Spaces:
Running
Running
File size: 37,856 Bytes
31c7d49 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 | #!/usr/bin/env python3
"""Add tiny block-boundary outputs to a DiT graph and record CPU ORT references."""
from __future__ import annotations
import argparse
import hashlib
import json
import os
from pathlib import Path
from typing import Any
INPUT_SHAPES = {
"latent": (1, 8192, 16),
"camera": (1, 1, 5),
"t": (1,),
"feature1": (1, 4101, 1280),
"feature2": (1, 4101, 128),
}
FIXTURE_FILES = (*INPUT_SHAPES, "pred_latent", "pred_camera")
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--graph", type=Path, required=True)
parser.add_argument("--sidecar", type=Path, required=True)
parser.add_argument("--input-fixture-dir", type=Path, required=True)
parser.add_argument("--output-graph", type=Path, required=True)
parser.add_argument("--output-fixture-dir", type=Path, required=True)
parser.add_argument(
"--probe-set",
choices=("blocks", "noise0", "context0"),
default="blocks",
help=(
"Block progression, detailed first noise-refiner probes, or bounded "
"first context-refiner attention probes (default: %(default)s)."
),
)
parser.add_argument(
"--inline-node-prefix",
action="append",
default=[],
help=(
"Inline external initializers consumed by nodes whose names start with this prefix. "
"May be repeated; intended for diagnosing browser external-data boundaries."
),
)
parser.add_argument("--session-threads", type=int, default=8)
return parser.parse_args()
def sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as stream:
while block := stream.read(8 * 1024 * 1024):
digest.update(block)
return digest.hexdigest()
def replace_with_hard_link(source: Path, destination: Path) -> None:
destination.parent.mkdir(parents=True, exist_ok=True)
if destination.exists():
if os.path.samefile(source, destination):
return
destination.unlink()
os.link(source, destination)
def probe_sources(
probe_set: str,
available: set[str],
shapes: dict[str, list[int]],
) -> list[tuple[str, str, list[int] | None]]:
vector = [1, 4, 16]
if probe_set == "context0":
context_vector = [1, 4, 1024]
attention_rows = [1, 16, 4, 4101]
return [
(
"context0_block_input",
"/flow_model/context_refiner.0/norm1/Cast_output_0",
context_vector,
),
(
"context0_attention_input",
"/flow_model/context_refiner.0/norm1/Cast_1_output_0",
context_vector,
),
(
"context0_qkv",
"/flow_model/context_refiner.0/attn/Reshape_output_0",
None,
),
(
"context0_q_after_rope",
"/flow_model/context_refiner.0/attn/Reshape_2_output_0",
None,
),
(
"context0_k_after_rope",
"/flow_model/context_refiner.0/attn/Reshape_4_output_0",
None,
),
(
"context0_v",
"/flow_model/context_refiner.0/attn/Squeeze_2_output_0",
None,
),
(
"context0_q_normalized",
"/flow_model/context_refiner.0/attn/q_norm/Slice_1_output_0",
None,
),
(
"context0_k_normalized",
"/flow_model/context_refiner.0/attn/k_norm/Slice_1_output_0",
None,
),
(
"context0_q_transposed",
"/flow_model/context_refiner.0/attn/Transpose_output_0",
attention_rows[:-1] + [64],
),
(
"context0_k_transposed",
"/flow_model/context_refiner.0/attn/Transpose_1_output_0",
[1, 16, 4101, 64],
),
(
"context0_v_transposed",
"/flow_model/context_refiner.0/attn/Transpose_2_output_0",
[1, 16, 4101, 64],
),
(
"context0_scaled_logits_rows",
"/flow_model/context_refiner.0/attn/MatMul_output_0",
attention_rows,
),
(
"context0_probabilities_rows",
"/flow_model/context_refiner.0/attn/Softmax_output_0",
attention_rows,
),
(
"context0_weighted_value_rows",
"/flow_model/context_refiner.0/attn/MatMul_1_output_0",
[1, 16, 4, 64],
),
(
"context0_pre_projection",
"/flow_model/context_refiner.0/attn/Reshape_5_output_0",
context_vector,
),
(
"context0_post_projection",
"/flow_model/context_refiner.0/attn/out/Reshape_output_0",
context_vector,
),
(
"context0_residual",
"/flow_model/context_refiner.0/Add_output_0",
context_vector,
),
]
if probe_set == "noise0":
attention = [1, 2, 4, 8]
full_attention_candidates = [
name
for name, shape in shapes.items()
if name.startswith("/flow_model/noise_refiner.0/attn/Concat")
and len(shape) == 4
and shape[1:] == [16, 8192, 64]
]
if len(full_attention_candidates) != 1:
raise ValueError(
"Expected exactly one full 16-head attention concatenation, found "
f"{full_attention_candidates}"
)
full_attention = full_attention_candidates[0]
attention_head_groups = [
name
for name, shape in shapes.items()
if name.startswith("/flow_model/noise_refiner.0/attn/Concat")
and len(shape) == 4
and shape[1:] == [8, 8192, 64]
]
projected = next(
candidate
for candidate in (
"/flow_model/noise_refiner.0/attn/out/Reshape_1_output_0",
"/flow_model/noise_refiner.0/attn/out/Reshape_output_0",
"/flow_model/noise_refiner.0/attn/out/Add_output_0",
)
if candidate in available
)
result = [
("input_positioned", "/flow_model/Add_1_output_0", vector),
("noise0_modulated", "/flow_model/noise_refiner.0/Add_3_output_0", vector),
("noise0_qkv", "/flow_model/noise_refiner.0/attn/qkv/Add_output_0", vector),
("noise0_qkv_full", "/flow_model/noise_refiner.0/attn/qkv/Add_output_0", None),
("noise0_q_norm", "/flow_model/noise_refiner.0/attn/q_norm/Cast_1_output_0", attention),
("noise0_q_norm_full", "/flow_model/noise_refiner.0/attn/q_norm/Cast_1_output_0", None),
("noise0_k_norm", "/flow_model/noise_refiner.0/attn/k_norm/Cast_1_output_0", attention),
("noise0_k_norm_full", "/flow_model/noise_refiner.0/attn/k_norm/Cast_1_output_0", None),
("noise0_q_transposed", "/flow_model/noise_refiner.0/attn/Transpose_output_0", attention),
("noise0_q_transposed_full", "/flow_model/noise_refiner.0/attn/Transpose_output_0", None),
("noise0_k_transposed_full", "/flow_model/noise_refiner.0/attn/Transpose_1_output_0", None),
("noise0_v_transposed_full", "/flow_model/noise_refiner.0/attn/Transpose_2_output_0", None),
("noise0_q_chunk0", "/flow_model/noise_refiner.0/attn/Cast_6_output_0", attention),
("noise0_score0", "/flow_model/noise_refiner.0/attn/MatMul_output_0", attention),
("noise0_softmax0", "/flow_model/noise_refiner.0/attn/Softmax_output_0", attention),
("noise0_value0", "/flow_model/noise_refiner.0/attn/MatMul_1_output_0", attention),
("noise0_attention_all", full_attention, attention),
("noise0_attention_reshaped", "/flow_model/noise_refiner.0/attn/Reshape_5_output_0", vector),
("noise0_attention_projected", projected, vector),
("noise0_attention_gated", "/flow_model/noise_refiner.0/Mul_7_output_0", vector),
("noise0_attention_residual", "/flow_model/noise_refiner.0/Add_4_output_0", vector),
("noise0_mlp_normalized", "/flow_model/noise_refiner.0/norm2/Cast_1_output_0", vector),
("noise0_mlp_hidden", "/flow_model/noise_refiner.0/mlp/mlp/mlp.1/Gelu_output_0", vector),
("noise0_output", "/flow_model/noise_refiner.0/Add_7_output_0", vector),
]
if (
"/flow_model/noise_refiner.0/attn/out/MatMul_output_0" in available
and "/flow_model/noise_refiner.0/attn/out/Reshape_1_output_0" in available
):
projection_steps = [
(
f"noise0_projection_input_tile_{index}",
(
"/flow_model/noise_refiner.0/attn/out/Slice_output_0"
if index == 0
else f"/flow_model/noise_refiner.0/attn/out/Slice_{index}_output_0"
),
[4, 8],
)
for index in range(4)
]
projection_steps.extend(
(
f"noise0_projection_partial_{index}",
(
"/flow_model/noise_refiner.0/attn/out/MatMul_output_0"
if index == 0
else f"/flow_model/noise_refiner.0/attn/out/MatMul_{index}_output_0"
),
[4, 8],
)
for index in range(4)
)
projection_steps.extend(
(
name,
source,
[4, 8],
)
for name, source in (
("noise0_projection_sum_01", "/flow_model/noise_refiner.0/attn/out/Add_output_0"),
("noise0_projection_sum_23", "/flow_model/noise_refiner.0/attn/out/Add_1_output_0"),
("noise0_projection_sum", "/flow_model/noise_refiner.0/attn/out/Add_2_output_0"),
("noise0_projection_bias", "/flow_model/noise_refiner.0/attn/out/Add_3_output_0"),
("noise0_projection_cast", "/flow_model/noise_refiner.0/attn/out/Cast_4_output_0"),
)
)
result[12:12] = projection_steps
elif "/flow_model/noise_refiner.0/attn/out/MatMul_output_0" in available:
result[12:12] = [
(
"noise0_projection_channels_first",
"/flow_model/noise_refiner.0/attn/out/Slice_output_0",
[1, 8, 4],
),
(
"noise0_projection_weight_left",
"/flow_model/noise_refiner.0/attn/out/MatMul_output_0",
[1, 8, 4],
),
(
"noise0_projection_weight_left_transposed",
"/flow_model/noise_refiner.0/attn/out/Transpose_1_output_0",
[1, 4, 8],
),
]
elif shapes.get("/flow_model/noise_refiner.0/attn/out/Add_9_output_0") == [1, 1, 256, 256]:
fixed_256_tree = [
*[
(
f"noise0_attention_head_group_{index}_full",
source,
None,
)
for index, source in enumerate(attention_head_groups)
],
(
"noise0_attention_all_full",
full_attention,
None,
),
(
"noise0_attention_reshaped_full",
"/flow_model/noise_refiner.0/attn/Reshape_5_output_0",
None,
),
(
"noise0_projection_first_token_full",
"/flow_model/noise_refiner.0/attn/out/Slice_output_0",
None,
),
(
"noise0_projection_special_products",
"/flow_model/noise_refiner.0/attn/out/Mul_output_0",
[1, 1, 4, 8],
),
(
"noise0_projection_special_products_full",
"/flow_model/noise_refiner.0/attn/out/Mul_output_0",
None,
),
(
"noise0_projection_products_transposed",
"/flow_model/noise_refiner.0/attn/out/Transpose_2_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_products_transposed_full",
"/flow_model/noise_refiner.0/attn/out/Transpose_2_output_0",
None,
),
*[
(
f"noise0_projection_tree_{index}",
(
"/flow_model/noise_refiner.0/attn/out/Add_output_0"
if index == 0
else f"/flow_model/noise_refiner.0/attn/out/Add_{index}_output_0"
),
[1, 1, 4, 16],
)
for index in range(10)
],
(
"noise0_projection_tree_0_full",
"/flow_model/noise_refiner.0/attn/out/Add_output_0",
None,
),
(
"noise0_projection_tree_1_full",
"/flow_model/noise_refiner.0/attn/out/Add_1_output_0",
None,
),
(
"noise0_projection_fixed_gather_left",
"/flow_model/noise_refiner.0/attn/out/Gather_1_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_fixed_gather_right",
"/flow_model/noise_refiner.0/attn/out/Gather_2_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_special_sum",
"/flow_model/noise_refiner.0/attn/out/Gather_17_output_0",
[1, 1, 16],
),
(
"noise0_projection_special_bias",
"/flow_model/noise_refiner.0/attn/out/Add_10_output_0",
[1, 1, 16],
),
(
"noise0_projection_bulk_tail",
"/flow_model/noise_refiner.0/attn/out/Gather_output_0",
[1, 4, 16],
),
(
"noise0_projection_corrected_chunk",
"/flow_model/noise_refiner.0/attn/out/Concat_output_0",
[1, 4, 16],
),
]
result[12:12] = fixed_256_tree
elif shapes.get("/flow_model/noise_refiner.0/attn/out/Add_9_output_0") == [1, 1, 128, 256]:
fixed_tree = [
(
"noise0_projection_special_products",
"/flow_model/noise_refiner.0/attn/out/Mul_output_0",
[1, 1, 4, 8],
),
(
"noise0_projection_products_transposed",
"/flow_model/noise_refiner.0/attn/out/Transpose_2_output_0",
[1, 1, 4, 16],
),
*[
(
f"noise0_projection_tree_{index}",
(
"/flow_model/noise_refiner.0/attn/out/Add_output_0"
if index == 0
else f"/flow_model/noise_refiner.0/attn/out/Add_{index}_output_0"
),
[1, 1, 4, 16],
)
for index in range(10)
],
(
"noise0_projection_tree_2_full",
"/flow_model/noise_refiner.0/attn/out/Add_2_output_0",
None,
),
(
"noise0_projection_fixed_gather_left",
"/flow_model/noise_refiner.0/attn/out/Gather_1_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_fixed_gather_right",
"/flow_model/noise_refiner.0/attn/out/Gather_2_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_fixed_masked_left",
"/flow_model/noise_refiner.0/attn/out/Mul_1_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_fixed_masked_right",
"/flow_model/noise_refiner.0/attn/out/Mul_2_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_special_sum",
"/flow_model/noise_refiner.0/attn/out/Gather_15_output_0",
[1, 1, 16],
),
(
"noise0_projection_special_bias",
"/flow_model/noise_refiner.0/attn/out/Add_10_output_0",
[1, 1, 16],
),
(
"noise0_projection_bulk_tail",
"/flow_model/noise_refiner.0/attn/out/Gather_output_0",
[1, 4, 16],
),
(
"noise0_projection_corrected_chunk",
"/flow_model/noise_refiner.0/attn/out/Concat_output_0",
[1, 4, 16],
),
]
result[12:12] = fixed_tree
elif "/flow_model/noise_refiner.0/attn/out/Gather_15_output_0" in available:
gathered_tree = [
(
"noise0_projection_special_products",
"/flow_model/noise_refiner.0/attn/out/Mul_output_0",
[1, 1, 4, 8],
),
*[
(
f"noise0_projection_tree_{index}",
(
"/flow_model/noise_refiner.0/attn/out/Add_output_0"
if index == 0
else f"/flow_model/noise_refiner.0/attn/out/Add_{index}_output_0"
),
[1, 1, 4, min(8, 512 >> index)] if index < 3 else [1, 1, 4, 128],
)
for index in range(10)
],
(
"noise0_projection_gathered_left",
"/flow_model/noise_refiner.0/attn/out/Gather_1_output_0",
[1, 1, 4, 8],
),
(
"noise0_projection_gathered_right",
"/flow_model/noise_refiner.0/attn/out/Gather_2_output_0",
[1, 1, 4, 8],
),
(
"noise0_projection_padded_left",
"/flow_model/noise_refiner.0/attn/out/Concat_output_0",
[1, 1, 4, 128],
),
(
"noise0_projection_padded_right",
"/flow_model/noise_refiner.0/attn/out/Concat_1_output_0",
[1, 1, 4, 128],
),
(
"noise0_projection_special_sum",
"/flow_model/noise_refiner.0/attn/out/Gather_15_output_0",
[1, 1, 16],
),
(
"noise0_projection_special_bias",
"/flow_model/noise_refiner.0/attn/out/Add_10_output_0",
[1, 1, 16],
),
(
"noise0_projection_bulk_tail",
"/flow_model/noise_refiner.0/attn/out/Gather_output_0",
[1, 4, 16],
),
(
"noise0_projection_corrected_chunk",
"/flow_model/noise_refiner.0/attn/out/Concat_14_output_0",
[1, 4, 16],
),
]
result[12:12] = gathered_tree
elif "/flow_model/noise_refiner.0/attn/out/Concat_14_output_0" in available:
padded_tree = [
(
"noise0_projection_special_products",
"/flow_model/noise_refiner.0/attn/out/Mul_output_0",
[1, 1, 4, 8],
),
*[
(
f"noise0_projection_tree_{index}",
(
"/flow_model/noise_refiner.0/attn/out/Add_output_0"
if index == 0
else f"/flow_model/noise_refiner.0/attn/out/Add_{index}_output_0"
),
[1, 1, 4, min(8, 512 >> index)] if index < 3 else [1, 1, 4, 128],
)
for index in range(10)
],
(
"noise0_projection_padded_left",
"/flow_model/noise_refiner.0/attn/out/Concat_output_0",
[1, 1, 4, 128],
),
(
"noise0_projection_padded_right",
"/flow_model/noise_refiner.0/attn/out/Concat_1_output_0",
[1, 1, 4, 128],
),
(
"noise0_projection_special_sum",
"/flow_model/noise_refiner.0/attn/out/Gather_1_output_0",
[1, 1, 16],
),
(
"noise0_projection_special_bias",
"/flow_model/noise_refiner.0/attn/out/Add_10_output_0",
[1, 1, 16],
),
(
"noise0_projection_bulk_tail",
"/flow_model/noise_refiner.0/attn/out/Gather_output_0",
[1, 4, 16],
),
(
"noise0_projection_corrected_chunk",
"/flow_model/noise_refiner.0/attn/out/Concat_14_output_0",
[1, 4, 16],
),
]
result[12:12] = padded_tree
elif shapes.get("/flow_model/noise_refiner.0/attn/out/Add_9_output_0") == [1, 1, 1, 256]:
transposed_tree = [
(
"noise0_projection_special_products",
"/flow_model/noise_refiner.0/attn/out/Mul_output_0",
[1, 1, 4, 8],
),
(
"noise0_projection_products_transposed",
"/flow_model/noise_refiner.0/attn/out/Transpose_2_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_first_left",
"/flow_model/noise_refiner.0/attn/out/Slice_1_output_0",
[1, 1, 4, 16],
),
(
"noise0_projection_first_right",
"/flow_model/noise_refiner.0/attn/out/Slice_2_output_0",
[1, 1, 4, 16],
),
*[
(
f"noise0_projection_tree_{index}",
(
"/flow_model/noise_refiner.0/attn/out/Add_output_0"
if index == 0
else f"/flow_model/noise_refiner.0/attn/out/Add_{index}_output_0"
),
[1, 1, min(4, 512 >> index), 16],
)
for index in range(10)
],
(
"noise0_projection_special_sum",
"/flow_model/noise_refiner.0/attn/out/Squeeze_1_output_0",
[1, 1, 16],
),
(
"noise0_projection_special_bias",
"/flow_model/noise_refiner.0/attn/out/Add_10_output_0",
[1, 1, 16],
),
(
"noise0_projection_bulk_tail",
"/flow_model/noise_refiner.0/attn/out/Gather_output_0",
[1, 4, 16],
),
(
"noise0_projection_corrected_chunk",
"/flow_model/noise_refiner.0/attn/out/Concat_output_0",
[1, 4, 16],
),
]
result[12:12] = transposed_tree
elif "/flow_model/noise_refiner.0/attn/out/Add_9_output_0" in available:
tree_levels = [
(
f"noise0_projection_tree_{index}",
(
"/flow_model/noise_refiner.0/attn/out/Add_output_0"
if index == 0
else f"/flow_model/noise_refiner.0/attn/out/Add_{index}_output_0"
),
[1, 1, 4, min(8, 512 >> index)],
)
for index in range(10)
]
result[12:12] = [
(
"noise0_projection_special_products",
"/flow_model/noise_refiner.0/attn/out/Mul_output_0",
[1, 1, 4, 8],
),
*tree_levels,
(
"noise0_projection_special_sum",
"/flow_model/noise_refiner.0/attn/out/Squeeze_1_output_0",
[1, 1, 16],
),
(
"noise0_projection_special_bias",
"/flow_model/noise_refiner.0/attn/out/Add_10_output_0",
[1, 1, 16],
),
(
"noise0_projection_bulk_tail",
"/flow_model/noise_refiner.0/attn/out/Gather_output_0",
[1, 4, 16],
),
(
"noise0_projection_corrected_chunk",
"/flow_model/noise_refiner.0/attn/out/Concat_output_0",
[1, 4, 16],
),
]
elif "/flow_model/noise_refiner.0/attn/out/ReduceSum_output_0" in available:
result[12:12] = [
(
"noise0_projection_special_products",
"/flow_model/noise_refiner.0/attn/out/Mul_output_0",
[1, 1, 4, 8],
),
(
"noise0_projection_special_sum",
"/flow_model/noise_refiner.0/attn/out/ReduceSum_output_0",
[1, 1, 16],
),
(
"noise0_projection_special_bias",
"/flow_model/noise_refiner.0/attn/out/Add_output_0",
[1, 1, 16],
),
(
"noise0_projection_bulk_tail",
"/flow_model/noise_refiner.0/attn/out/Gather_output_0",
[1, 4, 16],
),
(
"noise0_projection_corrected_chunk",
"/flow_model/noise_refiner.0/attn/out/Concat_output_0",
[1, 4, 16],
),
]
return result
result = [
("input_linear", "/flow_model/input_layer/Add_output_0", vector),
("input_positioned", "/flow_model/Add_1_output_0", vector),
("condition_dino", "/flow_model/cond_embedder/Add_output_0", vector),
("condition_vae", "/flow_model/cond_embedder2/Add_output_0", vector),
("condition_combined", "/flow_model/Add_output_0", vector),
(
"noise_refiner_00_attention_residual",
"/flow_model/noise_refiner.0/Add_4_output_0",
vector,
),
("noise_refiner_00", "/flow_model/noise_refiner.0/Add_7_output_0", vector),
(
"noise_refiner_01_attention_residual",
"/flow_model/noise_refiner.1/Add_4_output_0",
vector,
),
("noise_refiner_01", "/flow_model/noise_refiner.1/Add_7_output_0", vector),
(
"context_refiner_00_attention_residual",
"/flow_model/context_refiner.0/Add_output_0",
vector,
),
("context_refiner_00", "/flow_model/context_refiner.0/Add_1_output_0", vector),
(
"context_refiner_01_attention_residual",
"/flow_model/context_refiner.1/Add_output_0",
vector,
),
("context_refiner_01", "/flow_model/context_refiner.1/Add_1_output_0", vector),
("combined_tokens", "/flow_model/Concat_output_0", vector),
(
"block_00_attention_residual",
"/flow_model/blocks.0/Add_4_output_0",
vector,
),
]
result.extend(
(f"block_{index:02d}", f"/flow_model/blocks.{index}/Add_7_output_0", vector)
for index in range(24)
)
return result
def external_data_fields(tensor: Any) -> dict[str, str]:
return {entry.key: entry.value for entry in tensor.external_data}
def inline_selected_initializers(model: Any, sidecar: Path, node_prefixes: list[str]) -> list[str]:
if not node_prefixes:
return []
selected_inputs = {
input_name
for node in model.graph.node
if any(node.name.startswith(prefix) for prefix in node_prefixes)
for input_name in node.input
}
inlined: list[str] = []
with sidecar.open("rb") as stream:
for tensor in model.graph.initializer:
if tensor.name not in selected_inputs or not tensor.external_data:
continue
fields = external_data_fields(tensor)
location = fields.get("location")
if location and Path(location).name != sidecar.name:
raise ValueError(
f"Initializer {tensor.name!r} points to {location!r}, not {sidecar.name!r}."
)
offset = int(fields.get("offset", "0"))
length = int(fields["length"])
stream.seek(offset)
raw_data = stream.read(length)
if len(raw_data) != length:
raise EOFError(
f"Initializer {tensor.name!r} read {len(raw_data)} bytes at {offset}; expected {length}."
)
tensor.raw_data = raw_data
tensor.ClearField("external_data")
tensor.data_location = 0
inlined.append(tensor.name)
if not inlined:
raise ValueError(f"No external initializers matched node prefixes: {node_prefixes}")
return inlined
def main() -> None:
try:
import numpy as np
import onnx
import onnxruntime as ort
from onnx import helper
except ImportError as exc:
raise SystemExit(f"ONNX, ONNX Runtime, and NumPy are required: {exc}") from exc
args = parse_args()
graph = args.graph.expanduser().resolve()
sidecar = args.sidecar.expanduser().resolve()
fixture_dir = args.input_fixture_dir.expanduser().resolve()
output_graph = args.output_graph.expanduser().resolve()
output_fixture = args.output_fixture_dir.expanduser().resolve()
for path in (graph, sidecar):
if not path.is_file():
raise FileNotFoundError(path)
model = onnx.load_model(str(graph), load_external_data=False)
available = {value for node in model.graph.node for value in node.output}
inferred_model = onnx.shape_inference.infer_shapes(model)
shapes = {
value.name: [dimension.dim_value for dimension in value.type.tensor_type.shape.dim]
for value in (*inferred_model.graph.value_info, *inferred_model.graph.output)
if value.type.HasField("tensor_type")
}
# ONNX shape inference loses the fixed batch value at attention
# concatenations created by the statically unrolled chunk loops. All public
# inputs are fixed to batch one, so make that contract explicit on probes.
for shape in shapes.values():
if shape and shape[0] == 0:
shape[0] = 1
sources = probe_sources(args.probe_set, available, shapes)
missing = [source for _, source, _ in sources if source not in available]
if missing:
raise KeyError(f"Graph is missing probe values: {missing}")
inlined_initializers = inline_selected_initializers(model, sidecar, args.inline_node_prefix)
include_final_outputs = args.probe_set == "blocks"
if not include_final_outputs:
del model.graph.output[:]
for index, (name, source, ends) in enumerate(sources):
prefix = f"triposplat_probe_{index:02d}"
output_shape = shapes[source] if ends is None else ends
cast_input = source
if ends is not None:
constants = {
"starts": [0] * len(ends),
"ends": ends,
"axes": list(range(len(ends))),
"steps": [1] * len(ends),
}
inputs = []
for suffix, values in constants.items():
initializer_name = f"{prefix}_{suffix}"
model.graph.initializer.append(
helper.make_tensor(
initializer_name,
onnx.TensorProto.INT64,
[len(values)],
values,
)
)
inputs.append(initializer_name)
slice_output = f"{prefix}_slice"
model.graph.node.append(
helper.make_node(
"Slice",
[source, *inputs],
[slice_output],
name=f"/{prefix}/Slice",
)
)
cast_input = slice_output
output_name = f"probe_{name}"
model.graph.node.append(
helper.make_node(
"Cast",
[cast_input],
[output_name],
name=f"/{prefix}/Cast",
to=onnx.TensorProto.FLOAT,
)
)
model.graph.output.append(
helper.make_tensor_value_info(
output_name,
onnx.TensorProto.FLOAT,
output_shape,
)
)
sidecar_name = output_graph.name + ".data"
for tensor in model.graph.initializer:
for entry in tensor.external_data:
if entry.key == "location":
entry.value = sidecar_name
metadata = {entry.key: entry.value for entry in model.metadata_props}
metadata["triposplat.diagnostic_probes"] = str(len(sources))
if inlined_initializers:
metadata["triposplat.diagnostic_inlined_initializers"] = ",".join(inlined_initializers)
del model.metadata_props[:]
for key, value in sorted(metadata.items()):
entry = model.metadata_props.add()
entry.key = key
entry.value = value
output_graph.parent.mkdir(parents=True, exist_ok=True)
onnx.save_model(model, str(output_graph), save_as_external_data=False)
output_sidecar = output_graph.parent / sidecar_name
replace_with_hard_link(sidecar, output_sidecar)
output_fixture.mkdir(parents=True, exist_ok=True)
for name in FIXTURE_FILES:
replace_with_hard_link(fixture_dir / f"{name}.f32", output_fixture / f"{name}.f32")
feeds = {
name: np.fromfile(fixture_dir / f"{name}.f32", dtype="<f4").reshape(shape)
for name, shape in INPUT_SHAPES.items()
}
options = ort.SessionOptions()
options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
options.log_severity_level = 3
options.intra_op_num_threads = args.session_threads
session = ort.InferenceSession(
str(output_graph),
sess_options=options,
providers=["CPUExecutionProvider"],
)
output_names = [f"probe_{name}" for name, _, _ in sources]
outputs = session.run(output_names, feeds)
probes: list[dict[str, Any]] = []
for output_name, value in zip(output_names, outputs):
array = np.ascontiguousarray(value, dtype="<f4")
path = output_fixture / f"{output_name}.f32"
array.tofile(path)
probes.append(
{
"name": output_name,
"path": path.name,
"shape": list(array.shape),
"elements": int(array.size),
"sha256": sha256(path),
}
)
manifest = {
"format": "triposplat-dit-probes-v1",
"graph": output_graph.name,
"graphSha256": sha256(output_graph),
"externalData": output_sidecar.name,
"externalDataSha256": sha256(output_sidecar),
"includeFinalOutputs": include_final_outputs,
"outputs": probes,
}
manifest_path = output_fixture / "probes.json"
manifest_path.write_text(json.dumps(manifest, indent=2, sort_keys=True) + "\n")
print(f"Wrote {output_graph}, {output_sidecar}, and {manifest_path}")
if __name__ == "__main__":
main()
|