File size: 101,708 Bytes
bd7bc12 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 | program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.14.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
func main<ios17>(tensor<int32, [1, ?]> input_ids) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>>>((("DefaultShapes", {{"input_ids", [1, 16]}}), ("EnumeratedShapes", {{"input_ids_1_1_1_1_128_", {{"input_ids", [1, 128]}}}, {"input_ids_1_1_1_1_16_", {{"input_ids", [1, 16]}}}, {"input_ids_1_1_1_1_256_", {{"input_ids", [1, 256]}}}, {"input_ids_1_1_1_1_32_", {{"input_ids", [1, 32]}}}, {"input_ids_1_1_1_1_512_", {{"input_ids", [1, 512]}}}, {"input_ids_1_1_1_1_64_", {{"input_ids", [1, 64]}}}})))] {
tensor<int32, [1, 512]> m_embeddings_position_ids = const()[name = tensor<string, []>("m_embeddings_position_ids"), val = tensor<int32, [1, 512]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299, 300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 312, 313, 314, 315, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431, 432, 433, 434, 435, 436, 437, 438, 439, 440, 441, 442, 443, 444, 445, 446, 447, 448, 449, 450, 451, 452, 453, 454, 455, 456, 457, 458, 459, 460, 461, 462, 463, 464, 465, 466, 467, 468, 469, 470, 471, 472, 473, 474, 475, 476, 477, 478, 479, 480, 481, 482, 483, 484, 485, 486, 487, 488, 489, 490, 491, 492, 493, 494, 495, 496, 497, 498, 499, 500, 501, 502, 503, 504, 505, 506, 507, 508, 509, 510, 511]])];
tensor<int32, []> var_5 = const()[name = tensor<string, []>("op_5"), val = tensor<int32, []>(0)];
tensor<bool, [1, ?]> var_6 = not_equal(x = input_ids, y = var_5)[name = tensor<string, []>("op_6")];
tensor<string, []> mask_1_dtype_0 = const()[name = tensor<string, []>("mask_1_dtype_0"), val = tensor<string, []>("int32")];
tensor<int32, [1, ?]> input_1 = sub(x = input_ids, y = input_ids)[name = tensor<string, []>("sub_0")];
tensor<int32, []> var_38 = const()[name = tensor<string, []>("op_38"), val = tensor<int32, []>(1)];
tensor<int32, [2]> var_41_shape = shape(x = input_ids)[name = tensor<string, []>("op_41_shape")];
tensor<int32, []> gather_0_axis_0 = const()[name = tensor<string, []>("gather_0_axis_0"), val = tensor<int32, []>(0)];
tensor<int32, []> gather_0_batch_dims_0 = const()[name = tensor<string, []>("gather_0_batch_dims_0"), val = tensor<int32, []>(0)];
tensor<bool, []> gather_0_validate_indices_0 = const()[name = tensor<string, []>("gather_0_validate_indices_0"), val = tensor<bool, []>(false)];
tensor<string, []> var_41_shape_to_int16_dtype_0 = const()[name = tensor<string, []>("op_41_shape_to_int16_dtype_0"), val = tensor<string, []>("int16")];
tensor<uint16, []> gather_0_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_0_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
tensor<int16, [2]> var_41_shape_to_int16 = cast(dtype = var_41_shape_to_int16_dtype_0, x = var_41_shape)[name = tensor<string, []>("cast_47")];
tensor<int16, []> gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_41_shape_to_int16)[name = tensor<string, []>("gather_0_cast_uint16")];
tensor<string, []> gather_0_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_0_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
tensor<int32, []> concat_0_values0_0 = const()[name = tensor<string, []>("concat_0_values0_0"), val = tensor<int32, []>(1)];
tensor<int32, []> concat_0_axis_0 = const()[name = tensor<string, []>("concat_0_axis_0"), val = tensor<int32, []>(0)];
tensor<bool, []> concat_0_interleave_0 = const()[name = tensor<string, []>("concat_0_interleave_0"), val = tensor<bool, []>(false)];
tensor<int32, []> gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = tensor<string, []>("cast_46")];
tensor<int32, [2]> concat_0 = concat(axis = concat_0_axis_0, interleave = concat_0_interleave_0, values = (concat_0_values0_0, gather_0_cast_uint16_to_int32))[name = tensor<string, []>("concat_0")];
tensor<int32, [2]> input_3_begin_0 = const()[name = tensor<string, []>("input_3_begin_0"), val = tensor<int32, [2]>([0, 0])];
tensor<bool, [2]> input_3_end_mask_0 = const()[name = tensor<string, []>("input_3_end_mask_0"), val = tensor<bool, [2]>([true, false])];
tensor<int32, [1, ?]> input_3 = slice_by_index(begin = input_3_begin_0, end = concat_0, end_mask = input_3_end_mask_0, x = m_embeddings_position_ids)[name = tensor<string, []>("input_3")];
tensor<int32, []> inputs_embeds_axis_0 = const()[name = tensor<string, []>("inputs_embeds_axis_0"), val = tensor<int32, []>(0)];
tensor<int32, []> inputs_embeds_batch_dims_0 = const()[name = tensor<string, []>("inputs_embeds_batch_dims_0"), val = tensor<int32, []>(0)];
tensor<bool, []> inputs_embeds_validate_indices_0 = const()[name = tensor<string, []>("inputs_embeds_validate_indices_0"), val = tensor<bool, []>(false)];
tensor<fp16, [30522, 384]> m_embeddings_word_embeddings_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [30522, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
tensor<string, []> input_ids_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_ids_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
tensor<uint16, [1, ?]> input_ids_to_uint16 = cast(dtype = input_ids_to_uint16_dtype_0, x = input_ids)[name = tensor<string, []>("cast_45")];
tensor<fp16, [1, ?, 384]> inputs_embeds_cast_fp16_cast_uint16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids_to_uint16, validate_indices = inputs_embeds_validate_indices_0, x = m_embeddings_word_embeddings_weight_to_fp16)[name = tensor<string, []>("inputs_embeds_cast_fp16_cast_uint16")];
tensor<int32, []> token_type_embeddings_1_axis_0 = const()[name = tensor<string, []>("token_type_embeddings_1_axis_0"), val = tensor<int32, []>(0)];
tensor<int32, []> token_type_embeddings_1_batch_dims_0 = const()[name = tensor<string, []>("token_type_embeddings_1_batch_dims_0"), val = tensor<int32, []>(0)];
tensor<bool, []> token_type_embeddings_1_validate_indices_0 = const()[name = tensor<string, []>("token_type_embeddings_1_validate_indices_0"), val = tensor<bool, []>(false)];
tensor<fp16, [2, 384]> m_embeddings_token_type_embeddings_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_token_type_embeddings_weight_to_fp16"), val = tensor<fp16, [2, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23441024)))];
tensor<string, []> input_1_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_1_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
tensor<uint16, [1, ?]> input_1_to_uint16 = cast(dtype = input_1_to_uint16_dtype_0, x = input_1)[name = tensor<string, []>("cast_44")];
tensor<fp16, [1, ?, 384]> token_type_embeddings_1_cast_fp16_cast_uint16 = gather(axis = token_type_embeddings_1_axis_0, batch_dims = token_type_embeddings_1_batch_dims_0, indices = input_1_to_uint16, validate_indices = token_type_embeddings_1_validate_indices_0, x = m_embeddings_token_type_embeddings_weight_to_fp16)[name = tensor<string, []>("token_type_embeddings_1_cast_fp16_cast_uint16")];
tensor<fp16, [1, ?, 384]> embeddings_1_cast_fp16 = add(x = inputs_embeds_cast_fp16_cast_uint16, y = token_type_embeddings_1_cast_fp16_cast_uint16)[name = tensor<string, []>("embeddings_1_cast_fp16")];
tensor<int32, []> position_embeddings_1_axis_0 = const()[name = tensor<string, []>("position_embeddings_1_axis_0"), val = tensor<int32, []>(0)];
tensor<int32, []> position_embeddings_1_batch_dims_0 = const()[name = tensor<string, []>("position_embeddings_1_batch_dims_0"), val = tensor<int32, []>(0)];
tensor<bool, []> position_embeddings_1_validate_indices_0 = const()[name = tensor<string, []>("position_embeddings_1_validate_indices_0"), val = tensor<bool, []>(false)];
tensor<fp16, [512, 384]> m_embeddings_position_embeddings_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_position_embeddings_weight_to_fp16"), val = tensor<fp16, [512, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23442624)))];
tensor<string, []> input_3_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_3_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
tensor<uint16, [1, ?]> input_3_to_uint16 = cast(dtype = input_3_to_uint16_dtype_0, x = input_3)[name = tensor<string, []>("cast_43")];
tensor<fp16, [1, ?, 384]> position_embeddings_1_cast_fp16_cast_uint16 = gather(axis = position_embeddings_1_axis_0, batch_dims = position_embeddings_1_batch_dims_0, indices = input_3_to_uint16, validate_indices = position_embeddings_1_validate_indices_0, x = m_embeddings_position_embeddings_weight_to_fp16)[name = tensor<string, []>("position_embeddings_1_cast_fp16_cast_uint16")];
tensor<fp16, [1, ?, 384]> input_5_cast_fp16 = add(x = embeddings_1_cast_fp16, y = position_embeddings_1_cast_fp16_cast_uint16)[name = tensor<string, []>("input_5_cast_fp16")];
tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_embeddings_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23835904)))];
tensor<fp16, [384]> m_embeddings_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23836736)))];
tensor<fp16, []> var_35_to_fp16 = const()[name = tensor<string, []>("op_35_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
tensor<fp16, [1, ?, 384]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = m_embeddings_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_embeddings_LayerNorm_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
tensor<int32, [1, ?]> mask_1 = cast(dtype = mask_1_dtype_0, x = var_6)[name = tensor<string, []>("cast_48")];
tensor<int32, [2]> var_67_shape = shape(x = mask_1)[name = tensor<string, []>("op_67_shape")];
tensor<int32, []> gather_2 = const()[name = tensor<string, []>("gather_2"), val = tensor<int32, []>(1)];
tensor<int32, []> gather_3_axis_0 = const()[name = tensor<string, []>("gather_3_axis_0"), val = tensor<int32, []>(0)];
tensor<int32, []> gather_3_batch_dims_0 = const()[name = tensor<string, []>("gather_3_batch_dims_0"), val = tensor<int32, []>(0)];
tensor<bool, []> gather_3_validate_indices_0 = const()[name = tensor<string, []>("gather_3_validate_indices_0"), val = tensor<bool, []>(false)];
tensor<string, []> var_67_shape_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_67_shape_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
tensor<uint16, []> gather_3_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_3_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
tensor<uint16, [2]> var_67_shape_to_uint16 = cast(dtype = var_67_shape_to_uint16_dtype_0, x = var_67_shape)[name = tensor<string, []>("cast_42")];
tensor<uint16, []> gather_3_cast_uint16 = gather(axis = gather_3_axis_0, batch_dims = gather_3_batch_dims_0, indices = gather_3_indices_0_to_uint16, validate_indices = gather_3_validate_indices_0, x = var_67_shape_to_uint16)[name = tensor<string, []>("gather_3_cast_uint16")];
tensor<string, []> gather_3_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_3_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
tensor<int32, [1]> var_70_axes_0 = const()[name = tensor<string, []>("op_70_axes_0"), val = tensor<int32, [1]>([1])];
tensor<int32, [1, 1, ?]> var_70 = expand_dims(axes = var_70_axes_0, x = mask_1)[name = tensor<string, []>("op_70")];
tensor<int32, [1]> var_71_axes_0 = const()[name = tensor<string, []>("op_71_axes_0"), val = tensor<int32, [1]>([2])];
tensor<int32, [1, 1, 1, ?]> var_71 = expand_dims(axes = var_71_axes_0, x = var_70)[name = tensor<string, []>("op_71")];
tensor<int32, []> concat_1_axis_0 = const()[name = tensor<string, []>("concat_1_axis_0"), val = tensor<int32, []>(0)];
tensor<bool, []> concat_1_interleave_0 = const()[name = tensor<string, []>("concat_1_interleave_0"), val = tensor<bool, []>(false)];
tensor<int32, []> gather_3_cast_uint16_to_int32 = cast(dtype = gather_3_cast_uint16_to_int32_dtype_0, x = gather_3_cast_uint16)[name = tensor<string, []>("cast_41")];
tensor<int32, [4]> concat_1 = concat(axis = concat_1_axis_0, interleave = concat_1_interleave_0, values = (gather_2, var_38, gather_0_cast_uint16_to_int32, gather_3_cast_uint16_to_int32))[name = tensor<string, []>("concat_1")];
tensor<int32, [4]> shape_1 = shape(x = var_71)[name = tensor<string, []>("shape_1")];
tensor<int32, []> equal_0_y_0 = const()[name = tensor<string, []>("equal_0_y_0"), val = tensor<int32, []>(-1)];
tensor<bool, [4]> equal_0 = equal(x = concat_1, y = equal_0_y_0)[name = tensor<string, []>("equal_0")];
tensor<int32, [4]> select_0 = select(a = shape_1, b = concat_1, cond = equal_0)[name = tensor<string, []>("select_0")];
tensor<int32, [4]> real_div_0 = real_div(x = select_0, y = shape_1)[name = tensor<string, []>("real_div_0")];
tensor<int32, [?, ?, ?, ?]> var_74 = tile(reps = real_div_0, x = var_71)[name = tensor<string, []>("op_74")];
tensor<fp16, []> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
tensor<string, []> expanded_mask_to_fp16_dtype_0 = const()[name = tensor<string, []>("expanded_mask_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
tensor<fp16, [?, ?, ?, ?]> var_74_to_fp16 = cast(dtype = expanded_mask_to_fp16_dtype_0, x = var_74)[name = tensor<string, []>("cast_40")];
tensor<fp16, [?, ?, ?, ?]> inverted_mask_cast_fp16 = sub(x = const_0_to_fp16, y = var_74_to_fp16)[name = tensor<string, []>("inverted_mask_cast_fp16")];
tensor<string, []> var_79_dtype_0 = const()[name = tensor<string, []>("op_79_dtype_0"), val = tensor<string, []>("bool")];
tensor<fp16, []> var_22_to_fp16 = const()[name = tensor<string, []>("op_22_to_fp16"), val = tensor<fp16, []>(-inf)];
tensor<bool, [?, ?, ?, ?]> inverted_mask_cast_fp16_to_bool = cast(dtype = var_79_dtype_0, x = inverted_mask_cast_fp16)[name = tensor<string, []>("cast_39")];
tensor<fp16, [?, ?, ?, ?]> attention_mask_cast_fp16 = select(a = var_22_to_fp16, b = inverted_mask_cast_fp16, cond = inverted_mask_cast_fp16_to_bool)[name = tensor<string, []>("attention_mask_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_0_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_query_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23837568)))];
tensor<fp16, [384]> m_encoder_layer_0_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24132544)))];
tensor<fp16, [1, ?, 384]> linear_0_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_query_bias_to_fp16, weight = m_encoder_layer_0_attention_self_query_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
tensor<int32, [4]> var_106 = const()[name = tensor<string, []>("op_106"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_107_cast_fp16 = reshape(shape = var_106, x = linear_0_cast_fp16)[name = tensor<string, []>("op_107_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_0_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_key_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24133376)))];
tensor<fp16, [384]> m_encoder_layer_0_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24428352)))];
tensor<fp16, [1, ?, 384]> linear_1_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_key_bias_to_fp16, weight = m_encoder_layer_0_attention_self_key_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
tensor<int32, [4]> var_112 = const()[name = tensor<string, []>("op_112"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_113_cast_fp16 = reshape(shape = var_112, x = linear_1_cast_fp16)[name = tensor<string, []>("op_113_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_0_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_value_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24429184)))];
tensor<fp16, [384]> m_encoder_layer_0_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24724160)))];
tensor<fp16, [1, ?, 384]> linear_2_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_value_bias_to_fp16, weight = m_encoder_layer_0_attention_self_value_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
tensor<int32, [4]> var_118 = const()[name = tensor<string, []>("op_118"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_119_cast_fp16 = reshape(shape = var_118, x = linear_2_cast_fp16)[name = tensor<string, []>("op_119_cast_fp16")];
tensor<int32, [4]> value_layer_1_perm_0 = const()[name = tensor<string, []>("value_layer_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, []> mul_0_y_0_to_fp16 = const()[name = tensor<string, []>("mul_0_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
tensor<fp16, [1, ?, 12, 32]> mul_0_cast_fp16 = mul(x = var_107_cast_fp16, y = mul_0_y_0_to_fp16)[name = tensor<string, []>("mul_0_cast_fp16")];
tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<int32, [4]> transpose_24_perm_0 = const()[name = tensor<string, []>("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_25_perm_0 = const()[name = tensor<string, []>("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<fp16, [1, 12, ?, 32]> transpose_25 = transpose(perm = transpose_25_perm_0, x = var_113_cast_fp16)[name = tensor<string, []>("transpose_57")];
tensor<fp16, [1, 12, ?, 32]> transpose_24 = transpose(perm = transpose_24_perm_0, x = mul_0_cast_fp16)[name = tensor<string, []>("transpose_58")];
tensor<fp16, [1, 12, ?, ?]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor<string, []>("matmul_0_cast_fp16")];
tensor<fp16, [?, 12, ?, ?]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("add_0_cast_fp16")];
tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)];
tensor<fp16, [?, 12, ?, ?]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = tensor<string, []>("softmax_0_cast_fp16")];
tensor<bool, []> attn_output_1_transpose_x_0 = const()[name = tensor<string, []>("attn_output_1_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_output_1_transpose_y_0 = const()[name = tensor<string, []>("attn_output_1_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 12, ?, 32]> value_layer_1_cast_fp16 = transpose(perm = value_layer_1_perm_0, x = var_119_cast_fp16)[name = tensor<string, []>("transpose_59")];
tensor<fp16, [?, 12, ?, 32]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_layer_1_cast_fp16)[name = tensor<string, []>("attn_output_1_cast_fp16")];
tensor<int32, [4]> attn_output_3_perm_0 = const()[name = tensor<string, []>("attn_output_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [3]>([1, -1, 384])];
tensor<fp16, [?, ?, 12, 32]> attn_output_3_cast_fp16 = transpose(perm = attn_output_3_perm_0, x = attn_output_1_cast_fp16)[name = tensor<string, []>("transpose_56")];
tensor<fp16, [1, ?, 384]> input_9_cast_fp16 = reshape(shape = concat_2x, x = attn_output_3_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_0_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24724992)))];
tensor<fp16, [384]> m_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25019968)))];
tensor<fp16, [1, ?, 384]> linear_3_cast_fp16 = linear(bias = m_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_13_cast_fp16 = add(x = linear_3_cast_fp16, y = input_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
tensor<int32, [1]> input_15_axes_0 = const()[name = tensor<string, []>("input_15_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25020800)))];
tensor<fp16, [384]> m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25021632)))];
tensor<fp16, [1, ?, 384]> input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
tensor<fp16, [1536, 384]> m_encoder_layer_0_intermediate_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25022464)))];
tensor<fp16, [1536]> m_encoder_layer_0_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26202176)))];
tensor<fp16, [1, ?, 1536]> linear_4_cast_fp16 = linear(bias = m_encoder_layer_0_intermediate_dense_bias_to_fp16, weight = m_encoder_layer_0_intermediate_dense_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, ?, 1536]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_4_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
tensor<fp16, [384, 1536]> m_encoder_layer_0_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26205312)))];
tensor<fp16, [384]> m_encoder_layer_0_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27385024)))];
tensor<fp16, [1, ?, 384]> linear_5_cast_fp16 = linear(bias = m_encoder_layer_0_output_dense_bias_to_fp16, weight = m_encoder_layer_0_output_dense_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_23_cast_fp16 = add(x = linear_5_cast_fp16, y = input_15_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
tensor<int32, [1]> hidden_states_7_axes_0 = const()[name = tensor<string, []>("hidden_states_7_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_0_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27385856)))];
tensor<fp16, [384]> m_encoder_layer_0_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27386688)))];
tensor<fp16, [1, ?, 384]> hidden_states_7_cast_fp16 = layer_norm(axes = hidden_states_7_axes_0, beta = m_encoder_layer_0_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_0_output_LayerNorm_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_1_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_query_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27387520)))];
tensor<fp16, [384]> m_encoder_layer_1_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27682496)))];
tensor<fp16, [1, ?, 384]> linear_6_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_query_bias_to_fp16, weight = m_encoder_layer_1_attention_self_query_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
tensor<int32, [4]> var_165 = const()[name = tensor<string, []>("op_165"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_166_cast_fp16 = reshape(shape = var_165, x = linear_6_cast_fp16)[name = tensor<string, []>("op_166_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_1_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_key_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27683328)))];
tensor<fp16, [384]> m_encoder_layer_1_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27978304)))];
tensor<fp16, [1, ?, 384]> linear_7_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_key_bias_to_fp16, weight = m_encoder_layer_1_attention_self_key_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
tensor<int32, [4]> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_172_cast_fp16 = reshape(shape = var_171, x = linear_7_cast_fp16)[name = tensor<string, []>("op_172_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_1_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_value_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27979136)))];
tensor<fp16, [384]> m_encoder_layer_1_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28274112)))];
tensor<fp16, [1, ?, 384]> linear_8_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_value_bias_to_fp16, weight = m_encoder_layer_1_attention_self_value_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
tensor<int32, [4]> var_177 = const()[name = tensor<string, []>("op_177"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_178_cast_fp16 = reshape(shape = var_177, x = linear_8_cast_fp16)[name = tensor<string, []>("op_178_cast_fp16")];
tensor<int32, [4]> value_layer_3_perm_0 = const()[name = tensor<string, []>("value_layer_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, []> mul_1_y_0_to_fp16 = const()[name = tensor<string, []>("mul_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
tensor<fp16, [1, ?, 12, 32]> mul_1_cast_fp16 = mul(x = var_166_cast_fp16, y = mul_1_y_0_to_fp16)[name = tensor<string, []>("mul_1_cast_fp16")];
tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<int32, [4]> transpose_26_perm_0 = const()[name = tensor<string, []>("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_27_perm_0 = const()[name = tensor<string, []>("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<fp16, [1, 12, ?, 32]> transpose_27 = transpose(perm = transpose_27_perm_0, x = var_172_cast_fp16)[name = tensor<string, []>("transpose_53")];
tensor<fp16, [1, 12, ?, 32]> transpose_26 = transpose(perm = transpose_26_perm_0, x = mul_1_cast_fp16)[name = tensor<string, []>("transpose_54")];
tensor<fp16, [1, 12, ?, ?]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor<string, []>("matmul_1_cast_fp16")];
tensor<fp16, [?, 12, ?, ?]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)];
tensor<fp16, [?, 12, ?, ?]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = tensor<string, []>("softmax_1_cast_fp16")];
tensor<bool, []> attn_output_5_transpose_x_0 = const()[name = tensor<string, []>("attn_output_5_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_output_5_transpose_y_0 = const()[name = tensor<string, []>("attn_output_5_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 12, ?, 32]> value_layer_3_cast_fp16 = transpose(perm = value_layer_3_perm_0, x = var_178_cast_fp16)[name = tensor<string, []>("transpose_55")];
tensor<fp16, [?, 12, ?, 32]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_layer_3_cast_fp16)[name = tensor<string, []>("attn_output_5_cast_fp16")];
tensor<int32, [4]> attn_output_7_perm_0 = const()[name = tensor<string, []>("attn_output_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> concat_3x = const()[name = tensor<string, []>("concat_3x"), val = tensor<int32, [3]>([1, -1, 384])];
tensor<fp16, [?, ?, 12, 32]> attn_output_7_cast_fp16 = transpose(perm = attn_output_7_perm_0, x = attn_output_5_cast_fp16)[name = tensor<string, []>("transpose_52")];
tensor<fp16, [1, ?, 384]> input_25_cast_fp16 = reshape(shape = concat_3x, x = attn_output_7_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_1_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28274944)))];
tensor<fp16, [384]> m_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28569920)))];
tensor<fp16, [1, ?, 384]> linear_9_cast_fp16 = linear(bias = m_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_29_cast_fp16 = add(x = linear_9_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
tensor<int32, [1]> input_31_axes_0 = const()[name = tensor<string, []>("input_31_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28570752)))];
tensor<fp16, [384]> m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28571584)))];
tensor<fp16, [1, ?, 384]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
tensor<fp16, [1536, 384]> m_encoder_layer_1_intermediate_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28572416)))];
tensor<fp16, [1536]> m_encoder_layer_1_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29752128)))];
tensor<fp16, [1, ?, 1536]> linear_10_cast_fp16 = linear(bias = m_encoder_layer_1_intermediate_dense_bias_to_fp16, weight = m_encoder_layer_1_intermediate_dense_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
tensor<string, []> input_35_mode_0 = const()[name = tensor<string, []>("input_35_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, ?, 1536]> input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
tensor<fp16, [384, 1536]> m_encoder_layer_1_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29755264)))];
tensor<fp16, [384]> m_encoder_layer_1_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30934976)))];
tensor<fp16, [1, ?, 384]> linear_11_cast_fp16 = linear(bias = m_encoder_layer_1_output_dense_bias_to_fp16, weight = m_encoder_layer_1_output_dense_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_39_cast_fp16 = add(x = linear_11_cast_fp16, y = input_31_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
tensor<int32, [1]> hidden_states_13_axes_0 = const()[name = tensor<string, []>("hidden_states_13_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_1_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30935808)))];
tensor<fp16, [384]> m_encoder_layer_1_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30936640)))];
tensor<fp16, [1, ?, 384]> hidden_states_13_cast_fp16 = layer_norm(axes = hidden_states_13_axes_0, beta = m_encoder_layer_1_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_1_output_LayerNorm_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_2_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_query_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30937472)))];
tensor<fp16, [384]> m_encoder_layer_2_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31232448)))];
tensor<fp16, [1, ?, 384]> linear_12_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_query_bias_to_fp16, weight = m_encoder_layer_2_attention_self_query_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
tensor<int32, [4]> var_224 = const()[name = tensor<string, []>("op_224"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_225_cast_fp16 = reshape(shape = var_224, x = linear_12_cast_fp16)[name = tensor<string, []>("op_225_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_2_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_key_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31233280)))];
tensor<fp16, [384]> m_encoder_layer_2_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31528256)))];
tensor<fp16, [1, ?, 384]> linear_13_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_key_bias_to_fp16, weight = m_encoder_layer_2_attention_self_key_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
tensor<int32, [4]> var_230 = const()[name = tensor<string, []>("op_230"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_231_cast_fp16 = reshape(shape = var_230, x = linear_13_cast_fp16)[name = tensor<string, []>("op_231_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_2_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_value_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31529088)))];
tensor<fp16, [384]> m_encoder_layer_2_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31824064)))];
tensor<fp16, [1, ?, 384]> linear_14_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_value_bias_to_fp16, weight = m_encoder_layer_2_attention_self_value_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
tensor<int32, [4]> var_236 = const()[name = tensor<string, []>("op_236"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_237_cast_fp16 = reshape(shape = var_236, x = linear_14_cast_fp16)[name = tensor<string, []>("op_237_cast_fp16")];
tensor<int32, [4]> value_layer_5_perm_0 = const()[name = tensor<string, []>("value_layer_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, []> mul_2_y_0_to_fp16 = const()[name = tensor<string, []>("mul_2_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
tensor<fp16, [1, ?, 12, 32]> mul_2_cast_fp16 = mul(x = var_225_cast_fp16, y = mul_2_y_0_to_fp16)[name = tensor<string, []>("mul_2_cast_fp16")];
tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<int32, [4]> transpose_28_perm_0 = const()[name = tensor<string, []>("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_29_perm_0 = const()[name = tensor<string, []>("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<fp16, [1, 12, ?, 32]> transpose_29 = transpose(perm = transpose_29_perm_0, x = var_231_cast_fp16)[name = tensor<string, []>("transpose_49")];
tensor<fp16, [1, 12, ?, 32]> transpose_28 = transpose(perm = transpose_28_perm_0, x = mul_2_cast_fp16)[name = tensor<string, []>("transpose_50")];
tensor<fp16, [1, 12, ?, ?]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor<string, []>("matmul_2_cast_fp16")];
tensor<fp16, [?, 12, ?, ?]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("add_2_cast_fp16")];
tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)];
tensor<fp16, [?, 12, ?, ?]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = tensor<string, []>("softmax_2_cast_fp16")];
tensor<bool, []> attn_output_9_transpose_x_0 = const()[name = tensor<string, []>("attn_output_9_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_output_9_transpose_y_0 = const()[name = tensor<string, []>("attn_output_9_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 12, ?, 32]> value_layer_5_cast_fp16 = transpose(perm = value_layer_5_perm_0, x = var_237_cast_fp16)[name = tensor<string, []>("transpose_51")];
tensor<fp16, [?, 12, ?, 32]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_layer_5_cast_fp16)[name = tensor<string, []>("attn_output_9_cast_fp16")];
tensor<int32, [4]> attn_output_11_perm_0 = const()[name = tensor<string, []>("attn_output_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> concat_4x = const()[name = tensor<string, []>("concat_4x"), val = tensor<int32, [3]>([1, -1, 384])];
tensor<fp16, [?, ?, 12, 32]> attn_output_11_cast_fp16 = transpose(perm = attn_output_11_perm_0, x = attn_output_9_cast_fp16)[name = tensor<string, []>("transpose_48")];
tensor<fp16, [1, ?, 384]> input_41_cast_fp16 = reshape(shape = concat_4x, x = attn_output_11_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_2_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31824896)))];
tensor<fp16, [384]> m_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32119872)))];
tensor<fp16, [1, ?, 384]> linear_15_cast_fp16 = linear(bias = m_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_45_cast_fp16 = add(x = linear_15_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
tensor<int32, [1]> input_47_axes_0 = const()[name = tensor<string, []>("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32120704)))];
tensor<fp16, [384]> m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32121536)))];
tensor<fp16, [1, ?, 384]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
tensor<fp16, [1536, 384]> m_encoder_layer_2_intermediate_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32122368)))];
tensor<fp16, [1536]> m_encoder_layer_2_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33302080)))];
tensor<fp16, [1, ?, 1536]> linear_16_cast_fp16 = linear(bias = m_encoder_layer_2_intermediate_dense_bias_to_fp16, weight = m_encoder_layer_2_intermediate_dense_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
tensor<string, []> input_51_mode_0 = const()[name = tensor<string, []>("input_51_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, ?, 1536]> input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_16_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
tensor<fp16, [384, 1536]> m_encoder_layer_2_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33305216)))];
tensor<fp16, [384]> m_encoder_layer_2_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34484928)))];
tensor<fp16, [1, ?, 384]> linear_17_cast_fp16 = linear(bias = m_encoder_layer_2_output_dense_bias_to_fp16, weight = m_encoder_layer_2_output_dense_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_55_cast_fp16 = add(x = linear_17_cast_fp16, y = input_47_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
tensor<int32, [1]> hidden_states_19_axes_0 = const()[name = tensor<string, []>("hidden_states_19_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_2_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34485760)))];
tensor<fp16, [384]> m_encoder_layer_2_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34486592)))];
tensor<fp16, [1, ?, 384]> hidden_states_19_cast_fp16 = layer_norm(axes = hidden_states_19_axes_0, beta = m_encoder_layer_2_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_2_output_LayerNorm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("hidden_states_19_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_3_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_query_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34487424)))];
tensor<fp16, [384]> m_encoder_layer_3_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34782400)))];
tensor<fp16, [1, ?, 384]> linear_18_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_query_bias_to_fp16, weight = m_encoder_layer_3_attention_self_query_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
tensor<int32, [4]> var_283 = const()[name = tensor<string, []>("op_283"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_284_cast_fp16 = reshape(shape = var_283, x = linear_18_cast_fp16)[name = tensor<string, []>("op_284_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_3_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_key_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34783232)))];
tensor<fp16, [384]> m_encoder_layer_3_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35078208)))];
tensor<fp16, [1, ?, 384]> linear_19_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_key_bias_to_fp16, weight = m_encoder_layer_3_attention_self_key_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
tensor<int32, [4]> var_289 = const()[name = tensor<string, []>("op_289"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_290_cast_fp16 = reshape(shape = var_289, x = linear_19_cast_fp16)[name = tensor<string, []>("op_290_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_3_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_value_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35079040)))];
tensor<fp16, [384]> m_encoder_layer_3_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35374016)))];
tensor<fp16, [1, ?, 384]> linear_20_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_value_bias_to_fp16, weight = m_encoder_layer_3_attention_self_value_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
tensor<int32, [4]> var_295 = const()[name = tensor<string, []>("op_295"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_296_cast_fp16 = reshape(shape = var_295, x = linear_20_cast_fp16)[name = tensor<string, []>("op_296_cast_fp16")];
tensor<int32, [4]> value_layer_7_perm_0 = const()[name = tensor<string, []>("value_layer_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, []> mul_3_y_0_to_fp16 = const()[name = tensor<string, []>("mul_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
tensor<fp16, [1, ?, 12, 32]> mul_3_cast_fp16 = mul(x = var_284_cast_fp16, y = mul_3_y_0_to_fp16)[name = tensor<string, []>("mul_3_cast_fp16")];
tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<int32, [4]> transpose_30_perm_0 = const()[name = tensor<string, []>("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_31_perm_0 = const()[name = tensor<string, []>("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<fp16, [1, 12, ?, 32]> transpose_31 = transpose(perm = transpose_31_perm_0, x = var_290_cast_fp16)[name = tensor<string, []>("transpose_45")];
tensor<fp16, [1, 12, ?, 32]> transpose_30 = transpose(perm = transpose_30_perm_0, x = mul_3_cast_fp16)[name = tensor<string, []>("transpose_46")];
tensor<fp16, [1, 12, ?, ?]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = transpose_30, y = transpose_31)[name = tensor<string, []>("matmul_3_cast_fp16")];
tensor<fp16, [?, 12, ?, ?]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("add_3_cast_fp16")];
tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)];
tensor<fp16, [?, 12, ?, ?]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = tensor<string, []>("softmax_3_cast_fp16")];
tensor<bool, []> attn_output_13_transpose_x_0 = const()[name = tensor<string, []>("attn_output_13_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_output_13_transpose_y_0 = const()[name = tensor<string, []>("attn_output_13_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 12, ?, 32]> value_layer_7_cast_fp16 = transpose(perm = value_layer_7_perm_0, x = var_296_cast_fp16)[name = tensor<string, []>("transpose_47")];
tensor<fp16, [?, 12, ?, 32]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_layer_7_cast_fp16)[name = tensor<string, []>("attn_output_13_cast_fp16")];
tensor<int32, [4]> attn_output_15_perm_0 = const()[name = tensor<string, []>("attn_output_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> concat_5x = const()[name = tensor<string, []>("concat_5x"), val = tensor<int32, [3]>([1, -1, 384])];
tensor<fp16, [?, ?, 12, 32]> attn_output_15_cast_fp16 = transpose(perm = attn_output_15_perm_0, x = attn_output_13_cast_fp16)[name = tensor<string, []>("transpose_44")];
tensor<fp16, [1, ?, 384]> input_57_cast_fp16 = reshape(shape = concat_5x, x = attn_output_15_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_3_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35374848)))];
tensor<fp16, [384]> m_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35669824)))];
tensor<fp16, [1, ?, 384]> linear_21_cast_fp16 = linear(bias = m_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_61_cast_fp16 = add(x = linear_21_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
tensor<int32, [1]> input_63_axes_0 = const()[name = tensor<string, []>("input_63_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35670656)))];
tensor<fp16, [384]> m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35671488)))];
tensor<fp16, [1, ?, 384]> input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
tensor<fp16, [1536, 384]> m_encoder_layer_3_intermediate_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35672320)))];
tensor<fp16, [1536]> m_encoder_layer_3_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36852032)))];
tensor<fp16, [1, ?, 1536]> linear_22_cast_fp16 = linear(bias = m_encoder_layer_3_intermediate_dense_bias_to_fp16, weight = m_encoder_layer_3_intermediate_dense_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
tensor<string, []> input_67_mode_0 = const()[name = tensor<string, []>("input_67_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, ?, 1536]> input_67_cast_fp16 = gelu(mode = input_67_mode_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
tensor<fp16, [384, 1536]> m_encoder_layer_3_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36855168)))];
tensor<fp16, [384]> m_encoder_layer_3_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38034880)))];
tensor<fp16, [1, ?, 384]> linear_23_cast_fp16 = linear(bias = m_encoder_layer_3_output_dense_bias_to_fp16, weight = m_encoder_layer_3_output_dense_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_71_cast_fp16 = add(x = linear_23_cast_fp16, y = input_63_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
tensor<int32, [1]> hidden_states_25_axes_0 = const()[name = tensor<string, []>("hidden_states_25_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_3_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38035712)))];
tensor<fp16, [384]> m_encoder_layer_3_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38036544)))];
tensor<fp16, [1, ?, 384]> hidden_states_25_cast_fp16 = layer_norm(axes = hidden_states_25_axes_0, beta = m_encoder_layer_3_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_3_output_LayerNorm_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("hidden_states_25_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_4_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_self_query_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38037376)))];
tensor<fp16, [384]> m_encoder_layer_4_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38332352)))];
tensor<fp16, [1, ?, 384]> linear_24_cast_fp16 = linear(bias = m_encoder_layer_4_attention_self_query_bias_to_fp16, weight = m_encoder_layer_4_attention_self_query_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
tensor<int32, [4]> var_342 = const()[name = tensor<string, []>("op_342"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_343_cast_fp16 = reshape(shape = var_342, x = linear_24_cast_fp16)[name = tensor<string, []>("op_343_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_4_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_self_key_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38333184)))];
tensor<fp16, [384]> m_encoder_layer_4_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38628160)))];
tensor<fp16, [1, ?, 384]> linear_25_cast_fp16 = linear(bias = m_encoder_layer_4_attention_self_key_bias_to_fp16, weight = m_encoder_layer_4_attention_self_key_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
tensor<int32, [4]> var_348 = const()[name = tensor<string, []>("op_348"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_349_cast_fp16 = reshape(shape = var_348, x = linear_25_cast_fp16)[name = tensor<string, []>("op_349_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_4_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_self_value_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38628992)))];
tensor<fp16, [384]> m_encoder_layer_4_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38923968)))];
tensor<fp16, [1, ?, 384]> linear_26_cast_fp16 = linear(bias = m_encoder_layer_4_attention_self_value_bias_to_fp16, weight = m_encoder_layer_4_attention_self_value_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
tensor<int32, [4]> var_354 = const()[name = tensor<string, []>("op_354"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_355_cast_fp16 = reshape(shape = var_354, x = linear_26_cast_fp16)[name = tensor<string, []>("op_355_cast_fp16")];
tensor<int32, [4]> value_layer_9_perm_0 = const()[name = tensor<string, []>("value_layer_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, []> mul_4_y_0_to_fp16 = const()[name = tensor<string, []>("mul_4_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
tensor<fp16, [1, ?, 12, 32]> mul_4_cast_fp16 = mul(x = var_343_cast_fp16, y = mul_4_y_0_to_fp16)[name = tensor<string, []>("mul_4_cast_fp16")];
tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<int32, [4]> transpose_32_perm_0 = const()[name = tensor<string, []>("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_33_perm_0 = const()[name = tensor<string, []>("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<fp16, [1, 12, ?, 32]> transpose_33 = transpose(perm = transpose_33_perm_0, x = var_349_cast_fp16)[name = tensor<string, []>("transpose_41")];
tensor<fp16, [1, 12, ?, 32]> transpose_32 = transpose(perm = transpose_32_perm_0, x = mul_4_cast_fp16)[name = tensor<string, []>("transpose_42")];
tensor<fp16, [1, 12, ?, ?]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = transpose_32, y = transpose_33)[name = tensor<string, []>("matmul_4_cast_fp16")];
tensor<fp16, [?, 12, ?, ?]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("add_4_cast_fp16")];
tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)];
tensor<fp16, [?, 12, ?, ?]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = tensor<string, []>("softmax_4_cast_fp16")];
tensor<bool, []> attn_output_17_transpose_x_0 = const()[name = tensor<string, []>("attn_output_17_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_output_17_transpose_y_0 = const()[name = tensor<string, []>("attn_output_17_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 12, ?, 32]> value_layer_9_cast_fp16 = transpose(perm = value_layer_9_perm_0, x = var_355_cast_fp16)[name = tensor<string, []>("transpose_43")];
tensor<fp16, [?, 12, ?, 32]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_layer_9_cast_fp16)[name = tensor<string, []>("attn_output_17_cast_fp16")];
tensor<int32, [4]> attn_output_19_perm_0 = const()[name = tensor<string, []>("attn_output_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> concat_6x = const()[name = tensor<string, []>("concat_6x"), val = tensor<int32, [3]>([1, -1, 384])];
tensor<fp16, [?, ?, 12, 32]> attn_output_19_cast_fp16 = transpose(perm = attn_output_19_perm_0, x = attn_output_17_cast_fp16)[name = tensor<string, []>("transpose_40")];
tensor<fp16, [1, ?, 384]> input_73_cast_fp16 = reshape(shape = concat_6x, x = attn_output_19_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_4_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38924800)))];
tensor<fp16, [384]> m_encoder_layer_4_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39219776)))];
tensor<fp16, [1, ?, 384]> linear_27_cast_fp16 = linear(bias = m_encoder_layer_4_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_4_attention_output_dense_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_77_cast_fp16 = add(x = linear_27_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
tensor<int32, [1]> input_79_axes_0 = const()[name = tensor<string, []>("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39220608)))];
tensor<fp16, [384]> m_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39221440)))];
tensor<fp16, [1, ?, 384]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = m_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
tensor<fp16, [1536, 384]> m_encoder_layer_4_intermediate_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39222272)))];
tensor<fp16, [1536]> m_encoder_layer_4_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40401984)))];
tensor<fp16, [1, ?, 1536]> linear_28_cast_fp16 = linear(bias = m_encoder_layer_4_intermediate_dense_bias_to_fp16, weight = m_encoder_layer_4_intermediate_dense_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
tensor<string, []> input_83_mode_0 = const()[name = tensor<string, []>("input_83_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, ?, 1536]> input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_28_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
tensor<fp16, [384, 1536]> m_encoder_layer_4_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40405120)))];
tensor<fp16, [384]> m_encoder_layer_4_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41584832)))];
tensor<fp16, [1, ?, 384]> linear_29_cast_fp16 = linear(bias = m_encoder_layer_4_output_dense_bias_to_fp16, weight = m_encoder_layer_4_output_dense_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_87_cast_fp16 = add(x = linear_29_cast_fp16, y = input_79_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
tensor<int32, [1]> hidden_states_31_axes_0 = const()[name = tensor<string, []>("hidden_states_31_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_4_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41585664)))];
tensor<fp16, [384]> m_encoder_layer_4_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_4_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41586496)))];
tensor<fp16, [1, ?, 384]> hidden_states_31_cast_fp16 = layer_norm(axes = hidden_states_31_axes_0, beta = m_encoder_layer_4_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_4_output_LayerNorm_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("hidden_states_31_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_5_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_self_query_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41587328)))];
tensor<fp16, [384]> m_encoder_layer_5_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41882304)))];
tensor<fp16, [1, ?, 384]> linear_30_cast_fp16 = linear(bias = m_encoder_layer_5_attention_self_query_bias_to_fp16, weight = m_encoder_layer_5_attention_self_query_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
tensor<int32, [4]> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_402_cast_fp16 = reshape(shape = var_401, x = linear_30_cast_fp16)[name = tensor<string, []>("op_402_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_5_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_self_key_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41883136)))];
tensor<fp16, [384]> m_encoder_layer_5_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42178112)))];
tensor<fp16, [1, ?, 384]> linear_31_cast_fp16 = linear(bias = m_encoder_layer_5_attention_self_key_bias_to_fp16, weight = m_encoder_layer_5_attention_self_key_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
tensor<int32, [4]> var_407 = const()[name = tensor<string, []>("op_407"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_408_cast_fp16 = reshape(shape = var_407, x = linear_31_cast_fp16)[name = tensor<string, []>("op_408_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_5_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_self_value_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42178944)))];
tensor<fp16, [384]> m_encoder_layer_5_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42473920)))];
tensor<fp16, [1, ?, 384]> linear_32_cast_fp16 = linear(bias = m_encoder_layer_5_attention_self_value_bias_to_fp16, weight = m_encoder_layer_5_attention_self_value_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
tensor<int32, [4]> var_413 = const()[name = tensor<string, []>("op_413"), val = tensor<int32, [4]>([1, -1, 12, 32])];
tensor<fp16, [1, ?, 12, 32]> var_414_cast_fp16 = reshape(shape = var_413, x = linear_32_cast_fp16)[name = tensor<string, []>("op_414_cast_fp16")];
tensor<int32, [4]> value_layer_perm_0 = const()[name = tensor<string, []>("value_layer_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, []> mul_5_y_0_to_fp16 = const()[name = tensor<string, []>("mul_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
tensor<fp16, [1, ?, 12, 32]> mul_5_cast_fp16 = mul(x = var_402_cast_fp16, y = mul_5_y_0_to_fp16)[name = tensor<string, []>("mul_5_cast_fp16")];
tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<int32, [4]> transpose_34_perm_0 = const()[name = tensor<string, []>("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_35_perm_0 = const()[name = tensor<string, []>("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<fp16, [1, 12, ?, 32]> transpose_35 = transpose(perm = transpose_35_perm_0, x = var_408_cast_fp16)[name = tensor<string, []>("transpose_37")];
tensor<fp16, [1, 12, ?, 32]> transpose_34 = transpose(perm = transpose_34_perm_0, x = mul_5_cast_fp16)[name = tensor<string, []>("transpose_38")];
tensor<fp16, [1, 12, ?, ?]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = transpose_34, y = transpose_35)[name = tensor<string, []>("matmul_5_cast_fp16")];
tensor<fp16, [?, 12, ?, ?]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("add_5_cast_fp16")];
tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)];
tensor<fp16, [?, 12, ?, ?]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = tensor<string, []>("softmax_5_cast_fp16")];
tensor<bool, []> attn_output_21_transpose_x_0 = const()[name = tensor<string, []>("attn_output_21_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_output_21_transpose_y_0 = const()[name = tensor<string, []>("attn_output_21_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 12, ?, 32]> value_layer_cast_fp16 = transpose(perm = value_layer_perm_0, x = var_414_cast_fp16)[name = tensor<string, []>("transpose_39")];
tensor<fp16, [?, 12, ?, 32]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_layer_cast_fp16)[name = tensor<string, []>("attn_output_21_cast_fp16")];
tensor<int32, [4]> attn_output_perm_0 = const()[name = tensor<string, []>("attn_output_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> concat_7x = const()[name = tensor<string, []>("concat_7x"), val = tensor<int32, [3]>([1, -1, 384])];
tensor<fp16, [?, ?, 12, 32]> attn_output_cast_fp16 = transpose(perm = attn_output_perm_0, x = attn_output_21_cast_fp16)[name = tensor<string, []>("transpose_36")];
tensor<fp16, [1, ?, 384]> input_89_cast_fp16 = reshape(shape = concat_7x, x = attn_output_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
tensor<fp16, [384, 384]> m_encoder_layer_5_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42474752)))];
tensor<fp16, [384]> m_encoder_layer_5_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42769728)))];
tensor<fp16, [1, ?, 384]> linear_33_cast_fp16 = linear(bias = m_encoder_layer_5_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_5_attention_output_dense_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_93_cast_fp16 = add(x = linear_33_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
tensor<int32, [1]> input_95_axes_0 = const()[name = tensor<string, []>("input_95_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42770560)))];
tensor<fp16, [384]> m_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42771392)))];
tensor<fp16, [1, ?, 384]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = m_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
tensor<fp16, [1536, 384]> m_encoder_layer_5_intermediate_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42772224)))];
tensor<fp16, [1536]> m_encoder_layer_5_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43951936)))];
tensor<fp16, [1, ?, 1536]> linear_34_cast_fp16 = linear(bias = m_encoder_layer_5_intermediate_dense_bias_to_fp16, weight = m_encoder_layer_5_intermediate_dense_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
tensor<string, []> input_99_mode_0 = const()[name = tensor<string, []>("input_99_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, ?, 1536]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_34_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
tensor<fp16, [384, 1536]> m_encoder_layer_5_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_output_dense_weight_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43955072)))];
tensor<fp16, [384]> m_encoder_layer_5_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45134784)))];
tensor<fp16, [1, ?, 384]> linear_35_cast_fp16 = linear(bias = m_encoder_layer_5_output_dense_bias_to_fp16, weight = m_encoder_layer_5_output_dense_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
tensor<fp16, [1, ?, 384]> input_103_cast_fp16 = add(x = linear_35_cast_fp16, y = input_95_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
tensor<int32, [1]> token_embeddings_axes_0 = const()[name = tensor<string, []>("token_embeddings_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [384]> m_encoder_layer_5_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45135616)))];
tensor<fp16, [384]> m_encoder_layer_5_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_5_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45136448)))];
tensor<fp16, [1, ?, 384]> token_embeddings_cast_fp16 = layer_norm(axes = token_embeddings_axes_0, beta = m_encoder_layer_5_output_LayerNorm_bias_to_fp16, epsilon = var_35_to_fp16, gamma = m_encoder_layer_5_output_LayerNorm_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("token_embeddings_cast_fp16")];
tensor<int32, []> var_452 = const()[name = tensor<string, []>("op_452"), val = tensor<int32, []>(-1)];
tensor<int32, [1]> var_453_axes_0 = const()[name = tensor<string, []>("op_453_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<int32, [1, ?, 1]> var_453 = expand_dims(axes = var_453_axes_0, x = mask_1)[name = tensor<string, []>("op_453")];
tensor<int32, [3]> shape_2_cast_fp16 = shape(x = token_embeddings_cast_fp16)[name = tensor<string, []>("shape_2_cast_fp16")];
tensor<int32, [3]> shape_3 = shape(x = var_453)[name = tensor<string, []>("shape_3")];
tensor<int32, []> equal_1_y_0 = const()[name = tensor<string, []>("equal_1_y_0"), val = tensor<int32, []>(-1)];
tensor<bool, [3]> equal_1 = equal(x = shape_2_cast_fp16, y = equal_1_y_0)[name = tensor<string, []>("equal_1")];
tensor<int32, [3]> select_1 = select(a = shape_3, b = shape_2_cast_fp16, cond = equal_1)[name = tensor<string, []>("select_1")];
tensor<int32, [3]> real_div_1 = real_div(x = select_1, y = shape_3)[name = tensor<string, []>("real_div_1")];
tensor<int32, [?, ?, ?]> var_454 = tile(reps = real_div_1, x = var_453)[name = tensor<string, []>("op_454")];
tensor<string, []> mask_to_fp16_dtype_0 = const()[name = tensor<string, []>("mask_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
tensor<fp16, [?, ?, ?]> var_454_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_454)[name = tensor<string, []>("cast_38")];
tensor<fp16, [?, ?, 384]> var_456_cast_fp16 = mul(x = token_embeddings_cast_fp16, y = var_454_to_fp16)[name = tensor<string, []>("op_456_cast_fp16")];
tensor<int32, [1]> mean_sum_axes_0 = const()[name = tensor<string, []>("mean_sum_axes_0"), val = tensor<int32, [1]>([1])];
tensor<bool, []> mean_sum_keep_dims_0 = const()[name = tensor<string, []>("mean_sum_keep_dims_0"), val = tensor<bool, []>(false)];
tensor<fp16, [?, 384]> mean_sum_cast_fp16 = reduce_sum(axes = mean_sum_axes_0, keep_dims = mean_sum_keep_dims_0, x = var_456_cast_fp16)[name = tensor<string, []>("mean_sum_cast_fp16")];
tensor<int32, [1]> mean_mask_1_axes_0 = const()[name = tensor<string, []>("mean_mask_1_axes_0"), val = tensor<int32, [1]>([1])];
tensor<bool, []> mean_mask_1_keep_dims_0 = const()[name = tensor<string, []>("mean_mask_1_keep_dims_0"), val = tensor<bool, []>(false)];
tensor<fp16, [?, ?]> mean_mask_1_cast_fp16 = reduce_sum(axes = mean_mask_1_axes_0, keep_dims = mean_mask_1_keep_dims_0, x = var_454_to_fp16)[name = tensor<string, []>("mean_mask_1_cast_fp16")];
tensor<fp16, []> var_447_to_fp16 = const()[name = tensor<string, []>("op_447_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
tensor<fp16, []> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, []>(inf)];
tensor<fp16, [?, ?]> clip_0_cast_fp16 = clip(alpha = var_447_to_fp16, beta = const_1_to_fp16, x = mean_mask_1_cast_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
tensor<fp16, [?, 384]> var_462_cast_fp16 = real_div(x = mean_sum_cast_fp16, y = clip_0_cast_fp16)[name = tensor<string, []>("op_462_cast_fp16")];
tensor<bool, []> input_105_interleave_0 = const()[name = tensor<string, []>("input_105_interleave_0"), val = tensor<bool, []>(false)];
tensor<fp16, [?, 384]> input_105_cast_fp16 = concat(axis = var_452, interleave = input_105_interleave_0, values = var_462_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
tensor<fp16, [768, 384]> rest_1_linear_weight_to_fp16 = const()[name = tensor<string, []>("rest_1_linear_weight_to_fp16"), val = tensor<fp16, [768, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45137280)))];
tensor<fp16, [768]> rest_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("rest_1_linear_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45727168)))];
tensor<fp16, [?, 768]> linear_36_cast_fp16 = linear(bias = rest_1_linear_bias_to_fp16, weight = rest_1_linear_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
tensor<bool, []> var_471 = const()[name = tensor<string, []>("op_471"), val = tensor<bool, []>(true)];
tensor<int32, [1]> var_474 = const()[name = tensor<string, []>("op_474"), val = tensor<int32, [1]>([1])];
tensor<fp16, [?, 1]> var_475_cast_fp16 = reduce_l2_norm(axes = var_474, keep_dims = var_471, x = linear_36_cast_fp16)[name = tensor<string, []>("op_475_cast_fp16")];
tensor<fp16, []> var_469_to_fp16 = const()[name = tensor<string, []>("op_469_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
tensor<fp16, [?, 1]> var_476_cast_fp16 = maximum(x = var_475_cast_fp16, y = var_469_to_fp16)[name = tensor<string, []>("op_476_cast_fp16")];
tensor<int32, [2]> shape_4_cast_fp16 = shape(x = linear_36_cast_fp16)[name = tensor<string, []>("shape_4_cast_fp16")];
tensor<int32, [2]> shape_5_cast_fp16 = shape(x = var_476_cast_fp16)[name = tensor<string, []>("shape_5_cast_fp16")];
tensor<int32, []> equal_2_y_0 = const()[name = tensor<string, []>("equal_2_y_0"), val = tensor<int32, []>(-1)];
tensor<bool, [2]> equal_2 = equal(x = shape_4_cast_fp16, y = equal_2_y_0)[name = tensor<string, []>("equal_2")];
tensor<int32, [2]> select_2 = select(a = shape_5_cast_fp16, b = shape_4_cast_fp16, cond = equal_2)[name = tensor<string, []>("select_2")];
tensor<int32, [2]> real_div_2 = real_div(x = select_2, y = shape_5_cast_fp16)[name = tensor<string, []>("real_div_2")];
tensor<fp16, [?, ?]> denom_1_cast_fp16 = tile(reps = real_div_2, x = var_476_cast_fp16)[name = tensor<string, []>("denom_1_cast_fp16")];
tensor<fp16, [?, 768]> input_cast_fp16 = real_div(x = linear_36_cast_fp16, y = denom_1_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
tensor<int32, [1]> var_481 = const()[name = tensor<string, []>("op_481"), val = tensor<int32, [1]>([-1])];
tensor<bool, []> var_482 = const()[name = tensor<string, []>("op_482"), val = tensor<bool, []>(true)];
tensor<fp16, [?, 1]> var_484_cast_fp16 = reduce_l2_norm(axes = var_481, keep_dims = var_482, x = input_cast_fp16)[name = tensor<string, []>("op_484_cast_fp16")];
tensor<fp16, []> var_485_to_fp16 = const()[name = tensor<string, []>("op_485_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
tensor<fp16, [?, 1]> var_486_cast_fp16 = maximum(x = var_484_cast_fp16, y = var_485_to_fp16)[name = tensor<string, []>("op_486_cast_fp16")];
tensor<int32, [2]> shape_6_cast_fp16 = shape(x = input_cast_fp16)[name = tensor<string, []>("shape_6_cast_fp16")];
tensor<int32, [2]> shape_7_cast_fp16 = shape(x = var_486_cast_fp16)[name = tensor<string, []>("shape_7_cast_fp16")];
tensor<int32, []> equal_3_y_0 = const()[name = tensor<string, []>("equal_3_y_0"), val = tensor<int32, []>(-1)];
tensor<bool, [2]> equal_3 = equal(x = shape_6_cast_fp16, y = equal_3_y_0)[name = tensor<string, []>("equal_3")];
tensor<int32, [2]> select_3 = select(a = shape_7_cast_fp16, b = shape_6_cast_fp16, cond = equal_3)[name = tensor<string, []>("select_3")];
tensor<int32, [2]> real_div_3 = real_div(x = select_3, y = shape_7_cast_fp16)[name = tensor<string, []>("real_div_3")];
tensor<fp16, [?, ?]> denom_cast_fp16 = tile(reps = real_div_3, x = var_486_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
tensor<fp16, [?, 768]> var_488_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("op_488_cast_fp16")];
tensor<string, []> var_488_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_488_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
tensor<fp32, [?, 768]> embedding = cast(dtype = var_488_cast_fp16_to_fp32_dtype_0, x = var_488_cast_fp16)[name = tensor<string, []>("cast_37")];
} -> (embedding);
} |