File size: 96,246 Bytes
9ebc396 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 | program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3405.2.1"}})]
{
func main<ios18>(tensor<fp16, [?, 64, 1, 2]> x) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, dict<string, tensor<int32, [?]>>>>>((("DefaultShapes", {{"x", [12, 64, 1, 2]}}), ("EnumeratedShapes", {{"25a592c0", {{"x", [28, 64, 1, 2]}}}, {"38088a33", {{"x", [18, 64, 1, 2]}}}, {"42bba9e4", {{"x", [22, 64, 1, 2]}}}, {"59316af0", {{"x", [12, 64, 1, 2]}}}, {"6bca635d", {{"x", [16, 64, 1, 2]}}}, {"84aa3ba0", {{"x", [8, 64, 1, 2]}}}, {"981a1dc8", {{"x", [32, 64, 1, 2]}}}, {"9d44cf0c", {{"x", [24, 64, 1, 2]}}}, {"9e64e8ea", {{"x", [10, 64, 1, 2]}}}, {"bcc527a0", {{"x", [26, 64, 1, 2]}}}, {"ce7d4a38", {{"x", [20, 64, 1, 2]}}}, {"d8a49046", {{"x", [14, 64, 1, 2]}}}, {"efb221f6", {{"x", [30, 64, 1, 2]}}}, {"f0e65740", {{"x", [1, 64, 1, 2]}}}})))] {
string x_3_pad_type_0 = const()[name = string("x_3_pad_type_0"), val = string("valid")];
tensor<int32, [2]> x_3_strides_0 = const()[name = string("x_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> x_3_pad_0 = const()[name = string("x_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> x_3_dilations_0 = const()[name = string("x_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 x_3_groups_0 = const()[name = string("x_3_groups_0"), val = int32(1)];
tensor<fp16, [1024, 64, 1, 1]> var_53_to_fp16 = const()[name = string("op_53_to_fp16"), val = tensor<fp16, [1024, 64, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
tensor<fp16, [1024]> layer_in_proj_bias_to_fp16 = const()[name = string("layer_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131200)))];
tensor<fp16, [?, 1024, 1, 2]> x_3_cast_fp16 = conv(bias = layer_in_proj_bias_to_fp16, dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = var_53_to_fp16, x = x)[name = string("x_3_cast_fp16")];
fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
tensor<fp16, [?, 1024, 1, 2]> fill_like_0_cast_fp16 = fill_like(ref_tensor = x_3_cast_fp16, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
tensor<int32, [4]> var_70_begin_0 = const()[name = string("op_70_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> var_70_end_0 = const()[name = string("op_70_end_0"), val = tensor<int32, [4]>([0, 1024, 1, 1])];
tensor<bool, [4]> var_70_end_mask_0 = const()[name = string("op_70_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [?, 1024, 1, 1]> var_70_cast_fp16 = slice_by_index(begin = var_70_begin_0, end = var_70_end_0, end_mask = var_70_end_mask_0, x = fill_like_0_cast_fp16)[name = string("op_70_cast_fp16")];
tensor<fp16, [1, 1024, 1, 1]> var_71_to_fp16 = const()[name = string("op_71_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312)))];
tensor<fp16, [?, 1024, 1, 1]> special_tokens_cast_fp16 = mul(x = var_70_cast_fp16, y = var_71_to_fp16)[name = string("special_tokens_cast_fp16")];
int32 var_74 = const()[name = string("op_74"), val = int32(3)];
bool x_5_interleave_0 = const()[name = string("x_5_interleave_0"), val = bool(false)];
tensor<fp16, [?, 1024, 1, 3]> x_5_cast_fp16 = concat(axis = var_74, interleave = x_5_interleave_0, values = (special_tokens_cast_fp16, x_3_cast_fp16))[name = string("x_5_cast_fp16")];
int32 var_86 = const()[name = string("op_86"), val = int32(-2)];
int32 var_90 = const()[name = string("op_90"), val = int32(1)];
int32 var_95 = const()[name = string("op_95"), val = int32(2)];
fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_100_cast_fp16 = mul(x = x_5_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_100_cast_fp16")];
bool x_7_interleave_0 = const()[name = string("x_7_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_7_cast_fp16 = concat(axis = var_90, interleave = x_7_interleave_0, values = (x_5_cast_fp16, var_100_cast_fp16))[name = string("x_7_cast_fp16")];
tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_110_to_fp16 = const()[name = string("op_110_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_110_to_fp16, x = x_7_cast_fp16)[name = string("out_1_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135424)))];
tensor<fp16, [?, 2048, 1, 3]> out_3_cast_fp16 = mul(x = out_1_cast_fp16, y = layer_encoder_layers_0_input_layernorm_weight_to_fp16)[name = string("out_3_cast_fp16")];
tensor<int32, [2]> var_116_split_sizes_0 = const()[name = string("op_116_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_116_axis_0 = const()[name = string("op_116_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_116_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_116_cast_fp16_1 = split(axis = var_116_axis_0, split_sizes = var_116_split_sizes_0, x = out_3_cast_fp16)[name = string("op_116_cast_fp16")];
string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_81_to_fp16 = const()[name = string("op_81_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139584)))];
tensor<fp16, [?, 1024, 1, 3]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = var_81_to_fp16, x = var_116_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
tensor<fp16, [128, 1024, 1, 1]> var_82_to_fp16 = const()[name = string("op_82_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2236800)))];
tensor<fp16, [?, 128, 1, 3]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = var_82_to_fp16, x = var_116_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
tensor<fp16, [128, 1024, 1, 1]> var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2499008)))];
tensor<fp16, [?, 128, 1, 3]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = var_83_to_fp16, x = var_116_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([-1, 16, 64, 3])];
tensor<fp16, [?, 16, 64, 3]> embed_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("embed_1_cast_fp16")];
tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([-1, 2, 64, 3])];
tensor<fp16, [?, 2, 64, 3]> embed_3_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("embed_3_cast_fp16")];
tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([-1, 2, 64, 3])];
tensor<fp16, [?, 2, 64, 3]> value_states_3_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("value_states_3_cast_fp16")];
tensor<fp16, [64, 3]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [64, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2761216)))];
tensor<fp16, [?, 16, 64, 3]> var_142_cast_fp16 = mul(x = embed_1_cast_fp16, y = cos_to_fp16)[name = string("op_142_cast_fp16")];
tensor<int32, [2]> var_143_split_sizes_0 = const()[name = string("op_143_split_sizes_0"), val = tensor<int32, [2]>([32, 32])];
int32 var_143_axis_0 = const()[name = string("op_143_axis_0"), val = int32(-2)];
tensor<fp16, [?, 16, 32, 3]> var_143_cast_fp16_0, tensor<fp16, [?, 16, 32, 3]> var_143_cast_fp16_1 = split(axis = var_143_axis_0, split_sizes = var_143_split_sizes_0, x = embed_1_cast_fp16)[name = string("op_143_cast_fp16")];
fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 16, 32, 3]> var_145_cast_fp16 = mul(x = var_143_cast_fp16_1, y = const_2_promoted_to_fp16)[name = string("op_145_cast_fp16")];
bool var_147_interleave_0 = const()[name = string("op_147_interleave_0"), val = bool(false)];
tensor<fp16, [?, 16, 64, 3]> var_147_cast_fp16 = concat(axis = var_86, interleave = var_147_interleave_0, values = (var_145_cast_fp16, var_143_cast_fp16_0))[name = string("op_147_cast_fp16")];
tensor<fp16, [64, 3]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [64, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2761664)))];
tensor<fp16, [?, 16, 64, 3]> var_148_cast_fp16 = mul(x = var_147_cast_fp16, y = sin_to_fp16)[name = string("op_148_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> query_states_3_cast_fp16 = add(x = var_142_cast_fp16, y = var_148_cast_fp16)[name = string("query_states_3_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> var_150_cast_fp16 = mul(x = embed_3_cast_fp16, y = cos_to_fp16)[name = string("op_150_cast_fp16")];
tensor<int32, [2]> var_151_split_sizes_0 = const()[name = string("op_151_split_sizes_0"), val = tensor<int32, [2]>([32, 32])];
int32 var_151_axis_0 = const()[name = string("op_151_axis_0"), val = int32(-2)];
tensor<fp16, [?, 2, 32, 3]> var_151_cast_fp16_0, tensor<fp16, [?, 2, 32, 3]> var_151_cast_fp16_1 = split(axis = var_151_axis_0, split_sizes = var_151_split_sizes_0, x = embed_3_cast_fp16)[name = string("op_151_cast_fp16")];
fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 2, 32, 3]> var_153_cast_fp16 = mul(x = var_151_cast_fp16_1, y = const_3_promoted_to_fp16)[name = string("op_153_cast_fp16")];
bool var_155_interleave_0 = const()[name = string("op_155_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2, 64, 3]> var_155_cast_fp16 = concat(axis = var_86, interleave = var_155_interleave_0, values = (var_153_cast_fp16, var_151_cast_fp16_0))[name = string("op_155_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> var_156_cast_fp16 = mul(x = var_155_cast_fp16, y = sin_to_fp16)[name = string("op_156_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> key_states_3_cast_fp16 = add(x = var_150_cast_fp16, y = var_156_cast_fp16)[name = string("key_states_3_cast_fp16")];
tensor<int32, [2]> var_161_split_sizes_0 = const()[name = string("op_161_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
int32 var_161_axis_0 = const()[name = string("op_161_axis_0"), val = int32(1)];
tensor<fp16, [?, 8, 64, 3]> var_161_cast_fp16_0, tensor<fp16, [?, 8, 64, 3]> var_161_cast_fp16_1 = split(axis = var_161_axis_0, split_sizes = var_161_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_161_cast_fp16")];
tensor<int32, [2]> var_163_split_sizes_0 = const()[name = string("op_163_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_163_axis_0 = const()[name = string("op_163_axis_0"), val = int32(1)];
tensor<fp16, [?, 1, 64, 3]> var_163_cast_fp16_0, tensor<fp16, [?, 1, 64, 3]> var_163_cast_fp16_1 = split(axis = var_163_axis_0, split_sizes = var_163_split_sizes_0, x = key_states_3_cast_fp16)[name = string("op_163_cast_fp16")];
tensor<int32, [2]> var_165_split_sizes_0 = const()[name = string("op_165_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_165_axis_0 = const()[name = string("op_165_axis_0"), val = int32(1)];
tensor<fp16, [?, 1, 64, 3]> var_165_cast_fp16_0, tensor<fp16, [?, 1, 64, 3]> var_165_cast_fp16_1 = split(axis = var_165_axis_0, split_sizes = var_165_split_sizes_0, x = value_states_3_cast_fp16)[name = string("op_165_cast_fp16")];
bool attn_weights_1_transpose_x_1 = const()[name = string("attn_weights_1_transpose_x_1"), val = bool(true)];
bool attn_weights_1_transpose_y_1 = const()[name = string("attn_weights_1_transpose_y_1"), val = bool(false)];
tensor<fp16, [?, 8, 3, 3]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_1, transpose_y = attn_weights_1_transpose_y_1, x = var_163_cast_fp16_0, y = var_161_cast_fp16_0)[name = string("attn_weights_1_cast_fp16")];
fp16 _inversed_attn_weights_3_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_3_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [?, 8, 3, 3]> _inversed_attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = _inversed_attn_weights_3_y_0_to_fp16)[name = string("_inversed_attn_weights_3_cast_fp16")];
tensor<fp16, [?, 8, 3, 3]> attn_weights_5_cast_fp16 = softmax(axis = var_95, x = _inversed_attn_weights_3_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
bool var_172_transpose_x_0 = const()[name = string("op_172_transpose_x_0"), val = bool(false)];
bool var_172_transpose_y_0 = const()[name = string("op_172_transpose_y_0"), val = bool(false)];
tensor<fp16, [?, 8, 64, 3]> var_172_cast_fp16 = matmul(transpose_x = var_172_transpose_x_0, transpose_y = var_172_transpose_y_0, x = var_165_cast_fp16_0, y = attn_weights_5_cast_fp16)[name = string("op_172_cast_fp16")];
bool attn_weights_7_transpose_x_1 = const()[name = string("attn_weights_7_transpose_x_1"), val = bool(true)];
bool attn_weights_7_transpose_y_1 = const()[name = string("attn_weights_7_transpose_y_1"), val = bool(false)];
tensor<fp16, [?, 8, 3, 3]> attn_weights_7_cast_fp16 = matmul(transpose_x = attn_weights_7_transpose_x_1, transpose_y = attn_weights_7_transpose_y_1, x = var_163_cast_fp16_1, y = var_161_cast_fp16_1)[name = string("attn_weights_7_cast_fp16")];
fp16 _inversed_attn_weights_9_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_9_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [?, 8, 3, 3]> _inversed_attn_weights_9_cast_fp16 = mul(x = attn_weights_7_cast_fp16, y = _inversed_attn_weights_9_y_0_to_fp16)[name = string("_inversed_attn_weights_9_cast_fp16")];
tensor<fp16, [?, 8, 3, 3]> attn_weights_11_cast_fp16 = softmax(axis = var_95, x = _inversed_attn_weights_9_cast_fp16)[name = string("attn_weights_11_cast_fp16")];
bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
tensor<fp16, [?, 8, 64, 3]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_165_cast_fp16_1, y = attn_weights_11_cast_fp16)[name = string("attn_output_1_cast_fp16")];
bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
tensor<fp16, [?, 16, 64, 3]> attn_output_3_cast_fp16 = concat(axis = var_90, interleave = attn_output_3_interleave_0, values = (var_172_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
tensor<int32, [4]> concat_3x = const()[name = string("concat_3x"), val = tensor<int32, [4]>([-1, 1024, 1, 3])];
tensor<fp16, [?, 1024, 1, 3]> x_11_cast_fp16 = reshape(shape = concat_3x, x = attn_output_3_cast_fp16)[name = string("x_11_cast_fp16")];
string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_89_to_fp16 = const()[name = string("op_89_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2762112)))];
tensor<fp16, [?, 1024, 1, 3]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = var_89_to_fp16, x = x_11_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
tensor<fp16, [?, 1024, 1, 3]> x_13_cast_fp16 = add(x = x_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = string("x_13_cast_fp16")];
fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_191_cast_fp16 = mul(x = x_13_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_191_cast_fp16")];
bool x_15_interleave_0 = const()[name = string("x_15_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_15_cast_fp16 = concat(axis = var_90, interleave = x_15_interleave_0, values = (x_13_cast_fp16, var_191_cast_fp16))[name = string("x_15_cast_fp16")];
tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_201_to_fp16 = const()[name = string("op_201_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_201_to_fp16, x = x_15_cast_fp16)[name = string("out_7_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4859328)))];
tensor<fp16, [?, 2048, 1, 3]> out_9_cast_fp16 = mul(x = out_7_cast_fp16, y = layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("out_9_cast_fp16")];
tensor<int32, [2]> var_207_split_sizes_0 = const()[name = string("op_207_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_207_axis_0 = const()[name = string("op_207_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_207_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_207_cast_fp16_1 = split(axis = var_207_axis_0, split_sizes = var_207_split_sizes_0, x = out_9_cast_fp16)[name = string("op_207_cast_fp16")];
string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
tensor<fp16, [4096, 1024, 1, 1]> var_76_to_fp16 = const()[name = string("op_76_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4863488)))];
tensor<fp16, [?, 4096, 1, 3]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = var_76_to_fp16, x = var_207_cast_fp16_0)[name = string("input_1_cast_fp16")];
tensor<fp16, [?, 4096, 1, 3]> var_215_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_215_cast_fp16")];
string var_220_pad_type_0 = const()[name = string("op_220_pad_type_0"), val = string("valid")];
tensor<int32, [2]> var_220_strides_0 = const()[name = string("op_220_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> var_220_pad_0 = const()[name = string("op_220_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> var_220_dilations_0 = const()[name = string("op_220_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_220_groups_0 = const()[name = string("op_220_groups_0"), val = int32(1)];
tensor<fp16, [4096, 1024, 1, 1]> var_77_to_fp16 = const()[name = string("op_77_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13252160)))];
tensor<fp16, [?, 4096, 1, 3]> var_220_cast_fp16 = conv(dilations = var_220_dilations_0, groups = var_220_groups_0, pad = var_220_pad_0, pad_type = var_220_pad_type_0, strides = var_220_strides_0, weight = var_77_to_fp16, x = var_207_cast_fp16_0)[name = string("op_220_cast_fp16")];
tensor<fp16, [?, 4096, 1, 3]> x_21_cast_fp16 = mul(x = var_215_cast_fp16, y = var_220_cast_fp16)[name = string("x_21_cast_fp16")];
string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")];
tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)];
tensor<fp16, [1024, 4096, 1, 1]> var_78_to_fp16 = const()[name = string("op_78_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21640832)))];
tensor<fp16, [?, 1024, 1, 3]> hidden_states_5_cast_fp16 = conv(dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = var_78_to_fp16, x = x_21_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
tensor<fp16, [?, 1024, 1, 3]> x_23_cast_fp16 = add(x = x_13_cast_fp16, y = hidden_states_5_cast_fp16)[name = string("x_23_cast_fp16")];
int32 var_238 = const()[name = string("op_238"), val = int32(-2)];
int32 var_242 = const()[name = string("op_242"), val = int32(1)];
int32 var_247 = const()[name = string("op_247"), val = int32(2)];
fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_252_cast_fp16 = mul(x = x_23_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_252_cast_fp16")];
bool x_25_interleave_0 = const()[name = string("x_25_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_25_cast_fp16 = concat(axis = var_242, interleave = x_25_interleave_0, values = (x_23_cast_fp16, var_252_cast_fp16))[name = string("x_25_cast_fp16")];
tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_262_to_fp16 = const()[name = string("op_262_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_262_to_fp16, x = x_25_cast_fp16)[name = string("out_13_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30029504)))];
tensor<fp16, [?, 2048, 1, 3]> out_15_cast_fp16 = mul(x = out_13_cast_fp16, y = layer_encoder_layers_1_input_layernorm_weight_to_fp16)[name = string("out_15_cast_fp16")];
tensor<int32, [2]> var_268_split_sizes_0 = const()[name = string("op_268_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_268_axis_0 = const()[name = string("op_268_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_268_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_268_cast_fp16_1 = split(axis = var_268_axis_0, split_sizes = var_268_split_sizes_0, x = out_15_cast_fp16)[name = string("op_268_cast_fp16")];
string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_233_to_fp16 = const()[name = string("op_233_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30033664)))];
tensor<fp16, [?, 1024, 1, 3]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = var_233_to_fp16, x = var_268_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
string key_states_7_pad_type_0 = const()[name = string("key_states_7_pad_type_0"), val = string("valid")];
tensor<int32, [2]> key_states_7_strides_0 = const()[name = string("key_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_states_7_pad_0 = const()[name = string("key_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_states_7_dilations_0 = const()[name = string("key_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 key_states_7_groups_0 = const()[name = string("key_states_7_groups_0"), val = int32(1)];
tensor<fp16, [128, 1024, 1, 1]> var_234_to_fp16 = const()[name = string("op_234_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32130880)))];
tensor<fp16, [?, 128, 1, 3]> key_states_7_cast_fp16 = conv(dilations = key_states_7_dilations_0, groups = key_states_7_groups_0, pad = key_states_7_pad_0, pad_type = key_states_7_pad_type_0, strides = key_states_7_strides_0, weight = var_234_to_fp16, x = var_268_cast_fp16_0)[name = string("key_states_7_cast_fp16")];
string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
tensor<fp16, [128, 1024, 1, 1]> var_235_to_fp16 = const()[name = string("op_235_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32393088)))];
tensor<fp16, [?, 128, 1, 3]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = var_235_to_fp16, x = var_268_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
tensor<int32, [4]> concat_4x = const()[name = string("concat_4x"), val = tensor<int32, [4]>([-1, 16, 64, 3])];
tensor<fp16, [?, 16, 64, 3]> embed_5_cast_fp16 = reshape(shape = concat_4x, x = query_states_7_cast_fp16)[name = string("embed_5_cast_fp16")];
tensor<int32, [4]> concat_5x = const()[name = string("concat_5x"), val = tensor<int32, [4]>([-1, 2, 64, 3])];
tensor<fp16, [?, 2, 64, 3]> embed_7_cast_fp16 = reshape(shape = concat_5x, x = key_states_7_cast_fp16)[name = string("embed_7_cast_fp16")];
tensor<int32, [4]> concat_6x = const()[name = string("concat_6x"), val = tensor<int32, [4]>([-1, 2, 64, 3])];
tensor<fp16, [?, 2, 64, 3]> value_states_9_cast_fp16 = reshape(shape = concat_6x, x = value_states_7_cast_fp16)[name = string("value_states_9_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> var_294_cast_fp16 = mul(x = embed_5_cast_fp16, y = cos_to_fp16)[name = string("op_294_cast_fp16")];
tensor<int32, [2]> var_295_split_sizes_0 = const()[name = string("op_295_split_sizes_0"), val = tensor<int32, [2]>([32, 32])];
int32 var_295_axis_0 = const()[name = string("op_295_axis_0"), val = int32(-2)];
tensor<fp16, [?, 16, 32, 3]> var_295_cast_fp16_0, tensor<fp16, [?, 16, 32, 3]> var_295_cast_fp16_1 = split(axis = var_295_axis_0, split_sizes = var_295_split_sizes_0, x = embed_5_cast_fp16)[name = string("op_295_cast_fp16")];
fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 16, 32, 3]> var_297_cast_fp16 = mul(x = var_295_cast_fp16_1, y = const_6_promoted_to_fp16)[name = string("op_297_cast_fp16")];
bool var_299_interleave_0 = const()[name = string("op_299_interleave_0"), val = bool(false)];
tensor<fp16, [?, 16, 64, 3]> var_299_cast_fp16 = concat(axis = var_238, interleave = var_299_interleave_0, values = (var_297_cast_fp16, var_295_cast_fp16_0))[name = string("op_299_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> var_300_cast_fp16 = mul(x = var_299_cast_fp16, y = sin_to_fp16)[name = string("op_300_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> query_states_9_cast_fp16 = add(x = var_294_cast_fp16, y = var_300_cast_fp16)[name = string("query_states_9_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> var_302_cast_fp16 = mul(x = embed_7_cast_fp16, y = cos_to_fp16)[name = string("op_302_cast_fp16")];
tensor<int32, [2]> var_303_split_sizes_0 = const()[name = string("op_303_split_sizes_0"), val = tensor<int32, [2]>([32, 32])];
int32 var_303_axis_0 = const()[name = string("op_303_axis_0"), val = int32(-2)];
tensor<fp16, [?, 2, 32, 3]> var_303_cast_fp16_0, tensor<fp16, [?, 2, 32, 3]> var_303_cast_fp16_1 = split(axis = var_303_axis_0, split_sizes = var_303_split_sizes_0, x = embed_7_cast_fp16)[name = string("op_303_cast_fp16")];
fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 2, 32, 3]> var_305_cast_fp16 = mul(x = var_303_cast_fp16_1, y = const_7_promoted_to_fp16)[name = string("op_305_cast_fp16")];
bool var_307_interleave_0 = const()[name = string("op_307_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2, 64, 3]> var_307_cast_fp16 = concat(axis = var_238, interleave = var_307_interleave_0, values = (var_305_cast_fp16, var_303_cast_fp16_0))[name = string("op_307_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> var_308_cast_fp16 = mul(x = var_307_cast_fp16, y = sin_to_fp16)[name = string("op_308_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> key_states_9_cast_fp16 = add(x = var_302_cast_fp16, y = var_308_cast_fp16)[name = string("key_states_9_cast_fp16")];
tensor<int32, [2]> var_313_split_sizes_0 = const()[name = string("op_313_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
int32 var_313_axis_0 = const()[name = string("op_313_axis_0"), val = int32(1)];
tensor<fp16, [?, 8, 64, 3]> var_313_cast_fp16_0, tensor<fp16, [?, 8, 64, 3]> var_313_cast_fp16_1 = split(axis = var_313_axis_0, split_sizes = var_313_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_313_cast_fp16")];
tensor<int32, [2]> var_315_split_sizes_0 = const()[name = string("op_315_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_315_axis_0 = const()[name = string("op_315_axis_0"), val = int32(1)];
tensor<fp16, [?, 1, 64, 3]> var_315_cast_fp16_0, tensor<fp16, [?, 1, 64, 3]> var_315_cast_fp16_1 = split(axis = var_315_axis_0, split_sizes = var_315_split_sizes_0, x = key_states_9_cast_fp16)[name = string("op_315_cast_fp16")];
tensor<int32, [2]> var_317_split_sizes_0 = const()[name = string("op_317_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_317_axis_0 = const()[name = string("op_317_axis_0"), val = int32(1)];
tensor<fp16, [?, 1, 64, 3]> var_317_cast_fp16_0, tensor<fp16, [?, 1, 64, 3]> var_317_cast_fp16_1 = split(axis = var_317_axis_0, split_sizes = var_317_split_sizes_0, x = value_states_9_cast_fp16)[name = string("op_317_cast_fp16")];
bool attn_weights_13_transpose_x_1 = const()[name = string("attn_weights_13_transpose_x_1"), val = bool(true)];
bool attn_weights_13_transpose_y_1 = const()[name = string("attn_weights_13_transpose_y_1"), val = bool(false)];
tensor<fp16, [?, 8, 3, 3]> attn_weights_13_cast_fp16 = matmul(transpose_x = attn_weights_13_transpose_x_1, transpose_y = attn_weights_13_transpose_y_1, x = var_315_cast_fp16_0, y = var_313_cast_fp16_0)[name = string("attn_weights_13_cast_fp16")];
fp16 _inversed_attn_weights_15_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_15_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [?, 8, 3, 3]> _inversed_attn_weights_15_cast_fp16 = mul(x = attn_weights_13_cast_fp16, y = _inversed_attn_weights_15_y_0_to_fp16)[name = string("_inversed_attn_weights_15_cast_fp16")];
tensor<fp16, [?, 8, 3, 3]> attn_weights_17_cast_fp16 = softmax(axis = var_247, x = _inversed_attn_weights_15_cast_fp16)[name = string("attn_weights_17_cast_fp16")];
bool var_324_transpose_x_0 = const()[name = string("op_324_transpose_x_0"), val = bool(false)];
bool var_324_transpose_y_0 = const()[name = string("op_324_transpose_y_0"), val = bool(false)];
tensor<fp16, [?, 8, 64, 3]> var_324_cast_fp16 = matmul(transpose_x = var_324_transpose_x_0, transpose_y = var_324_transpose_y_0, x = var_317_cast_fp16_0, y = attn_weights_17_cast_fp16)[name = string("op_324_cast_fp16")];
bool attn_weights_19_transpose_x_1 = const()[name = string("attn_weights_19_transpose_x_1"), val = bool(true)];
bool attn_weights_19_transpose_y_1 = const()[name = string("attn_weights_19_transpose_y_1"), val = bool(false)];
tensor<fp16, [?, 8, 3, 3]> attn_weights_19_cast_fp16 = matmul(transpose_x = attn_weights_19_transpose_x_1, transpose_y = attn_weights_19_transpose_y_1, x = var_315_cast_fp16_1, y = var_313_cast_fp16_1)[name = string("attn_weights_19_cast_fp16")];
fp16 _inversed_attn_weights_21_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_21_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [?, 8, 3, 3]> _inversed_attn_weights_21_cast_fp16 = mul(x = attn_weights_19_cast_fp16, y = _inversed_attn_weights_21_y_0_to_fp16)[name = string("_inversed_attn_weights_21_cast_fp16")];
tensor<fp16, [?, 8, 3, 3]> attn_weights_23_cast_fp16 = softmax(axis = var_247, x = _inversed_attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
tensor<fp16, [?, 8, 64, 3]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = var_317_cast_fp16_1, y = attn_weights_23_cast_fp16)[name = string("attn_output_5_cast_fp16")];
bool attn_output_7_interleave_0 = const()[name = string("attn_output_7_interleave_0"), val = bool(false)];
tensor<fp16, [?, 16, 64, 3]> attn_output_7_cast_fp16 = concat(axis = var_242, interleave = attn_output_7_interleave_0, values = (var_324_cast_fp16, attn_output_5_cast_fp16))[name = string("attn_output_7_cast_fp16")];
tensor<int32, [4]> concat_7x = const()[name = string("concat_7x"), val = tensor<int32, [4]>([-1, 1024, 1, 3])];
tensor<fp16, [?, 1024, 1, 3]> x_29_cast_fp16 = reshape(shape = concat_7x, x = attn_output_7_cast_fp16)[name = string("x_29_cast_fp16")];
string hidden_states_9_pad_type_0 = const()[name = string("hidden_states_9_pad_type_0"), val = string("valid")];
tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = string("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = string("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = string("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 hidden_states_9_groups_0 = const()[name = string("hidden_states_9_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_241_to_fp16 = const()[name = string("op_241_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32655296)))];
tensor<fp16, [?, 1024, 1, 3]> hidden_states_9_cast_fp16 = conv(dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = var_241_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
tensor<fp16, [?, 1024, 1, 3]> x_31_cast_fp16 = add(x = x_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = string("x_31_cast_fp16")];
fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_343_cast_fp16 = mul(x = x_31_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_343_cast_fp16")];
bool x_33_interleave_0 = const()[name = string("x_33_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_33_cast_fp16 = concat(axis = var_242, interleave = x_33_interleave_0, values = (x_31_cast_fp16, var_343_cast_fp16))[name = string("x_33_cast_fp16")];
tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_353_to_fp16 = const()[name = string("op_353_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_353_to_fp16, x = x_33_cast_fp16)[name = string("out_19_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34752512)))];
tensor<fp16, [?, 2048, 1, 3]> out_21_cast_fp16 = mul(x = out_19_cast_fp16, y = layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("out_21_cast_fp16")];
tensor<int32, [2]> var_359_split_sizes_0 = const()[name = string("op_359_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_359_axis_0 = const()[name = string("op_359_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_359_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_359_cast_fp16_1 = split(axis = var_359_axis_0, split_sizes = var_359_split_sizes_0, x = out_21_cast_fp16)[name = string("op_359_cast_fp16")];
string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
tensor<fp16, [4096, 1024, 1, 1]> var_228_to_fp16 = const()[name = string("op_228_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34756672)))];
tensor<fp16, [?, 4096, 1, 3]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = var_228_to_fp16, x = var_359_cast_fp16_0)[name = string("input_3_cast_fp16")];
tensor<fp16, [?, 4096, 1, 3]> var_367_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_367_cast_fp16")];
string var_372_pad_type_0 = const()[name = string("op_372_pad_type_0"), val = string("valid")];
tensor<int32, [2]> var_372_strides_0 = const()[name = string("op_372_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> var_372_pad_0 = const()[name = string("op_372_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> var_372_dilations_0 = const()[name = string("op_372_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_372_groups_0 = const()[name = string("op_372_groups_0"), val = int32(1)];
tensor<fp16, [4096, 1024, 1, 1]> var_229_to_fp16 = const()[name = string("op_229_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43145344)))];
tensor<fp16, [?, 4096, 1, 3]> var_372_cast_fp16 = conv(dilations = var_372_dilations_0, groups = var_372_groups_0, pad = var_372_pad_0, pad_type = var_372_pad_type_0, strides = var_372_strides_0, weight = var_229_to_fp16, x = var_359_cast_fp16_0)[name = string("op_372_cast_fp16")];
tensor<fp16, [?, 4096, 1, 3]> x_39_cast_fp16 = mul(x = var_367_cast_fp16, y = var_372_cast_fp16)[name = string("x_39_cast_fp16")];
string hidden_states_11_pad_type_0 = const()[name = string("hidden_states_11_pad_type_0"), val = string("valid")];
tensor<int32, [2]> hidden_states_11_strides_0 = const()[name = string("hidden_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = string("hidden_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_11_dilations_0 = const()[name = string("hidden_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 hidden_states_11_groups_0 = const()[name = string("hidden_states_11_groups_0"), val = int32(1)];
tensor<fp16, [1024, 4096, 1, 1]> var_230_to_fp16 = const()[name = string("op_230_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51534016)))];
tensor<fp16, [?, 1024, 1, 3]> hidden_states_11_cast_fp16 = conv(dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = var_230_to_fp16, x = x_39_cast_fp16)[name = string("hidden_states_11_cast_fp16")];
tensor<fp16, [?, 1024, 1, 3]> x_41_cast_fp16 = add(x = x_31_cast_fp16, y = hidden_states_11_cast_fp16)[name = string("x_41_cast_fp16")];
int32 var_390 = const()[name = string("op_390"), val = int32(-2)];
int32 var_394 = const()[name = string("op_394"), val = int32(1)];
int32 var_399 = const()[name = string("op_399"), val = int32(2)];
fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_404_cast_fp16 = mul(x = x_41_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_404_cast_fp16")];
bool x_43_interleave_0 = const()[name = string("x_43_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_43_cast_fp16 = concat(axis = var_394, interleave = x_43_interleave_0, values = (x_41_cast_fp16, var_404_cast_fp16))[name = string("x_43_cast_fp16")];
tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_414_to_fp16 = const()[name = string("op_414_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_414_to_fp16, x = x_43_cast_fp16)[name = string("out_25_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59922688)))];
tensor<fp16, [?, 2048, 1, 3]> out_27_cast_fp16 = mul(x = out_25_cast_fp16, y = layer_encoder_layers_2_input_layernorm_weight_to_fp16)[name = string("out_27_cast_fp16")];
tensor<int32, [2]> var_420_split_sizes_0 = const()[name = string("op_420_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_420_axis_0 = const()[name = string("op_420_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_420_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_420_cast_fp16_1 = split(axis = var_420_axis_0, split_sizes = var_420_split_sizes_0, x = out_27_cast_fp16)[name = string("op_420_cast_fp16")];
string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_385_to_fp16 = const()[name = string("op_385_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59926848)))];
tensor<fp16, [?, 1024, 1, 3]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = var_385_to_fp16, x = var_420_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
string key_states_13_pad_type_0 = const()[name = string("key_states_13_pad_type_0"), val = string("valid")];
tensor<int32, [2]> key_states_13_strides_0 = const()[name = string("key_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_states_13_pad_0 = const()[name = string("key_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_states_13_dilations_0 = const()[name = string("key_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 key_states_13_groups_0 = const()[name = string("key_states_13_groups_0"), val = int32(1)];
tensor<fp16, [128, 1024, 1, 1]> var_386_to_fp16 = const()[name = string("op_386_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62024064)))];
tensor<fp16, [?, 128, 1, 3]> key_states_13_cast_fp16 = conv(dilations = key_states_13_dilations_0, groups = key_states_13_groups_0, pad = key_states_13_pad_0, pad_type = key_states_13_pad_type_0, strides = key_states_13_strides_0, weight = var_386_to_fp16, x = var_420_cast_fp16_0)[name = string("key_states_13_cast_fp16")];
string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
tensor<fp16, [128, 1024, 1, 1]> var_387_to_fp16 = const()[name = string("op_387_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62286272)))];
tensor<fp16, [?, 128, 1, 3]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = var_387_to_fp16, x = var_420_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
tensor<int32, [4]> concat_8x = const()[name = string("concat_8x"), val = tensor<int32, [4]>([-1, 16, 64, 3])];
tensor<fp16, [?, 16, 64, 3]> embed_9_cast_fp16 = reshape(shape = concat_8x, x = query_states_13_cast_fp16)[name = string("embed_9_cast_fp16")];
tensor<int32, [4]> concat_9x = const()[name = string("concat_9x"), val = tensor<int32, [4]>([-1, 2, 64, 3])];
tensor<fp16, [?, 2, 64, 3]> embed_11_cast_fp16 = reshape(shape = concat_9x, x = key_states_13_cast_fp16)[name = string("embed_11_cast_fp16")];
tensor<int32, [4]> concat_10x = const()[name = string("concat_10x"), val = tensor<int32, [4]>([-1, 2, 64, 3])];
tensor<fp16, [?, 2, 64, 3]> value_states_15_cast_fp16 = reshape(shape = concat_10x, x = value_states_13_cast_fp16)[name = string("value_states_15_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> var_446_cast_fp16 = mul(x = embed_9_cast_fp16, y = cos_to_fp16)[name = string("op_446_cast_fp16")];
tensor<int32, [2]> var_447_split_sizes_0 = const()[name = string("op_447_split_sizes_0"), val = tensor<int32, [2]>([32, 32])];
int32 var_447_axis_0 = const()[name = string("op_447_axis_0"), val = int32(-2)];
tensor<fp16, [?, 16, 32, 3]> var_447_cast_fp16_0, tensor<fp16, [?, 16, 32, 3]> var_447_cast_fp16_1 = split(axis = var_447_axis_0, split_sizes = var_447_split_sizes_0, x = embed_9_cast_fp16)[name = string("op_447_cast_fp16")];
fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 16, 32, 3]> var_449_cast_fp16 = mul(x = var_447_cast_fp16_1, y = const_10_promoted_to_fp16)[name = string("op_449_cast_fp16")];
bool var_451_interleave_0 = const()[name = string("op_451_interleave_0"), val = bool(false)];
tensor<fp16, [?, 16, 64, 3]> var_451_cast_fp16 = concat(axis = var_390, interleave = var_451_interleave_0, values = (var_449_cast_fp16, var_447_cast_fp16_0))[name = string("op_451_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> var_452_cast_fp16 = mul(x = var_451_cast_fp16, y = sin_to_fp16)[name = string("op_452_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> query_states_15_cast_fp16 = add(x = var_446_cast_fp16, y = var_452_cast_fp16)[name = string("query_states_15_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> var_454_cast_fp16 = mul(x = embed_11_cast_fp16, y = cos_to_fp16)[name = string("op_454_cast_fp16")];
tensor<int32, [2]> var_455_split_sizes_0 = const()[name = string("op_455_split_sizes_0"), val = tensor<int32, [2]>([32, 32])];
int32 var_455_axis_0 = const()[name = string("op_455_axis_0"), val = int32(-2)];
tensor<fp16, [?, 2, 32, 3]> var_455_cast_fp16_0, tensor<fp16, [?, 2, 32, 3]> var_455_cast_fp16_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = embed_11_cast_fp16)[name = string("op_455_cast_fp16")];
fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 2, 32, 3]> var_457_cast_fp16 = mul(x = var_455_cast_fp16_1, y = const_11_promoted_to_fp16)[name = string("op_457_cast_fp16")];
bool var_459_interleave_0 = const()[name = string("op_459_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2, 64, 3]> var_459_cast_fp16 = concat(axis = var_390, interleave = var_459_interleave_0, values = (var_457_cast_fp16, var_455_cast_fp16_0))[name = string("op_459_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> var_460_cast_fp16 = mul(x = var_459_cast_fp16, y = sin_to_fp16)[name = string("op_460_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> key_states_15_cast_fp16 = add(x = var_454_cast_fp16, y = var_460_cast_fp16)[name = string("key_states_15_cast_fp16")];
tensor<int32, [2]> var_465_split_sizes_0 = const()[name = string("op_465_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
int32 var_465_axis_0 = const()[name = string("op_465_axis_0"), val = int32(1)];
tensor<fp16, [?, 8, 64, 3]> var_465_cast_fp16_0, tensor<fp16, [?, 8, 64, 3]> var_465_cast_fp16_1 = split(axis = var_465_axis_0, split_sizes = var_465_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_465_cast_fp16")];
tensor<int32, [2]> var_467_split_sizes_0 = const()[name = string("op_467_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_467_axis_0 = const()[name = string("op_467_axis_0"), val = int32(1)];
tensor<fp16, [?, 1, 64, 3]> var_467_cast_fp16_0, tensor<fp16, [?, 1, 64, 3]> var_467_cast_fp16_1 = split(axis = var_467_axis_0, split_sizes = var_467_split_sizes_0, x = key_states_15_cast_fp16)[name = string("op_467_cast_fp16")];
tensor<int32, [2]> var_469_split_sizes_0 = const()[name = string("op_469_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_469_axis_0 = const()[name = string("op_469_axis_0"), val = int32(1)];
tensor<fp16, [?, 1, 64, 3]> var_469_cast_fp16_0, tensor<fp16, [?, 1, 64, 3]> var_469_cast_fp16_1 = split(axis = var_469_axis_0, split_sizes = var_469_split_sizes_0, x = value_states_15_cast_fp16)[name = string("op_469_cast_fp16")];
bool attn_weights_25_transpose_x_1 = const()[name = string("attn_weights_25_transpose_x_1"), val = bool(true)];
bool attn_weights_25_transpose_y_1 = const()[name = string("attn_weights_25_transpose_y_1"), val = bool(false)];
tensor<fp16, [?, 8, 3, 3]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_1, transpose_y = attn_weights_25_transpose_y_1, x = var_467_cast_fp16_0, y = var_465_cast_fp16_0)[name = string("attn_weights_25_cast_fp16")];
fp16 _inversed_attn_weights_27_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_27_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [?, 8, 3, 3]> _inversed_attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = _inversed_attn_weights_27_y_0_to_fp16)[name = string("_inversed_attn_weights_27_cast_fp16")];
tensor<fp16, [?, 8, 3, 3]> attn_weights_29_cast_fp16 = softmax(axis = var_399, x = _inversed_attn_weights_27_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
bool var_476_transpose_x_0 = const()[name = string("op_476_transpose_x_0"), val = bool(false)];
bool var_476_transpose_y_0 = const()[name = string("op_476_transpose_y_0"), val = bool(false)];
tensor<fp16, [?, 8, 64, 3]> var_476_cast_fp16 = matmul(transpose_x = var_476_transpose_x_0, transpose_y = var_476_transpose_y_0, x = var_469_cast_fp16_0, y = attn_weights_29_cast_fp16)[name = string("op_476_cast_fp16")];
bool attn_weights_31_transpose_x_1 = const()[name = string("attn_weights_31_transpose_x_1"), val = bool(true)];
bool attn_weights_31_transpose_y_1 = const()[name = string("attn_weights_31_transpose_y_1"), val = bool(false)];
tensor<fp16, [?, 8, 3, 3]> attn_weights_31_cast_fp16 = matmul(transpose_x = attn_weights_31_transpose_x_1, transpose_y = attn_weights_31_transpose_y_1, x = var_467_cast_fp16_1, y = var_465_cast_fp16_1)[name = string("attn_weights_31_cast_fp16")];
fp16 _inversed_attn_weights_33_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_33_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [?, 8, 3, 3]> _inversed_attn_weights_33_cast_fp16 = mul(x = attn_weights_31_cast_fp16, y = _inversed_attn_weights_33_y_0_to_fp16)[name = string("_inversed_attn_weights_33_cast_fp16")];
tensor<fp16, [?, 8, 3, 3]> attn_weights_35_cast_fp16 = softmax(axis = var_399, x = _inversed_attn_weights_33_cast_fp16)[name = string("attn_weights_35_cast_fp16")];
bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
tensor<fp16, [?, 8, 64, 3]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = var_469_cast_fp16_1, y = attn_weights_35_cast_fp16)[name = string("attn_output_9_cast_fp16")];
bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
tensor<fp16, [?, 16, 64, 3]> attn_output_11_cast_fp16 = concat(axis = var_394, interleave = attn_output_11_interleave_0, values = (var_476_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([-1, 1024, 1, 3])];
tensor<fp16, [?, 1024, 1, 3]> x_47_cast_fp16 = reshape(shape = concat_11x, x = attn_output_11_cast_fp16)[name = string("x_47_cast_fp16")];
string hidden_states_15_pad_type_0 = const()[name = string("hidden_states_15_pad_type_0"), val = string("valid")];
tensor<int32, [2]> hidden_states_15_strides_0 = const()[name = string("hidden_states_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = string("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_15_dilations_0 = const()[name = string("hidden_states_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 hidden_states_15_groups_0 = const()[name = string("hidden_states_15_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_393_to_fp16 = const()[name = string("op_393_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62548480)))];
tensor<fp16, [?, 1024, 1, 3]> hidden_states_15_cast_fp16 = conv(dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = var_393_to_fp16, x = x_47_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
tensor<fp16, [?, 1024, 1, 3]> x_49_cast_fp16 = add(x = x_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = string("x_49_cast_fp16")];
fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_495_cast_fp16 = mul(x = x_49_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_495_cast_fp16")];
bool x_51_interleave_0 = const()[name = string("x_51_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_51_cast_fp16 = concat(axis = var_394, interleave = x_51_interleave_0, values = (x_49_cast_fp16, var_495_cast_fp16))[name = string("x_51_cast_fp16")];
tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_505_to_fp16 = const()[name = string("op_505_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_505_to_fp16, x = x_51_cast_fp16)[name = string("out_31_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64645696)))];
tensor<fp16, [?, 2048, 1, 3]> out_33_cast_fp16 = mul(x = out_31_cast_fp16, y = layer_encoder_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("out_33_cast_fp16")];
tensor<int32, [2]> var_511_split_sizes_0 = const()[name = string("op_511_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_511_axis_0 = const()[name = string("op_511_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_511_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_511_cast_fp16_1 = split(axis = var_511_axis_0, split_sizes = var_511_split_sizes_0, x = out_33_cast_fp16)[name = string("op_511_cast_fp16")];
string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
tensor<fp16, [4096, 1024, 1, 1]> var_380_to_fp16 = const()[name = string("op_380_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64649856)))];
tensor<fp16, [?, 4096, 1, 3]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = var_380_to_fp16, x = var_511_cast_fp16_0)[name = string("input_5_cast_fp16")];
tensor<fp16, [?, 4096, 1, 3]> var_519_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_519_cast_fp16")];
string var_524_pad_type_0 = const()[name = string("op_524_pad_type_0"), val = string("valid")];
tensor<int32, [2]> var_524_strides_0 = const()[name = string("op_524_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> var_524_pad_0 = const()[name = string("op_524_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> var_524_dilations_0 = const()[name = string("op_524_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_524_groups_0 = const()[name = string("op_524_groups_0"), val = int32(1)];
tensor<fp16, [4096, 1024, 1, 1]> var_381_to_fp16 = const()[name = string("op_381_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73038528)))];
tensor<fp16, [?, 4096, 1, 3]> var_524_cast_fp16 = conv(dilations = var_524_dilations_0, groups = var_524_groups_0, pad = var_524_pad_0, pad_type = var_524_pad_type_0, strides = var_524_strides_0, weight = var_381_to_fp16, x = var_511_cast_fp16_0)[name = string("op_524_cast_fp16")];
tensor<fp16, [?, 4096, 1, 3]> x_57_cast_fp16 = mul(x = var_519_cast_fp16, y = var_524_cast_fp16)[name = string("x_57_cast_fp16")];
string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
tensor<fp16, [1024, 4096, 1, 1]> var_382_to_fp16 = const()[name = string("op_382_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81427200)))];
tensor<fp16, [?, 1024, 1, 3]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = var_382_to_fp16, x = x_57_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
tensor<fp16, [?, 1024, 1, 3]> x_59_cast_fp16 = add(x = x_49_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("x_59_cast_fp16")];
int32 var_542 = const()[name = string("op_542"), val = int32(-2)];
int32 var_546 = const()[name = string("op_546"), val = int32(1)];
int32 var_551 = const()[name = string("op_551"), val = int32(2)];
fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_556_cast_fp16 = mul(x = x_59_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_556_cast_fp16")];
bool x_61_interleave_0 = const()[name = string("x_61_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_61_cast_fp16 = concat(axis = var_546, interleave = x_61_interleave_0, values = (x_59_cast_fp16, var_556_cast_fp16))[name = string("x_61_cast_fp16")];
tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_566_to_fp16 = const()[name = string("op_566_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_566_to_fp16, x = x_61_cast_fp16)[name = string("out_37_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89815872)))];
tensor<fp16, [?, 2048, 1, 3]> out_39_cast_fp16 = mul(x = out_37_cast_fp16, y = layer_encoder_layers_3_input_layernorm_weight_to_fp16)[name = string("out_39_cast_fp16")];
tensor<int32, [2]> var_572_split_sizes_0 = const()[name = string("op_572_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_572_axis_0 = const()[name = string("op_572_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_572_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_572_cast_fp16_1 = split(axis = var_572_axis_0, split_sizes = var_572_split_sizes_0, x = out_39_cast_fp16)[name = string("op_572_cast_fp16")];
string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_537_to_fp16 = const()[name = string("op_537_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89820032)))];
tensor<fp16, [?, 1024, 1, 3]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = var_537_to_fp16, x = var_572_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
string key_states_19_pad_type_0 = const()[name = string("key_states_19_pad_type_0"), val = string("valid")];
tensor<int32, [2]> key_states_19_strides_0 = const()[name = string("key_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_states_19_pad_0 = const()[name = string("key_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_states_19_dilations_0 = const()[name = string("key_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 key_states_19_groups_0 = const()[name = string("key_states_19_groups_0"), val = int32(1)];
tensor<fp16, [128, 1024, 1, 1]> var_538_to_fp16 = const()[name = string("op_538_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91917248)))];
tensor<fp16, [?, 128, 1, 3]> key_states_19_cast_fp16 = conv(dilations = key_states_19_dilations_0, groups = key_states_19_groups_0, pad = key_states_19_pad_0, pad_type = key_states_19_pad_type_0, strides = key_states_19_strides_0, weight = var_538_to_fp16, x = var_572_cast_fp16_0)[name = string("key_states_19_cast_fp16")];
string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
tensor<fp16, [128, 1024, 1, 1]> var_539_to_fp16 = const()[name = string("op_539_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92179456)))];
tensor<fp16, [?, 128, 1, 3]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = var_539_to_fp16, x = var_572_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([-1, 16, 64, 3])];
tensor<fp16, [?, 16, 64, 3]> embed_13_cast_fp16 = reshape(shape = concat_12x, x = query_states_19_cast_fp16)[name = string("embed_13_cast_fp16")];
tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([-1, 2, 64, 3])];
tensor<fp16, [?, 2, 64, 3]> embed_cast_fp16 = reshape(shape = concat_13x, x = key_states_19_cast_fp16)[name = string("embed_cast_fp16")];
tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([-1, 2, 64, 3])];
tensor<fp16, [?, 2, 64, 3]> value_states_21_cast_fp16 = reshape(shape = concat_14x, x = value_states_19_cast_fp16)[name = string("value_states_21_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> var_598_cast_fp16 = mul(x = embed_13_cast_fp16, y = cos_to_fp16)[name = string("op_598_cast_fp16")];
tensor<int32, [2]> var_599_split_sizes_0 = const()[name = string("op_599_split_sizes_0"), val = tensor<int32, [2]>([32, 32])];
int32 var_599_axis_0 = const()[name = string("op_599_axis_0"), val = int32(-2)];
tensor<fp16, [?, 16, 32, 3]> var_599_cast_fp16_0, tensor<fp16, [?, 16, 32, 3]> var_599_cast_fp16_1 = split(axis = var_599_axis_0, split_sizes = var_599_split_sizes_0, x = embed_13_cast_fp16)[name = string("op_599_cast_fp16")];
fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 16, 32, 3]> var_601_cast_fp16 = mul(x = var_599_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_601_cast_fp16")];
bool var_603_interleave_0 = const()[name = string("op_603_interleave_0"), val = bool(false)];
tensor<fp16, [?, 16, 64, 3]> var_603_cast_fp16 = concat(axis = var_542, interleave = var_603_interleave_0, values = (var_601_cast_fp16, var_599_cast_fp16_0))[name = string("op_603_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> var_604_cast_fp16 = mul(x = var_603_cast_fp16, y = sin_to_fp16)[name = string("op_604_cast_fp16")];
tensor<fp16, [?, 16, 64, 3]> query_states_21_cast_fp16 = add(x = var_598_cast_fp16, y = var_604_cast_fp16)[name = string("query_states_21_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> var_606_cast_fp16 = mul(x = embed_cast_fp16, y = cos_to_fp16)[name = string("op_606_cast_fp16")];
tensor<int32, [2]> var_607_split_sizes_0 = const()[name = string("op_607_split_sizes_0"), val = tensor<int32, [2]>([32, 32])];
int32 var_607_axis_0 = const()[name = string("op_607_axis_0"), val = int32(-2)];
tensor<fp16, [?, 2, 32, 3]> var_607_cast_fp16_0, tensor<fp16, [?, 2, 32, 3]> var_607_cast_fp16_1 = split(axis = var_607_axis_0, split_sizes = var_607_split_sizes_0, x = embed_cast_fp16)[name = string("op_607_cast_fp16")];
fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 2, 32, 3]> var_609_cast_fp16 = mul(x = var_607_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_609_cast_fp16")];
bool var_611_interleave_0 = const()[name = string("op_611_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2, 64, 3]> var_611_cast_fp16 = concat(axis = var_542, interleave = var_611_interleave_0, values = (var_609_cast_fp16, var_607_cast_fp16_0))[name = string("op_611_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> var_612_cast_fp16 = mul(x = var_611_cast_fp16, y = sin_to_fp16)[name = string("op_612_cast_fp16")];
tensor<fp16, [?, 2, 64, 3]> key_states_21_cast_fp16 = add(x = var_606_cast_fp16, y = var_612_cast_fp16)[name = string("key_states_21_cast_fp16")];
tensor<int32, [2]> var_617_split_sizes_0 = const()[name = string("op_617_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
int32 var_617_axis_0 = const()[name = string("op_617_axis_0"), val = int32(1)];
tensor<fp16, [?, 8, 64, 3]> var_617_cast_fp16_0, tensor<fp16, [?, 8, 64, 3]> var_617_cast_fp16_1 = split(axis = var_617_axis_0, split_sizes = var_617_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_617_cast_fp16")];
tensor<int32, [2]> var_619_split_sizes_0 = const()[name = string("op_619_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_619_axis_0 = const()[name = string("op_619_axis_0"), val = int32(1)];
tensor<fp16, [?, 1, 64, 3]> var_619_cast_fp16_0, tensor<fp16, [?, 1, 64, 3]> var_619_cast_fp16_1 = split(axis = var_619_axis_0, split_sizes = var_619_split_sizes_0, x = key_states_21_cast_fp16)[name = string("op_619_cast_fp16")];
tensor<int32, [2]> var_621_split_sizes_0 = const()[name = string("op_621_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_621_axis_0 = const()[name = string("op_621_axis_0"), val = int32(1)];
tensor<fp16, [?, 1, 64, 3]> var_621_cast_fp16_0, tensor<fp16, [?, 1, 64, 3]> var_621_cast_fp16_1 = split(axis = var_621_axis_0, split_sizes = var_621_split_sizes_0, x = value_states_21_cast_fp16)[name = string("op_621_cast_fp16")];
bool attn_weights_37_transpose_x_1 = const()[name = string("attn_weights_37_transpose_x_1"), val = bool(true)];
bool attn_weights_37_transpose_y_1 = const()[name = string("attn_weights_37_transpose_y_1"), val = bool(false)];
tensor<fp16, [?, 8, 3, 3]> attn_weights_37_cast_fp16 = matmul(transpose_x = attn_weights_37_transpose_x_1, transpose_y = attn_weights_37_transpose_y_1, x = var_619_cast_fp16_0, y = var_617_cast_fp16_0)[name = string("attn_weights_37_cast_fp16")];
fp16 _inversed_attn_weights_39_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_39_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [?, 8, 3, 3]> _inversed_attn_weights_39_cast_fp16 = mul(x = attn_weights_37_cast_fp16, y = _inversed_attn_weights_39_y_0_to_fp16)[name = string("_inversed_attn_weights_39_cast_fp16")];
tensor<fp16, [?, 8, 3, 3]> attn_weights_41_cast_fp16 = softmax(axis = var_551, x = _inversed_attn_weights_39_cast_fp16)[name = string("attn_weights_41_cast_fp16")];
bool var_628_transpose_x_0 = const()[name = string("op_628_transpose_x_0"), val = bool(false)];
bool var_628_transpose_y_0 = const()[name = string("op_628_transpose_y_0"), val = bool(false)];
tensor<fp16, [?, 8, 64, 3]> var_628_cast_fp16 = matmul(transpose_x = var_628_transpose_x_0, transpose_y = var_628_transpose_y_0, x = var_621_cast_fp16_0, y = attn_weights_41_cast_fp16)[name = string("op_628_cast_fp16")];
bool attn_weights_43_transpose_x_1 = const()[name = string("attn_weights_43_transpose_x_1"), val = bool(true)];
bool attn_weights_43_transpose_y_1 = const()[name = string("attn_weights_43_transpose_y_1"), val = bool(false)];
tensor<fp16, [?, 8, 3, 3]> attn_weights_43_cast_fp16 = matmul(transpose_x = attn_weights_43_transpose_x_1, transpose_y = attn_weights_43_transpose_y_1, x = var_619_cast_fp16_1, y = var_617_cast_fp16_1)[name = string("attn_weights_43_cast_fp16")];
fp16 _inversed_attn_weights_45_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_45_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [?, 8, 3, 3]> _inversed_attn_weights_45_cast_fp16 = mul(x = attn_weights_43_cast_fp16, y = _inversed_attn_weights_45_y_0_to_fp16)[name = string("_inversed_attn_weights_45_cast_fp16")];
tensor<fp16, [?, 8, 3, 3]> attn_weights_cast_fp16 = softmax(axis = var_551, x = _inversed_attn_weights_45_cast_fp16)[name = string("attn_weights_cast_fp16")];
bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
tensor<fp16, [?, 8, 64, 3]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = var_621_cast_fp16_1, y = attn_weights_cast_fp16)[name = string("attn_output_13_cast_fp16")];
bool attn_output_interleave_0 = const()[name = string("attn_output_interleave_0"), val = bool(false)];
tensor<fp16, [?, 16, 64, 3]> attn_output_cast_fp16 = concat(axis = var_546, interleave = attn_output_interleave_0, values = (var_628_cast_fp16, attn_output_13_cast_fp16))[name = string("attn_output_cast_fp16")];
tensor<int32, [4]> concat_15x = const()[name = string("concat_15x"), val = tensor<int32, [4]>([-1, 1024, 1, 3])];
tensor<fp16, [?, 1024, 1, 3]> x_65_cast_fp16 = reshape(shape = concat_15x, x = attn_output_cast_fp16)[name = string("x_65_cast_fp16")];
string hidden_states_21_pad_type_0 = const()[name = string("hidden_states_21_pad_type_0"), val = string("valid")];
tensor<int32, [2]> hidden_states_21_strides_0 = const()[name = string("hidden_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_21_pad_0 = const()[name = string("hidden_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_21_dilations_0 = const()[name = string("hidden_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 hidden_states_21_groups_0 = const()[name = string("hidden_states_21_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_545_to_fp16 = const()[name = string("op_545_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92441664)))];
tensor<fp16, [?, 1024, 1, 3]> hidden_states_21_cast_fp16 = conv(dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = var_545_to_fp16, x = x_65_cast_fp16)[name = string("hidden_states_21_cast_fp16")];
tensor<fp16, [?, 1024, 1, 3]> x_67_cast_fp16 = add(x = x_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = string("x_67_cast_fp16")];
fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_647_cast_fp16 = mul(x = x_67_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_647_cast_fp16")];
bool x_69_interleave_0 = const()[name = string("x_69_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_69_cast_fp16 = concat(axis = var_546, interleave = x_69_interleave_0, values = (x_67_cast_fp16, var_647_cast_fp16))[name = string("x_69_cast_fp16")];
tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_657_to_fp16 = const()[name = string("op_657_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_657_to_fp16, x = x_69_cast_fp16)[name = string("out_43_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94538880)))];
tensor<fp16, [?, 2048, 1, 3]> out_45_cast_fp16 = mul(x = out_43_cast_fp16, y = layer_encoder_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("out_45_cast_fp16")];
tensor<int32, [2]> var_663_split_sizes_0 = const()[name = string("op_663_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_663_axis_0 = const()[name = string("op_663_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_663_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_663_cast_fp16_1 = split(axis = var_663_axis_0, split_sizes = var_663_split_sizes_0, x = out_45_cast_fp16)[name = string("op_663_cast_fp16")];
string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
tensor<fp16, [4096, 1024, 1, 1]> var_532_to_fp16 = const()[name = string("op_532_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94543040)))];
tensor<fp16, [?, 4096, 1, 3]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = var_532_to_fp16, x = var_663_cast_fp16_0)[name = string("input_cast_fp16")];
tensor<fp16, [?, 4096, 1, 3]> var_671_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_671_cast_fp16")];
string var_676_pad_type_0 = const()[name = string("op_676_pad_type_0"), val = string("valid")];
tensor<int32, [2]> var_676_strides_0 = const()[name = string("op_676_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> var_676_pad_0 = const()[name = string("op_676_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> var_676_dilations_0 = const()[name = string("op_676_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_676_groups_0 = const()[name = string("op_676_groups_0"), val = int32(1)];
tensor<fp16, [4096, 1024, 1, 1]> var_533_to_fp16 = const()[name = string("op_533_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102931712)))];
tensor<fp16, [?, 4096, 1, 3]> var_676_cast_fp16 = conv(dilations = var_676_dilations_0, groups = var_676_groups_0, pad = var_676_pad_0, pad_type = var_676_pad_type_0, strides = var_676_strides_0, weight = var_533_to_fp16, x = var_663_cast_fp16_0)[name = string("op_676_cast_fp16")];
tensor<fp16, [?, 4096, 1, 3]> x_75_cast_fp16 = mul(x = var_671_cast_fp16, y = var_676_cast_fp16)[name = string("x_75_cast_fp16")];
string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
tensor<fp16, [1024, 4096, 1, 1]> var_534_to_fp16 = const()[name = string("op_534_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111320384)))];
tensor<fp16, [?, 1024, 1, 3]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = var_534_to_fp16, x = x_75_cast_fp16)[name = string("hidden_states_cast_fp16")];
tensor<fp16, [?, 1024, 1, 3]> x_77_cast_fp16 = add(x = x_67_cast_fp16, y = hidden_states_cast_fp16)[name = string("x_77_cast_fp16")];
int32 var_688 = const()[name = string("op_688"), val = int32(1)];
fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
tensor<fp16, [?, 1024, 1, 3]> var_691_cast_fp16 = mul(x = x_77_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_691_cast_fp16")];
bool x_79_interleave_0 = const()[name = string("x_79_interleave_0"), val = bool(false)];
tensor<fp16, [?, 2048, 1, 3]> x_79_cast_fp16 = concat(axis = var_688, interleave = x_79_interleave_0, values = (x_77_cast_fp16, var_691_cast_fp16))[name = string("x_79_cast_fp16")];
tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
fp16 var_701_to_fp16 = const()[name = string("op_701_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [?, 2048, 1, 3]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_701_to_fp16, x = x_79_cast_fp16)[name = string("out_49_cast_fp16")];
tensor<fp16, [1, 2048, 1, 1]> layer_encoder_norm_weight_to_fp16 = const()[name = string("layer_encoder_norm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119709056)))];
tensor<fp16, [?, 2048, 1, 3]> out_51_cast_fp16 = mul(x = out_49_cast_fp16, y = layer_encoder_norm_weight_to_fp16)[name = string("out_51_cast_fp16")];
tensor<int32, [2]> var_707_split_sizes_0 = const()[name = string("op_707_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
int32 var_707_axis_0 = const()[name = string("op_707_axis_0"), val = int32(1)];
tensor<fp16, [?, 1024, 1, 3]> var_707_cast_fp16_0, tensor<fp16, [?, 1024, 1, 3]> var_707_cast_fp16_1 = split(axis = var_707_axis_0, split_sizes = var_707_split_sizes_0, x = out_51_cast_fp16)[name = string("op_707_cast_fp16")];
tensor<int32, [4]> x_begin_0 = const()[name = string("x_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> x_end_0 = const()[name = string("x_end_0"), val = tensor<int32, [4]>([0, 1024, 1, 1])];
tensor<bool, [4]> x_end_mask_0 = const()[name = string("x_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [?, 1024, 1, 1]> x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, x = var_707_cast_fp16_0)[name = string("x_cast_fp16")];
string var_725_pad_type_0 = const()[name = string("op_725_pad_type_0"), val = string("valid")];
tensor<int32, [2]> var_725_strides_0 = const()[name = string("op_725_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> var_725_pad_0 = const()[name = string("op_725_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> var_725_dilations_0 = const()[name = string("op_725_dilations_0"), val = tensor<int32, [2]>([1, 1])];
int32 var_725_groups_0 = const()[name = string("op_725_groups_0"), val = int32(1)];
tensor<fp16, [1024, 1024, 1, 1]> var_719_to_fp16 = const()[name = string("op_719_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119713216)))];
tensor<fp16, [1024]> enc_to_lm_proj_bias_to_fp16 = const()[name = string("enc_to_lm_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121810432)))];
tensor<fp16, [?, 1024, 1, 1]> output = conv(bias = enc_to_lm_proj_bias_to_fp16, dilations = var_725_dilations_0, groups = var_725_groups_0, pad = var_725_pad_0, pad_type = var_725_pad_type_0, strides = var_725_strides_0, weight = var_719_to_fp16, x = x_cast_fp16)[name = string("op_725_cast_fp16")];
} -> (output);
} |