File size: 141,035 Bytes
03ef4b6 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 | program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}})]
{
func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, tensor<fp16, [1, 384, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 1536, 1, 448]> key_cache, tensor<fp16, [1, 448]> kv_cache_update_mask, tensor<fp16, [1, 1536, 1, 448]> value_cache) {
tensor<int32, []> var_24_axis_0 = const()[name = tensor<string, []>("op_24_axis_0"), val = tensor<int32, []>(0)];
tensor<int32, []> var_24_batch_dims_0 = const()[name = tensor<string, []>("op_24_batch_dims_0"), val = tensor<int32, []>(0)];
tensor<fp16, [51865, 384]> embed_tokens_weight_to_fp16 = const()[name = tensor<string, []>("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51865, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
tensor<fp16, [1, 384]> var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_24_cast_fp16")];
tensor<int32, []> var_28_axis_0 = const()[name = tensor<string, []>("op_28_axis_0"), val = tensor<int32, []>(0)];
tensor<int32, []> var_28_batch_dims_0 = const()[name = tensor<string, []>("op_28_batch_dims_0"), val = tensor<int32, []>(0)];
tensor<fp16, [448, 384]> embed_positions_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39832448)))];
tensor<fp16, [1, 384]> var_28_cast_fp16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
tensor<fp16, [1, 384]> hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_28_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([2])];
tensor<fp16, [1, 384, 1]> var_42_cast_fp16 = expand_dims(axes = var_42_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_42_cast_fp16")];
tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])];
tensor<fp16, [1, 384, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_42_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
tensor<int32, [4]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [4]>([384, 384, 384, 384])];
tensor<int32, []> var_47_axis_0 = const()[name = tensor<string, []>("op_47_axis_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1, 384, 1, 448]> var_47_cast_fp16_0, tensor<fp16, [1, 384, 1, 448]> var_47_cast_fp16_1, tensor<fp16, [1, 384, 1, 448]> var_47_cast_fp16_2, tensor<fp16, [1, 384, 1, 448]> var_47_cast_fp16_3 = split(axis = var_47_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_47_cast_fp16")];
tensor<int32, [4]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [4]>([384, 384, 384, 384])];
tensor<int32, []> var_54_axis_0 = const()[name = tensor<string, []>("op_54_axis_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1, 384, 1, 448]> var_54_cast_fp16_0, tensor<fp16, [1, 384, 1, 448]> var_54_cast_fp16_1, tensor<fp16, [1, 384, 1, 448]> var_54_cast_fp16_2, tensor<fp16, [1, 384, 1, 448]> var_54_cast_fp16_3 = split(axis = var_54_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_54_cast_fp16")];
tensor<int32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_89_to_fp16 = const()[name = tensor<string, []>("op_89_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_89_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
tensor<fp16, [384]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40176576)))];
tensor<fp16, [384]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40177408)))];
tensor<fp16, [384]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40178240)))];
tensor<fp16, [384]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40179072)))];
tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_1_strides_0 = const()[name = tensor<string, []>("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_1_dilations_0 = const()[name = tensor<string, []>("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_1_groups_0 = const()[name = tensor<string, []>("query_1_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40179904))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40327424))), name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40328000)))];
tensor<fp16, [1, 384, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
tensor<string, []> current_key_1_pad_type_0 = const()[name = tensor<string, []>("current_key_1_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> current_key_1_strides_0 = const()[name = tensor<string, []>("current_key_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> current_key_1_pad_0 = const()[name = tensor<string, []>("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> current_key_1_dilations_0 = const()[name = tensor<string, []>("current_key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> current_key_1_groups_0 = const()[name = tensor<string, []>("current_key_1_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40328832))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40476352))), name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")];
tensor<string, []> current_value_1_pad_type_0 = const()[name = tensor<string, []>("current_value_1_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> current_value_1_strides_0 = const()[name = tensor<string, []>("current_value_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> current_value_1_pad_0 = const()[name = tensor<string, []>("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> current_value_1_dilations_0 = const()[name = tensor<string, []>("current_value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> current_value_1_groups_0 = const()[name = tensor<string, []>("current_value_1_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40476928))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40624448))), name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40625024)))];
tensor<fp16, [1, 384, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")];
tensor<int32, [1]> var_124_axes_0 = const()[name = tensor<string, []>("op_124_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, [1, 1, 448]> var_124_cast_fp16 = expand_dims(axes = var_124_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_124_cast_fp16")];
tensor<int32, [1]> var_125_axes_0 = const()[name = tensor<string, []>("op_125_axes_0"), val = tensor<int32, [1]>([2])];
tensor<fp16, [1, 1, 1, 448]> var_125_cast_fp16 = expand_dims(axes = var_125_axes_0, x = var_124_cast_fp16)[name = tensor<string, []>("op_125_cast_fp16")];
tensor<fp16, []> var_65_to_fp16 = const()[name = tensor<string, []>("op_65_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
tensor<fp16, [1, 1, 1, 448]> var_127_cast_fp16 = sub(x = var_65_to_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_127_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_128_cast_fp16 = mul(x = var_47_cast_fp16_0, y = var_127_cast_fp16)[name = tensor<string, []>("op_128_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_129_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_129_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> key_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_129_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_132_cast_fp16 = mul(x = var_54_cast_fp16_0, y = var_127_cast_fp16)[name = tensor<string, []>("op_132_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_133_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_133_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> value_1_cast_fp16 = add(x = var_132_cast_fp16, y = var_133_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
tensor<int32, [4]> var_137 = const()[name = tensor<string, []>("op_137"), val = tensor<int32, [4]>([1, 6, 64, 1])];
tensor<fp16, [1, 6, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_137, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")];
tensor<fp16, []> var_139_to_fp16 = const()[name = tensor<string, []>("op_139_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 6, 64, 1]> var_140_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_139_to_fp16)[name = tensor<string, []>("op_140_cast_fp16")];
tensor<int32, [4]> var_143 = const()[name = tensor<string, []>("op_143"), val = tensor<int32, [4]>([1, 6, 64, 448])];
tensor<fp16, [1, 6, 64, 448]> var_144_cast_fp16 = reshape(shape = var_143, x = key_1_cast_fp16)[name = tensor<string, []>("op_144_cast_fp16")];
tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_140_cast_fp16, y = var_144_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
tensor<int32, [1]> var_148_axes_0 = const()[name = tensor<string, []>("op_148_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, [1, 1, 448]> var_148_cast_fp16 = expand_dims(axes = var_148_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_148_cast_fp16")];
tensor<int32, [1]> var_149_axes_0 = const()[name = tensor<string, []>("op_149_axes_0"), val = tensor<int32, [1]>([2])];
tensor<fp16, [1, 1, 1, 448]> var_149_cast_fp16 = expand_dims(axes = var_149_axes_0, x = var_148_cast_fp16)[name = tensor<string, []>("op_149_cast_fp16")];
tensor<fp16, [1, 6, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_149_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
tensor<fp16, [1, 6, 1, 448]> var_152_cast_fp16 = softmax(axis = var_64, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_152_cast_fp16")];
tensor<int32, [4]> var_153 = const()[name = tensor<string, []>("op_153"), val = tensor<int32, [4]>([1, 6, 64, 448])];
tensor<fp16, [1, 6, 64, 448]> var_154_cast_fp16 = reshape(shape = var_153, x = value_1_cast_fp16)[name = tensor<string, []>("op_154_cast_fp16")];
tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 6, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_154_cast_fp16, y = var_152_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
tensor<int32, [4]> var_157 = const()[name = tensor<string, []>("op_157"), val = tensor<int32, [4]>([1, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> input_1_cast_fp16 = reshape(shape = var_157, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_7_strides_0 = const()[name = tensor<string, []>("obj_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_7_dilations_0 = const()[name = tensor<string, []>("obj_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_7_groups_0 = const()[name = tensor<string, []>("obj_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40625856))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40773376))), name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40773952)))];
tensor<fp16, [1, 384, 1, 1]> obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_179_to_fp16 = const()[name = tensor<string, []>("op_179_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_179_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
tensor<fp16, [384]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40774784)))];
tensor<fp16, [384]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40775616)))];
tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_3_strides_0 = const()[name = tensor<string, []>("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_3_dilations_0 = const()[name = tensor<string, []>("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_3_groups_0 = const()[name = tensor<string, []>("query_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40776448))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40923968))), name = tensor<string, []>("layers_0_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40924544)))];
tensor<fp16, [1, 384, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_3_strides_0 = const()[name = tensor<string, []>("key_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_3_dilations_0 = const()[name = tensor<string, []>("key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_3_groups_0 = const()[name = tensor<string, []>("key_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_0_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40925376))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41072896))), name = tensor<string, []>("layers_0_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1500]> key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_0_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_3_cast_fp16")];
tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_3_strides_0 = const()[name = tensor<string, []>("value_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_3_dilations_0 = const()[name = tensor<string, []>("value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_3_groups_0 = const()[name = tensor<string, []>("value_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_0_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41073472))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41220992))), name = tensor<string, []>("layers_0_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41221568)))];
tensor<fp16, [1, 384, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_0_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_3_cast_fp16")];
tensor<int32, [4]> var_215 = const()[name = tensor<string, []>("op_215"), val = tensor<int32, [4]>([1, 6, 64, 1])];
tensor<fp16, [1, 6, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_215, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")];
tensor<fp16, []> var_217_to_fp16 = const()[name = tensor<string, []>("op_217_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 6, 64, 1]> var_218_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_217_to_fp16)[name = tensor<string, []>("op_218_cast_fp16")];
tensor<int32, [4]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [4]>([1, 6, 64, 1500])];
tensor<fp16, [1, 6, 64, 1500]> var_222_cast_fp16 = reshape(shape = var_221, x = key_3_cast_fp16)[name = tensor<string, []>("op_222_cast_fp16")];
tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_218_cast_fp16, y = var_222_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
tensor<fp16, [1, 6, 1, 1500]> obj_13_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
tensor<int32, [4]> var_226 = const()[name = tensor<string, []>("op_226"), val = tensor<int32, [4]>([1, 6, 64, 1500])];
tensor<fp16, [1, 6, 64, 1500]> var_227_cast_fp16 = reshape(shape = var_226, x = value_3_cast_fp16)[name = tensor<string, []>("op_227_cast_fp16")];
tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 6, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_227_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
tensor<int32, [4]> var_230 = const()[name = tensor<string, []>("op_230"), val = tensor<int32, [4]>([1, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> input_3_cast_fp16 = reshape(shape = var_230, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_11_strides_0 = const()[name = tensor<string, []>("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_11_dilations_0 = const()[name = tensor<string, []>("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_11_groups_0 = const()[name = tensor<string, []>("obj_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41222400))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41369920))), name = tensor<string, []>("layers_0_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41370496)))];
tensor<fp16, [1, 384, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_248_to_fp16 = const()[name = tensor<string, []>("op_248_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_248_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
tensor<fp16, [384]> input_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_5_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41371328)))];
tensor<fp16, [384]> input_5_beta_0_to_fp16 = const()[name = tensor<string, []>("input_5_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41372160)))];
tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_7_strides_0 = const()[name = tensor<string, []>("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_7_dilations_0 = const()[name = tensor<string, []>("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_7_groups_0 = const()[name = tensor<string, []>("input_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1536, 384, 1, 1]> layers_0_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [589824]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41372992))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41962880))), name = tensor<string, []>("layers_0_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1536, 384, 1, 1])];
tensor<fp16, [1536]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41963456)))];
tensor<fp16, [1, 1536, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16_palettized, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 1536, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
tensor<string, []> hidden_states_3_pad_type_0 = const()[name = tensor<string, []>("hidden_states_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = tensor<string, []>("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = tensor<string, []>("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = tensor<string, []>("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_3_groups_0 = const()[name = tensor<string, []>("hidden_states_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 1536, 1, 1]> layers_0_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [589824]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41966592))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42556480))), name = tensor<string, []>("layers_0_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 1536, 1, 1])];
tensor<fp16, [384]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42557056)))];
tensor<fp16, [1, 384, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
tensor<int32, []> var_283 = const()[name = tensor<string, []>("op_283"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_308_to_fp16 = const()[name = tensor<string, []>("op_308_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_308_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
tensor<fp16, [384]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42557888)))];
tensor<fp16, [384]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42558720)))];
tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_5_strides_0 = const()[name = tensor<string, []>("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_5_dilations_0 = const()[name = tensor<string, []>("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_5_groups_0 = const()[name = tensor<string, []>("query_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42559552))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42707072))), name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42707648)))];
tensor<fp16, [1, 384, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
tensor<string, []> current_key_3_pad_type_0 = const()[name = tensor<string, []>("current_key_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> current_key_3_strides_0 = const()[name = tensor<string, []>("current_key_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> current_key_3_pad_0 = const()[name = tensor<string, []>("current_key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> current_key_3_dilations_0 = const()[name = tensor<string, []>("current_key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> current_key_3_groups_0 = const()[name = tensor<string, []>("current_key_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42708480))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42856000))), name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> current_key_3_cast_fp16 = conv(dilations = current_key_3_dilations_0, groups = current_key_3_groups_0, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = current_key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")];
tensor<string, []> current_value_3_pad_type_0 = const()[name = tensor<string, []>("current_value_3_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> current_value_3_strides_0 = const()[name = tensor<string, []>("current_value_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> current_value_3_pad_0 = const()[name = tensor<string, []>("current_value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> current_value_3_dilations_0 = const()[name = tensor<string, []>("current_value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> current_value_3_groups_0 = const()[name = tensor<string, []>("current_value_3_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42856576))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43004096))), name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43004672)))];
tensor<fp16, [1, 384, 1, 1]> current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_3_dilations_0, groups = current_value_3_groups_0, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = current_value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_347_cast_fp16 = mul(x = var_47_cast_fp16_1, y = var_127_cast_fp16)[name = tensor<string, []>("op_347_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_348_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_348_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> key_5_cast_fp16 = add(x = var_347_cast_fp16, y = var_348_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_351_cast_fp16 = mul(x = var_54_cast_fp16_1, y = var_127_cast_fp16)[name = tensor<string, []>("op_351_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_352_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_352_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> value_5_cast_fp16 = add(x = var_351_cast_fp16, y = var_352_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
tensor<int32, [4]> var_356 = const()[name = tensor<string, []>("op_356"), val = tensor<int32, [4]>([1, 6, 64, 1])];
tensor<fp16, [1, 6, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_356, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")];
tensor<fp16, []> var_358_to_fp16 = const()[name = tensor<string, []>("op_358_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 6, 64, 1]> var_359_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_358_to_fp16)[name = tensor<string, []>("op_359_cast_fp16")];
tensor<int32, [4]> var_362 = const()[name = tensor<string, []>("op_362"), val = tensor<int32, [4]>([1, 6, 64, 448])];
tensor<fp16, [1, 6, 64, 448]> var_363_cast_fp16 = reshape(shape = var_362, x = key_5_cast_fp16)[name = tensor<string, []>("op_363_cast_fp16")];
tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1, 448]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_359_cast_fp16, y = var_363_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
tensor<fp16, [1, 6, 1, 448]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_149_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
tensor<fp16, [1, 6, 1, 448]> var_371_cast_fp16 = softmax(axis = var_283, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_371_cast_fp16")];
tensor<int32, [4]> var_372 = const()[name = tensor<string, []>("op_372"), val = tensor<int32, [4]>([1, 6, 64, 448])];
tensor<fp16, [1, 6, 64, 448]> var_373_cast_fp16 = reshape(shape = var_372, x = value_5_cast_fp16)[name = tensor<string, []>("op_373_cast_fp16")];
tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 6, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_373_cast_fp16, y = var_371_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
tensor<int32, [4]> var_376 = const()[name = tensor<string, []>("op_376"), val = tensor<int32, [4]>([1, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> input_11_cast_fp16 = reshape(shape = var_376, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
tensor<string, []> obj_21_pad_type_0 = const()[name = tensor<string, []>("obj_21_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_21_strides_0 = const()[name = tensor<string, []>("obj_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_21_pad_0 = const()[name = tensor<string, []>("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_21_dilations_0 = const()[name = tensor<string, []>("obj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_21_groups_0 = const()[name = tensor<string, []>("obj_21_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43005504))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43153024))), name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43153600)))];
tensor<fp16, [1, 384, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_398_to_fp16 = const()[name = tensor<string, []>("op_398_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_398_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
tensor<fp16, [384]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43154432)))];
tensor<fp16, [384]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43155264)))];
tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
tensor<string, []> query_7_pad_type_0 = const()[name = tensor<string, []>("query_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_7_strides_0 = const()[name = tensor<string, []>("query_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_7_pad_0 = const()[name = tensor<string, []>("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_7_dilations_0 = const()[name = tensor<string, []>("query_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_7_groups_0 = const()[name = tensor<string, []>("query_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43156096))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43303616))), name = tensor<string, []>("layers_1_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43304192)))];
tensor<fp16, [1, 384, 1, 1]> query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_7_strides_0 = const()[name = tensor<string, []>("key_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_7_pad_0 = const()[name = tensor<string, []>("key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_7_dilations_0 = const()[name = tensor<string, []>("key_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_7_groups_0 = const()[name = tensor<string, []>("key_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_1_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43305024))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43452544))), name = tensor<string, []>("layers_1_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1500]> key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_1_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_7_cast_fp16")];
tensor<string, []> value_7_pad_type_0 = const()[name = tensor<string, []>("value_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_7_strides_0 = const()[name = tensor<string, []>("value_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_7_pad_0 = const()[name = tensor<string, []>("value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_7_dilations_0 = const()[name = tensor<string, []>("value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_7_groups_0 = const()[name = tensor<string, []>("value_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_1_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43453120))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43600640))), name = tensor<string, []>("layers_1_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43601216)))];
tensor<fp16, [1, 384, 1, 1500]> value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_1_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_7_cast_fp16")];
tensor<int32, [4]> var_434 = const()[name = tensor<string, []>("op_434"), val = tensor<int32, [4]>([1, 6, 64, 1])];
tensor<fp16, [1, 6, 64, 1]> mh_q_7_cast_fp16 = reshape(shape = var_434, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")];
tensor<fp16, []> var_436_to_fp16 = const()[name = tensor<string, []>("op_436_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 6, 64, 1]> var_437_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_436_to_fp16)[name = tensor<string, []>("op_437_cast_fp16")];
tensor<int32, [4]> var_440 = const()[name = tensor<string, []>("op_440"), val = tensor<int32, [4]>([1, 6, 64, 1500])];
tensor<fp16, [1, 6, 64, 1500]> var_441_cast_fp16 = reshape(shape = var_440, x = key_7_cast_fp16)[name = tensor<string, []>("op_441_cast_fp16")];
tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_437_cast_fp16, y = var_441_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
tensor<fp16, [1, 6, 1, 1500]> obj_27_cast_fp16 = softmax(axis = var_283, x = mh_w_11_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
tensor<int32, [4]> var_445 = const()[name = tensor<string, []>("op_445"), val = tensor<int32, [4]>([1, 6, 64, 1500])];
tensor<fp16, [1, 6, 64, 1500]> var_446_cast_fp16 = reshape(shape = var_445, x = value_7_cast_fp16)[name = tensor<string, []>("op_446_cast_fp16")];
tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 6, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_446_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
tensor<int32, [4]> var_449 = const()[name = tensor<string, []>("op_449"), val = tensor<int32, [4]>([1, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> input_13_cast_fp16 = reshape(shape = var_449, x = attn_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
tensor<string, []> obj_25_pad_type_0 = const()[name = tensor<string, []>("obj_25_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_25_strides_0 = const()[name = tensor<string, []>("obj_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_25_pad_0 = const()[name = tensor<string, []>("obj_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_25_dilations_0 = const()[name = tensor<string, []>("obj_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_25_groups_0 = const()[name = tensor<string, []>("obj_25_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43602048))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43749568))), name = tensor<string, []>("layers_1_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43750144)))];
tensor<fp16, [1, 384, 1, 1]> obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_467_to_fp16 = const()[name = tensor<string, []>("op_467_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_467_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
tensor<fp16, [384]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43750976)))];
tensor<fp16, [384]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43751808)))];
tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_17_strides_0 = const()[name = tensor<string, []>("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_17_dilations_0 = const()[name = tensor<string, []>("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_17_groups_0 = const()[name = tensor<string, []>("input_17_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1536, 384, 1, 1]> layers_1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [589824]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43752640))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44342528))), name = tensor<string, []>("layers_1_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1536, 384, 1, 1])];
tensor<fp16, [1536]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44343104)))];
tensor<fp16, [1, 1536, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 1536, 1, 1]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = tensor<string, []>("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = tensor<string, []>("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_5_groups_0 = const()[name = tensor<string, []>("hidden_states_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 1536, 1, 1]> layers_1_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [589824]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44346240))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44936128))), name = tensor<string, []>("layers_1_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 1536, 1, 1])];
tensor<fp16, [384]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44936704)))];
tensor<fp16, [1, 384, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
tensor<int32, []> var_502 = const()[name = tensor<string, []>("op_502"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_527_to_fp16 = const()[name = tensor<string, []>("op_527_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_527_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
tensor<fp16, [384]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44937536)))];
tensor<fp16, [384]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44938368)))];
tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_9_strides_0 = const()[name = tensor<string, []>("query_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_9_dilations_0 = const()[name = tensor<string, []>("query_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_9_groups_0 = const()[name = tensor<string, []>("query_9_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44939200))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45086720))), name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45087296)))];
tensor<fp16, [1, 384, 1, 1]> query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
tensor<string, []> current_key_5_pad_type_0 = const()[name = tensor<string, []>("current_key_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> current_key_5_strides_0 = const()[name = tensor<string, []>("current_key_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> current_key_5_pad_0 = const()[name = tensor<string, []>("current_key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> current_key_5_dilations_0 = const()[name = tensor<string, []>("current_key_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> current_key_5_groups_0 = const()[name = tensor<string, []>("current_key_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45088128))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45235648))), name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> current_key_5_cast_fp16 = conv(dilations = current_key_5_dilations_0, groups = current_key_5_groups_0, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = current_key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")];
tensor<string, []> current_value_5_pad_type_0 = const()[name = tensor<string, []>("current_value_5_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> current_value_5_strides_0 = const()[name = tensor<string, []>("current_value_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> current_value_5_pad_0 = const()[name = tensor<string, []>("current_value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> current_value_5_dilations_0 = const()[name = tensor<string, []>("current_value_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> current_value_5_groups_0 = const()[name = tensor<string, []>("current_value_5_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45236224))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45383744))), name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45384320)))];
tensor<fp16, [1, 384, 1, 1]> current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = current_value_5_dilations_0, groups = current_value_5_groups_0, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = current_value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_566_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_127_cast_fp16)[name = tensor<string, []>("op_566_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_567_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_567_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> key_9_cast_fp16 = add(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_570_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_127_cast_fp16)[name = tensor<string, []>("op_570_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_571_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_571_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> value_9_cast_fp16 = add(x = var_570_cast_fp16, y = var_571_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
tensor<int32, [4]> var_575 = const()[name = tensor<string, []>("op_575"), val = tensor<int32, [4]>([1, 6, 64, 1])];
tensor<fp16, [1, 6, 64, 1]> mh_q_9_cast_fp16 = reshape(shape = var_575, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")];
tensor<fp16, []> var_577_to_fp16 = const()[name = tensor<string, []>("op_577_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 6, 64, 1]> var_578_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_577_to_fp16)[name = tensor<string, []>("op_578_cast_fp16")];
tensor<int32, [4]> var_581 = const()[name = tensor<string, []>("op_581"), val = tensor<int32, [4]>([1, 6, 64, 448])];
tensor<fp16, [1, 6, 64, 448]> var_582_cast_fp16 = reshape(shape = var_581, x = key_9_cast_fp16)[name = tensor<string, []>("op_582_cast_fp16")];
tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_578_cast_fp16, y = var_582_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
tensor<fp16, [1, 6, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_149_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
tensor<fp16, [1, 6, 1, 448]> var_590_cast_fp16 = softmax(axis = var_502, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_590_cast_fp16")];
tensor<int32, [4]> var_591 = const()[name = tensor<string, []>("op_591"), val = tensor<int32, [4]>([1, 6, 64, 448])];
tensor<fp16, [1, 6, 64, 448]> var_592_cast_fp16 = reshape(shape = var_591, x = value_9_cast_fp16)[name = tensor<string, []>("op_592_cast_fp16")];
tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 6, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_592_cast_fp16, y = var_590_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
tensor<int32, [4]> var_595 = const()[name = tensor<string, []>("op_595"), val = tensor<int32, [4]>([1, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> input_21_cast_fp16 = reshape(shape = var_595, x = attn_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_35_strides_0 = const()[name = tensor<string, []>("obj_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_35_dilations_0 = const()[name = tensor<string, []>("obj_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_35_groups_0 = const()[name = tensor<string, []>("obj_35_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45385152))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45532672))), name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45533248)))];
tensor<fp16, [1, 384, 1, 1]> obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_617_to_fp16 = const()[name = tensor<string, []>("op_617_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_617_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
tensor<fp16, [384]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45534080)))];
tensor<fp16, [384]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45534912)))];
tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
tensor<string, []> query_11_pad_type_0 = const()[name = tensor<string, []>("query_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_11_strides_0 = const()[name = tensor<string, []>("query_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_11_pad_0 = const()[name = tensor<string, []>("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_11_dilations_0 = const()[name = tensor<string, []>("query_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_11_groups_0 = const()[name = tensor<string, []>("query_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_2_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45535744))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45683264))), name = tensor<string, []>("layers_2_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45683840)))];
tensor<fp16, [1, 384, 1, 1]> query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_2_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_11_strides_0 = const()[name = tensor<string, []>("key_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_11_dilations_0 = const()[name = tensor<string, []>("key_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_11_groups_0 = const()[name = tensor<string, []>("key_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_2_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45684672))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45832192))), name = tensor<string, []>("layers_2_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1500]> key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_2_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_11_cast_fp16")];
tensor<string, []> value_11_pad_type_0 = const()[name = tensor<string, []>("value_11_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_11_strides_0 = const()[name = tensor<string, []>("value_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_11_pad_0 = const()[name = tensor<string, []>("value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_11_dilations_0 = const()[name = tensor<string, []>("value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_11_groups_0 = const()[name = tensor<string, []>("value_11_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_2_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45832768))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45980288))), name = tensor<string, []>("layers_2_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45980864)))];
tensor<fp16, [1, 384, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_2_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_11_cast_fp16")];
tensor<int32, [4]> var_653 = const()[name = tensor<string, []>("op_653"), val = tensor<int32, [4]>([1, 6, 64, 1])];
tensor<fp16, [1, 6, 64, 1]> mh_q_11_cast_fp16 = reshape(shape = var_653, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")];
tensor<fp16, []> var_655_to_fp16 = const()[name = tensor<string, []>("op_655_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 6, 64, 1]> var_656_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_655_to_fp16)[name = tensor<string, []>("op_656_cast_fp16")];
tensor<int32, [4]> var_659 = const()[name = tensor<string, []>("op_659"), val = tensor<int32, [4]>([1, 6, 64, 1500])];
tensor<fp16, [1, 6, 64, 1500]> var_660_cast_fp16 = reshape(shape = var_659, x = key_11_cast_fp16)[name = tensor<string, []>("op_660_cast_fp16")];
tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_656_cast_fp16, y = var_660_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
tensor<fp16, [1, 6, 1, 1500]> obj_41_cast_fp16 = softmax(axis = var_502, x = mh_w_17_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
tensor<int32, [4]> var_664 = const()[name = tensor<string, []>("op_664"), val = tensor<int32, [4]>([1, 6, 64, 1500])];
tensor<fp16, [1, 6, 64, 1500]> var_665_cast_fp16 = reshape(shape = var_664, x = value_11_cast_fp16)[name = tensor<string, []>("op_665_cast_fp16")];
tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 6, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_665_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
tensor<int32, [4]> var_668 = const()[name = tensor<string, []>("op_668"), val = tensor<int32, [4]>([1, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> input_23_cast_fp16 = reshape(shape = var_668, x = attn_11_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
tensor<string, []> obj_39_pad_type_0 = const()[name = tensor<string, []>("obj_39_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_39_strides_0 = const()[name = tensor<string, []>("obj_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_39_pad_0 = const()[name = tensor<string, []>("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_39_dilations_0 = const()[name = tensor<string, []>("obj_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_39_groups_0 = const()[name = tensor<string, []>("obj_39_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_2_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45981696))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46129216))), name = tensor<string, []>("layers_2_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46129792)))];
tensor<fp16, [1, 384, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_2_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_689_to_fp16 = const()[name = tensor<string, []>("op_689_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_689_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
tensor<fp16, [384]> input_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_25_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46130624)))];
tensor<fp16, [384]> input_25_beta_0_to_fp16 = const()[name = tensor<string, []>("input_25_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46131456)))];
tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
tensor<string, []> input_27_pad_type_0 = const()[name = tensor<string, []>("input_27_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_27_strides_0 = const()[name = tensor<string, []>("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_27_pad_0 = const()[name = tensor<string, []>("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_27_dilations_0 = const()[name = tensor<string, []>("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_27_groups_0 = const()[name = tensor<string, []>("input_27_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1536, 384, 1, 1]> layers_2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [589824]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46132288))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46722176))), name = tensor<string, []>("layers_2_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1536, 384, 1, 1])];
tensor<fp16, [1536]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46722752)))];
tensor<fp16, [1, 1536, 1, 1]> input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_2_fc1_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 1536, 1, 1]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = tensor<string, []>("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = tensor<string, []>("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_7_groups_0 = const()[name = tensor<string, []>("hidden_states_7_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 1536, 1, 1]> layers_2_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [589824]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46725888))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47315776))), name = tensor<string, []>("layers_2_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 1536, 1, 1])];
tensor<fp16, [384]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47316352)))];
tensor<fp16, [1, 384, 1, 1]> hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_2_fc2_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
tensor<int32, []> var_725 = const()[name = tensor<string, []>("op_725"), val = tensor<int32, []>(3)];
tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_750_to_fp16 = const()[name = tensor<string, []>("op_750_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_750_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
tensor<fp16, [384]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47317184)))];
tensor<fp16, [384]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47318016)))];
tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_13_strides_0 = const()[name = tensor<string, []>("query_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_13_dilations_0 = const()[name = tensor<string, []>("query_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_13_groups_0 = const()[name = tensor<string, []>("query_13_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47318848))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47466368))), name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47466944)))];
tensor<fp16, [1, 384, 1, 1]> query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
tensor<string, []> current_key_pad_type_0 = const()[name = tensor<string, []>("current_key_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> current_key_strides_0 = const()[name = tensor<string, []>("current_key_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> current_key_pad_0 = const()[name = tensor<string, []>("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> current_key_dilations_0 = const()[name = tensor<string, []>("current_key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> current_key_groups_0 = const()[name = tensor<string, []>("current_key_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47467776))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47615296))), name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")];
tensor<string, []> current_value_pad_type_0 = const()[name = tensor<string, []>("current_value_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> current_value_strides_0 = const()[name = tensor<string, []>("current_value_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> current_value_pad_0 = const()[name = tensor<string, []>("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> current_value_dilations_0 = const()[name = tensor<string, []>("current_value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> current_value_groups_0 = const()[name = tensor<string, []>("current_value_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47615872))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47763392))), name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47763968)))];
tensor<fp16, [1, 384, 1, 1]> current_value_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_789_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_127_cast_fp16)[name = tensor<string, []>("op_789_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_790_cast_fp16 = mul(x = current_key_cast_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_790_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> key_13_cast_fp16 = add(x = var_789_cast_fp16, y = var_790_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_793_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_127_cast_fp16)[name = tensor<string, []>("op_793_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> var_794_cast_fp16 = mul(x = current_value_cast_fp16, y = var_125_cast_fp16)[name = tensor<string, []>("op_794_cast_fp16")];
tensor<fp16, [1, 384, 1, 448]> value_13_cast_fp16 = add(x = var_793_cast_fp16, y = var_794_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
tensor<int32, [4]> var_798 = const()[name = tensor<string, []>("op_798"), val = tensor<int32, [4]>([1, 6, 64, 1])];
tensor<fp16, [1, 6, 64, 1]> mh_q_13_cast_fp16 = reshape(shape = var_798, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")];
tensor<fp16, []> var_800_to_fp16 = const()[name = tensor<string, []>("op_800_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 6, 64, 1]> var_801_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_800_to_fp16)[name = tensor<string, []>("op_801_cast_fp16")];
tensor<int32, [4]> var_804 = const()[name = tensor<string, []>("op_804"), val = tensor<int32, [4]>([1, 6, 64, 448])];
tensor<fp16, [1, 6, 64, 448]> var_805_cast_fp16 = reshape(shape = var_804, x = key_13_cast_fp16)[name = tensor<string, []>("op_805_cast_fp16")];
tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_801_cast_fp16, y = var_805_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
tensor<fp16, [1, 6, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_149_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
tensor<fp16, [1, 6, 1, 448]> var_813_cast_fp16 = softmax(axis = var_725, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_813_cast_fp16")];
tensor<int32, [4]> var_814 = const()[name = tensor<string, []>("op_814"), val = tensor<int32, [4]>([1, 6, 64, 448])];
tensor<fp16, [1, 6, 64, 448]> var_815_cast_fp16 = reshape(shape = var_814, x = value_13_cast_fp16)[name = tensor<string, []>("op_815_cast_fp16")];
tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 6, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_815_cast_fp16, y = var_813_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
tensor<int32, [4]> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, [4]>([1, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> input_31_cast_fp16 = reshape(shape = var_818, x = attn_13_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
tensor<string, []> obj_49_pad_type_0 = const()[name = tensor<string, []>("obj_49_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_49_strides_0 = const()[name = tensor<string, []>("obj_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_49_pad_0 = const()[name = tensor<string, []>("obj_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_49_dilations_0 = const()[name = tensor<string, []>("obj_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_49_groups_0 = const()[name = tensor<string, []>("obj_49_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47764800))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47912320))), name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47912896)))];
tensor<fp16, [1, 384, 1, 1]> obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_840_to_fp16 = const()[name = tensor<string, []>("op_840_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_840_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
tensor<fp16, [384]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47913728)))];
tensor<fp16, [384]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47914560)))];
tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")];
tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> query_strides_0 = const()[name = tensor<string, []>("query_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> query_dilations_0 = const()[name = tensor<string, []>("query_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> query_groups_0 = const()[name = tensor<string, []>("query_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_3_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47915392))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48062912))), name = tensor<string, []>("layers_3_encoder_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48063488)))];
tensor<fp16, [1, 384, 1, 1]> query_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_3_encoder_attn_q_proj_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> key_strides_0 = const()[name = tensor<string, []>("key_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> key_dilations_0 = const()[name = tensor<string, []>("key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> key_groups_0 = const()[name = tensor<string, []>("key_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_3_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48064320))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48211840))), name = tensor<string, []>("layers_3_encoder_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1500]> key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_3_encoder_attn_k_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")];
tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> value_strides_0 = const()[name = tensor<string, []>("value_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> value_dilations_0 = const()[name = tensor<string, []>("value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> value_groups_0 = const()[name = tensor<string, []>("value_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_3_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48212416))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48359936))), name = tensor<string, []>("layers_3_encoder_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48360512)))];
tensor<fp16, [1, 384, 1, 1500]> value_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_3_encoder_attn_v_proj_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")];
tensor<int32, [4]> var_876 = const()[name = tensor<string, []>("op_876"), val = tensor<int32, [4]>([1, 6, 64, 1])];
tensor<fp16, [1, 6, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_876, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")];
tensor<fp16, []> var_878_to_fp16 = const()[name = tensor<string, []>("op_878_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
tensor<fp16, [1, 6, 64, 1]> var_879_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_878_to_fp16)[name = tensor<string, []>("op_879_cast_fp16")];
tensor<int32, [4]> var_882 = const()[name = tensor<string, []>("op_882"), val = tensor<int32, [4]>([1, 6, 64, 1500])];
tensor<fp16, [1, 6, 64, 1500]> var_883_cast_fp16 = reshape(shape = var_882, x = key_cast_fp16)[name = tensor<string, []>("op_883_cast_fp16")];
tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_879_cast_fp16, y = var_883_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
tensor<fp16, [1, 6, 1, 1500]> obj_55_cast_fp16 = softmax(axis = var_725, x = mh_w_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
tensor<int32, [4]> var_887 = const()[name = tensor<string, []>("op_887"), val = tensor<int32, [4]>([1, 6, 64, 1500])];
tensor<fp16, [1, 6, 64, 1500]> var_888_cast_fp16 = reshape(shape = var_887, x = value_cast_fp16)[name = tensor<string, []>("op_888_cast_fp16")];
tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
tensor<fp16, [1, 6, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_888_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
tensor<int32, [4]> var_891 = const()[name = tensor<string, []>("op_891"), val = tensor<int32, [4]>([1, 384, 1, 1])];
tensor<fp16, [1, 384, 1, 1]> input_33_cast_fp16 = reshape(shape = var_891, x = attn_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
tensor<string, []> obj_53_pad_type_0 = const()[name = tensor<string, []>("obj_53_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> obj_53_strides_0 = const()[name = tensor<string, []>("obj_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> obj_53_pad_0 = const()[name = tensor<string, []>("obj_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> obj_53_dilations_0 = const()[name = tensor<string, []>("obj_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> obj_53_groups_0 = const()[name = tensor<string, []>("obj_53_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 384, 1, 1]> layers_3_encoder_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [147456]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48361344))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48508864))), name = tensor<string, []>("layers_3_encoder_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 384, 1, 1])];
tensor<fp16, [384]> layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48509440)))];
tensor<fp16, [1, 384, 1, 1]> obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_3_encoder_attn_o_proj_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_912_to_fp16 = const()[name = tensor<string, []>("op_912_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_912_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
tensor<fp16, [384]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48510272)))];
tensor<fp16, [384]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48511104)))];
tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> input_37_strides_0 = const()[name = tensor<string, []>("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> input_37_dilations_0 = const()[name = tensor<string, []>("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> input_37_groups_0 = const()[name = tensor<string, []>("input_37_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [1536, 384, 1, 1]> layers_3_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [589824]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48511936))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49101824))), name = tensor<string, []>("layers_3_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1536, 384, 1, 1])];
tensor<fp16, [1536]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49102400)))];
tensor<fp16, [1, 1536, 1, 1]> input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_3_fc1_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
tensor<fp16, [1, 1536, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("valid")];
tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = tensor<string, []>("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = tensor<string, []>("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
tensor<int32, []> hidden_states_9_groups_0 = const()[name = tensor<string, []>("hidden_states_9_groups_0"), val = tensor<int32, []>(1)];
tensor<fp16, [384, 1536, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49105536)))];
tensor<fp16, [384]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50285248)))];
tensor<fp16, [1, 384, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
tensor<fp16, [1, 384, 1, 1]> inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, []> var_955_to_fp16 = const()[name = tensor<string, []>("op_955_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_955_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
tensor<fp16, [384]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50286080)))];
tensor<fp16, [384]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50286912)))];
tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
tensor<fp16, [1, 384, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
tensor<int32, [1]> var_966_axes_0 = const()[name = tensor<string, []>("op_966_axes_0"), val = tensor<int32, [1]>([2])];
tensor<fp16, [1, 384, 1]> var_966_cast_fp16 = squeeze(axes = var_966_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_966_cast_fp16")];
tensor<int32, [3]> var_969_perm_0 = const()[name = tensor<string, []>("op_969_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
tensor<fp16, [51865]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51865]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50287744)))];
tensor<fp16, [1, 1, 384]> var_969_cast_fp16 = transpose(perm = var_969_perm_0, x = var_966_cast_fp16)[name = tensor<string, []>("transpose_0")];
tensor<fp16, [1, 1, 51865]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_969_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
tensor<int32, []> var_973 = const()[name = tensor<string, []>("op_973"), val = tensor<int32, []>(1)];
tensor<bool, []> obj_59_interleave_0 = const()[name = tensor<string, []>("obj_59_interleave_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 1536, 1, 1]> key_cache_updates = concat(axis = var_973, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_59_cast_fp16")];
tensor<int32, []> var_976 = const()[name = tensor<string, []>("op_976"), val = tensor<int32, []>(1)];
tensor<bool, []> obj_61_interleave_0 = const()[name = tensor<string, []>("obj_61_interleave_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 1536, 1, 1]> value_cache_updates = concat(axis = var_976, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_61_cast_fp16")];
tensor<int32, [4]> var_987_begin_0 = const()[name = tensor<string, []>("op_987_begin_0"), val = tensor<int32, [4]>([0, 2, 0, 0])];
tensor<int32, [4]> var_987_end_0 = const()[name = tensor<string, []>("op_987_end_0"), val = tensor<int32, [4]>([1, 3, 1, 1500])];
tensor<bool, [4]> var_987_end_mask_0 = const()[name = tensor<string, []>("op_987_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
tensor<fp16, [1, 1, 1, 1500]> var_987_cast_fp16 = slice_by_index(begin = var_987_begin_0, end = var_987_end_0, end_mask = var_987_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_987_cast_fp16")];
tensor<int32, [4]> var_990_begin_0 = const()[name = tensor<string, []>("op_990_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> var_990_end_0 = const()[name = tensor<string, []>("op_990_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
tensor<bool, [4]> var_990_end_mask_0 = const()[name = tensor<string, []>("op_990_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
tensor<bool, [4]> var_990_squeeze_mask_0 = const()[name = tensor<string, []>("op_990_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
tensor<fp16, [1, 1, 1500]> var_990_cast_fp16 = slice_by_index(begin = var_990_begin_0, end = var_990_end_0, end_mask = var_990_end_mask_0, squeeze_mask = var_990_squeeze_mask_0, x = var_987_cast_fp16)[name = tensor<string, []>("op_990_cast_fp16")];
tensor<int32, [4]> var_1005_begin_0 = const()[name = tensor<string, []>("op_1005_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> var_1005_end_0 = const()[name = tensor<string, []>("op_1005_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
tensor<bool, [4]> var_1005_end_mask_0 = const()[name = tensor<string, []>("op_1005_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
tensor<fp16, [1, 1, 1, 1500]> var_1005_cast_fp16 = slice_by_index(begin = var_1005_begin_0, end = var_1005_end_0, end_mask = var_1005_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1005_cast_fp16")];
tensor<int32, [4]> var_1008_begin_0 = const()[name = tensor<string, []>("op_1008_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> var_1008_end_0 = const()[name = tensor<string, []>("op_1008_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
tensor<bool, [4]> var_1008_end_mask_0 = const()[name = tensor<string, []>("op_1008_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
tensor<bool, [4]> var_1008_squeeze_mask_0 = const()[name = tensor<string, []>("op_1008_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
tensor<fp16, [1, 1, 1500]> var_1008_cast_fp16 = slice_by_index(begin = var_1008_begin_0, end = var_1008_end_0, end_mask = var_1008_end_mask_0, squeeze_mask = var_1008_squeeze_mask_0, x = var_1005_cast_fp16)[name = tensor<string, []>("op_1008_cast_fp16")];
tensor<int32, [4]> var_1023_begin_0 = const()[name = tensor<string, []>("op_1023_begin_0"), val = tensor<int32, [4]>([0, 2, 0, 0])];
tensor<int32, [4]> var_1023_end_0 = const()[name = tensor<string, []>("op_1023_end_0"), val = tensor<int32, [4]>([1, 3, 1, 1500])];
tensor<bool, [4]> var_1023_end_mask_0 = const()[name = tensor<string, []>("op_1023_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
tensor<fp16, [1, 1, 1, 1500]> var_1023_cast_fp16 = slice_by_index(begin = var_1023_begin_0, end = var_1023_end_0, end_mask = var_1023_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1023_cast_fp16")];
tensor<int32, [4]> var_1026_begin_0 = const()[name = tensor<string, []>("op_1026_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> var_1026_end_0 = const()[name = tensor<string, []>("op_1026_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
tensor<bool, [4]> var_1026_end_mask_0 = const()[name = tensor<string, []>("op_1026_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
tensor<bool, [4]> var_1026_squeeze_mask_0 = const()[name = tensor<string, []>("op_1026_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
tensor<fp16, [1, 1, 1500]> var_1026_cast_fp16 = slice_by_index(begin = var_1026_begin_0, end = var_1026_end_0, end_mask = var_1026_end_mask_0, squeeze_mask = var_1026_squeeze_mask_0, x = var_1023_cast_fp16)[name = tensor<string, []>("op_1026_cast_fp16")];
tensor<int32, [4]> var_1041_begin_0 = const()[name = tensor<string, []>("op_1041_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])];
tensor<int32, [4]> var_1041_end_0 = const()[name = tensor<string, []>("op_1041_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1500])];
tensor<bool, [4]> var_1041_end_mask_0 = const()[name = tensor<string, []>("op_1041_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
tensor<fp16, [1, 1, 1, 1500]> var_1041_cast_fp16 = slice_by_index(begin = var_1041_begin_0, end = var_1041_end_0, end_mask = var_1041_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1041_cast_fp16")];
tensor<int32, [4]> var_1044_begin_0 = const()[name = tensor<string, []>("op_1044_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> var_1044_end_0 = const()[name = tensor<string, []>("op_1044_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
tensor<bool, [4]> var_1044_end_mask_0 = const()[name = tensor<string, []>("op_1044_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
tensor<bool, [4]> var_1044_squeeze_mask_0 = const()[name = tensor<string, []>("op_1044_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
tensor<fp16, [1, 1, 1500]> var_1044_cast_fp16 = slice_by_index(begin = var_1044_begin_0, end = var_1044_end_0, end_mask = var_1044_end_mask_0, squeeze_mask = var_1044_squeeze_mask_0, x = var_1041_cast_fp16)[name = tensor<string, []>("op_1044_cast_fp16")];
tensor<int32, [4]> var_1059_begin_0 = const()[name = tensor<string, []>("op_1059_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
tensor<int32, [4]> var_1059_end_0 = const()[name = tensor<string, []>("op_1059_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])];
tensor<bool, [4]> var_1059_end_mask_0 = const()[name = tensor<string, []>("op_1059_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
tensor<fp16, [1, 1, 1, 1500]> var_1059_cast_fp16 = slice_by_index(begin = var_1059_begin_0, end = var_1059_end_0, end_mask = var_1059_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1059_cast_fp16")];
tensor<int32, [4]> var_1062_begin_0 = const()[name = tensor<string, []>("op_1062_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> var_1062_end_0 = const()[name = tensor<string, []>("op_1062_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
tensor<bool, [4]> var_1062_end_mask_0 = const()[name = tensor<string, []>("op_1062_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
tensor<bool, [4]> var_1062_squeeze_mask_0 = const()[name = tensor<string, []>("op_1062_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
tensor<fp16, [1, 1, 1500]> var_1062_cast_fp16 = slice_by_index(begin = var_1062_begin_0, end = var_1062_end_0, end_mask = var_1062_end_mask_0, squeeze_mask = var_1062_squeeze_mask_0, x = var_1059_cast_fp16)[name = tensor<string, []>("op_1062_cast_fp16")];
tensor<int32, [4]> var_1077_begin_0 = const()[name = tensor<string, []>("op_1077_begin_0"), val = tensor<int32, [4]>([0, 5, 0, 0])];
tensor<int32, [4]> var_1077_end_0 = const()[name = tensor<string, []>("op_1077_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
tensor<bool, [4]> var_1077_end_mask_0 = const()[name = tensor<string, []>("op_1077_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 1, 1, 1500]> var_1077_cast_fp16 = slice_by_index(begin = var_1077_begin_0, end = var_1077_end_0, end_mask = var_1077_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1077_cast_fp16")];
tensor<int32, [4]> var_1080_begin_0 = const()[name = tensor<string, []>("op_1080_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> var_1080_end_0 = const()[name = tensor<string, []>("op_1080_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
tensor<bool, [4]> var_1080_end_mask_0 = const()[name = tensor<string, []>("op_1080_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
tensor<bool, [4]> var_1080_squeeze_mask_0 = const()[name = tensor<string, []>("op_1080_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
tensor<fp16, [1, 1, 1500]> var_1080_cast_fp16 = slice_by_index(begin = var_1080_begin_0, end = var_1080_end_0, end_mask = var_1080_end_mask_0, squeeze_mask = var_1080_squeeze_mask_0, x = var_1077_cast_fp16)[name = tensor<string, []>("op_1080_cast_fp16")];
tensor<int32, []> var_1087 = const()[name = tensor<string, []>("op_1087"), val = tensor<int32, []>(1)];
tensor<bool, []> var_1088_interleave_0 = const()[name = tensor<string, []>("op_1088_interleave_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 6, 1500]> var_1088_cast_fp16 = concat(axis = var_1087, interleave = var_1088_interleave_0, values = (var_990_cast_fp16, var_1008_cast_fp16, var_1026_cast_fp16, var_1044_cast_fp16, var_1062_cast_fp16, var_1080_cast_fp16))[name = tensor<string, []>("op_1088_cast_fp16")];
tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([1])];
tensor<bool, []> obj_keep_dims_0 = const()[name = tensor<string, []>("obj_keep_dims_0"), val = tensor<bool, []>(false)];
tensor<fp16, [1, 1500]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = obj_keep_dims_0, x = var_1088_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
} -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights);
} |