File size: 39,352 Bytes
2c44909 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 | #!/usr/bin/env python3
"""Automatic adjacent-pair selection via configurable scoring metrics."""
import copy
import math
from contextlib import contextmanager
from typing import Dict, List, Optional, Set, Tuple
import torch
import torch.nn.functional as F
from fuse_layers_model import (
build_head_permutation,
compute_fisher,
compute_head_means,
find_attention_module,
find_layer_container,
merge_layers,
permute_attention_heads,
)
_DWCE_GRAD_CACHE_MAX_BYTES = 1 << 30
class _DwceGradCacheOverflow(RuntimeError):
"""Raised when shared-backward DWCE caching exceeds the configured budget."""
def _get_hidden_size(model) -> int:
hidden_size = getattr(model.config, "hidden_size", None)
if hidden_size is None:
hidden_size = getattr(model.config, "n_embd", None)
if hidden_size is None:
raise SystemExit("Model config missing hidden_size/n_embd")
return int(hidden_size)
def _detach_arg(arg):
if torch.is_tensor(arg):
return arg.detach()
if isinstance(arg, (list, tuple)):
return type(arg)(_detach_arg(x) for x in arg)
if isinstance(arg, dict):
return {k: _detach_arg(v) for k, v in arg.items()}
return arg
def _register_forward_hook(layer, hook):
try:
def wrapper(module, inputs, kwargs, output):
return hook(module, inputs, output, kwargs)
handle = layer.register_forward_hook(wrapper, with_kwargs=True)
return handle, True
except TypeError:
def wrapper(module, inputs, output):
return hook(module, inputs, output, None)
handle = layer.register_forward_hook(wrapper)
return handle, False
@contextmanager
def _temporary_layers(parent: object, name: str, new_layers: object):
original = getattr(parent, name)
setattr(parent, name, new_layers)
try:
yield
finally:
setattr(parent, name, original)
def _extract_hidden(output):
if torch.is_tensor(output):
return output
if isinstance(output, (tuple, list)):
if output and all(torch.is_tensor(item) for item in output):
return output[0]
for item in output:
hidden = _extract_hidden(item)
if hidden is not None:
return hidden
return None
if isinstance(output, dict):
for key in ("hidden_states", "last_hidden_state", "hidden_state"):
if key in output:
value = output[key]
if isinstance(value, (tuple, list)) and value and all(
torch.is_tensor(item) for item in value
):
return value[-1]
hidden = _extract_hidden(value)
if hidden is not None:
return hidden
for value in output.values():
hidden = _extract_hidden(value)
if hidden is not None:
return hidden
return None
for attr in ("hidden_states", "last_hidden_state"):
if hasattr(output, attr):
value = getattr(output, attr)
if isinstance(value, (tuple, list)) and value and all(
torch.is_tensor(item) for item in value
):
return value[-1]
hidden = _extract_hidden(value)
if hidden is not None:
return hidden
return None
def _build_fused_layer_for_pair(
model,
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
dataloader,
device: str,
fisher_mode: str,
eps: float,
hidden_size: int,
enable_head_permute: bool = True,
) -> Tuple[torch.nn.Module, Dict[str, float]]:
attn_a = find_attention_module(layer_a)
attn_b = find_attention_module(layer_b)
perm = None
inv_perm = None
num_heads = None
num_kv_heads = None
head_dim = None
if enable_head_permute:
mean_a, mean_b, num_heads, num_kv_heads, head_dim = compute_head_means(
model,
attn_a,
attn_b,
dataloader,
device,
hidden_size,
)
perm = build_head_permutation(
mean_a,
mean_b,
num_heads=num_heads,
num_kv_heads=num_kv_heads,
eps=eps,
)
layer_a_copy = copy.deepcopy(layer_a)
layer_b_copy = copy.deepcopy(layer_b)
attn_b_copy = find_attention_module(layer_b_copy)
if perm is not None:
permute_attention_heads(
attn_b_copy, perm, num_heads, num_kv_heads, head_dim=head_dim
)
inv_perm = [0] * len(perm)
for idx, mapped in enumerate(perm):
inv_perm[mapped] = idx
permute_attention_heads(attn_b, perm, num_heads, num_kv_heads, head_dim=head_dim)
try:
fisher_sums, num_batches, param_numels = compute_fisher(
model,
layer_a,
layer_b,
dataloader,
fisher_mode=fisher_mode,
device=device,
)
finally:
if inv_perm is not None:
permute_attention_heads(
attn_b, inv_perm, num_heads, num_kv_heads, head_dim=head_dim
)
merge_layers(
layer_a_copy,
layer_b_copy,
fisher_sums[0],
fisher_sums[1],
num_batches,
param_numels[0],
param_numels[1],
fisher_mode=fisher_mode,
eps=eps,
)
# Scalar mixing coefficients per parameter tensor; used by pressure redistribution
# to simulate future fusions without running another Fisher pass.
fuse_priors: Dict[str, float] = {}
params_b = {name: param for name, param in layer_b.named_parameters()}
clamp_eps = 1e-4
for name, param_a in layer_a.named_parameters():
param_b = params_b.get(name)
if param_b is None or param_b.shape != param_a.shape:
continue
if fisher_mode == "param":
fa = fisher_sums[0][name] / max(num_batches, 1)
fb = fisher_sums[1][name] / max(num_batches, 1)
if isinstance(fa, torch.Tensor):
fa_val = float(fa.mean().item())
else:
fa_val = float(fa)
if isinstance(fb, torch.Tensor):
fb_val = float(fb.mean().item())
else:
fb_val = float(fb)
else:
fa_val = float(
fisher_sums[0][name]
/ (max(num_batches, 1) * max(param_numels[0].get(name, 1), 1))
)
fb_val = float(
fisher_sums[1][name]
/ (max(num_batches, 1) * max(param_numels[1].get(name, 1), 1))
)
denom = fa_val + fb_val
if denom <= eps:
lam = 0.5
else:
lam = fa_val / (denom + eps)
lam = min(max(lam, clamp_eps), 1.0 - clamp_eps)
fuse_priors[name] = lam
layer_a_copy.eval()
return layer_a_copy, fuse_priors
def _init_fisher_accumulators(
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
fisher_mode: str,
device: str,
) -> Tuple[List[Dict[str, object]], List[Dict[str, int]]]:
fisher_sums: List[Dict[str, object]] = []
param_numels: List[Dict[str, int]] = []
for layer in (layer_a, layer_b):
layer_sums: Dict[str, object] = {}
layer_numels: Dict[str, int] = {}
for name, param in layer.named_parameters():
if not param.requires_grad:
continue
if fisher_mode == "param":
layer_sums[name] = torch.zeros_like(
param, dtype=torch.float32, device="cpu"
)
else:
layer_sums[name] = torch.zeros((), dtype=torch.float32, device=device)
layer_numels[name] = param.numel()
fisher_sums.append(layer_sums)
param_numels.append(layer_numels)
return fisher_sums, param_numels
def _accumulate_fisher_from_grads(
layer: torch.nn.Module,
layer_sums: Dict[str, object],
fisher_mode: str,
) -> None:
for name, param in layer.named_parameters():
if not param.requires_grad or param.grad is None:
continue
grad_sq = param.grad.detach().float().pow(2)
if fisher_mode == "param":
layer_sums[name] += grad_sq.cpu()
else:
layer_sums[name] += grad_sq.sum()
def _finalize_fisher_sums(
fisher_sums: List[Dict[str, object]],
fisher_mode: str,
) -> List[Dict[str, object]]:
if fisher_mode == "param":
return fisher_sums
finalized: List[Dict[str, object]] = []
for layer_sums in fisher_sums:
finalized_layer: Dict[str, object] = {}
for name, value in layer_sums.items():
if isinstance(value, torch.Tensor):
finalized_layer[name] = float(value.detach().cpu().item())
else:
finalized_layer[name] = float(value)
finalized.append(finalized_layer)
return finalized
def _compute_fuse_priors(
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
fisher_sums: List[Dict[str, object]],
num_batches: int,
param_numels: List[Dict[str, int]],
fisher_mode: str,
eps: float,
) -> Dict[str, float]:
fuse_priors: Dict[str, float] = {}
params_b = {name: param for name, param in layer_b.named_parameters()}
clamp_eps = 1e-4
for name, param_a in layer_a.named_parameters():
param_b = params_b.get(name)
if param_b is None or param_b.shape != param_a.shape:
continue
if fisher_mode == "param":
fa = fisher_sums[0][name] / max(num_batches, 1)
fb = fisher_sums[1][name] / max(num_batches, 1)
fa_val = float(fa.mean().item()) if isinstance(fa, torch.Tensor) else float(fa)
fb_val = float(fb.mean().item()) if isinstance(fb, torch.Tensor) else float(fb)
else:
fa_val = float(
fisher_sums[0][name]
/ (max(num_batches, 1) * max(param_numels[0].get(name, 1), 1))
)
fb_val = float(
fisher_sums[1][name]
/ (max(num_batches, 1) * max(param_numels[1].get(name, 1), 1))
)
denom = fa_val + fb_val
lam = 0.5 if denom <= eps else fa_val / (denom + eps)
fuse_priors[name] = min(max(lam, clamp_eps), 1.0 - clamp_eps)
return fuse_priors
def _score_dwce_with_shared_backward(
model,
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
dataloader,
device: str,
fisher_mode: str,
max_batches: int,
eps: float,
norm: str,
hidden_size: int,
enable_head_permute: bool = True,
) -> Tuple[float, Dict[str, object]]:
attn_a = find_attention_module(layer_a)
attn_b = find_attention_module(layer_b)
perm = None
inv_perm = None
num_heads = None
num_kv_heads = None
head_dim = None
if enable_head_permute:
mean_a, mean_b, num_heads, num_kv_heads, head_dim = compute_head_means(
model,
attn_a,
attn_b,
dataloader,
device,
hidden_size,
)
perm = build_head_permutation(
mean_a,
mean_b,
num_heads=num_heads,
num_kv_heads=num_kv_heads,
eps=eps,
)
layer_a_copy = copy.deepcopy(layer_a)
layer_b_copy = copy.deepcopy(layer_b)
attn_b_copy = find_attention_module(layer_b_copy)
if perm is not None:
permute_attention_heads(
attn_b_copy, perm, num_heads, num_kv_heads, head_dim=head_dim
)
inv_perm = [0] * len(perm)
for idx, mapped in enumerate(perm):
inv_perm[mapped] = idx
cache: Dict[str, Optional[torch.Tensor]] = {"teacher": None}
grad_sq_cache: List[torch.Tensor] = []
cached_bytes = 0
def hook_b(_module, _inputs, output, _kwargs=None):
teacher_hidden = _extract_hidden(output)
if teacher_hidden is None:
raise RuntimeError("Failed to extract teacher hidden state output.")
cache["teacher"] = teacher_hidden
if teacher_hidden.requires_grad:
teacher_hidden.retain_grad()
return output
handle_b, _ = _register_forward_hook(layer_b, hook_b)
for param in model.parameters():
param.requires_grad_(False)
for layer in (layer_a, layer_b):
for param in layer.parameters():
param.requires_grad_(True)
fisher_sums, param_numels = _init_fisher_accumulators(
layer_a, layer_b, fisher_mode, device
)
num_batches = 0
if perm is not None:
permute_attention_heads(attn_b, perm, num_heads, num_kv_heads, head_dim=head_dim)
try:
model.eval()
for batch_idx, batch in enumerate(dataloader):
if max_batches and batch_idx >= max_batches:
break
cache["teacher"] = None
input_ids = batch[0].to(device)
attention_mask = batch[1].to(device) if len(batch) > 1 else None
model.zero_grad(set_to_none=True)
outputs = model(
input_ids=input_ids,
attention_mask=attention_mask,
labels=input_ids,
)
outputs.loss.backward()
teacher = cache["teacher"]
grad = None if teacher is None else teacher.grad
if teacher is None or grad is None:
raise RuntimeError(
"Auto selection hooks failed to capture outputs/gradients. "
"Try updating PyTorch or run with --layer <index>."
)
grad_sq = grad.detach().pow(2).to(device=device, dtype=torch.float16)
cached_bytes += grad_sq.numel() * grad_sq.element_size()
if cached_bytes > _DWCE_GRAD_CACHE_MAX_BYTES:
raise _DwceGradCacheOverflow(
"DWCE grad cache exceeded device-memory budget during shared-backward scoring."
)
grad_sq_cache.append(grad_sq)
_accumulate_fisher_from_grads(layer_a, fisher_sums[0], fisher_mode)
_accumulate_fisher_from_grads(layer_b, fisher_sums[1], fisher_mode)
model.zero_grad(set_to_none=True)
num_batches += 1
finally:
handle_b.remove()
if inv_perm is not None:
permute_attention_heads(
attn_b, inv_perm, num_heads, num_kv_heads, head_dim=head_dim
)
for param in model.parameters():
param.requires_grad_(True)
if num_batches == 0:
raise RuntimeError("No batches processed; check dataset or text inputs.")
fisher_sums = _finalize_fisher_sums(fisher_sums, fisher_mode)
merge_layers(
layer_a_copy,
layer_b_copy,
fisher_sums[0],
fisher_sums[1],
num_batches,
param_numels[0],
param_numels[1],
fisher_mode=fisher_mode,
eps=eps,
)
fuse_priors = _compute_fuse_priors(
layer_a,
layer_b,
fisher_sums,
num_batches,
param_numels,
fisher_mode,
eps,
)
fused_layer = layer_a_copy
fused_layer.eval()
phase2_cache = {"teacher": None, "fused": None}
def hook_a(_module, inputs, output, kwargs=None):
with torch.no_grad():
detached_inputs = tuple(_detach_arg(arg) for arg in inputs)
if kwargs:
detached_kwargs = {k: _detach_arg(v) for k, v in kwargs.items()}
fused_out = fused_layer(*detached_inputs, **detached_kwargs)
else:
fused_out = fused_layer(*detached_inputs)
fused_hidden = _extract_hidden(fused_out)
if fused_hidden is None:
raise RuntimeError("Failed to extract fused hidden state output.")
phase2_cache["fused"] = fused_hidden
return output
def hook_b_eval(_module, _inputs, output, _kwargs=None):
teacher_hidden = _extract_hidden(output)
if teacher_hidden is None:
raise RuntimeError("Failed to extract teacher hidden state output.")
phase2_cache["teacher"] = teacher_hidden
return output
handle_a, has_kwargs_a = _register_forward_hook(layer_a, hook_a)
handle_b_eval, has_kwargs_b = _register_forward_hook(layer_b, hook_b_eval)
supports_kwargs = has_kwargs_a and has_kwargs_b
score_num = 0.0
score_den = 0.0
token_count = 0.0
try:
model.eval()
for batch_idx, batch in enumerate(dataloader):
if batch_idx >= num_batches:
break
phase2_cache["teacher"] = None
phase2_cache["fused"] = None
input_ids = batch[0].to(device)
attention_mask = batch[1].to(device) if len(batch) > 1 else None
with torch.no_grad():
model(
input_ids=input_ids,
attention_mask=attention_mask,
use_cache=False,
)
teacher = phase2_cache["teacher"]
fused = phase2_cache["fused"]
if teacher is None or fused is None:
raise RuntimeError(
"Auto selection hooks failed to capture outputs during DWCE replay."
)
grad_sq = grad_sq_cache[batch_idx].to(dtype=torch.float32)
if attention_mask is not None:
mask = attention_mask.to(dtype=torch.float32).unsqueeze(-1)
batch_tokens = float(mask.sum().item())
grad_sq = grad_sq * mask
else:
mask = None
batch_tokens = float(input_ids.numel())
token_count += batch_tokens
delta = fused - teacher
if mask is not None:
delta = delta * mask
score_num += (delta.float().pow(2) * grad_sq).sum().item()
score_den += (teacher.float().pow(2) * grad_sq).sum().item()
finally:
handle_a.remove()
handle_b_eval.remove()
score = (
score_num / (score_den + eps)
if norm == "relative"
else score_num / max(token_count, 1.0)
)
meta = {
"num_batches": num_batches,
"token_count": token_count,
"norm": norm,
"supports_kwargs": supports_kwargs,
"fuse_priors": fuse_priors,
"metric": "dwce",
"dwce_mode": "shared",
}
return score, meta
def _compute_dwce_for_pair(
model,
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
fused_layer: torch.nn.Module,
dataloader,
device: str,
max_batches: int,
eps: float,
norm: str,
) -> Tuple[float, Dict[str, object]]:
cache = {"teacher": None, "fused": None}
supports_kwargs = True
def hook_a(_module, inputs, output, kwargs=None):
with torch.no_grad():
detached_inputs = tuple(_detach_arg(arg) for arg in inputs)
if kwargs is not None and len(kwargs) > 0:
detached_kwargs = {k: _detach_arg(v) for k, v in kwargs.items()}
fused_out = fused_layer(*detached_inputs, **detached_kwargs)
else:
fused_out = fused_layer(*detached_inputs)
fused_hidden = _extract_hidden(fused_out)
if fused_hidden is None:
raise RuntimeError("Failed to extract fused hidden state output.")
cache["fused"] = fused_hidden
return output
def hook_b(_module, _inputs, output, _kwargs=None):
teacher_hidden = _extract_hidden(output)
if teacher_hidden is None:
raise RuntimeError("Failed to extract teacher hidden state output.")
cache["teacher"] = teacher_hidden
if teacher_hidden.requires_grad:
teacher_hidden.retain_grad()
return output
handle_a, has_kwargs_a = _register_forward_hook(layer_a, hook_a)
handle_b, has_kwargs_b = _register_forward_hook(layer_b, hook_b)
supports_kwargs = has_kwargs_a and has_kwargs_b
score_num = 0.0
score_den = 0.0
token_count = 0.0
num_batches = 0
model.eval()
for batch_idx, batch in enumerate(dataloader):
if max_batches and batch_idx >= max_batches:
break
cache["teacher"] = None
cache["fused"] = None
input_ids = batch[0].to(device)
attention_mask = batch[1].to(device) if len(batch) > 1 else None
model.zero_grad(set_to_none=True)
outputs = model(
input_ids=input_ids,
attention_mask=attention_mask,
labels=input_ids,
)
loss = outputs.loss
loss.backward()
teacher = cache["teacher"]
fused = cache["fused"]
grad = None if teacher is None else teacher.grad
if teacher is None or fused is None or grad is None:
raise RuntimeError(
"Auto selection hooks failed to capture outputs/gradients. "
"Try updating PyTorch or run with --layer <index>."
)
if not teacher.requires_grad:
raise RuntimeError(
"Teacher hidden state does not require grad. "
"Ensure model parameters require grad for DWCE."
)
with torch.no_grad():
if attention_mask is not None:
mask = attention_mask.to(dtype=torch.float32).unsqueeze(-1)
batch_tokens = float(mask.sum().item())
else:
mask = None
batch_tokens = float(input_ids.numel())
token_count += batch_tokens
delta = fused - teacher
grad_sq = grad.pow(2)
if mask is not None:
delta = delta * mask
grad_sq = grad_sq * mask
score_num += (delta.pow(2) * grad_sq).sum().item()
score_den += (teacher.pow(2) * grad_sq).sum().item()
num_batches += 1
handle_a.remove()
handle_b.remove()
if norm == "relative":
score = score_num / (score_den + eps)
else:
denom = token_count if token_count > 0 else 1.0
score = score_num / denom
meta = {
"num_batches": num_batches,
"token_count": token_count,
"norm": norm,
"supports_kwargs": supports_kwargs,
}
return score, meta
def _compute_cosine_for_pair(
model,
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
dataloader,
device: str,
max_batches: int,
eps: float,
) -> Tuple[float, Dict[str, object]]:
cache = {"a": None, "b": None}
supports_kwargs = True
def hook_a(_module, _inputs, output, _kwargs=None):
hidden = _extract_hidden(output)
if hidden is None:
raise RuntimeError("Failed to extract layer_a hidden state output.")
cache["a"] = hidden
return output
def hook_b(_module, _inputs, output, _kwargs=None):
hidden = _extract_hidden(output)
if hidden is None:
raise RuntimeError("Failed to extract layer_b hidden state output.")
cache["b"] = hidden
return output
handle_a, has_kwargs_a = _register_forward_hook(layer_a, hook_a)
handle_b, has_kwargs_b = _register_forward_hook(layer_b, hook_b)
supports_kwargs = has_kwargs_a and has_kwargs_b
score_sum = 0.0
token_count = 0.0
num_batches = 0
model.eval()
for batch_idx, batch in enumerate(dataloader):
if max_batches and batch_idx >= max_batches:
break
cache["a"] = None
cache["b"] = None
input_ids = batch[0].to(device)
attention_mask = batch[1].to(device) if len(batch) > 1 else None
with torch.no_grad():
model(
input_ids=input_ids,
attention_mask=attention_mask,
use_cache=False,
)
hidden_a = cache["a"]
hidden_b = cache["b"]
if hidden_a is None or hidden_b is None:
raise RuntimeError(
"Auto selection hooks failed to capture outputs for cosine scoring."
)
with torch.no_grad():
a = hidden_a.float()
b = hidden_b.float()
cos = F.cosine_similarity(a, b, dim=-1, eps=eps)
distance = 1.0 - cos
if attention_mask is not None:
mask = attention_mask.to(dtype=torch.float32)
batch_tokens = float(mask.sum().item())
distance = distance * mask
else:
batch_tokens = float(distance.numel())
token_count += batch_tokens
score_sum += float(distance.sum().item())
num_batches += 1
handle_a.remove()
handle_b.remove()
denom = token_count if token_count > 0 else 1.0
score = score_sum / denom
meta = {
"num_batches": num_batches,
"token_count": token_count,
"metric": "cosine",
"supports_kwargs": supports_kwargs,
}
return score, meta
def _compute_global_rel_change_for_pair(
model,
layers: List[torch.nn.Module],
pair_idx: int,
dataloader,
args,
max_batches: int,
eps: float,
) -> Tuple[float, Dict[str, object]]:
hidden_size = _get_hidden_size(model)
head_permute_select = not bool(getattr(args, "no_head_permute_select", False))
layer_a = layers[pair_idx]
layer_b = layers[pair_idx + 1]
fused_layer, fuse_priors = _build_fused_layer_for_pair(
model,
layer_a,
layer_b,
dataloader,
device=args.device,
fisher_mode=args.fisher_mode,
eps=eps,
hidden_size=hidden_size,
enable_head_permute=head_permute_select,
)
fused_layer.to(args.device)
fused_layer.eval()
parent, name, container = find_layer_container(model, getattr(args, "layer_path", None))
if len(list(container)) != len(layers):
raise RuntimeError("Layer container changed during auto-selection; aborting rerank.")
virtual_layers = list(layers)
virtual_layers[pair_idx] = fused_layer
del virtual_layers[pair_idx + 1]
if isinstance(container, torch.nn.ModuleList):
virtual_container = torch.nn.ModuleList(virtual_layers)
elif isinstance(container, list):
virtual_container = virtual_layers
else:
raise TypeError("Layer container must be ModuleList or list")
teacher_cache = {"pair": None, "final": None}
supports_kwargs = True
def hook_pair(_module, _inputs, output, _kwargs=None):
hidden = _extract_hidden(output)
if hidden is None:
raise RuntimeError("Failed to extract pair output for global relation rerank.")
teacher_cache["pair"] = hidden
return output
handle_pair, has_kwargs_pair = _register_forward_hook(layer_b, hook_pair)
supports_kwargs = supports_kwargs and has_kwargs_pair
score_sum = 0.0
token_count = 0.0
num_batches = 0
model.eval()
for batch_idx, batch in enumerate(dataloader):
if max_batches and batch_idx >= max_batches:
break
teacher_cache["pair"] = None
input_ids = batch[0].to(args.device)
attention_mask = batch[1].to(args.device) if len(batch) > 1 else None
with torch.no_grad():
teacher_outputs = model(
input_ids=input_ids,
attention_mask=attention_mask,
output_hidden_states=True,
use_cache=False,
)
teacher_hidden_states = getattr(teacher_outputs, "hidden_states", None)
if not teacher_hidden_states:
raise RuntimeError("Teacher forward did not return hidden_states.")
teacher_final = teacher_hidden_states[-1]
teacher_pair = teacher_cache["pair"]
if teacher_pair is None or teacher_final is None:
raise RuntimeError(
"Failed to capture teacher pair/final hidden states for global rerank."
)
with torch.no_grad(), _temporary_layers(parent, name, virtual_container):
fused_outputs = model(
input_ids=input_ids,
attention_mask=attention_mask,
output_hidden_states=True,
use_cache=False,
)
fused_hidden_states = getattr(fused_outputs, "hidden_states", None)
if not fused_hidden_states:
raise RuntimeError("Fused forward did not return hidden_states.")
fused_final = fused_hidden_states[-1]
if fused_final is None:
raise RuntimeError("Failed to capture fused final hidden state for global rerank.")
with torch.no_grad():
teacher_pair_f = teacher_pair.float()
teacher_final_f = teacher_final.float()
fused_final_f = fused_final.float()
teacher_rel = F.cosine_similarity(
teacher_pair_f, teacher_final_f, dim=-1, eps=eps
)
fused_rel = F.cosine_similarity(
teacher_pair_f, fused_final_f, dim=-1, eps=eps
)
rel_change = (teacher_rel - fused_rel).abs()
if attention_mask is not None:
mask = attention_mask.to(dtype=torch.float32)
batch_tokens = float(mask.sum().item())
rel_change = rel_change * mask
else:
batch_tokens = float(rel_change.numel())
token_count += batch_tokens
score_sum += float(rel_change.sum().item())
num_batches += 1
handle_pair.remove()
del fused_layer
if torch.cuda.is_available():
torch.cuda.empty_cache()
denom = token_count if token_count > 0 else 1.0
score = score_sum / denom
meta = {
"num_batches": num_batches,
"token_count": token_count,
"metric": "global_rel_change",
"supports_kwargs": supports_kwargs,
"fuse_priors": fuse_priors,
}
return score, meta
def select_layer_auto(
model,
layers: List[torch.nn.Module],
dataloader,
args,
previous_scores: Optional[List[float]] = None,
start_index: int = 0,
exclude_pairs: Optional[Set[int]] = None,
) -> Tuple[int, List[float], Dict[str, object]]:
num_layers = len(layers)
if num_layers < 2:
raise SystemExit("Model must have at least 2 layers for auto selection.")
hidden_size = _get_hidden_size(model)
num_pairs = num_layers - 1
scores: List[float] = [float("inf")] * num_pairs
meta_per_pair: List[Optional[Dict[str, object]]] = [None] * num_pairs
supports_kwargs_all = True
head_permute_select = not bool(getattr(args, "no_head_permute_select", False))
exclude_set: Set[int] = {
int(idx)
for idx in (exclude_pairs or set())
if isinstance(idx, int) and 0 <= int(idx) < num_pairs
}
max_batches = args.auto_max_batches
start_index = max(0, min(start_index, num_pairs))
auto_metric = str(getattr(args, "auto_metric", "dwce")).strip().lower()
if auto_metric == "hybrid":
auto_metric = "hybrid_cosine"
if auto_metric not in {
"dwce",
"cosine",
"hybrid_cosine",
"hybrid_global_rel",
}:
raise SystemExit(
"--auto_metric must be one of: dwce, cosine, hybrid, "
"hybrid_cosine, hybrid_global_rel"
)
auto_cosine_topk = int(getattr(args, "auto_cosine_topk", 3))
if auto_cosine_topk <= 0:
raise SystemExit("--auto_cosine_topk must be >= 1")
print(
f"[auto] metric={auto_metric}; using "
f"{('all' if max_batches == 0 else max_batches)} batches "
"from calibration samples."
)
reuse_upto = 0
allow_reuse = auto_metric == "dwce"
if previous_scores:
reuse_upto = min(start_index, len(previous_scores), num_pairs) if allow_reuse else 0
for idx in range(reuse_upto):
if idx in exclude_set:
scores[idx] = float("inf")
meta_per_pair[idx] = {"excluded": True}
print(f"[auto] skipped excluded pair {idx}-{idx+1}.")
continue
scores[idx] = previous_scores[idx]
meta_per_pair[idx] = (
{
"num_batches": 0,
"token_count": 0.0,
"norm": args.auto_norm,
"metric": auto_metric,
"supports_kwargs": True,
"reused": True,
}
)
print(f"[auto] reused pair {idx}-{idx+1}: {scores[idx]:.6e}")
compute_start = start_index if reuse_upto == start_index else reuse_upto
pairs_to_score: List[int] = []
for idx in range(compute_start, num_pairs):
if idx in exclude_set:
scores[idx] = float("inf")
meta_per_pair[idx] = {"excluded": True}
print(f"[auto] skipped excluded pair {idx}-{idx+1}.")
continue
pairs_to_score.append(idx)
def _score_dwce_for_pair(idx: int) -> Tuple[float, Dict[str, object]]:
print(f"[auto] building fused pair {idx}-{idx+1} for DWCE...")
layer_a = layers[idx]
layer_b = layers[idx + 1]
dwce_mode = str(getattr(args, "auto_dwce_mode", "separate")).strip().lower()
if dwce_mode == "shared":
try:
return _score_dwce_with_shared_backward(
model,
layer_a,
layer_b,
dataloader,
device=args.device,
fisher_mode=args.fisher_mode,
max_batches=max_batches,
eps=args.eps,
norm=args.auto_norm,
hidden_size=hidden_size,
enable_head_permute=head_permute_select,
)
except _DwceGradCacheOverflow:
print(
"[auto] shared-backward DWCE cache exceeded budget; "
"falling back to separate mode."
)
fused, fuse_priors = _build_fused_layer_for_pair(
model,
layer_a,
layer_b,
dataloader,
device=args.device,
fisher_mode=args.fisher_mode,
eps=args.eps,
hidden_size=hidden_size,
enable_head_permute=head_permute_select,
)
fused.to(args.device)
fused.eval()
for param in model.parameters():
param.requires_grad_(True)
score, meta = _compute_dwce_for_pair(
model,
layer_a,
layer_b,
fused,
dataloader,
device=args.device,
max_batches=max_batches,
eps=args.eps,
norm=args.auto_norm,
)
meta["fuse_priors"] = fuse_priors
meta["metric"] = "dwce"
del fused
if torch.cuda.is_available():
torch.cuda.empty_cache()
return score, meta
def _score_cosine_for_pair(idx: int) -> Tuple[float, Dict[str, object]]:
print(f"[auto] scoring cosine for pair {idx}-{idx+1}...")
layer_a = layers[idx]
layer_b = layers[idx + 1]
return _compute_cosine_for_pair(
model,
layer_a,
layer_b,
dataloader,
device=args.device,
max_batches=max_batches,
eps=args.eps,
)
def _score_global_rel_for_pair(idx: int) -> Tuple[float, Dict[str, object]]:
print(f"[auto] scoring global relation change for pair {idx}-{idx+1}...")
return _compute_global_rel_change_for_pair(
model,
layers,
idx,
dataloader,
args=args,
max_batches=max_batches,
eps=args.eps,
)
if auto_metric in {"dwce", "cosine"}:
for idx in pairs_to_score:
if auto_metric == "dwce":
score, meta = _score_dwce_for_pair(idx)
else:
score, meta = _score_cosine_for_pair(idx)
supports_kwargs_all = supports_kwargs_all and meta.get("supports_kwargs", True)
scores[idx] = score
meta_per_pair[idx] = meta
print(f"[auto] {auto_metric} pair {idx}-{idx+1}: {score:.6e}")
else:
dwce_prefilter: Dict[int, float] = {}
for idx in pairs_to_score:
score, meta = _score_dwce_for_pair(idx)
dwce_prefilter[idx] = score
supports_kwargs_all = supports_kwargs_all and meta.get("supports_kwargs", True)
meta_per_pair[idx] = {
"prefilter_dwce": score,
"dwce_meta": meta,
"metric": "hybrid",
}
print(f"[auto] hybrid prefilter DWCE pair {idx}-{idx+1}: {score:.6e}")
ranked = sorted(pairs_to_score, key=lambda i: float(dwce_prefilter[i]))
shortlist = ranked[: min(auto_cosine_topk, len(ranked))]
print(f"[auto] hybrid shortlist (dwce top-{len(shortlist)}): {shortlist}")
for idx in shortlist:
if auto_metric == "hybrid_global_rel":
score, rerank_meta = _score_global_rel_for_pair(idx)
score_metric = "global_rel_change"
else:
score, rerank_meta = _score_cosine_for_pair(idx)
score_metric = "cosine"
supports_kwargs_all = supports_kwargs_all and rerank_meta.get(
"supports_kwargs", True
)
scores[idx] = score
pair_meta = meta_per_pair[idx] or {}
pair_meta["rerank_meta"] = rerank_meta
pair_meta["score_metric"] = score_metric
meta_per_pair[idx] = pair_meta
print(f"[auto] hybrid {score_metric} pair {idx}-{idx+1}: {score:.6e}")
if not supports_kwargs_all:
print(
"[auto] Warning: forward hooks did not capture kwargs; "
"fused-layer calls may be approximate."
)
print(f"[auto] score summary (metric={auto_metric}, norm={args.auto_norm}):")
for idx, score in enumerate(scores):
if idx in exclude_set:
print(f"[auto] pair {idx}-{idx+1}: excluded")
elif math.isfinite(float(score)):
print(f"[auto] pair {idx}-{idx+1}: {score:.6e}")
else:
print(f"[auto] pair {idx}-{idx+1}: {score}")
candidates = [i for i in range(num_pairs) if i not in exclude_set]
if not candidates:
raise SystemExit("All pairs are excluded; cannot auto-select a fusion layer.")
best_idx = min(candidates, key=lambda i: scores[i])
best_score = float(scores[best_idx])
if not math.isfinite(best_score):
raise SystemExit(
"Auto selection failed: all candidate pairs have non-finite scores "
"(check --exclude_pairs and data)."
)
print(f"[auto] Selected layer {best_idx} (score={best_score:.6e})")
meta = {
"per_pair": meta_per_pair,
"supports_kwargs": supports_kwargs_all,
"max_batches": max_batches,
"norm": args.auto_norm,
"metric": auto_metric,
"cosine_topk": auto_cosine_topk,
"start_index": start_index,
"excluded_pairs": sorted(exclude_set),
}
return best_idx, scores, meta
|