0, pad_type = value_103_pad_type_0, strides = var_5668, weight = layers_25_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_103_cast_fp16")];
            tensor<int32, [4]> var_5674 = const()[name = tensor<string, []>("op_5674"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5675_cast_fp16 = reshape(shape = var_5674, x = query_103_cast_fp16)[name = tensor<string, []>("op_5675_cast_fp16")];
            tensor<fp16, []> var_5676_to_fp16 = const()[name = tensor<string, []>("op_5676_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5677_cast_fp16 = mul(x = var_5675_cast_fp16, y = var_5676_to_fp16)[name = tensor<string, []>("op_5677_cast_fp16")];
            tensor<int32, [4]> var_5678 = const()[name = tensor<string, []>("op_5678"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5679_cast_fp16 = reshape(shape = var_5678, x = key_103_cast_fp16)[name = tensor<string, []>("op_5679_cast_fp16")];
            tensor<bool, []> mh_w_155_transpose_x_0 = const()[name = tensor<string, []>("mh_w_155_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_155_transpose_y_0 = const()[name = tensor<string, []>("mh_w_155_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_155_cast_fp16 = matmul(transpose_x = mh_w_155_transpose_x_0, transpose_y = mh_w_155_transpose_y_0, x = var_5677_cast_fp16, y = var_5679_cast_fp16)[name = tensor<string, []>("mh_w_155_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_5682_cast_fp16 = softmax(axis = var_5526, x = mh_w_155_cast_fp16)[name = tensor<string, []>("op_5682_cast_fp16")];
            tensor<int32, [4]> var_5683 = const()[name = tensor<string, []>("op_5683"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5684_cast_fp16 = reshape(shape = var_5683, x = value_103_cast_fp16)[name = tensor<string, []>("op_5684_cast_fp16")];
            tensor<bool, []> attn_103_transpose_x_0 = const()[name = tensor<string, []>("attn_103_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_103_transpose_y_0 = const()[name = tensor<string, []>("attn_103_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_103_cast_fp16 = matmul(transpose_x = attn_103_transpose_x_0, transpose_y = attn_103_transpose_y_0, x = var_5684_cast_fp16, y = var_5682_cast_fp16)[name = tensor<string, []>("attn_103_cast_fp16")];
            tensor<int32, [4]> var_5687 = const()[name = tensor<string, []>("op_5687"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_253_cast_fp16 = reshape(shape = var_5687, x = attn_103_cast_fp16)[name = tensor<string, []>("input_253_cast_fp16")];
            tensor<int32, [2]> var_5691 = const()[name = tensor<string, []>("op_5691"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5693 = const()[name = tensor<string, []>("op_5693"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_311_pad_type_0 = const()[name = tensor<string, []>("obj_311_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_311_pad_0 = const()[name = tensor<string, []>("obj_311_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_25_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(684238976))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(685467840))), name = tensor<string, []>("layers_25_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_25_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(685468032)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_311_cast_fp16 = conv(bias = layers_25_encoder_attn_o_proj_bias_to_fp16, dilations = var_5693, groups = var_5533, pad = obj_311_pad_0, pad_type = obj_311_pad_type_0, strides = var_5691, weight = layers_25_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_253_cast_fp16)[name = tensor<string, []>("obj_311_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_155_cast_fp16 = add(x = inputs_153_cast_fp16, y = obj_311_cast_fp16)[name = tensor<string, []>("inputs_155_cast_fp16")];
            tensor<int32, [1]> var_5699 = const()[name = tensor<string, []>("op_5699"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_155_cast_fp16 = reduce_mean(axes = var_5699, keep_dims = var_5534, x = inputs_155_cast_fp16)[name = tensor<string, []>("channels_mean_155_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_155_cast_fp16 = sub(x = inputs_155_cast_fp16, y = channels_mean_155_cast_fp16)[name = tensor<string, []>("zero_mean_155_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = zero_mean_155_cast_fp16)[name = tensor<string, []>("zero_mean_sq_155_cast_fp16")];
            tensor<int32, [1]> var_5703 = const()[name = tensor<string, []>("op_5703"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5704_cast_fp16 = reduce_mean(axes = var_5703, keep_dims = var_5534, x = zero_mean_sq_155_cast_fp16)[name = tensor<string, []>("op_5704_cast_fp16")];
            tensor<fp16, []> var_5705_to_fp16 = const()[name = tensor<string, []>("op_5705_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5706_cast_fp16 = add(x = var_5704_cast_fp16, y = var_5705_to_fp16)[name = tensor<string, []>("op_5706_cast_fp16")];
            tensor<fp32, []> denom_155_epsilon_0 = const()[name = tensor<string, []>("denom_155_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_155_cast_fp16 = rsqrt(epsilon = denom_155_epsilon_0, x = var_5706_cast_fp16)[name = tensor<string, []>("denom_155_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = denom_155_cast_fp16)[name = tensor<string, []>("out_155_cast_fp16")];
            tensor<fp16, [1280]> input_255_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_255_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(685470656)))];
            tensor<fp16, [1280]> input_255_beta_0_to_fp16 = const()[name = tensor<string, []>("input_255_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(685473280)))];
            tensor<fp16, []> input_255_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_255_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_255_cast_fp16 = batch_norm(beta = input_255_beta_0_to_fp16, epsilon = input_255_epsilon_0_to_fp16, gamma = input_255_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
            tensor<int32, [2]> var_5717 = const()[name = tensor<string, []>("op_5717"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5719 = const()[name = tensor<string, []>("op_5719"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_257_pad_type_0 = const()[name = tensor<string, []>("input_257_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_257_pad_0 = const()[name = tensor<string, []>("input_257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_25_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(685475904))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688752768))), name = tensor<string, []>("layers_25_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_25_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688752896)))];
            tensor<fp16, [1, 5120, 1, 1]> input_257_cast_fp16 = conv(bias = layers_25_fc1_bias_to_fp16, dilations = var_5719, groups = var_5533, pad = input_257_pad_0, pad_type = input_257_pad_type_0, strides = var_5717, weight = layers_25_fc1_weight_to_fp16_palettized, x = input_255_cast_fp16)[name = tensor<string, []>("input_257_cast_fp16")];
            tensor<string, []> input_259_mode_0 = const()[name = tensor<string, []>("input_259_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_259_cast_fp16 = gelu(mode = input_259_mode_0, x = input_257_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
            tensor<int32, [2]> var_5725 = const()[name = tensor<string, []>("op_5725"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5727 = const()[name = tensor<string, []>("op_5727"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_53_pad_type_0 = const()[name = tensor<string, []>("hidden_states_53_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = tensor<string, []>("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_25_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688763200))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693678464))), name = tensor<string, []>("layers_25_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_25_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693678656)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_53_cast_fp16 = conv(bias = layers_25_fc2_bias_to_fp16, dilations = var_5727, groups = var_5533, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = var_5725, weight = layers_25_fc2_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = tensor<string, []>("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_157_cast_fp16 = add(x = inputs_155_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor<string, []>("inputs_157_cast_fp16")];
            tensor<int32, []> var_5740 = const()[name = tensor<string, []>("op_5740"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_5747 = const()[name = tensor<string, []>("op_5747"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_5748 = const()[name = tensor<string, []>("op_5748"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_5760 = const()[name = tensor<string, []>("op_5760"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_157_cast_fp16 = reduce_mean(axes = var_5760, keep_dims = var_5748, x = inputs_157_cast_fp16)[name = tensor<string, []>("channels_mean_157_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_157_cast_fp16 = sub(x = inputs_157_cast_fp16, y = channels_mean_157_cast_fp16)[name = tensor<string, []>("zero_mean_157_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = zero_mean_157_cast_fp16)[name = tensor<string, []>("zero_mean_sq_157_cast_fp16")];
            tensor<int32, [1]> var_5764 = const()[name = tensor<string, []>("op_5764"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5765_cast_fp16 = reduce_mean(axes = var_5764, keep_dims = var_5748, x = zero_mean_sq_157_cast_fp16)[name = tensor<string, []>("op_5765_cast_fp16")];
            tensor<fp16, []> var_5766_to_fp16 = const()[name = tensor<string, []>("op_5766_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5767_cast_fp16 = add(x = var_5765_cast_fp16, y = var_5766_to_fp16)[name = tensor<string, []>("op_5767_cast_fp16")];
            tensor<fp32, []> denom_157_epsilon_0 = const()[name = tensor<string, []>("denom_157_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_157_cast_fp16 = rsqrt(epsilon = denom_157_epsilon_0, x = var_5767_cast_fp16)[name = tensor<string, []>("denom_157_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = denom_157_cast_fp16)[name = tensor<string, []>("out_157_cast_fp16")];
            tensor<fp16, [1280]> obj_313_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_313_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693681280)))];
            tensor<fp16, [1280]> obj_313_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_313_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693683904)))];
            tensor<fp16, []> obj_313_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_313_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_313_cast_fp16 = batch_norm(beta = obj_313_beta_0_to_fp16, epsilon = obj_313_epsilon_0_to_fp16, gamma = obj_313_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor<string, []>("obj_313_cast_fp16")];
            tensor<int32, [2]> var_5782 = const()[name = tensor<string, []>("op_5782"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5784 = const()[name = tensor<string, []>("op_5784"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_105_pad_type_0 = const()[name = tensor<string, []>("query_105_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_105_pad_0 = const()[name = tensor<string, []>("query_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693686528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(694505792))), name = tensor<string, []>("layers_26_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(694505920)))];
            tensor<fp16, [1, 1280, 1, 1]> query_105_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_bias_to_fp16, dilations = var_5784, groups = var_5747, pad = query_105_pad_0, pad_type = query_105_pad_type_0, strides = var_5782, weight = layers_26_self_attn_q_proj_weight_to_fp16_palettized, x = obj_313_cast_fp16)[name = tensor<string, []>("query_105_cast_fp16")];
            tensor<int32, [2]> var_5788 = const()[name = tensor<string, []>("op_5788"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5790 = const()[name = tensor<string, []>("op_5790"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_53_pad_type_0 = const()[name = tensor<string, []>("current_key_53_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_53_pad_0 = const()[name = tensor<string, []>("current_key_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(694508544))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696147008))), name = tensor<string, []>("layers_26_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_53_cast_fp16 = conv(dilations = var_5790, groups = var_5747, pad = current_key_53_pad_0, pad_type = current_key_53_pad_type_0, strides = var_5788, weight = layers_26_self_attn_k_proj_weight_to_fp16_palettized, x = obj_313_cast_fp16)[name = tensor<string, []>("current_key_53_cast_fp16")];
            tensor<int32, [2]> var_5795 = const()[name = tensor<string, []>("op_5795"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5797 = const()[name = tensor<string, []>("op_5797"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_53_pad_type_0 = const()[name = tensor<string, []>("current_value_53_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_53_pad_0 = const()[name = tensor<string, []>("current_value_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696147584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696966848))), name = tensor<string, []>("layers_26_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696966976)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_53_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_bias_to_fp16, dilations = var_5797, groups = var_5747, pad = current_value_53_pad_0, pad_type = current_value_53_pad_type_0, strides = var_5795, weight = layers_26_self_attn_v_proj_weight_to_fp16_palettized, x = obj_313_cast_fp16)[name = tensor<string, []>("current_value_53_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5804_cast_fp16 = mul(x = current_key_53_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_5804_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5806_cast_fp16 = mul(x = var_103_cast_fp16_26, y = var_241_cast_fp16)[name = tensor<string, []>("op_5806_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_105_cast_fp16 = add(x = var_5804_cast_fp16, y = var_5806_cast_fp16)[name = tensor<string, []>("key_105_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5808_cast_fp16 = mul(x = current_value_53_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_5808_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_5810_cast_fp16 = mul(x = var_138_cast_fp16_26, y = var_241_cast_fp16)[name = tensor<string, []>("op_5810_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_105_cast_fp16 = add(x = var_5808_cast_fp16, y = var_5810_cast_fp16)[name = tensor<string, []>("value_105_cast_fp16")];
            tensor<int32, [4]> var_5813 = const()[name = tensor<string, []>("op_5813"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5814_cast_fp16 = reshape(shape = var_5813, x = query_105_cast_fp16)[name = tensor<string, []>("op_5814_cast_fp16")];
            tensor<fp16, []> var_5815_to_fp16 = const()[name = tensor<string, []>("op_5815_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5816_cast_fp16 = mul(x = var_5814_cast_fp16, y = var_5815_to_fp16)[name = tensor<string, []>("op_5816_cast_fp16")];
            tensor<int32, [4]> var_5817 = const()[name = tensor<string, []>("op_5817"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_5818_cast_fp16 = reshape(shape = var_5817, x = key_105_cast_fp16)[name = tensor<string, []>("op_5818_cast_fp16")];
            tensor<bool, []> mh_w_157_transpose_x_0 = const()[name = tensor<string, []>("mh_w_157_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_157_transpose_y_0 = const()[name = tensor<string, []>("mh_w_157_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_157_cast_fp16 = matmul(transpose_x = mh_w_157_transpose_x_0, transpose_y = mh_w_157_transpose_y_0, x = var_5816_cast_fp16, y = var_5818_cast_fp16)[name = tensor<string, []>("mh_w_157_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_159_cast_fp16 = add(x = mh_w_157_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_159_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_5826_cast_fp16 = softmax(axis = var_5740, x = mh_w_159_cast_fp16)[name = tensor<string, []>("op_5826_cast_fp16")];
            tensor<int32, [4]> var_5827 = const()[name = tensor<string, []>("op_5827"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_5828_cast_fp16 = reshape(shape = var_5827, x = value_105_cast_fp16)[name = tensor<string, []>("op_5828_cast_fp16")];
            tensor<bool, []> attn_105_transpose_x_0 = const()[name = tensor<string, []>("attn_105_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_105_transpose_y_0 = const()[name = tensor<string, []>("attn_105_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_105_cast_fp16 = matmul(transpose_x = attn_105_transpose_x_0, transpose_y = attn_105_transpose_y_0, x = var_5828_cast_fp16, y = var_5826_cast_fp16)[name = tensor<string, []>("attn_105_cast_fp16")];
            tensor<int32, [4]> var_5831 = const()[name = tensor<string, []>("op_5831"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_261_cast_fp16 = reshape(shape = var_5831, x = attn_105_cast_fp16)[name = tensor<string, []>("input_261_cast_fp16")];
            tensor<int32, [2]> var_5835 = const()[name = tensor<string, []>("op_5835"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5837 = const()[name = tensor<string, []>("op_5837"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_319_pad_type_0 = const()[name = tensor<string, []>("obj_319_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_319_pad_0 = const()[name = tensor<string, []>("obj_319_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(696969600))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697788864))), name = tensor<string, []>("layers_26_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_26_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697788992)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_319_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_bias_to_fp16, dilations = var_5837, groups = var_5747, pad = obj_319_pad_0, pad_type = obj_319_pad_type_0, strides = var_5835, weight = layers_26_self_attn_o_proj_weight_to_fp16_palettized, x = input_261_cast_fp16)[name = tensor<string, []>("obj_319_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = obj_319_cast_fp16)[name = tensor<string, []>("inputs_159_cast_fp16")];
            tensor<int32, [1]> var_5847 = const()[name = tensor<string, []>("op_5847"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_159_cast_fp16 = reduce_mean(axes = var_5847, keep_dims = var_5748, x = inputs_159_cast_fp16)[name = tensor<string, []>("channels_mean_159_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_159_cast_fp16 = sub(x = inputs_159_cast_fp16, y = channels_mean_159_cast_fp16)[name = tensor<string, []>("zero_mean_159_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = zero_mean_159_cast_fp16)[name = tensor<string, []>("zero_mean_sq_159_cast_fp16")];
            tensor<int32, [1]> var_5851 = const()[name = tensor<string, []>("op_5851"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5852_cast_fp16 = reduce_mean(axes = var_5851, keep_dims = var_5748, x = zero_mean_sq_159_cast_fp16)[name = tensor<string, []>("op_5852_cast_fp16")];
            tensor<fp16, []> var_5853_to_fp16 = const()[name = tensor<string, []>("op_5853_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5854_cast_fp16 = add(x = var_5852_cast_fp16, y = var_5853_to_fp16)[name = tensor<string, []>("op_5854_cast_fp16")];
            tensor<fp32, []> denom_159_epsilon_0 = const()[name = tensor<string, []>("denom_159_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_159_cast_fp16 = rsqrt(epsilon = denom_159_epsilon_0, x = var_5854_cast_fp16)[name = tensor<string, []>("denom_159_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = denom_159_cast_fp16)[name = tensor<string, []>("out_159_cast_fp16")];
            tensor<fp16, [1280]> obj_321_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_321_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697791616)))];
            tensor<fp16, [1280]> obj_321_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_321_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697794240)))];
            tensor<fp16, []> obj_321_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_321_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_321_cast_fp16 = batch_norm(beta = obj_321_beta_0_to_fp16, epsilon = obj_321_epsilon_0_to_fp16, gamma = obj_321_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor<string, []>("obj_321_cast_fp16")];
            tensor<int32, [2]> var_5869 = const()[name = tensor<string, []>("op_5869"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5871 = const()[name = tensor<string, []>("op_5871"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_107_pad_type_0 = const()[name = tensor<string, []>("query_107_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_107_pad_0 = const()[name = tensor<string, []>("query_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697796864))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(699025728))), name = tensor<string, []>("layers_26_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_26_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(699025920)))];
            tensor<fp16, [1, 1280, 1, 1]> query_107_cast_fp16 = conv(bias = layers_26_encoder_attn_q_proj_bias_to_fp16, dilations = var_5871, groups = var_5747, pad = query_107_pad_0, pad_type = query_107_pad_type_0, strides = var_5869, weight = layers_26_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_321_cast_fp16)[name = tensor<string, []>("query_107_cast_fp16")];
            tensor<int32, [2]> var_5875 = const()[name = tensor<string, []>("op_5875"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5877 = const()[name = tensor<string, []>("op_5877"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_107_pad_type_0 = const()[name = tensor<string, []>("key_107_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_107_pad_0 = const()[name = tensor<string, []>("key_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(699028544))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(699847808))), name = tensor<string, []>("layers_26_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_107_cast_fp16 = conv(dilations = var_5877, groups = var_5747, pad = key_107_pad_0, pad_type = key_107_pad_type_0, strides = var_5875, weight = layers_26_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_107_cast_fp16")];
            tensor<int32, [2]> var_5882 = const()[name = tensor<string, []>("op_5882"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5884 = const()[name = tensor<string, []>("op_5884"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_107_pad_type_0 = const()[name = tensor<string, []>("value_107_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_107_pad_0 = const()[name = tensor<string, []>("value_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(699847936))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(701076800))), name = tensor<string, []>("layers_26_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_26_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(701076992)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_107_cast_fp16 = conv(bias = layers_26_encoder_attn_v_proj_bias_to_fp16, dilations = var_5884, groups = var_5747, pad = value_107_pad_0, pad_type = value_107_pad_type_0, strides = var_5882, weight = layers_26_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_107_cast_fp16")];
            tensor<int32, [4]> var_5888 = const()[name = tensor<string, []>("op_5888"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_5889_cast_fp16 = reshape(shape = var_5888, x = query_107_cast_fp16)[name = tensor<string, []>("op_5889_cast_fp16")];
            tensor<fp16, []> var_5890_to_fp16 = const()[name = tensor<string, []>("op_5890_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_5891_cast_fp16 = mul(x = var_5889_cast_fp16, y = var_5890_to_fp16)[name = tensor<string, []>("op_5891_cast_fp16")];
            tensor<int32, [4]> var_5892 = const()[name = tensor<string, []>("op_5892"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5893_cast_fp16 = reshape(shape = var_5892, x = key_107_cast_fp16)[name = tensor<string, []>("op_5893_cast_fp16")];
            tensor<bool, []> mh_w_161_transpose_x_0 = const()[name = tensor<string, []>("mh_w_161_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_161_transpose_y_0 = const()[name = tensor<string, []>("mh_w_161_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_161_cast_fp16 = matmul(transpose_x = mh_w_161_transpose_x_0, transpose_y = mh_w_161_transpose_y_0, x = var_5891_cast_fp16, y = var_5893_cast_fp16)[name = tensor<string, []>("mh_w_161_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_5896_cast_fp16 = softmax(axis = var_5740, x = mh_w_161_cast_fp16)[name = tensor<string, []>("op_5896_cast_fp16")];
            tensor<int32, [4]> var_5897 = const()[name = tensor<string, []>("op_5897"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_5898_cast_fp16 = reshape(shape = var_5897, x = value_107_cast_fp16)[name = tensor<string, []>("op_5898_cast_fp16")];
            tensor<bool, []> attn_107_transpose_x_0 = const()[name = tensor<string, []>("attn_107_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_107_transpose_y_0 = const()[name = tensor<string, []>("attn_107_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_107_cast_fp16 = matmul(transpose_x = attn_107_transpose_x_0, transpose_y = attn_107_transpose_y_0, x = var_5898_cast_fp16, y = var_5896_cast_fp16)[name = tensor<string, []>("attn_107_cast_fp16")];
            tensor<int32, [4]> var_5901 = const()[name = tensor<string, []>("op_5901"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_263_cast_fp16 = reshape(shape = var_5901, x = attn_107_cast_fp16)[name = tensor<string, []>("input_263_cast_fp16")];
            tensor<int32, [2]> var_5905 = const()[name = tensor<string, []>("op_5905"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5907 = const()[name = tensor<string, []>("op_5907"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_323_pad_type_0 = const()[name = tensor<string, []>("obj_323_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_323_pad_0 = const()[name = tensor<string, []>("obj_323_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_26_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(701079616))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(702308480))), name = tensor<string, []>("layers_26_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_26_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(702308672)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_323_cast_fp16 = conv(bias = layers_26_encoder_attn_o_proj_bias_to_fp16, dilations = var_5907, groups = var_5747, pad = obj_323_pad_0, pad_type = obj_323_pad_type_0, strides = var_5905, weight = layers_26_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_263_cast_fp16)[name = tensor<string, []>("obj_323_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_161_cast_fp16 = add(x = inputs_159_cast_fp16, y = obj_323_cast_fp16)[name = tensor<string, []>("inputs_161_cast_fp16")];
            tensor<int32, [1]> var_5913 = const()[name = tensor<string, []>("op_5913"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_161_cast_fp16 = reduce_mean(axes = var_5913, keep_dims = var_5748, x = inputs_161_cast_fp16)[name = tensor<string, []>("channels_mean_161_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_161_cast_fp16 = sub(x = inputs_161_cast_fp16, y = channels_mean_161_cast_fp16)[name = tensor<string, []>("zero_mean_161_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = zero_mean_161_cast_fp16)[name = tensor<string, []>("zero_mean_sq_161_cast_fp16")];
            tensor<int32, [1]> var_5917 = const()[name = tensor<string, []>("op_5917"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5918_cast_fp16 = reduce_mean(axes = var_5917, keep_dims = var_5748, x = zero_mean_sq_161_cast_fp16)[name = tensor<string, []>("op_5918_cast_fp16")];
            tensor<fp16, []> var_5919_to_fp16 = const()[name = tensor<string, []>("op_5919_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5920_cast_fp16 = add(x = var_5918_cast_fp16, y = var_5919_to_fp16)[name = tensor<string, []>("op_5920_cast_fp16")];
            tensor<fp32, []> denom_161_epsilon_0 = const()[name = tensor<string, []>("denom_161_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_161_cast_fp16 = rsqrt(epsilon = denom_161_epsilon_0, x = var_5920_cast_fp16)[name = tensor<string, []>("denom_161_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = denom_161_cast_fp16)[name = tensor<string, []>("out_161_cast_fp16")];
            tensor<fp16, [1280]> input_265_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_265_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(702311296)))];
            tensor<fp16, [1280]> input_265_beta_0_to_fp16 = const()[name = tensor<string, []>("input_265_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(702313920)))];
            tensor<fp16, []> input_265_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_265_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_265_cast_fp16 = batch_norm(beta = input_265_beta_0_to_fp16, epsilon = input_265_epsilon_0_to_fp16, gamma = input_265_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor<string, []>("input_265_cast_fp16")];
            tensor<int32, [2]> var_5931 = const()[name = tensor<string, []>("op_5931"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5933 = const()[name = tensor<string, []>("op_5933"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_267_pad_type_0 = const()[name = tensor<string, []>("input_267_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_267_pad_0 = const()[name = tensor<string, []>("input_267_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_26_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(702316544))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(705593408))), name = tensor<string, []>("layers_26_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_26_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(705593536)))];
            tensor<fp16, [1, 5120, 1, 1]> input_267_cast_fp16 = conv(bias = layers_26_fc1_bias_to_fp16, dilations = var_5933, groups = var_5747, pad = input_267_pad_0, pad_type = input_267_pad_type_0, strides = var_5931, weight = layers_26_fc1_weight_to_fp16_palettized, x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
            tensor<string, []> input_269_mode_0 = const()[name = tensor<string, []>("input_269_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_269_cast_fp16 = gelu(mode = input_269_mode_0, x = input_267_cast_fp16)[name = tensor<string, []>("input_269_cast_fp16")];
            tensor<int32, [2]> var_5939 = const()[name = tensor<string, []>("op_5939"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5941 = const()[name = tensor<string, []>("op_5941"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_55_pad_type_0 = const()[name = tensor<string, []>("hidden_states_55_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_55_pad_0 = const()[name = tensor<string, []>("hidden_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_26_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(705603840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(710519104))), name = tensor<string, []>("layers_26_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_26_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(710519296)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_55_cast_fp16 = conv(bias = layers_26_fc2_bias_to_fp16, dilations = var_5941, groups = var_5747, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = var_5939, weight = layers_26_fc2_weight_to_fp16_palettized, x = input_269_cast_fp16)[name = tensor<string, []>("hidden_states_55_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor<string, []>("inputs_163_cast_fp16")];
            tensor<int32, []> var_5954 = const()[name = tensor<string, []>("op_5954"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_5961 = const()[name = tensor<string, []>("op_5961"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_5962 = const()[name = tensor<string, []>("op_5962"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_5974 = const()[name = tensor<string, []>("op_5974"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_163_cast_fp16 = reduce_mean(axes = var_5974, keep_dims = var_5962, x = inputs_163_cast_fp16)[name = tensor<string, []>("channels_mean_163_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_163_cast_fp16 = sub(x = inputs_163_cast_fp16, y = channels_mean_163_cast_fp16)[name = tensor<string, []>("zero_mean_163_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = zero_mean_163_cast_fp16)[name = tensor<string, []>("zero_mean_sq_163_cast_fp16")];
            tensor<int32, [1]> var_5978 = const()[name = tensor<string, []>("op_5978"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_5979_cast_fp16 = reduce_mean(axes = var_5978, keep_dims = var_5962, x = zero_mean_sq_163_cast_fp16)[name = tensor<string, []>("op_5979_cast_fp16")];
            tensor<fp16, []> var_5980_to_fp16 = const()[name = tensor<string, []>("op_5980_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_5981_cast_fp16 = add(x = var_5979_cast_fp16, y = var_5980_to_fp16)[name = tensor<string, []>("op_5981_cast_fp16")];
            tensor<fp32, []> denom_163_epsilon_0 = const()[name = tensor<string, []>("denom_163_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_163_cast_fp16 = rsqrt(epsilon = denom_163_epsilon_0, x = var_5981_cast_fp16)[name = tensor<string, []>("denom_163_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = denom_163_cast_fp16)[name = tensor<string, []>("out_163_cast_fp16")];
            tensor<fp16, [1280]> obj_325_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_325_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(710521920)))];
            tensor<fp16, [1280]> obj_325_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_325_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(710524544)))];
            tensor<fp16, []> obj_325_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_325_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_325_cast_fp16 = batch_norm(beta = obj_325_beta_0_to_fp16, epsilon = obj_325_epsilon_0_to_fp16, gamma = obj_325_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_163_cast_fp16)[name = tensor<string, []>("obj_325_cast_fp16")];
            tensor<int32, [2]> var_5996 = const()[name = tensor<string, []>("op_5996"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5998 = const()[name = tensor<string, []>("op_5998"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_109_pad_type_0 = const()[name = tensor<string, []>("query_109_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_109_pad_0 = const()[name = tensor<string, []>("query_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(710527168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(711756032))), name = tensor<string, []>("layers_27_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_27_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(711756224)))];
            tensor<fp16, [1, 1280, 1, 1]> query_109_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_bias_to_fp16, dilations = var_5998, groups = var_5961, pad = query_109_pad_0, pad_type = query_109_pad_type_0, strides = var_5996, weight = layers_27_self_attn_q_proj_weight_to_fp16_palettized, x = obj_325_cast_fp16)[name = tensor<string, []>("query_109_cast_fp16")];
            tensor<int32, [2]> var_6002 = const()[name = tensor<string, []>("op_6002"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6004 = const()[name = tensor<string, []>("op_6004"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_55_pad_type_0 = const()[name = tensor<string, []>("current_key_55_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_55_pad_0 = const()[name = tensor<string, []>("current_key_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(711758848))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712987712))), name = tensor<string, []>("layers_27_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_55_cast_fp16 = conv(dilations = var_6004, groups = var_5961, pad = current_key_55_pad_0, pad_type = current_key_55_pad_type_0, strides = var_6002, weight = layers_27_self_attn_k_proj_weight_to_fp16_palettized, x = obj_325_cast_fp16)[name = tensor<string, []>("current_key_55_cast_fp16")];
            tensor<int32, [2]> var_6009 = const()[name = tensor<string, []>("op_6009"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6011 = const()[name = tensor<string, []>("op_6011"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_55_pad_type_0 = const()[name = tensor<string, []>("current_value_55_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_55_pad_0 = const()[name = tensor<string, []>("current_value_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712987904))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713807168))), name = tensor<string, []>("layers_27_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_27_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713807296)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_55_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_bias_to_fp16, dilations = var_6011, groups = var_5961, pad = current_value_55_pad_0, pad_type = current_value_55_pad_type_0, strides = var_6009, weight = layers_27_self_attn_v_proj_weight_to_fp16_palettized, x = obj_325_cast_fp16)[name = tensor<string, []>("current_value_55_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6018_cast_fp16 = mul(x = current_key_55_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6018_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6020_cast_fp16 = mul(x = var_103_cast_fp16_27, y = var_241_cast_fp16)[name = tensor<string, []>("op_6020_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_109_cast_fp16 = add(x = var_6018_cast_fp16, y = var_6020_cast_fp16)[name = tensor<string, []>("key_109_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6022_cast_fp16 = mul(x = current_value_55_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6022_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6024_cast_fp16 = mul(x = var_138_cast_fp16_27, y = var_241_cast_fp16)[name = tensor<string, []>("op_6024_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_109_cast_fp16 = add(x = var_6022_cast_fp16, y = var_6024_cast_fp16)[name = tensor<string, []>("value_109_cast_fp16")];
            tensor<int32, [4]> var_6027 = const()[name = tensor<string, []>("op_6027"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6028_cast_fp16 = reshape(shape = var_6027, x = query_109_cast_fp16)[name = tensor<string, []>("op_6028_cast_fp16")];
            tensor<fp16, []> var_6029_to_fp16 = const()[name = tensor<string, []>("op_6029_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6030_cast_fp16 = mul(x = var_6028_cast_fp16, y = var_6029_to_fp16)[name = tensor<string, []>("op_6030_cast_fp16")];
            tensor<int32, [4]> var_6031 = const()[name = tensor<string, []>("op_6031"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6032_cast_fp16 = reshape(shape = var_6031, x = key_109_cast_fp16)[name = tensor<string, []>("op_6032_cast_fp16")];
            tensor<bool, []> mh_w_163_transpose_x_0 = const()[name = tensor<string, []>("mh_w_163_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_163_transpose_y_0 = const()[name = tensor<string, []>("mh_w_163_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_163_cast_fp16 = matmul(transpose_x = mh_w_163_transpose_x_0, transpose_y = mh_w_163_transpose_y_0, x = var_6030_cast_fp16, y = var_6032_cast_fp16)[name = tensor<string, []>("mh_w_163_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_165_cast_fp16 = add(x = mh_w_163_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_165_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_6040_cast_fp16 = softmax(axis = var_5954, x = mh_w_165_cast_fp16)[name = tensor<string, []>("op_6040_cast_fp16")];
            tensor<int32, [4]> var_6041 = const()[name = tensor<string, []>("op_6041"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6042_cast_fp16 = reshape(shape = var_6041, x = value_109_cast_fp16)[name = tensor<string, []>("op_6042_cast_fp16")];
            tensor<bool, []> attn_109_transpose_x_0 = const()[name = tensor<string, []>("attn_109_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_109_transpose_y_0 = const()[name = tensor<string, []>("attn_109_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_109_cast_fp16 = matmul(transpose_x = attn_109_transpose_x_0, transpose_y = attn_109_transpose_y_0, x = var_6042_cast_fp16, y = var_6040_cast_fp16)[name = tensor<string, []>("attn_109_cast_fp16")];
            tensor<int32, [4]> var_6045 = const()[name = tensor<string, []>("op_6045"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_271_cast_fp16 = reshape(shape = var_6045, x = attn_109_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
            tensor<int32, [2]> var_6049 = const()[name = tensor<string, []>("op_6049"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6051 = const()[name = tensor<string, []>("op_6051"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_331_pad_type_0 = const()[name = tensor<string, []>("obj_331_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_331_pad_0 = const()[name = tensor<string, []>("obj_331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713809920))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(715038784))), name = tensor<string, []>("layers_27_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_27_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(715038976)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_331_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_bias_to_fp16, dilations = var_6051, groups = var_5961, pad = obj_331_pad_0, pad_type = obj_331_pad_type_0, strides = var_6049, weight = layers_27_self_attn_o_proj_weight_to_fp16_palettized, x = input_271_cast_fp16)[name = tensor<string, []>("obj_331_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_165_cast_fp16 = add(x = inputs_163_cast_fp16, y = obj_331_cast_fp16)[name = tensor<string, []>("inputs_165_cast_fp16")];
            tensor<int32, [1]> var_6061 = const()[name = tensor<string, []>("op_6061"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_165_cast_fp16 = reduce_mean(axes = var_6061, keep_dims = var_5962, x = inputs_165_cast_fp16)[name = tensor<string, []>("channels_mean_165_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_165_cast_fp16 = sub(x = inputs_165_cast_fp16, y = channels_mean_165_cast_fp16)[name = tensor<string, []>("zero_mean_165_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = zero_mean_165_cast_fp16)[name = tensor<string, []>("zero_mean_sq_165_cast_fp16")];
            tensor<int32, [1]> var_6065 = const()[name = tensor<string, []>("op_6065"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6066_cast_fp16 = reduce_mean(axes = var_6065, keep_dims = var_5962, x = zero_mean_sq_165_cast_fp16)[name = tensor<string, []>("op_6066_cast_fp16")];
            tensor<fp16, []> var_6067_to_fp16 = const()[name = tensor<string, []>("op_6067_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6068_cast_fp16 = add(x = var_6066_cast_fp16, y = var_6067_to_fp16)[name = tensor<string, []>("op_6068_cast_fp16")];
            tensor<fp32, []> denom_165_epsilon_0 = const()[name = tensor<string, []>("denom_165_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_165_cast_fp16 = rsqrt(epsilon = denom_165_epsilon_0, x = var_6068_cast_fp16)[name = tensor<string, []>("denom_165_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = denom_165_cast_fp16)[name = tensor<string, []>("out_165_cast_fp16")];
            tensor<fp16, [1280]> obj_333_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_333_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(715041600)))];
            tensor<fp16, [1280]> obj_333_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_333_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(715044224)))];
            tensor<fp16, []> obj_333_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_333_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_333_cast_fp16 = batch_norm(beta = obj_333_beta_0_to_fp16, epsilon = obj_333_epsilon_0_to_fp16, gamma = obj_333_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor<string, []>("obj_333_cast_fp16")];
            tensor<int32, [2]> var_6083 = const()[name = tensor<string, []>("op_6083"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6085 = const()[name = tensor<string, []>("op_6085"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_111_pad_type_0 = const()[name = tensor<string, []>("query_111_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_111_pad_0 = const()[name = tensor<string, []>("query_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(715046848))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(716275712))), name = tensor<string, []>("layers_27_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_27_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(716275904)))];
            tensor<fp16, [1, 1280, 1, 1]> query_111_cast_fp16 = conv(bias = layers_27_encoder_attn_q_proj_bias_to_fp16, dilations = var_6085, groups = var_5961, pad = query_111_pad_0, pad_type = query_111_pad_type_0, strides = var_6083, weight = layers_27_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_333_cast_fp16)[name = tensor<string, []>("query_111_cast_fp16")];
            tensor<int32, [2]> var_6089 = const()[name = tensor<string, []>("op_6089"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6091 = const()[name = tensor<string, []>("op_6091"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_111_pad_type_0 = const()[name = tensor<string, []>("key_111_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_111_pad_0 = const()[name = tensor<string, []>("key_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(716278528))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(717916992))), name = tensor<string, []>("layers_27_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_111_cast_fp16 = conv(dilations = var_6091, groups = var_5961, pad = key_111_pad_0, pad_type = key_111_pad_type_0, strides = var_6089, weight = layers_27_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_111_cast_fp16")];
            tensor<int32, [2]> var_6096 = const()[name = tensor<string, []>("op_6096"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6098 = const()[name = tensor<string, []>("op_6098"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_111_pad_type_0 = const()[name = tensor<string, []>("value_111_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_111_pad_0 = const()[name = tensor<string, []>("value_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(717917568)))];
            tensor<fp16, [1280]> layers_27_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721194432)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_111_cast_fp16 = conv(bias = layers_27_encoder_attn_v_proj_bias_to_fp16, dilations = var_6098, groups = var_5961, pad = value_111_pad_0, pad_type = value_111_pad_type_0, strides = var_6096, weight = layers_27_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_111_cast_fp16")];
            tensor<int32, [4]> var_6102 = const()[name = tensor<string, []>("op_6102"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6103_cast_fp16 = reshape(shape = var_6102, x = query_111_cast_fp16)[name = tensor<string, []>("op_6103_cast_fp16")];
            tensor<fp16, []> var_6104_to_fp16 = const()[name = tensor<string, []>("op_6104_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6105_cast_fp16 = mul(x = var_6103_cast_fp16, y = var_6104_to_fp16)[name = tensor<string, []>("op_6105_cast_fp16")];
            tensor<int32, [4]> var_6106 = const()[name = tensor<string, []>("op_6106"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6107_cast_fp16 = reshape(shape = var_6106, x = key_111_cast_fp16)[name = tensor<string, []>("op_6107_cast_fp16")];
            tensor<bool, []> mh_w_167_transpose_x_0 = const()[name = tensor<string, []>("mh_w_167_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_167_transpose_y_0 = const()[name = tensor<string, []>("mh_w_167_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_167_cast_fp16 = matmul(transpose_x = mh_w_167_transpose_x_0, transpose_y = mh_w_167_transpose_y_0, x = var_6105_cast_fp16, y = var_6107_cast_fp16)[name = tensor<string, []>("mh_w_167_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_6110_cast_fp16 = softmax(axis = var_5954, x = mh_w_167_cast_fp16)[name = tensor<string, []>("op_6110_cast_fp16")];
            tensor<int32, [4]> var_6111 = const()[name = tensor<string, []>("op_6111"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6112_cast_fp16 = reshape(shape = var_6111, x = value_111_cast_fp16)[name = tensor<string, []>("op_6112_cast_fp16")];
            tensor<bool, []> attn_111_transpose_x_0 = const()[name = tensor<string, []>("attn_111_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_111_transpose_y_0 = const()[name = tensor<string, []>("attn_111_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_111_cast_fp16 = matmul(transpose_x = attn_111_transpose_x_0, transpose_y = attn_111_transpose_y_0, x = var_6112_cast_fp16, y = var_6110_cast_fp16)[name = tensor<string, []>("attn_111_cast_fp16")];
            tensor<int32, [4]> var_6115 = const()[name = tensor<string, []>("op_6115"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_273_cast_fp16 = reshape(shape = var_6115, x = attn_111_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
            tensor<int32, [2]> var_6119 = const()[name = tensor<string, []>("op_6119"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6121 = const()[name = tensor<string, []>("op_6121"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_335_pad_type_0 = const()[name = tensor<string, []>("obj_335_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_335_pad_0 = const()[name = tensor<string, []>("obj_335_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_27_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721197056)))];
            tensor<fp16, [1280]> layers_27_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724473920)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_335_cast_fp16 = conv(bias = layers_27_encoder_attn_o_proj_bias_to_fp16, dilations = var_6121, groups = var_5961, pad = obj_335_pad_0, pad_type = obj_335_pad_type_0, strides = var_6119, weight = layers_27_encoder_attn_o_proj_weight_to_fp16, x = input_273_cast_fp16)[name = tensor<string, []>("obj_335_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_167_cast_fp16 = add(x = inputs_165_cast_fp16, y = obj_335_cast_fp16)[name = tensor<string, []>("inputs_167_cast_fp16")];
            tensor<int32, [1]> var_6127 = const()[name = tensor<string, []>("op_6127"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_167_cast_fp16 = reduce_mean(axes = var_6127, keep_dims = var_5962, x = inputs_167_cast_fp16)[name = tensor<string, []>("channels_mean_167_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_167_cast_fp16 = sub(x = inputs_167_cast_fp16, y = channels_mean_167_cast_fp16)[name = tensor<string, []>("zero_mean_167_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = zero_mean_167_cast_fp16)[name = tensor<string, []>("zero_mean_sq_167_cast_fp16")];
            tensor<int32, [1]> var_6131 = const()[name = tensor<string, []>("op_6131"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6132_cast_fp16 = reduce_mean(axes = var_6131, keep_dims = var_5962, x = zero_mean_sq_167_cast_fp16)[name = tensor<string, []>("op_6132_cast_fp16")];
            tensor<fp16, []> var_6133_to_fp16 = const()[name = tensor<string, []>("op_6133_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6134_cast_fp16 = add(x = var_6132_cast_fp16, y = var_6133_to_fp16)[name = tensor<string, []>("op_6134_cast_fp16")];
            tensor<fp32, []> denom_167_epsilon_0 = const()[name = tensor<string, []>("denom_167_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_167_cast_fp16 = rsqrt(epsilon = denom_167_epsilon_0, x = var_6134_cast_fp16)[name = tensor<string, []>("denom_167_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = denom_167_cast_fp16)[name = tensor<string, []>("out_167_cast_fp16")];
            tensor<fp16, [1280]> input_275_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_275_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724476544)))];
            tensor<fp16, [1280]> input_275_beta_0_to_fp16 = const()[name = tensor<string, []>("input_275_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724479168)))];
            tensor<fp16, []> input_275_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_275_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_275_cast_fp16 = batch_norm(beta = input_275_beta_0_to_fp16, epsilon = input_275_epsilon_0_to_fp16, gamma = input_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor<string, []>("input_275_cast_fp16")];
            tensor<int32, [2]> var_6145 = const()[name = tensor<string, []>("op_6145"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6147 = const()[name = tensor<string, []>("op_6147"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_277_pad_type_0 = const()[name = tensor<string, []>("input_277_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_277_pad_0 = const()[name = tensor<string, []>("input_277_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_27_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724481792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(729397056))), name = tensor<string, []>("layers_27_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_27_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(729397248)))];
            tensor<fp16, [1, 5120, 1, 1]> input_277_cast_fp16 = conv(bias = layers_27_fc1_bias_to_fp16, dilations = var_6147, groups = var_5961, pad = input_277_pad_0, pad_type = input_277_pad_type_0, strides = var_6145, weight = layers_27_fc1_weight_to_fp16_palettized, x = input_275_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
            tensor<string, []> input_279_mode_0 = const()[name = tensor<string, []>("input_279_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_279_cast_fp16 = gelu(mode = input_279_mode_0, x = input_277_cast_fp16)[name = tensor<string, []>("input_279_cast_fp16")];
            tensor<int32, [2]> var_6153 = const()[name = tensor<string, []>("op_6153"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6155 = const()[name = tensor<string, []>("op_6155"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_57_pad_type_0 = const()[name = tensor<string, []>("hidden_states_57_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = tensor<string, []>("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_27_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(729407552))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734322816))), name = tensor<string, []>("layers_27_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_27_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734323008)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_57_cast_fp16 = conv(bias = layers_27_fc2_bias_to_fp16, dilations = var_6155, groups = var_5961, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = var_6153, weight = layers_27_fc2_weight_to_fp16_palettized, x = input_279_cast_fp16)[name = tensor<string, []>("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor<string, []>("inputs_169_cast_fp16")];
            tensor<int32, []> var_6168 = const()[name = tensor<string, []>("op_6168"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_6175 = const()[name = tensor<string, []>("op_6175"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_6176 = const()[name = tensor<string, []>("op_6176"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_6188 = const()[name = tensor<string, []>("op_6188"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_169_cast_fp16 = reduce_mean(axes = var_6188, keep_dims = var_6176, x = inputs_169_cast_fp16)[name = tensor<string, []>("channels_mean_169_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_169_cast_fp16 = sub(x = inputs_169_cast_fp16, y = channels_mean_169_cast_fp16)[name = tensor<string, []>("zero_mean_169_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = zero_mean_169_cast_fp16)[name = tensor<string, []>("zero_mean_sq_169_cast_fp16")];
            tensor<int32, [1]> var_6192 = const()[name = tensor<string, []>("op_6192"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6193_cast_fp16 = reduce_mean(axes = var_6192, keep_dims = var_6176, x = zero_mean_sq_169_cast_fp16)[name = tensor<string, []>("op_6193_cast_fp16")];
            tensor<fp16, []> var_6194_to_fp16 = const()[name = tensor<string, []>("op_6194_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6195_cast_fp16 = add(x = var_6193_cast_fp16, y = var_6194_to_fp16)[name = tensor<string, []>("op_6195_cast_fp16")];
            tensor<fp32, []> denom_169_epsilon_0 = const()[name = tensor<string, []>("denom_169_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_169_cast_fp16 = rsqrt(epsilon = denom_169_epsilon_0, x = var_6195_cast_fp16)[name = tensor<string, []>("denom_169_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = denom_169_cast_fp16)[name = tensor<string, []>("out_169_cast_fp16")];
            tensor<fp16, [1280]> obj_337_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_337_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734325632)))];
            tensor<fp16, [1280]> obj_337_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_337_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734328256)))];
            tensor<fp16, []> obj_337_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_337_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_337_cast_fp16 = batch_norm(beta = obj_337_beta_0_to_fp16, epsilon = obj_337_epsilon_0_to_fp16, gamma = obj_337_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor<string, []>("obj_337_cast_fp16")];
            tensor<int32, [2]> var_6210 = const()[name = tensor<string, []>("op_6210"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6212 = const()[name = tensor<string, []>("op_6212"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_113_pad_type_0 = const()[name = tensor<string, []>("query_113_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_113_pad_0 = const()[name = tensor<string, []>("query_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734330880))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(735559744))), name = tensor<string, []>("layers_28_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_28_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(735559936)))];
            tensor<fp16, [1, 1280, 1, 1]> query_113_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_bias_to_fp16, dilations = var_6212, groups = var_6175, pad = query_113_pad_0, pad_type = query_113_pad_type_0, strides = var_6210, weight = layers_28_self_attn_q_proj_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor<string, []>("query_113_cast_fp16")];
            tensor<int32, [2]> var_6216 = const()[name = tensor<string, []>("op_6216"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6218 = const()[name = tensor<string, []>("op_6218"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_57_pad_type_0 = const()[name = tensor<string, []>("current_key_57_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_57_pad_0 = const()[name = tensor<string, []>("current_key_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(735562560))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736381824))), name = tensor<string, []>("layers_28_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_57_cast_fp16 = conv(dilations = var_6218, groups = var_6175, pad = current_key_57_pad_0, pad_type = current_key_57_pad_type_0, strides = var_6216, weight = layers_28_self_attn_k_proj_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor<string, []>("current_key_57_cast_fp16")];
            tensor<int32, [2]> var_6223 = const()[name = tensor<string, []>("op_6223"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6225 = const()[name = tensor<string, []>("op_6225"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_57_pad_type_0 = const()[name = tensor<string, []>("current_value_57_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_57_pad_0 = const()[name = tensor<string, []>("current_value_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736381952))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(737201216))), name = tensor<string, []>("layers_28_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_28_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(737201344)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_57_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_bias_to_fp16, dilations = var_6225, groups = var_6175, pad = current_value_57_pad_0, pad_type = current_value_57_pad_type_0, strides = var_6223, weight = layers_28_self_attn_v_proj_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor<string, []>("current_value_57_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6232_cast_fp16 = mul(x = current_key_57_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6232_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6234_cast_fp16 = mul(x = var_103_cast_fp16_28, y = var_241_cast_fp16)[name = tensor<string, []>("op_6234_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_113_cast_fp16 = add(x = var_6232_cast_fp16, y = var_6234_cast_fp16)[name = tensor<string, []>("key_113_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6236_cast_fp16 = mul(x = current_value_57_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6236_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6238_cast_fp16 = mul(x = var_138_cast_fp16_28, y = var_241_cast_fp16)[name = tensor<string, []>("op_6238_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_113_cast_fp16 = add(x = var_6236_cast_fp16, y = var_6238_cast_fp16)[name = tensor<string, []>("value_113_cast_fp16")];
            tensor<int32, [4]> var_6241 = const()[name = tensor<string, []>("op_6241"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6242_cast_fp16 = reshape(shape = var_6241, x = query_113_cast_fp16)[name = tensor<string, []>("op_6242_cast_fp16")];
            tensor<fp16, []> var_6243_to_fp16 = const()[name = tensor<string, []>("op_6243_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6244_cast_fp16 = mul(x = var_6242_cast_fp16, y = var_6243_to_fp16)[name = tensor<string, []>("op_6244_cast_fp16")];
            tensor<int32, [4]> var_6245 = const()[name = tensor<string, []>("op_6245"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6246_cast_fp16 = reshape(shape = var_6245, x = key_113_cast_fp16)[name = tensor<string, []>("op_6246_cast_fp16")];
            tensor<bool, []> mh_w_169_transpose_x_0 = const()[name = tensor<string, []>("mh_w_169_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_169_transpose_y_0 = const()[name = tensor<string, []>("mh_w_169_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_169_cast_fp16 = matmul(transpose_x = mh_w_169_transpose_x_0, transpose_y = mh_w_169_transpose_y_0, x = var_6244_cast_fp16, y = var_6246_cast_fp16)[name = tensor<string, []>("mh_w_169_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_171_cast_fp16 = add(x = mh_w_169_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_171_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_6254_cast_fp16 = softmax(axis = var_6168, x = mh_w_171_cast_fp16)[name = tensor<string, []>("op_6254_cast_fp16")];
            tensor<int32, [4]> var_6255 = const()[name = tensor<string, []>("op_6255"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6256_cast_fp16 = reshape(shape = var_6255, x = value_113_cast_fp16)[name = tensor<string, []>("op_6256_cast_fp16")];
            tensor<bool, []> attn_113_transpose_x_0 = const()[name = tensor<string, []>("attn_113_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_113_transpose_y_0 = const()[name = tensor<string, []>("attn_113_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_113_cast_fp16 = matmul(transpose_x = attn_113_transpose_x_0, transpose_y = attn_113_transpose_y_0, x = var_6256_cast_fp16, y = var_6254_cast_fp16)[name = tensor<string, []>("attn_113_cast_fp16")];
            tensor<int32, [4]> var_6259 = const()[name = tensor<string, []>("op_6259"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_281_cast_fp16 = reshape(shape = var_6259, x = attn_113_cast_fp16)[name = tensor<string, []>("input_281_cast_fp16")];
            tensor<int32, [2]> var_6263 = const()[name = tensor<string, []>("op_6263"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6265 = const()[name = tensor<string, []>("op_6265"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_343_pad_type_0 = const()[name = tensor<string, []>("obj_343_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_343_pad_0 = const()[name = tensor<string, []>("obj_343_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(737203968))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738023232))), name = tensor<string, []>("layers_28_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_28_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738023360)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_343_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_bias_to_fp16, dilations = var_6265, groups = var_6175, pad = obj_343_pad_0, pad_type = obj_343_pad_type_0, strides = var_6263, weight = layers_28_self_attn_o_proj_weight_to_fp16_palettized, x = input_281_cast_fp16)[name = tensor<string, []>("obj_343_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_171_cast_fp16 = add(x = inputs_169_cast_fp16, y = obj_343_cast_fp16)[name = tensor<string, []>("inputs_171_cast_fp16")];
            tensor<int32, [1]> var_6275 = const()[name = tensor<string, []>("op_6275"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_171_cast_fp16 = reduce_mean(axes = var_6275, keep_dims = var_6176, x = inputs_171_cast_fp16)[name = tensor<string, []>("channels_mean_171_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_171_cast_fp16 = sub(x = inputs_171_cast_fp16, y = channels_mean_171_cast_fp16)[name = tensor<string, []>("zero_mean_171_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = zero_mean_171_cast_fp16)[name = tensor<string, []>("zero_mean_sq_171_cast_fp16")];
            tensor<int32, [1]> var_6279 = const()[name = tensor<string, []>("op_6279"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6280_cast_fp16 = reduce_mean(axes = var_6279, keep_dims = var_6176, x = zero_mean_sq_171_cast_fp16)[name = tensor<string, []>("op_6280_cast_fp16")];
            tensor<fp16, []> var_6281_to_fp16 = const()[name = tensor<string, []>("op_6281_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6282_cast_fp16 = add(x = var_6280_cast_fp16, y = var_6281_to_fp16)[name = tensor<string, []>("op_6282_cast_fp16")];
            tensor<fp32, []> denom_171_epsilon_0 = const()[name = tensor<string, []>("denom_171_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_171_cast_fp16 = rsqrt(epsilon = denom_171_epsilon_0, x = var_6282_cast_fp16)[name = tensor<string, []>("denom_171_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = denom_171_cast_fp16)[name = tensor<string, []>("out_171_cast_fp16")];
            tensor<fp16, [1280]> obj_345_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_345_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738025984)))];
            tensor<fp16, [1280]> obj_345_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_345_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738028608)))];
            tensor<fp16, []> obj_345_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_345_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_345_cast_fp16 = batch_norm(beta = obj_345_beta_0_to_fp16, epsilon = obj_345_epsilon_0_to_fp16, gamma = obj_345_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor<string, []>("obj_345_cast_fp16")];
            tensor<int32, [2]> var_6297 = const()[name = tensor<string, []>("op_6297"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6299 = const()[name = tensor<string, []>("op_6299"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_115_pad_type_0 = const()[name = tensor<string, []>("query_115_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_115_pad_0 = const()[name = tensor<string, []>("query_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738031232))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(739260096))), name = tensor<string, []>("layers_28_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_28_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(739260288)))];
            tensor<fp16, [1, 1280, 1, 1]> query_115_cast_fp16 = conv(bias = layers_28_encoder_attn_q_proj_bias_to_fp16, dilations = var_6299, groups = var_6175, pad = query_115_pad_0, pad_type = query_115_pad_type_0, strides = var_6297, weight = layers_28_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_345_cast_fp16)[name = tensor<string, []>("query_115_cast_fp16")];
            tensor<int32, [2]> var_6303 = const()[name = tensor<string, []>("op_6303"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6305 = const()[name = tensor<string, []>("op_6305"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_115_pad_type_0 = const()[name = tensor<string, []>("key_115_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_115_pad_0 = const()[name = tensor<string, []>("key_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(739262912))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(740491776))), name = tensor<string, []>("layers_28_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_115_cast_fp16 = conv(dilations = var_6305, groups = var_6175, pad = key_115_pad_0, pad_type = key_115_pad_type_0, strides = var_6303, weight = layers_28_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_115_cast_fp16")];
            tensor<int32, [2]> var_6310 = const()[name = tensor<string, []>("op_6310"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6312 = const()[name = tensor<string, []>("op_6312"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_115_pad_type_0 = const()[name = tensor<string, []>("value_115_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_115_pad_0 = const()[name = tensor<string, []>("value_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1638400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(740491968))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(742130432))), name = tensor<string, []>("layers_28_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_28_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(742131008)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_115_cast_fp16 = conv(bias = layers_28_encoder_attn_v_proj_bias_to_fp16, dilations = var_6312, groups = var_6175, pad = value_115_pad_0, pad_type = value_115_pad_type_0, strides = var_6310, weight = layers_28_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_115_cast_fp16")];
            tensor<int32, [4]> var_6316 = const()[name = tensor<string, []>("op_6316"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6317_cast_fp16 = reshape(shape = var_6316, x = query_115_cast_fp16)[name = tensor<string, []>("op_6317_cast_fp16")];
            tensor<fp16, []> var_6318_to_fp16 = const()[name = tensor<string, []>("op_6318_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6319_cast_fp16 = mul(x = var_6317_cast_fp16, y = var_6318_to_fp16)[name = tensor<string, []>("op_6319_cast_fp16")];
            tensor<int32, [4]> var_6320 = const()[name = tensor<string, []>("op_6320"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6321_cast_fp16 = reshape(shape = var_6320, x = key_115_cast_fp16)[name = tensor<string, []>("op_6321_cast_fp16")];
            tensor<bool, []> mh_w_173_transpose_x_0 = const()[name = tensor<string, []>("mh_w_173_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_173_transpose_y_0 = const()[name = tensor<string, []>("mh_w_173_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_173_cast_fp16 = matmul(transpose_x = mh_w_173_transpose_x_0, transpose_y = mh_w_173_transpose_y_0, x = var_6319_cast_fp16, y = var_6321_cast_fp16)[name = tensor<string, []>("mh_w_173_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_6324_cast_fp16 = softmax(axis = var_6168, x = mh_w_173_cast_fp16)[name = tensor<string, []>("op_6324_cast_fp16")];
            tensor<int32, [4]> var_6325 = const()[name = tensor<string, []>("op_6325"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6326_cast_fp16 = reshape(shape = var_6325, x = value_115_cast_fp16)[name = tensor<string, []>("op_6326_cast_fp16")];
            tensor<bool, []> attn_115_transpose_x_0 = const()[name = tensor<string, []>("attn_115_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_115_transpose_y_0 = const()[name = tensor<string, []>("attn_115_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_115_cast_fp16 = matmul(transpose_x = attn_115_transpose_x_0, transpose_y = attn_115_transpose_y_0, x = var_6326_cast_fp16, y = var_6324_cast_fp16)[name = tensor<string, []>("attn_115_cast_fp16")];
            tensor<int32, [4]> var_6329 = const()[name = tensor<string, []>("op_6329"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_283_cast_fp16 = reshape(shape = var_6329, x = attn_115_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
            tensor<int32, [2]> var_6333 = const()[name = tensor<string, []>("op_6333"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6335 = const()[name = tensor<string, []>("op_6335"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_347_pad_type_0 = const()[name = tensor<string, []>("obj_347_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_347_pad_0 = const()[name = tensor<string, []>("obj_347_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_28_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(742133632))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(743362496))), name = tensor<string, []>("layers_28_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_28_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(743362688)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_347_cast_fp16 = conv(bias = layers_28_encoder_attn_o_proj_bias_to_fp16, dilations = var_6335, groups = var_6175, pad = obj_347_pad_0, pad_type = obj_347_pad_type_0, strides = var_6333, weight = layers_28_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_283_cast_fp16)[name = tensor<string, []>("obj_347_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = obj_347_cast_fp16)[name = tensor<string, []>("inputs_173_cast_fp16")];
            tensor<int32, [1]> var_6341 = const()[name = tensor<string, []>("op_6341"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_173_cast_fp16 = reduce_mean(axes = var_6341, keep_dims = var_6176, x = inputs_173_cast_fp16)[name = tensor<string, []>("channels_mean_173_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_173_cast_fp16 = sub(x = inputs_173_cast_fp16, y = channels_mean_173_cast_fp16)[name = tensor<string, []>("zero_mean_173_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = zero_mean_173_cast_fp16)[name = tensor<string, []>("zero_mean_sq_173_cast_fp16")];
            tensor<int32, [1]> var_6345 = const()[name = tensor<string, []>("op_6345"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6346_cast_fp16 = reduce_mean(axes = var_6345, keep_dims = var_6176, x = zero_mean_sq_173_cast_fp16)[name = tensor<string, []>("op_6346_cast_fp16")];
            tensor<fp16, []> var_6347_to_fp16 = const()[name = tensor<string, []>("op_6347_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6348_cast_fp16 = add(x = var_6346_cast_fp16, y = var_6347_to_fp16)[name = tensor<string, []>("op_6348_cast_fp16")];
            tensor<fp32, []> denom_173_epsilon_0 = const()[name = tensor<string, []>("denom_173_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_173_cast_fp16 = rsqrt(epsilon = denom_173_epsilon_0, x = var_6348_cast_fp16)[name = tensor<string, []>("denom_173_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = denom_173_cast_fp16)[name = tensor<string, []>("out_173_cast_fp16")];
            tensor<fp16, [1280]> input_285_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_285_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(743365312)))];
            tensor<fp16, [1280]> input_285_beta_0_to_fp16 = const()[name = tensor<string, []>("input_285_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(743367936)))];
            tensor<fp16, []> input_285_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_285_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_285_cast_fp16 = batch_norm(beta = input_285_beta_0_to_fp16, epsilon = input_285_epsilon_0_to_fp16, gamma = input_285_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_173_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
            tensor<int32, [2]> var_6359 = const()[name = tensor<string, []>("op_6359"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6361 = const()[name = tensor<string, []>("op_6361"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_287_pad_type_0 = const()[name = tensor<string, []>("input_287_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_287_pad_0 = const()[name = tensor<string, []>("input_287_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_28_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(743370560))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748285824))), name = tensor<string, []>("layers_28_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_28_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748286016)))];
            tensor<fp16, [1, 5120, 1, 1]> input_287_cast_fp16 = conv(bias = layers_28_fc1_bias_to_fp16, dilations = var_6361, groups = var_6175, pad = input_287_pad_0, pad_type = input_287_pad_type_0, strides = var_6359, weight = layers_28_fc1_weight_to_fp16_palettized, x = input_285_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
            tensor<string, []> input_289_mode_0 = const()[name = tensor<string, []>("input_289_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = input_287_cast_fp16)[name = tensor<string, []>("input_289_cast_fp16")];
            tensor<int32, [2]> var_6367 = const()[name = tensor<string, []>("op_6367"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6369 = const()[name = tensor<string, []>("op_6369"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_59_pad_type_0 = const()[name = tensor<string, []>("hidden_states_59_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_59_pad_0 = const()[name = tensor<string, []>("hidden_states_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_28_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748296320))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753211584))), name = tensor<string, []>("layers_28_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_28_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753211776)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_59_cast_fp16 = conv(bias = layers_28_fc2_bias_to_fp16, dilations = var_6369, groups = var_6175, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = var_6367, weight = layers_28_fc2_weight_to_fp16_palettized, x = input_289_cast_fp16)[name = tensor<string, []>("hidden_states_59_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_175_cast_fp16 = add(x = inputs_173_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor<string, []>("inputs_175_cast_fp16")];
            tensor<int32, []> var_6382 = const()[name = tensor<string, []>("op_6382"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_6389 = const()[name = tensor<string, []>("op_6389"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_6390 = const()[name = tensor<string, []>("op_6390"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_6402 = const()[name = tensor<string, []>("op_6402"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_175_cast_fp16 = reduce_mean(axes = var_6402, keep_dims = var_6390, x = inputs_175_cast_fp16)[name = tensor<string, []>("channels_mean_175_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_175_cast_fp16 = sub(x = inputs_175_cast_fp16, y = channels_mean_175_cast_fp16)[name = tensor<string, []>("zero_mean_175_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = zero_mean_175_cast_fp16)[name = tensor<string, []>("zero_mean_sq_175_cast_fp16")];
            tensor<int32, [1]> var_6406 = const()[name = tensor<string, []>("op_6406"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6407_cast_fp16 = reduce_mean(axes = var_6406, keep_dims = var_6390, x = zero_mean_sq_175_cast_fp16)[name = tensor<string, []>("op_6407_cast_fp16")];
            tensor<fp16, []> var_6408_to_fp16 = const()[name = tensor<string, []>("op_6408_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6409_cast_fp16 = add(x = var_6407_cast_fp16, y = var_6408_to_fp16)[name = tensor<string, []>("op_6409_cast_fp16")];
            tensor<fp32, []> denom_175_epsilon_0 = const()[name = tensor<string, []>("denom_175_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_175_cast_fp16 = rsqrt(epsilon = denom_175_epsilon_0, x = var_6409_cast_fp16)[name = tensor<string, []>("denom_175_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = denom_175_cast_fp16)[name = tensor<string, []>("out_175_cast_fp16")];
            tensor<fp16, [1280]> obj_349_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_349_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753214400)))];
            tensor<fp16, [1280]> obj_349_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_349_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753217024)))];
            tensor<fp16, []> obj_349_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_349_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_349_cast_fp16 = batch_norm(beta = obj_349_beta_0_to_fp16, epsilon = obj_349_epsilon_0_to_fp16, gamma = obj_349_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor<string, []>("obj_349_cast_fp16")];
            tensor<int32, [2]> var_6424 = const()[name = tensor<string, []>("op_6424"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6426 = const()[name = tensor<string, []>("op_6426"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_117_pad_type_0 = const()[name = tensor<string, []>("query_117_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_117_pad_0 = const()[name = tensor<string, []>("query_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753219648))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754038912))), name = tensor<string, []>("layers_29_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_29_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754039040)))];
            tensor<fp16, [1, 1280, 1, 1]> query_117_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_bias_to_fp16, dilations = var_6426, groups = var_6389, pad = query_117_pad_0, pad_type = query_117_pad_type_0, strides = var_6424, weight = layers_29_self_attn_q_proj_weight_to_fp16_palettized, x = obj_349_cast_fp16)[name = tensor<string, []>("query_117_cast_fp16")];
            tensor<int32, [2]> var_6430 = const()[name = tensor<string, []>("op_6430"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6432 = const()[name = tensor<string, []>("op_6432"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_59_pad_type_0 = const()[name = tensor<string, []>("current_key_59_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_59_pad_0 = const()[name = tensor<string, []>("current_key_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754041664))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754860928))), name = tensor<string, []>("layers_29_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_59_cast_fp16 = conv(dilations = var_6432, groups = var_6389, pad = current_key_59_pad_0, pad_type = current_key_59_pad_type_0, strides = var_6430, weight = layers_29_self_attn_k_proj_weight_to_fp16_palettized, x = obj_349_cast_fp16)[name = tensor<string, []>("current_key_59_cast_fp16")];
            tensor<int32, [2]> var_6437 = const()[name = tensor<string, []>("op_6437"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6439 = const()[name = tensor<string, []>("op_6439"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_59_pad_type_0 = const()[name = tensor<string, []>("current_value_59_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_59_pad_0 = const()[name = tensor<string, []>("current_value_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754861056))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(755680320))), name = tensor<string, []>("layers_29_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_29_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(755680448)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_59_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_bias_to_fp16, dilations = var_6439, groups = var_6389, pad = current_value_59_pad_0, pad_type = current_value_59_pad_type_0, strides = var_6437, weight = layers_29_self_attn_v_proj_weight_to_fp16_palettized, x = obj_349_cast_fp16)[name = tensor<string, []>("current_value_59_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6446_cast_fp16 = mul(x = current_key_59_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6446_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6448_cast_fp16 = mul(x = var_103_cast_fp16_29, y = var_241_cast_fp16)[name = tensor<string, []>("op_6448_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_117_cast_fp16 = add(x = var_6446_cast_fp16, y = var_6448_cast_fp16)[name = tensor<string, []>("key_117_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6450_cast_fp16 = mul(x = current_value_59_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6450_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6452_cast_fp16 = mul(x = var_138_cast_fp16_29, y = var_241_cast_fp16)[name = tensor<string, []>("op_6452_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_117_cast_fp16 = add(x = var_6450_cast_fp16, y = var_6452_cast_fp16)[name = tensor<string, []>("value_117_cast_fp16")];
            tensor<int32, [4]> var_6455 = const()[name = tensor<string, []>("op_6455"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6456_cast_fp16 = reshape(shape = var_6455, x = query_117_cast_fp16)[name = tensor<string, []>("op_6456_cast_fp16")];
            tensor<fp16, []> var_6457_to_fp16 = const()[name = tensor<string, []>("op_6457_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6458_cast_fp16 = mul(x = var_6456_cast_fp16, y = var_6457_to_fp16)[name = tensor<string, []>("op_6458_cast_fp16")];
            tensor<int32, [4]> var_6459 = const()[name = tensor<string, []>("op_6459"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6460_cast_fp16 = reshape(shape = var_6459, x = key_117_cast_fp16)[name = tensor<string, []>("op_6460_cast_fp16")];
            tensor<bool, []> mh_w_175_transpose_x_0 = const()[name = tensor<string, []>("mh_w_175_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_175_transpose_y_0 = const()[name = tensor<string, []>("mh_w_175_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_175_cast_fp16 = matmul(transpose_x = mh_w_175_transpose_x_0, transpose_y = mh_w_175_transpose_y_0, x = var_6458_cast_fp16, y = var_6460_cast_fp16)[name = tensor<string, []>("mh_w_175_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_177_cast_fp16 = add(x = mh_w_175_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_177_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_6468_cast_fp16 = softmax(axis = var_6382, x = mh_w_177_cast_fp16)[name = tensor<string, []>("op_6468_cast_fp16")];
            tensor<int32, [4]> var_6469 = const()[name = tensor<string, []>("op_6469"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6470_cast_fp16 = reshape(shape = var_6469, x = value_117_cast_fp16)[name = tensor<string, []>("op_6470_cast_fp16")];
            tensor<bool, []> attn_117_transpose_x_0 = const()[name = tensor<string, []>("attn_117_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_117_transpose_y_0 = const()[name = tensor<string, []>("attn_117_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_117_cast_fp16 = matmul(transpose_x = attn_117_transpose_x_0, transpose_y = attn_117_transpose_y_0, x = var_6470_cast_fp16, y = var_6468_cast_fp16)[name = tensor<string, []>("attn_117_cast_fp16")];
            tensor<int32, [4]> var_6473 = const()[name = tensor<string, []>("op_6473"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_291_cast_fp16 = reshape(shape = var_6473, x = attn_117_cast_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
            tensor<int32, [2]> var_6477 = const()[name = tensor<string, []>("op_6477"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6479 = const()[name = tensor<string, []>("op_6479"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_355_pad_type_0 = const()[name = tensor<string, []>("obj_355_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_355_pad_0 = const()[name = tensor<string, []>("obj_355_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(755683072))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756502336))), name = tensor<string, []>("layers_29_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_29_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756502464)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_355_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_bias_to_fp16, dilations = var_6479, groups = var_6389, pad = obj_355_pad_0, pad_type = obj_355_pad_type_0, strides = var_6477, weight = layers_29_self_attn_o_proj_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = tensor<string, []>("obj_355_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_177_cast_fp16 = add(x = inputs_175_cast_fp16, y = obj_355_cast_fp16)[name = tensor<string, []>("inputs_177_cast_fp16")];
            tensor<int32, [1]> var_6489 = const()[name = tensor<string, []>("op_6489"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_177_cast_fp16 = reduce_mean(axes = var_6489, keep_dims = var_6390, x = inputs_177_cast_fp16)[name = tensor<string, []>("channels_mean_177_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_177_cast_fp16 = sub(x = inputs_177_cast_fp16, y = channels_mean_177_cast_fp16)[name = tensor<string, []>("zero_mean_177_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = zero_mean_177_cast_fp16)[name = tensor<string, []>("zero_mean_sq_177_cast_fp16")];
            tensor<int32, [1]> var_6493 = const()[name = tensor<string, []>("op_6493"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6494_cast_fp16 = reduce_mean(axes = var_6493, keep_dims = var_6390, x = zero_mean_sq_177_cast_fp16)[name = tensor<string, []>("op_6494_cast_fp16")];
            tensor<fp16, []> var_6495_to_fp16 = const()[name = tensor<string, []>("op_6495_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6496_cast_fp16 = add(x = var_6494_cast_fp16, y = var_6495_to_fp16)[name = tensor<string, []>("op_6496_cast_fp16")];
            tensor<fp32, []> denom_177_epsilon_0 = const()[name = tensor<string, []>("denom_177_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_177_cast_fp16 = rsqrt(epsilon = denom_177_epsilon_0, x = var_6496_cast_fp16)[name = tensor<string, []>("denom_177_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = denom_177_cast_fp16)[name = tensor<string, []>("out_177_cast_fp16")];
            tensor<fp16, [1280]> obj_357_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_357_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756505088)))];
            tensor<fp16, [1280]> obj_357_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_357_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756507712)))];
            tensor<fp16, []> obj_357_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_357_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_357_cast_fp16 = batch_norm(beta = obj_357_beta_0_to_fp16, epsilon = obj_357_epsilon_0_to_fp16, gamma = obj_357_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor<string, []>("obj_357_cast_fp16")];
            tensor<int32, [2]> var_6511 = const()[name = tensor<string, []>("op_6511"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6513 = const()[name = tensor<string, []>("op_6513"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_119_pad_type_0 = const()[name = tensor<string, []>("query_119_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_119_pad_0 = const()[name = tensor<string, []>("query_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756510336))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757329600))), name = tensor<string, []>("layers_29_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_29_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757329728)))];
            tensor<fp16, [1, 1280, 1, 1]> query_119_cast_fp16 = conv(bias = layers_29_encoder_attn_q_proj_bias_to_fp16, dilations = var_6513, groups = var_6389, pad = query_119_pad_0, pad_type = query_119_pad_type_0, strides = var_6511, weight = layers_29_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_357_cast_fp16)[name = tensor<string, []>("query_119_cast_fp16")];
            tensor<int32, [2]> var_6517 = const()[name = tensor<string, []>("op_6517"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6519 = const()[name = tensor<string, []>("op_6519"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_119_pad_type_0 = const()[name = tensor<string, []>("key_119_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_119_pad_0 = const()[name = tensor<string, []>("key_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757332352))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(758151616))), name = tensor<string, []>("layers_29_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_119_cast_fp16 = conv(dilations = var_6519, groups = var_6389, pad = key_119_pad_0, pad_type = key_119_pad_type_0, strides = var_6517, weight = layers_29_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_119_cast_fp16")];
            tensor<int32, [2]> var_6524 = const()[name = tensor<string, []>("op_6524"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6526 = const()[name = tensor<string, []>("op_6526"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_119_pad_type_0 = const()[name = tensor<string, []>("value_119_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_119_pad_0 = const()[name = tensor<string, []>("value_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(758151744))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(759380608))), name = tensor<string, []>("layers_29_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_29_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(759380800)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_119_cast_fp16 = conv(bias = layers_29_encoder_attn_v_proj_bias_to_fp16, dilations = var_6526, groups = var_6389, pad = value_119_pad_0, pad_type = value_119_pad_type_0, strides = var_6524, weight = layers_29_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_119_cast_fp16")];
            tensor<int32, [4]> var_6530 = const()[name = tensor<string, []>("op_6530"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6531_cast_fp16 = reshape(shape = var_6530, x = query_119_cast_fp16)[name = tensor<string, []>("op_6531_cast_fp16")];
            tensor<fp16, []> var_6532_to_fp16 = const()[name = tensor<string, []>("op_6532_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6533_cast_fp16 = mul(x = var_6531_cast_fp16, y = var_6532_to_fp16)[name = tensor<string, []>("op_6533_cast_fp16")];
            tensor<int32, [4]> var_6534 = const()[name = tensor<string, []>("op_6534"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6535_cast_fp16 = reshape(shape = var_6534, x = key_119_cast_fp16)[name = tensor<string, []>("op_6535_cast_fp16")];
            tensor<bool, []> mh_w_179_transpose_x_0 = const()[name = tensor<string, []>("mh_w_179_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_179_transpose_y_0 = const()[name = tensor<string, []>("mh_w_179_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_179_cast_fp16 = matmul(transpose_x = mh_w_179_transpose_x_0, transpose_y = mh_w_179_transpose_y_0, x = var_6533_cast_fp16, y = var_6535_cast_fp16)[name = tensor<string, []>("mh_w_179_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_6538_cast_fp16 = softmax(axis = var_6382, x = mh_w_179_cast_fp16)[name = tensor<string, []>("op_6538_cast_fp16")];
            tensor<int32, [4]> var_6539 = const()[name = tensor<string, []>("op_6539"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6540_cast_fp16 = reshape(shape = var_6539, x = value_119_cast_fp16)[name = tensor<string, []>("op_6540_cast_fp16")];
            tensor<bool, []> attn_119_transpose_x_0 = const()[name = tensor<string, []>("attn_119_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_119_transpose_y_0 = const()[name = tensor<string, []>("attn_119_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_119_cast_fp16 = matmul(transpose_x = attn_119_transpose_x_0, transpose_y = attn_119_transpose_y_0, x = var_6540_cast_fp16, y = var_6538_cast_fp16)[name = tensor<string, []>("attn_119_cast_fp16")];
            tensor<int32, [4]> var_6543 = const()[name = tensor<string, []>("op_6543"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_293_cast_fp16 = reshape(shape = var_6543, x = attn_119_cast_fp16)[name = tensor<string, []>("input_293_cast_fp16")];
            tensor<int32, [2]> var_6547 = const()[name = tensor<string, []>("op_6547"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6549 = const()[name = tensor<string, []>("op_6549"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_359_pad_type_0 = const()[name = tensor<string, []>("obj_359_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_359_pad_0 = const()[name = tensor<string, []>("obj_359_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_29_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(759383424))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(760202688))), name = tensor<string, []>("layers_29_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_29_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(760202816)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_359_cast_fp16 = conv(bias = layers_29_encoder_attn_o_proj_bias_to_fp16, dilations = var_6549, groups = var_6389, pad = obj_359_pad_0, pad_type = obj_359_pad_type_0, strides = var_6547, weight = layers_29_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_293_cast_fp16)[name = tensor<string, []>("obj_359_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = obj_359_cast_fp16)[name = tensor<string, []>("inputs_179_cast_fp16")];
            tensor<int32, [1]> var_6555 = const()[name = tensor<string, []>("op_6555"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_179_cast_fp16 = reduce_mean(axes = var_6555, keep_dims = var_6390, x = inputs_179_cast_fp16)[name = tensor<string, []>("channels_mean_179_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_179_cast_fp16 = sub(x = inputs_179_cast_fp16, y = channels_mean_179_cast_fp16)[name = tensor<string, []>("zero_mean_179_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = zero_mean_179_cast_fp16)[name = tensor<string, []>("zero_mean_sq_179_cast_fp16")];
            tensor<int32, [1]> var_6559 = const()[name = tensor<string, []>("op_6559"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6560_cast_fp16 = reduce_mean(axes = var_6559, keep_dims = var_6390, x = zero_mean_sq_179_cast_fp16)[name = tensor<string, []>("op_6560_cast_fp16")];
            tensor<fp16, []> var_6561_to_fp16 = const()[name = tensor<string, []>("op_6561_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6562_cast_fp16 = add(x = var_6560_cast_fp16, y = var_6561_to_fp16)[name = tensor<string, []>("op_6562_cast_fp16")];
            tensor<fp32, []> denom_179_epsilon_0 = const()[name = tensor<string, []>("denom_179_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_179_cast_fp16 = rsqrt(epsilon = denom_179_epsilon_0, x = var_6562_cast_fp16)[name = tensor<string, []>("denom_179_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = denom_179_cast_fp16)[name = tensor<string, []>("out_179_cast_fp16")];
            tensor<fp16, [1280]> input_295_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_295_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(760205440)))];
            tensor<fp16, [1280]> input_295_beta_0_to_fp16 = const()[name = tensor<string, []>("input_295_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(760208064)))];
            tensor<fp16, []> input_295_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_295_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_295_cast_fp16 = batch_norm(beta = input_295_beta_0_to_fp16, epsilon = input_295_epsilon_0_to_fp16, gamma = input_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor<string, []>("input_295_cast_fp16")];
            tensor<int32, [2]> var_6573 = const()[name = tensor<string, []>("op_6573"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6575 = const()[name = tensor<string, []>("op_6575"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_297_pad_type_0 = const()[name = tensor<string, []>("input_297_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_297_pad_0 = const()[name = tensor<string, []>("input_297_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_29_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(760210688))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765125952))), name = tensor<string, []>("layers_29_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_29_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765126144)))];
            tensor<fp16, [1, 5120, 1, 1]> input_297_cast_fp16 = conv(bias = layers_29_fc1_bias_to_fp16, dilations = var_6575, groups = var_6389, pad = input_297_pad_0, pad_type = input_297_pad_type_0, strides = var_6573, weight = layers_29_fc1_weight_to_fp16_palettized, x = input_295_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
            tensor<string, []> input_299_mode_0 = const()[name = tensor<string, []>("input_299_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_299_cast_fp16 = gelu(mode = input_299_mode_0, x = input_297_cast_fp16)[name = tensor<string, []>("input_299_cast_fp16")];
            tensor<int32, [2]> var_6581 = const()[name = tensor<string, []>("op_6581"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6583 = const()[name = tensor<string, []>("op_6583"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_61_pad_type_0 = const()[name = tensor<string, []>("hidden_states_61_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_61_pad_0 = const()[name = tensor<string, []>("hidden_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_29_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6553600]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765136448))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(771690112))), name = tensor<string, []>("layers_29_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_29_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(771690688)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_61_cast_fp16 = conv(bias = layers_29_fc2_bias_to_fp16, dilations = var_6583, groups = var_6389, pad = hidden_states_61_pad_0, pad_type = hidden_states_61_pad_type_0, strides = var_6581, weight = layers_29_fc2_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = tensor<string, []>("hidden_states_61_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_181_cast_fp16 = add(x = inputs_179_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor<string, []>("inputs_181_cast_fp16")];
            tensor<int32, []> var_6596 = const()[name = tensor<string, []>("op_6596"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_6603 = const()[name = tensor<string, []>("op_6603"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_6604 = const()[name = tensor<string, []>("op_6604"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_6616 = const()[name = tensor<string, []>("op_6616"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_181_cast_fp16 = reduce_mean(axes = var_6616, keep_dims = var_6604, x = inputs_181_cast_fp16)[name = tensor<string, []>("channels_mean_181_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_181_cast_fp16 = sub(x = inputs_181_cast_fp16, y = channels_mean_181_cast_fp16)[name = tensor<string, []>("zero_mean_181_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = zero_mean_181_cast_fp16)[name = tensor<string, []>("zero_mean_sq_181_cast_fp16")];
            tensor<int32, [1]> var_6620 = const()[name = tensor<string, []>("op_6620"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6621_cast_fp16 = reduce_mean(axes = var_6620, keep_dims = var_6604, x = zero_mean_sq_181_cast_fp16)[name = tensor<string, []>("op_6621_cast_fp16")];
            tensor<fp16, []> var_6622_to_fp16 = const()[name = tensor<string, []>("op_6622_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6623_cast_fp16 = add(x = var_6621_cast_fp16, y = var_6622_to_fp16)[name = tensor<string, []>("op_6623_cast_fp16")];
            tensor<fp32, []> denom_181_epsilon_0 = const()[name = tensor<string, []>("denom_181_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_181_cast_fp16 = rsqrt(epsilon = denom_181_epsilon_0, x = var_6623_cast_fp16)[name = tensor<string, []>("denom_181_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = denom_181_cast_fp16)[name = tensor<string, []>("out_181_cast_fp16")];
            tensor<fp16, [1280]> obj_361_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_361_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(771693312)))];
            tensor<fp16, [1280]> obj_361_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_361_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(771695936)))];
            tensor<fp16, []> obj_361_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_361_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_361_cast_fp16 = batch_norm(beta = obj_361_beta_0_to_fp16, epsilon = obj_361_epsilon_0_to_fp16, gamma = obj_361_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor<string, []>("obj_361_cast_fp16")];
            tensor<int32, [2]> var_6638 = const()[name = tensor<string, []>("op_6638"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6640 = const()[name = tensor<string, []>("op_6640"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_121_pad_type_0 = const()[name = tensor<string, []>("query_121_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_121_pad_0 = const()[name = tensor<string, []>("query_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(771698560))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(772517824))), name = tensor<string, []>("layers_30_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_30_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(772517952)))];
            tensor<fp16, [1, 1280, 1, 1]> query_121_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_bias_to_fp16, dilations = var_6640, groups = var_6603, pad = query_121_pad_0, pad_type = query_121_pad_type_0, strides = var_6638, weight = layers_30_self_attn_q_proj_weight_to_fp16_palettized, x = obj_361_cast_fp16)[name = tensor<string, []>("query_121_cast_fp16")];
            tensor<int32, [2]> var_6644 = const()[name = tensor<string, []>("op_6644"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6646 = const()[name = tensor<string, []>("op_6646"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_61_pad_type_0 = const()[name = tensor<string, []>("current_key_61_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_61_pad_0 = const()[name = tensor<string, []>("current_key_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(772520576))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(773339840))), name = tensor<string, []>("layers_30_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_61_cast_fp16 = conv(dilations = var_6646, groups = var_6603, pad = current_key_61_pad_0, pad_type = current_key_61_pad_type_0, strides = var_6644, weight = layers_30_self_attn_k_proj_weight_to_fp16_palettized, x = obj_361_cast_fp16)[name = tensor<string, []>("current_key_61_cast_fp16")];
            tensor<int32, [2]> var_6651 = const()[name = tensor<string, []>("op_6651"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6653 = const()[name = tensor<string, []>("op_6653"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_61_pad_type_0 = const()[name = tensor<string, []>("current_value_61_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_61_pad_0 = const()[name = tensor<string, []>("current_value_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(773339968))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774159232))), name = tensor<string, []>("layers_30_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_30_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774159360)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_61_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_bias_to_fp16, dilations = var_6653, groups = var_6603, pad = current_value_61_pad_0, pad_type = current_value_61_pad_type_0, strides = var_6651, weight = layers_30_self_attn_v_proj_weight_to_fp16_palettized, x = obj_361_cast_fp16)[name = tensor<string, []>("current_value_61_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6660_cast_fp16 = mul(x = current_key_61_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6660_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6662_cast_fp16 = mul(x = var_103_cast_fp16_30, y = var_241_cast_fp16)[name = tensor<string, []>("op_6662_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_121_cast_fp16 = add(x = var_6660_cast_fp16, y = var_6662_cast_fp16)[name = tensor<string, []>("key_121_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6664_cast_fp16 = mul(x = current_value_61_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6664_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6666_cast_fp16 = mul(x = var_138_cast_fp16_30, y = var_241_cast_fp16)[name = tensor<string, []>("op_6666_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_121_cast_fp16 = add(x = var_6664_cast_fp16, y = var_6666_cast_fp16)[name = tensor<string, []>("value_121_cast_fp16")];
            tensor<int32, [4]> var_6669 = const()[name = tensor<string, []>("op_6669"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6670_cast_fp16 = reshape(shape = var_6669, x = query_121_cast_fp16)[name = tensor<string, []>("op_6670_cast_fp16")];
            tensor<fp16, []> var_6671_to_fp16 = const()[name = tensor<string, []>("op_6671_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6672_cast_fp16 = mul(x = var_6670_cast_fp16, y = var_6671_to_fp16)[name = tensor<string, []>("op_6672_cast_fp16")];
            tensor<int32, [4]> var_6673 = const()[name = tensor<string, []>("op_6673"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6674_cast_fp16 = reshape(shape = var_6673, x = key_121_cast_fp16)[name = tensor<string, []>("op_6674_cast_fp16")];
            tensor<bool, []> mh_w_181_transpose_x_0 = const()[name = tensor<string, []>("mh_w_181_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_181_transpose_y_0 = const()[name = tensor<string, []>("mh_w_181_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_181_cast_fp16 = matmul(transpose_x = mh_w_181_transpose_x_0, transpose_y = mh_w_181_transpose_y_0, x = var_6672_cast_fp16, y = var_6674_cast_fp16)[name = tensor<string, []>("mh_w_181_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_183_cast_fp16 = add(x = mh_w_181_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_183_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_6682_cast_fp16 = softmax(axis = var_6596, x = mh_w_183_cast_fp16)[name = tensor<string, []>("op_6682_cast_fp16")];
            tensor<int32, [4]> var_6683 = const()[name = tensor<string, []>("op_6683"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6684_cast_fp16 = reshape(shape = var_6683, x = value_121_cast_fp16)[name = tensor<string, []>("op_6684_cast_fp16")];
            tensor<bool, []> attn_121_transpose_x_0 = const()[name = tensor<string, []>("attn_121_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_121_transpose_y_0 = const()[name = tensor<string, []>("attn_121_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_121_cast_fp16 = matmul(transpose_x = attn_121_transpose_x_0, transpose_y = attn_121_transpose_y_0, x = var_6684_cast_fp16, y = var_6682_cast_fp16)[name = tensor<string, []>("attn_121_cast_fp16")];
            tensor<int32, [4]> var_6687 = const()[name = tensor<string, []>("op_6687"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_301_cast_fp16 = reshape(shape = var_6687, x = attn_121_cast_fp16)[name = tensor<string, []>("input_301_cast_fp16")];
            tensor<int32, [2]> var_6691 = const()[name = tensor<string, []>("op_6691"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6693 = const()[name = tensor<string, []>("op_6693"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_367_pad_type_0 = const()[name = tensor<string, []>("obj_367_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_367_pad_0 = const()[name = tensor<string, []>("obj_367_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774161984))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774981248))), name = tensor<string, []>("layers_30_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_30_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774981376)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_367_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_bias_to_fp16, dilations = var_6693, groups = var_6603, pad = obj_367_pad_0, pad_type = obj_367_pad_type_0, strides = var_6691, weight = layers_30_self_attn_o_proj_weight_to_fp16_palettized, x = input_301_cast_fp16)[name = tensor<string, []>("obj_367_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = obj_367_cast_fp16)[name = tensor<string, []>("inputs_183_cast_fp16")];
            tensor<int32, [1]> var_6703 = const()[name = tensor<string, []>("op_6703"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_183_cast_fp16 = reduce_mean(axes = var_6703, keep_dims = var_6604, x = inputs_183_cast_fp16)[name = tensor<string, []>("channels_mean_183_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_183_cast_fp16 = sub(x = inputs_183_cast_fp16, y = channels_mean_183_cast_fp16)[name = tensor<string, []>("zero_mean_183_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = zero_mean_183_cast_fp16)[name = tensor<string, []>("zero_mean_sq_183_cast_fp16")];
            tensor<int32, [1]> var_6707 = const()[name = tensor<string, []>("op_6707"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6708_cast_fp16 = reduce_mean(axes = var_6707, keep_dims = var_6604, x = zero_mean_sq_183_cast_fp16)[name = tensor<string, []>("op_6708_cast_fp16")];
            tensor<fp16, []> var_6709_to_fp16 = const()[name = tensor<string, []>("op_6709_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6710_cast_fp16 = add(x = var_6708_cast_fp16, y = var_6709_to_fp16)[name = tensor<string, []>("op_6710_cast_fp16")];
            tensor<fp32, []> denom_183_epsilon_0 = const()[name = tensor<string, []>("denom_183_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_183_cast_fp16 = rsqrt(epsilon = denom_183_epsilon_0, x = var_6710_cast_fp16)[name = tensor<string, []>("denom_183_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = denom_183_cast_fp16)[name = tensor<string, []>("out_183_cast_fp16")];
            tensor<fp16, [1280]> obj_369_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_369_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774984000)))];
            tensor<fp16, [1280]> obj_369_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_369_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774986624)))];
            tensor<fp16, []> obj_369_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_369_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_369_cast_fp16 = batch_norm(beta = obj_369_beta_0_to_fp16, epsilon = obj_369_epsilon_0_to_fp16, gamma = obj_369_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_183_cast_fp16)[name = tensor<string, []>("obj_369_cast_fp16")];
            tensor<int32, [2]> var_6725 = const()[name = tensor<string, []>("op_6725"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6727 = const()[name = tensor<string, []>("op_6727"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_123_pad_type_0 = const()[name = tensor<string, []>("query_123_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_123_pad_0 = const()[name = tensor<string, []>("query_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774989248)))];
            tensor<fp16, [1280]> layers_30_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778266112)))];
            tensor<fp16, [1, 1280, 1, 1]> query_123_cast_fp16 = conv(bias = layers_30_encoder_attn_q_proj_bias_to_fp16, dilations = var_6727, groups = var_6603, pad = query_123_pad_0, pad_type = query_123_pad_type_0, strides = var_6725, weight = layers_30_encoder_attn_q_proj_weight_to_fp16, x = obj_369_cast_fp16)[name = tensor<string, []>("query_123_cast_fp16")];
            tensor<int32, [2]> var_6731 = const()[name = tensor<string, []>("op_6731"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6733 = const()[name = tensor<string, []>("op_6733"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_123_pad_type_0 = const()[name = tensor<string, []>("key_123_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_123_pad_0 = const()[name = tensor<string, []>("key_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778268736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779497600))), name = tensor<string, []>("layers_30_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_123_cast_fp16 = conv(dilations = var_6733, groups = var_6603, pad = key_123_pad_0, pad_type = key_123_pad_type_0, strides = var_6731, weight = layers_30_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_123_cast_fp16")];
            tensor<int32, [2]> var_6738 = const()[name = tensor<string, []>("op_6738"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6740 = const()[name = tensor<string, []>("op_6740"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_123_pad_type_0 = const()[name = tensor<string, []>("value_123_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_123_pad_0 = const()[name = tensor<string, []>("value_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779497792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(780726656))), name = tensor<string, []>("layers_30_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_30_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(780726848)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_123_cast_fp16 = conv(bias = layers_30_encoder_attn_v_proj_bias_to_fp16, dilations = var_6740, groups = var_6603, pad = value_123_pad_0, pad_type = value_123_pad_type_0, strides = var_6738, weight = layers_30_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_123_cast_fp16")];
            tensor<int32, [4]> var_6744 = const()[name = tensor<string, []>("op_6744"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6745_cast_fp16 = reshape(shape = var_6744, x = query_123_cast_fp16)[name = tensor<string, []>("op_6745_cast_fp16")];
            tensor<fp16, []> var_6746_to_fp16 = const()[name = tensor<string, []>("op_6746_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6747_cast_fp16 = mul(x = var_6745_cast_fp16, y = var_6746_to_fp16)[name = tensor<string, []>("op_6747_cast_fp16")];
            tensor<int32, [4]> var_6748 = const()[name = tensor<string, []>("op_6748"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6749_cast_fp16 = reshape(shape = var_6748, x = key_123_cast_fp16)[name = tensor<string, []>("op_6749_cast_fp16")];
            tensor<bool, []> mh_w_185_transpose_x_0 = const()[name = tensor<string, []>("mh_w_185_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_185_transpose_y_0 = const()[name = tensor<string, []>("mh_w_185_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_185_cast_fp16 = matmul(transpose_x = mh_w_185_transpose_x_0, transpose_y = mh_w_185_transpose_y_0, x = var_6747_cast_fp16, y = var_6749_cast_fp16)[name = tensor<string, []>("mh_w_185_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_6752_cast_fp16 = softmax(axis = var_6596, x = mh_w_185_cast_fp16)[name = tensor<string, []>("op_6752_cast_fp16")];
            tensor<int32, [4]> var_6753 = const()[name = tensor<string, []>("op_6753"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6754_cast_fp16 = reshape(shape = var_6753, x = value_123_cast_fp16)[name = tensor<string, []>("op_6754_cast_fp16")];
            tensor<bool, []> attn_123_transpose_x_0 = const()[name = tensor<string, []>("attn_123_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_123_transpose_y_0 = const()[name = tensor<string, []>("attn_123_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_123_cast_fp16 = matmul(transpose_x = attn_123_transpose_x_0, transpose_y = attn_123_transpose_y_0, x = var_6754_cast_fp16, y = var_6752_cast_fp16)[name = tensor<string, []>("attn_123_cast_fp16")];
            tensor<int32, [4]> var_6757 = const()[name = tensor<string, []>("op_6757"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_303_cast_fp16 = reshape(shape = var_6757, x = attn_123_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
            tensor<int32, [2]> var_6761 = const()[name = tensor<string, []>("op_6761"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6763 = const()[name = tensor<string, []>("op_6763"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_371_pad_type_0 = const()[name = tensor<string, []>("obj_371_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_371_pad_0 = const()[name = tensor<string, []>("obj_371_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_30_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(780729472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781958336))), name = tensor<string, []>("layers_30_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_30_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781958528)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_371_cast_fp16 = conv(bias = layers_30_encoder_attn_o_proj_bias_to_fp16, dilations = var_6763, groups = var_6603, pad = obj_371_pad_0, pad_type = obj_371_pad_type_0, strides = var_6761, weight = layers_30_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = tensor<string, []>("obj_371_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_185_cast_fp16 = add(x = inputs_183_cast_fp16, y = obj_371_cast_fp16)[name = tensor<string, []>("inputs_185_cast_fp16")];
            tensor<int32, [1]> var_6769 = const()[name = tensor<string, []>("op_6769"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_185_cast_fp16 = reduce_mean(axes = var_6769, keep_dims = var_6604, x = inputs_185_cast_fp16)[name = tensor<string, []>("channels_mean_185_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_185_cast_fp16 = sub(x = inputs_185_cast_fp16, y = channels_mean_185_cast_fp16)[name = tensor<string, []>("zero_mean_185_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = zero_mean_185_cast_fp16)[name = tensor<string, []>("zero_mean_sq_185_cast_fp16")];
            tensor<int32, [1]> var_6773 = const()[name = tensor<string, []>("op_6773"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6774_cast_fp16 = reduce_mean(axes = var_6773, keep_dims = var_6604, x = zero_mean_sq_185_cast_fp16)[name = tensor<string, []>("op_6774_cast_fp16")];
            tensor<fp16, []> var_6775_to_fp16 = const()[name = tensor<string, []>("op_6775_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6776_cast_fp16 = add(x = var_6774_cast_fp16, y = var_6775_to_fp16)[name = tensor<string, []>("op_6776_cast_fp16")];
            tensor<fp32, []> denom_185_epsilon_0 = const()[name = tensor<string, []>("denom_185_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_185_cast_fp16 = rsqrt(epsilon = denom_185_epsilon_0, x = var_6776_cast_fp16)[name = tensor<string, []>("denom_185_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = denom_185_cast_fp16)[name = tensor<string, []>("out_185_cast_fp16")];
            tensor<fp16, [1280]> input_305_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_305_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781961152)))];
            tensor<fp16, [1280]> input_305_beta_0_to_fp16 = const()[name = tensor<string, []>("input_305_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781963776)))];
            tensor<fp16, []> input_305_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_305_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_305_cast_fp16 = batch_norm(beta = input_305_beta_0_to_fp16, epsilon = input_305_epsilon_0_to_fp16, gamma = input_305_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor<string, []>("input_305_cast_fp16")];
            tensor<int32, [2]> var_6787 = const()[name = tensor<string, []>("op_6787"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6789 = const()[name = tensor<string, []>("op_6789"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_307_pad_type_0 = const()[name = tensor<string, []>("input_307_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_307_pad_0 = const()[name = tensor<string, []>("input_307_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_30_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781966400))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(786881664))), name = tensor<string, []>("layers_30_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_30_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(786881856)))];
            tensor<fp16, [1, 5120, 1, 1]> input_307_cast_fp16 = conv(bias = layers_30_fc1_bias_to_fp16, dilations = var_6789, groups = var_6603, pad = input_307_pad_0, pad_type = input_307_pad_type_0, strides = var_6787, weight = layers_30_fc1_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
            tensor<string, []> input_309_mode_0 = const()[name = tensor<string, []>("input_309_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_309_cast_fp16 = gelu(mode = input_309_mode_0, x = input_307_cast_fp16)[name = tensor<string, []>("input_309_cast_fp16")];
            tensor<int32, [2]> var_6795 = const()[name = tensor<string, []>("op_6795"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6797 = const()[name = tensor<string, []>("op_6797"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_63_pad_type_0 = const()[name = tensor<string, []>("hidden_states_63_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = tensor<string, []>("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_30_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(786892160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(791807424))), name = tensor<string, []>("layers_30_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_30_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(791807616)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_63_cast_fp16 = conv(bias = layers_30_fc2_bias_to_fp16, dilations = var_6797, groups = var_6603, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = var_6795, weight = layers_30_fc2_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor<string, []>("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_187_cast_fp16 = add(x = inputs_185_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor<string, []>("inputs_187_cast_fp16")];
            tensor<int32, []> var_6810 = const()[name = tensor<string, []>("op_6810"), val = tensor<int32, []>(3)];
            tensor<int32, []> var_6817 = const()[name = tensor<string, []>("op_6817"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_6818 = const()[name = tensor<string, []>("op_6818"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_6830 = const()[name = tensor<string, []>("op_6830"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_187_cast_fp16 = reduce_mean(axes = var_6830, keep_dims = var_6818, x = inputs_187_cast_fp16)[name = tensor<string, []>("channels_mean_187_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_187_cast_fp16 = sub(x = inputs_187_cast_fp16, y = channels_mean_187_cast_fp16)[name = tensor<string, []>("zero_mean_187_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = zero_mean_187_cast_fp16)[name = tensor<string, []>("zero_mean_sq_187_cast_fp16")];
            tensor<int32, [1]> var_6834 = const()[name = tensor<string, []>("op_6834"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6835_cast_fp16 = reduce_mean(axes = var_6834, keep_dims = var_6818, x = zero_mean_sq_187_cast_fp16)[name = tensor<string, []>("op_6835_cast_fp16")];
            tensor<fp16, []> var_6836_to_fp16 = const()[name = tensor<string, []>("op_6836_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6837_cast_fp16 = add(x = var_6835_cast_fp16, y = var_6836_to_fp16)[name = tensor<string, []>("op_6837_cast_fp16")];
            tensor<fp32, []> denom_187_epsilon_0 = const()[name = tensor<string, []>("denom_187_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_187_cast_fp16 = rsqrt(epsilon = denom_187_epsilon_0, x = var_6837_cast_fp16)[name = tensor<string, []>("denom_187_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = denom_187_cast_fp16)[name = tensor<string, []>("out_187_cast_fp16")];
            tensor<fp16, [1280]> obj_373_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_373_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(791810240)))];
            tensor<fp16, [1280]> obj_373_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_373_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(791812864)))];
            tensor<fp16, []> obj_373_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_373_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_373_cast_fp16 = batch_norm(beta = obj_373_beta_0_to_fp16, epsilon = obj_373_epsilon_0_to_fp16, gamma = obj_373_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor<string, []>("obj_373_cast_fp16")];
            tensor<int32, [2]> var_6852 = const()[name = tensor<string, []>("op_6852"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6854 = const()[name = tensor<string, []>("op_6854"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_125_pad_type_0 = const()[name = tensor<string, []>("query_125_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_125_pad_0 = const()[name = tensor<string, []>("query_125_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(791815488))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(792634752))), name = tensor<string, []>("layers_31_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_31_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(792634880)))];
            tensor<fp16, [1, 1280, 1, 1]> query_125_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_bias_to_fp16, dilations = var_6854, groups = var_6817, pad = query_125_pad_0, pad_type = query_125_pad_type_0, strides = var_6852, weight = layers_31_self_attn_q_proj_weight_to_fp16_palettized, x = obj_373_cast_fp16)[name = tensor<string, []>("query_125_cast_fp16")];
            tensor<int32, [2]> var_6858 = const()[name = tensor<string, []>("op_6858"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6860 = const()[name = tensor<string, []>("op_6860"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_key_pad_type_0 = const()[name = tensor<string, []>("current_key_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_key_pad_0 = const()[name = tensor<string, []>("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(792637504))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(793456768))), name = tensor<string, []>("layers_31_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = conv(dilations = var_6860, groups = var_6817, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_6858, weight = layers_31_self_attn_k_proj_weight_to_fp16_palettized, x = obj_373_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")];
            tensor<int32, [2]> var_6865 = const()[name = tensor<string, []>("op_6865"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6867 = const()[name = tensor<string, []>("op_6867"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> current_value_pad_type_0 = const()[name = tensor<string, []>("current_value_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> current_value_pad_0 = const()[name = tensor<string, []>("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(793456896))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(794276160))), name = tensor<string, []>("layers_31_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_31_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(794276288)))];
            tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_bias_to_fp16, dilations = var_6867, groups = var_6817, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_6865, weight = layers_31_self_attn_v_proj_weight_to_fp16_palettized, x = obj_373_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6874_cast_fp16 = mul(x = current_key_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6874_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6876_cast_fp16 = mul(x = var_103_cast_fp16_31, y = var_241_cast_fp16)[name = tensor<string, []>("op_6876_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> key_125_cast_fp16 = add(x = var_6874_cast_fp16, y = var_6876_cast_fp16)[name = tensor<string, []>("key_125_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6878_cast_fp16 = mul(x = current_value_cast_fp16, y = var_238_cast_fp16)[name = tensor<string, []>("op_6878_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> var_6880_cast_fp16 = mul(x = var_138_cast_fp16_31, y = var_241_cast_fp16)[name = tensor<string, []>("op_6880_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 448]> value_125_cast_fp16 = add(x = var_6878_cast_fp16, y = var_6880_cast_fp16)[name = tensor<string, []>("value_125_cast_fp16")];
            tensor<int32, [4]> var_6883 = const()[name = tensor<string, []>("op_6883"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6884_cast_fp16 = reshape(shape = var_6883, x = query_125_cast_fp16)[name = tensor<string, []>("op_6884_cast_fp16")];
            tensor<fp16, []> var_6885_to_fp16 = const()[name = tensor<string, []>("op_6885_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6886_cast_fp16 = mul(x = var_6884_cast_fp16, y = var_6885_to_fp16)[name = tensor<string, []>("op_6886_cast_fp16")];
            tensor<int32, [4]> var_6887 = const()[name = tensor<string, []>("op_6887"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6888_cast_fp16 = reshape(shape = var_6887, x = key_125_cast_fp16)[name = tensor<string, []>("op_6888_cast_fp16")];
            tensor<bool, []> mh_w_187_transpose_x_0 = const()[name = tensor<string, []>("mh_w_187_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_187_transpose_y_0 = const()[name = tensor<string, []>("mh_w_187_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 448]> mh_w_187_cast_fp16 = matmul(transpose_x = mh_w_187_transpose_x_0, transpose_y = mh_w_187_transpose_y_0, x = var_6886_cast_fp16, y = var_6888_cast_fp16)[name = tensor<string, []>("mh_w_187_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> mh_w_189_cast_fp16 = add(x = mh_w_187_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("mh_w_189_cast_fp16")];
            tensor<fp16, [1, 20, 1, 448]> var_6896_cast_fp16 = softmax(axis = var_6810, x = mh_w_189_cast_fp16)[name = tensor<string, []>("op_6896_cast_fp16")];
            tensor<int32, [4]> var_6897 = const()[name = tensor<string, []>("op_6897"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 448]> var_6898_cast_fp16 = reshape(shape = var_6897, x = value_125_cast_fp16)[name = tensor<string, []>("op_6898_cast_fp16")];
            tensor<bool, []> attn_125_transpose_x_0 = const()[name = tensor<string, []>("attn_125_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_125_transpose_y_0 = const()[name = tensor<string, []>("attn_125_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_125_cast_fp16 = matmul(transpose_x = attn_125_transpose_x_0, transpose_y = attn_125_transpose_y_0, x = var_6898_cast_fp16, y = var_6896_cast_fp16)[name = tensor<string, []>("attn_125_cast_fp16")];
            tensor<int32, [4]> var_6901 = const()[name = tensor<string, []>("op_6901"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_311_cast_fp16 = reshape(shape = var_6901, x = attn_125_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
            tensor<int32, [2]> var_6905 = const()[name = tensor<string, []>("op_6905"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6907 = const()[name = tensor<string, []>("op_6907"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_379_pad_type_0 = const()[name = tensor<string, []>("obj_379_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_379_pad_0 = const()[name = tensor<string, []>("obj_379_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(794278912))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795098176))), name = tensor<string, []>("layers_31_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_31_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795098304)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_379_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_bias_to_fp16, dilations = var_6907, groups = var_6817, pad = obj_379_pad_0, pad_type = obj_379_pad_type_0, strides = var_6905, weight = layers_31_self_attn_o_proj_weight_to_fp16_palettized, x = input_311_cast_fp16)[name = tensor<string, []>("obj_379_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = obj_379_cast_fp16)[name = tensor<string, []>("inputs_189_cast_fp16")];
            tensor<int32, [1]> var_6917 = const()[name = tensor<string, []>("op_6917"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_189_cast_fp16 = reduce_mean(axes = var_6917, keep_dims = var_6818, x = inputs_189_cast_fp16)[name = tensor<string, []>("channels_mean_189_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_189_cast_fp16 = sub(x = inputs_189_cast_fp16, y = channels_mean_189_cast_fp16)[name = tensor<string, []>("zero_mean_189_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = zero_mean_189_cast_fp16)[name = tensor<string, []>("zero_mean_sq_189_cast_fp16")];
            tensor<int32, [1]> var_6921 = const()[name = tensor<string, []>("op_6921"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6922_cast_fp16 = reduce_mean(axes = var_6921, keep_dims = var_6818, x = zero_mean_sq_189_cast_fp16)[name = tensor<string, []>("op_6922_cast_fp16")];
            tensor<fp16, []> var_6923_to_fp16 = const()[name = tensor<string, []>("op_6923_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6924_cast_fp16 = add(x = var_6922_cast_fp16, y = var_6923_to_fp16)[name = tensor<string, []>("op_6924_cast_fp16")];
            tensor<fp32, []> denom_189_epsilon_0 = const()[name = tensor<string, []>("denom_189_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_189_cast_fp16 = rsqrt(epsilon = denom_189_epsilon_0, x = var_6924_cast_fp16)[name = tensor<string, []>("denom_189_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = denom_189_cast_fp16)[name = tensor<string, []>("out_189_cast_fp16")];
            tensor<fp16, [1280]> obj_381_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_381_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795100928)))];
            tensor<fp16, [1280]> obj_381_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_381_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795103552)))];
            tensor<fp16, []> obj_381_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_381_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> obj_381_cast_fp16 = batch_norm(beta = obj_381_beta_0_to_fp16, epsilon = obj_381_epsilon_0_to_fp16, gamma = obj_381_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor<string, []>("obj_381_cast_fp16")];
            tensor<int32, [2]> var_6939 = const()[name = tensor<string, []>("op_6939"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6941 = const()[name = tensor<string, []>("op_6941"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795106176))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(796335040))), name = tensor<string, []>("layers_31_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_31_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(796335232)))];
            tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = conv(bias = layers_31_encoder_attn_q_proj_bias_to_fp16, dilations = var_6941, groups = var_6817, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_6939, weight = layers_31_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_381_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
            tensor<int32, [2]> var_6945 = const()[name = tensor<string, []>("op_6945"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6947 = const()[name = tensor<string, []>("op_6947"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(796337856))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(797566720))), name = tensor<string, []>("layers_31_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = conv(dilations = var_6947, groups = var_6817, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_6945, weight = layers_31_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")];
            tensor<int32, [2]> var_6952 = const()[name = tensor<string, []>("op_6952"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6954 = const()[name = tensor<string, []>("op_6954"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(797566912))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798795776))), name = tensor<string, []>("layers_31_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_31_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798795968)))];
            tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = conv(bias = layers_31_encoder_attn_v_proj_bias_to_fp16, dilations = var_6954, groups = var_6817, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_6952, weight = layers_31_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")];
            tensor<int32, [4]> var_6958 = const()[name = tensor<string, []>("op_6958"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1]> var_6959_cast_fp16 = reshape(shape = var_6958, x = query_cast_fp16)[name = tensor<string, []>("op_6959_cast_fp16")];
            tensor<fp16, []> var_6960_to_fp16 = const()[name = tensor<string, []>("op_6960_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 20, 64, 1]> var_6961_cast_fp16 = mul(x = var_6959_cast_fp16, y = var_6960_to_fp16)[name = tensor<string, []>("op_6961_cast_fp16")];
            tensor<int32, [4]> var_6962 = const()[name = tensor<string, []>("op_6962"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6963_cast_fp16 = reshape(shape = var_6962, x = key_cast_fp16)[name = tensor<string, []>("op_6963_cast_fp16")];
            tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 20, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_6961_cast_fp16, y = var_6963_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
            tensor<fp16, [1, 20, 1, 1500]> var_6966_cast_fp16 = softmax(axis = var_6810, x = mh_w_cast_fp16)[name = tensor<string, []>("op_6966_cast_fp16")];
            tensor<int32, [4]> var_6967 = const()[name = tensor<string, []>("op_6967"), val = tensor<int32, [4]>([1, 20, 64, -1])];
            tensor<fp16, [1, 20, 64, 1500]> var_6968_cast_fp16 = reshape(shape = var_6967, x = value_cast_fp16)[name = tensor<string, []>("op_6968_cast_fp16")];
            tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_6968_cast_fp16, y = var_6966_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
            tensor<int32, [4]> var_6971 = const()[name = tensor<string, []>("op_6971"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
            tensor<fp16, [1, 1280, 1, 1]> input_313_cast_fp16 = reshape(shape = var_6971, x = attn_cast_fp16)[name = tensor<string, []>("input_313_cast_fp16")];
            tensor<int32, [2]> var_6975 = const()[name = tensor<string, []>("op_6975"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_6977 = const()[name = tensor<string, []>("op_6977"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> obj_383_pad_type_0 = const()[name = tensor<string, []>("obj_383_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> obj_383_pad_0 = const()[name = tensor<string, []>("obj_383_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 1280, 1, 1]> layers_31_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1228800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798798592))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800027456))), name = tensor<string, []>("layers_31_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
            tensor<fp16, [1280]> layers_31_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800027648)))];
            tensor<fp16, [1, 1280, 1, 1]> obj_383_cast_fp16 = conv(bias = layers_31_encoder_attn_o_proj_bias_to_fp16, dilations = var_6977, groups = var_6817, pad = obj_383_pad_0, pad_type = obj_383_pad_type_0, strides = var_6975, weight = layers_31_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_313_cast_fp16)[name = tensor<string, []>("obj_383_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_191_cast_fp16 = add(x = inputs_189_cast_fp16, y = obj_383_cast_fp16)[name = tensor<string, []>("inputs_191_cast_fp16")];
            tensor<int32, [1]> var_6983 = const()[name = tensor<string, []>("op_6983"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_191_cast_fp16 = reduce_mean(axes = var_6983, keep_dims = var_6818, x = inputs_191_cast_fp16)[name = tensor<string, []>("channels_mean_191_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_191_cast_fp16 = sub(x = inputs_191_cast_fp16, y = channels_mean_191_cast_fp16)[name = tensor<string, []>("zero_mean_191_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = zero_mean_191_cast_fp16)[name = tensor<string, []>("zero_mean_sq_191_cast_fp16")];
            tensor<int32, [1]> var_6987 = const()[name = tensor<string, []>("op_6987"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_6988_cast_fp16 = reduce_mean(axes = var_6987, keep_dims = var_6818, x = zero_mean_sq_191_cast_fp16)[name = tensor<string, []>("op_6988_cast_fp16")];
            tensor<fp16, []> var_6989_to_fp16 = const()[name = tensor<string, []>("op_6989_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_6990_cast_fp16 = add(x = var_6988_cast_fp16, y = var_6989_to_fp16)[name = tensor<string, []>("op_6990_cast_fp16")];
            tensor<fp32, []> denom_191_epsilon_0 = const()[name = tensor<string, []>("denom_191_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_191_cast_fp16 = rsqrt(epsilon = denom_191_epsilon_0, x = var_6990_cast_fp16)[name = tensor<string, []>("denom_191_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = denom_191_cast_fp16)[name = tensor<string, []>("out_191_cast_fp16")];
            tensor<fp16, [1280]> input_315_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_315_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800030272)))];
            tensor<fp16, [1280]> input_315_beta_0_to_fp16 = const()[name = tensor<string, []>("input_315_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800032896)))];
            tensor<fp16, []> input_315_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_315_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 1, 1]> input_315_cast_fp16 = batch_norm(beta = input_315_beta_0_to_fp16, epsilon = input_315_epsilon_0_to_fp16, gamma = input_315_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor<string, []>("input_315_cast_fp16")];
            tensor<int32, [2]> var_7001 = const()[name = tensor<string, []>("op_7001"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_7003 = const()[name = tensor<string, []>("op_7003"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_317_pad_type_0 = const()[name = tensor<string, []>("input_317_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_317_pad_0 = const()[name = tensor<string, []>("input_317_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [5120, 1280, 1, 1]> layers_31_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800035520))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(804950784))), name = tensor<string, []>("layers_31_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
            tensor<fp16, [5120]> layers_31_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(804950976)))];
            tensor<fp16, [1, 5120, 1, 1]> input_317_cast_fp16 = conv(bias = layers_31_fc1_bias_to_fp16, dilations = var_7003, groups = var_6817, pad = input_317_pad_0, pad_type = input_317_pad_type_0, strides = var_7001, weight = layers_31_fc1_weight_to_fp16_palettized, x = input_315_cast_fp16)[name = tensor<string, []>("input_317_cast_fp16")];
            tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 5120, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_317_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<int32, [2]> var_7009 = const()[name = tensor<string, []>("op_7009"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_7011 = const()[name = tensor<string, []>("op_7011"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> hidden_states_65_pad_type_0 = const()[name = tensor<string, []>("hidden_states_65_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> hidden_states_65_pad_0 = const()[name = tensor<string, []>("hidden_states_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [1280, 5120, 1, 1]> layers_31_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(804961280))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(809876544))), name = tensor<string, []>("layers_31_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
            tensor<fp16, [1280]> layers_31_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(809876736)))];
            tensor<fp16, [1, 1280, 1, 1]> hidden_states_65_cast_fp16 = conv(bias = layers_31_fc2_bias_to_fp16, dilations = var_7011, groups = var_6817, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = var_7009, weight = layers_31_fc2_weight_to_fp16_palettized, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_65_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> inputs_cast_fp16 = add(x = inputs_191_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
            tensor<bool, []> var_7021 = const()[name = tensor<string, []>("op_7021"), val = tensor<bool, []>(true)];
            tensor<int32, [1]> var_7025 = const()[name = tensor<string, []>("op_7025"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> channels_mean_cast_fp16 = reduce_mean(axes = var_7025, keep_dims = var_7021, x = inputs_cast_fp16)[name = tensor<string, []>("channels_mean_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor<string, []>("zero_mean_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor<string, []>("zero_mean_sq_cast_fp16")];
            tensor<int32, [1]> var_7029 = const()[name = tensor<string, []>("op_7029"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 1]> var_7030_cast_fp16 = reduce_mean(axes = var_7029, keep_dims = var_7021, x = zero_mean_sq_cast_fp16)[name = tensor<string, []>("op_7030_cast_fp16")];
            tensor<fp16, []> var_7031_to_fp16 = const()[name = tensor<string, []>("op_7031_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1, 1]> var_7032_cast_fp16 = add(x = var_7030_cast_fp16, y = var_7031_to_fp16)[name = tensor<string, []>("op_7032_cast_fp16")];
            tensor<fp32, []> denom_epsilon_0 = const()[name = tensor<string, []>("denom_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1]> denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0, x = var_7032_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
            tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
            tensor<fp16, [1280]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(809879360)))];
            tensor<fp16, [1280]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(809881984)))];
            tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1280, 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_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
            tensor<int32, [1]> var_7042_axes_0 = const()[name = tensor<string, []>("op_7042_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1280, 1]> var_7042_cast_fp16 = squeeze(axes = var_7042_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_7042_cast_fp16")];
            tensor<int32, [3]> var_7045_perm_0 = const()[name = tensor<string, []>("op_7045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [51866]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51866]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(809884608)))];
            tensor<fp16, [1, 1, 1280]> transpose_0 = transpose(perm = var_7045_perm_0, x = var_7042_cast_fp16)[name = tensor<string, []>("transpose_0")];
            tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<int32, []> var_7049 = const()[name = tensor<string, []>("op_7049"), val = tensor<int32, []>(1)];
            tensor<bool, []> obj_387_interleave_0 = const()[name = tensor<string, []>("obj_387_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 40960, 1, 1]> key_cache_updates = concat(axis = var_7049, interleave = obj_387_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_23_cast_fp16, current_key_25_cast_fp16, current_key_27_cast_fp16, current_key_29_cast_fp16, current_key_31_cast_fp16, current_key_33_cast_fp16, current_key_35_cast_fp16, current_key_37_cast_fp16, current_key_39_cast_fp16, current_key_41_cast_fp16, current_key_43_cast_fp16, current_key_45_cast_fp16, current_key_47_cast_fp16, current_key_49_cast_fp16, current_key_51_cast_fp16, current_key_53_cast_fp16, current_key_55_cast_fp16, current_key_57_cast_fp16, current_key_59_cast_fp16, current_key_61_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_387_cast_fp16")];
            tensor<int32, []> var_7052 = const()[name = tensor<string, []>("op_7052"), val = tensor<int32, []>(1)];
            tensor<bool, []> obj_interleave_0 = const()[name = tensor<string, []>("obj_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 40960, 1, 1]> value_cache_updates = concat(axis = var_7052, interleave = obj_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_23_cast_fp16, current_value_25_cast_fp16, current_value_27_cast_fp16, current_value_29_cast_fp16, current_value_31_cast_fp16, current_value_33_cast_fp16, current_value_35_cast_fp16, current_value_37_cast_fp16, current_value_39_cast_fp16, current_value_41_cast_fp16, current_value_43_cast_fp16, current_value_45_cast_fp16, current_value_47_cast_fp16, current_value_49_cast_fp16, current_value_51_cast_fp16, current_value_53_cast_fp16, current_value_55_cast_fp16, current_value_57_cast_fp16, current_value_59_cast_fp16, current_value_61_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_cast_fp16")];
        } -> (logits, key_cache_updates, value_cache_updates);
}