File size: 36,924 Bytes
2b534de | 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 | from ldm.modules.attention import *
import global_
import torch
import torch.nn as nn
import torch.nn.functional as F
from my_py_lib.torch_util import custom_repr_v3
from confs import *
import cv2, numpy as np
from lmk_util.lmk_extractor import lmkAll_2_lmkMain, get_lmkMain_indices
from MoE import *
from lora_layers import *
import json
import copy
"""
Global knobs for shared experts and routing (no argparse per user preference)
"""
NUM_SHARED_FFN = 8
GATE_TOPK = 2
# Sparse MoE FFN for all FFN blocks (in addition to shared orig + task LoRA)
# default off to keep behavior unchanged; enable by setting EXTRA_MoE_enable to True
EXTRA_MoE_enable :bool = 1
EXTRA_MoE_num_ep = 8 # number of sparse MoE experts (narrow FFN)
EXTRA_MoE_inner_divisor = 64 # each expert intermediate dim = original FFN intermediate dim * this ratio
EXTRA_MoE_topK = 2 # sparse routing selects top-k experts (k fixed to 2)
EXTRA_MoE_add_noise :bool = 1 # add random noise to routing scores for exploration
EXTRA_MoE_noise_std = 0.1 # noise strength (Gaussian standard deviation)
EXTRA_MoE_en_auxLoss :bool = 0 # compute load-balancing auxiliary loss
EXTRA_MoE_aux_coef = 1e-2 # coefficient for auxiliary loss when adding to total loss
EXTRA_MoE_routing_mode = 'sparse' # 'sparse' | 'dense'
LMK_PICK_IDX = None
NUM_lmk_pick = len(LMK_PICK_IDX) if LMK_PICK_IDX is not None else len(get_lmkMain_indices(include_face_oval=True))
print(f"{NUM_lmk_pick=}")
IMAGE_SIZE_FOR_LMK_NORM = 512.0
def _log2(orig_modules, lora_modules):
"""Calculate and log parameter statistics for original and LoRA modules"""
# Calculate original module stats
orig_params = sum(p.numel() for p in orig_modules.parameters())
orig_size = sum(p.numel() * p.element_size() for p in orig_modules.parameters())
# Calculate LoRA stats (handle both single module and tuple/list)
if isinstance(lora_modules, (list, tuple)):
lora_params = sum(p.numel() for m in lora_modules for p in m.parameters())
lora_size = sum(p.numel() * p.element_size() for m in lora_modules for p in m.parameters())
# Try to get rank from lora modules
ranks = []
for m in lora_modules:
if hasattr(m, 'rank'):
ranks.append(m.rank)
if len(ranks) == 2:
rank_str = f" (rank_in={ranks[0]} rank_out={ranks[1]})"
elif len(ranks) == 1:
rank_str = f" (rank={ranks[0]})"
else:
rank_str = ""
else:
lora_params = sum(p.numel() for p in lora_modules.parameters())
lora_size = sum(p.numel() * p.element_size() for p in lora_modules.parameters())
# Try to get rank from lora module
if hasattr(lora_modules, 'rank'):
rank_str = f" (rank={lora_modules.rank})"
else:
rank_str = ""
msg1 = f"orig: {orig_params:,} params, {orig_size/1024/1024:.2f}MB"
msg2 = f"LoRA: {lora_params:,} params, {lora_size/1024/1024:.2f}MB{rank_str}"
for msg in [msg1, msg2]:
print(msg)
continue
with open(_verify_log_file, 'a') as f:
f.write(msg + '\n')
def _log1(msg: str):
"""Print message and append to log file"""
print(msg)
return
with open(_verify_log_file, 'a') as f:
f.write(msg + '\n')
def build_ffn_gate_input_common(x: torch.Tensor, token_pos_grid__cur, tasks: list):
"""Build gate input for FFN routing (reusable across FFN classes)."""
b, n, d = x.shape
token_feat = x # token
avg_feat = x.mean(dim=1, keepdim=True).expand(-1, n, -1) # avg(all tokens)
len_task = len(tasks) # task one-hot
task_1h = x.new_zeros(b, len_task)
task_1h[:, global_.task] = 1
task_1h = task_1h.unsqueeze(1).expand(-1, n, -1)
token_pos = token_pos_grid__cur # token-position from global_.token_pos_grid__cur
assert token_pos.shape[:2] == (b, n), (token_pos.shape, (b, n), )
rel_flat = x.new_zeros(b, n, 2*NUM_lmk_pick) # token-relative-position to lmks
lmk = global_.lmk_
if 1:
lmk = lmk.to(x.device).float()# TODO to check is it normed already?
if LMK_PICK_IDX is None:
assert NUM_lmk_pick==lmk.shape[1]
else:
lmk = lmk[:, LMK_PICK_IDX, :]
rel = token_pos.unsqueeze(2) - lmk.unsqueeze(1) # [b,n,L,2]
rel_flat = rel.reshape(b, n, -1)
gate_in = torch.cat([token_feat, avg_feat, task_1h, token_pos, rel_flat], dim=-1)
ctx = {'token_feat': token_feat, 'avg_feat': avg_feat, 'task_1h': task_1h, 'token_pos': token_pos, 'lmk': lmk, 'rel': rel, 'rel_flat': rel_flat}
return gate_in, ctx
def replace_modules_lossless(
module: nn.Module,
src_modules: list,
l_task: list,
parent_name: str = "",
depth :int = 0,
for_refnet: bool = False,
):
"""
Apply policy:
- FFN: shared-plus-task (lossless upcycle)
- CrossAttention linear projections (to_q, to_k, to_v, to_out.0): shared-plus-task
- Conv2d: keep task-specific or wrap with shared-plus-task if desired
- Norms: keep task-specific (LayerNorm/GroupNorm)
"""
if depth==0:
CONV2D_PARAM_STATS.clear()
# Skip modules with no parameters
if len(list(module.parameters())) == 0:
# print(f'[replace_modules_lossless] Skipping module with no parameters: {module}')
return module
if len(list(module.named_children()))==0:
print('\n!!!! len(list(module.named_children()))==0',module)
assert 0
for name, child in module.named_children():
full_name = f"{parent_name}.{name}" if parent_name else f".{name}"
src_child_modules = [getattr(src_module, name) for src_module in src_modules]
if len({id(s) for s in src_child_modules}) < len(src_child_modules):
raise Exception('Duplicate source modules detected!')
# if sources are the same instance(s), clone to ensure distinct expert modules
src_child_modules = [copy.deepcopy(src_child_modules[0]) for _ in src_child_modules]
if isinstance(child, FeedForward):
if 0:
setattr(module, name, TaskSpecific_MoE([s for s in src_child_modules], tasks=l_task))
else:
# FFN -> shared average + per-task LoRA
setattr(module, name, upCycle_module(src_child_modules, l_task, module_name=full_name))
continue
if isinstance(child, CrossAttention):
# replace linear projections
# if for_refnet:
if 0:
for proj_name in ["to_q", "to_k", "to_v"]:
src_proj_list = [getattr(s, proj_name) for s in src_child_modules]
setattr(child, proj_name, upCycle_module(src_proj_list, l_task, module_name=f"{full_name}.{proj_name}"))
if hasattr(child.to_out, "__getitem__"):
src_linear0 = [s.to_out[0] for s in src_child_modules]
child.to_out[0] = upCycle_module(src_linear0, l_task, module_name=f"{full_name}.to_out.0")
else:
for proj_name in ["to_q", "to_k", "to_v"]:
src_proj_list = [getattr(s, proj_name) for s in src_child_modules]
setattr(child, proj_name, TaskSpecific_MoE([s for s in src_proj_list], tasks=l_task) )
if hasattr(child.to_out, "__getitem__"):
src_linear0 = [s.to_out[0] for s in src_child_modules]
child.to_out[0] = TaskSpecific_MoE([s for s in src_linear0], tasks=l_task)
continue
if isinstance(child, nn.Conv2d):
num_params = sum(p.numel() for p in child.parameters())
CONV2D_PARAM_STATS.append((num_params, full_name))
# if num_params > CONV2D_PARAM_MOE_THRES and (not any(full_name.startswith(p) for p in FORCE_TASKSPEC_PREFIXES)):
if 1:
printC(f"shared+LoRA Conv2d",f"{full_name}")
setattr(module, name, upCycle_module(src_child_modules, l_task, module_name=full_name))
else:
setattr(module, name, TaskSpecific_MoE([s for s in src_child_modules], tasks=l_task))
continue
elif isinstance(child, (nn.LayerNorm, nn.GroupNorm)):
setattr(module, name, TaskSpecific_MoE([s for s in src_child_modules], tasks=l_task))
continue
elif isinstance(child, nn.Linear):
# default linear: task-specific
setattr(module, name, TaskSpecific_MoE([s for s in src_child_modules], tasks=l_task))
continue
else:
replace_modules_lossless(child, src_child_modules, l_task, parent_name=full_name, depth=depth+1, for_refnet=for_refnet)
if depth==0:
stats_sorted = sorted(CONV2D_PARAM_STATS, key=lambda x: x[0], reverse=True)
if gate_("[Conv2d param stats] count, name (sorted desc):"):
for cnt, n in stats_sorted:
print(f" {cnt:12d} {n}")
return module
def upCycle_module(l_modules, l_task, module_name: str = None):
assert len( set( [type(m) for m in l_modules] ) ) == 1
m0 = l_modules[0]
if isinstance(m0, FeedForward):
obj = FFN_Shared_Plus_TaskLoRA(l_modules, l_task, module_name=module_name)
elif isinstance(m0, nn.Linear):
obj = Linear_Shared_Plus_TaskLoRA(l_modules, l_task, module_name=module_name)
elif isinstance(m0, nn.Conv2d):
obj = Conv_Shared_Plus_TaskLoRA(l_modules, l_task, module_name=module_name)
else:
raise Exception(module_name,m0)
return TaskSpecific_MoE([s for s in l_modules], tasks=l_task)
if obj.dont_lora:
return TaskSpecific_MoE([s for s in l_modules], tasks=l_task)
return obj
class ResidualAdapterLinearOnly(nn.Module):
"""
Full-rank residual adapter returning the linear delta (orig - shared).
"""
def __init__(self, in_features: int, out_features: int, scaling: float = 1.0, use_bias_delta: bool = True):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.rank = min(in_features, out_features)
self.scaling = scaling
self.use_bias_delta = use_bias_delta
self.delta_weight = nn.Parameter(torch.zeros(out_features, in_features))
if use_bias_delta:
self.delta_bias = nn.Parameter(torch.zeros(out_features))
else:
self.register_parameter('delta_bias', None)
@torch.no_grad()
def init_from_diff(self, weight_diff: torch.Tensor, bias_diff: torch.Tensor = None):
self.delta_weight.copy_(weight_diff)
if (self.delta_bias is not None) and (bias_diff is not None):
self.delta_bias.copy_(bias_diff)
def forward(self, x: torch.Tensor) -> torch.Tensor:
update = x @ self.delta_weight.T
if self.delta_bias is not None:
update = update + self.delta_bias
return update * self.scaling
class ResidualAdapterConv2dOnly(nn.Module):
"""
Full-rank residual adapter for Conv2d, returning the convolutional delta (orig - shared).
"""
def __init__(self, in_channels: int, out_channels: int, kernel_size: tuple, stride: tuple, padding: tuple, dilation: tuple, groups: int = 1, scaling: float = 1.0, use_bias_delta: bool = True):
super().__init__()
if isinstance(kernel_size, int):
kernel_size = (kernel_size, kernel_size)
if isinstance(stride, int):
stride = (stride, stride)
if isinstance(padding, int):
padding = (padding, padding)
if isinstance(dilation, int):
dilation = (dilation, dilation)
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.dilation = dilation
self.groups = groups
kH, kW = kernel_size
self.rank = min(out_channels, in_channels * kH * kW)
self.scaling = scaling
self.use_bias_delta = use_bias_delta
self.delta_weight = nn.Parameter(torch.zeros(out_channels, in_channels // groups, kH, kW))
if use_bias_delta:
self.delta_bias = nn.Parameter(torch.zeros(out_channels))
else:
self.register_parameter('delta_bias', None)
@torch.no_grad()
def init_from_diff(self, weight_diff: torch.Tensor, bias_diff: torch.Tensor = None):
self.delta_weight.copy_(weight_diff)
if (self.delta_bias is not None) and (bias_diff is not None):
self.delta_bias.copy_(bias_diff)
def forward(self, x: torch.Tensor) -> torch.Tensor:
u = F.conv2d(x, self.delta_weight, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups)
if self.delta_bias is not None:
u = u + self.delta_bias.view(1, -1, 1, 1)
return u * self.scaling
class Linear_TaskSpecific_Plus_Shared(nn.Module):
def __init__(self, l_proj: list, l_task: list):
super().__init__()
assert len(l_proj) >= 1
p0 = l_proj[0]
assert isinstance(p0, nn.Linear)
in_f, out_f = p0.in_features, p0.out_features
bias = p0.bias is not None
self.shared = nn.Linear(in_f, out_f, bias=bias)
self.shared = zero_module(self.shared)
self.tasks = l_task
self.task_proj = ModuleDict_W(l_proj, self.tasks)
def forward(self, x):
t = global_.task
return self.task_proj[t](x) + self.shared(x)
class Conv_TaskSpecific_Plus_Shared(nn.Module):
def __init__(self, l_conv: list, l_task: list):
super().__init__()
assert len(l_conv) >= 1
c0 = l_conv[0]
assert isinstance(c0, nn.Conv2d)
self.shared = nn.Conv2d(c0.in_channels, c0.out_channels, kernel_size=c0.kernel_size, stride=c0.stride, padding=c0.padding, dilation=c0.dilation, groups=c0.groups, bias=(c0.bias is not None), padding_mode=c0.padding_mode)
self.shared = zero_module(self.shared)
self.tasks = l_task
self.task_conv = ModuleDict_W(l_conv, self.tasks)
def forward(self, x):
t = global_.task
return self.task_conv[t](x) + self.shared(x)
def _average_state_dict(modules: list):
assert len(modules) > 0
sd0 = modules[0].state_dict()
avg = {k: torch.zeros_like(v) for k, v in sd0.items()}
for m in modules:
msd = m.state_dict()
for k in avg:
avg[k] += msd[k]
for k in avg:
avg[k] /= len(modules)
return avg
class FFN_Shared_Plus_TaskLoRA(nn.Module):
def __init__(self, l_ffn: list, l_task: list, module_name: str = None):
super().__init__()
self.module_name = module_name
# _log1(f"-------- {module_name} --------")
assert len(l_ffn) >= 1
self.tasks = l_task
self.num_tasks = len(l_task)
self.dont_lora = False
f0: FeedForward = l_ffn[0]
# build shared from f0 and load avg
self.shared_ffn: FeedForward = copy.deepcopy(f0)
if FOR_upcycle_ckpt_GEN_or_USE:
avg_sd = _average_state_dict(l_ffn)
self.shared_ffn.load_state_dict(avg_sd)
# freeze shared
for p in self.shared_ffn.parameters():
p.requires_grad = False
# discover inner layers
self.is_glu = isinstance(self.shared_ffn.net[0], GEGLU)
if self.is_glu:
in_linear: nn.Linear = self.shared_ffn.net[0].proj
else:
assert isinstance(self.shared_ffn.net[0], nn.Sequential)
in_linear: nn.Linear = self.shared_ffn.net[0][0]
out_linear: nn.Linear = self.shared_ffn.net[2]
self.in_features = in_linear.in_features
self.mid_features = in_linear.out_features
self.out_features = out_linear.out_features
if 1: # cal/read adaptive rank across tasks
if FOR_upcycle_ckpt_GEN_or_USE:
w_diff_in_list = []
w_diff_out_list = []
for f in l_ffn:
if self.is_glu:
tin: nn.Linear = f.net[0].proj
else:
tin: nn.Linear = f.net[0][0]
tout: nn.Linear = f.net[2]
w_diff_in_list.append(tin.weight.data - in_linear.weight.data)
w_diff_out_list.append(tout.weight.data - out_linear.weight.data)
if FORCE_SAME_RANK_ACROSS_TASKS:
rank_in = compute_adaptive_rank_for_linear_diffs(w_diff_in_list)
rank_out = compute_adaptive_rank_for_linear_diffs(w_diff_out_list)
global_.moduleName_2_adaRank[module_name] = [rank_in, rank_out]
else:
ranks_in = compute_adaptive_rank_for_linear_diffs(w_diff_in_list, per_task=True)
ranks_out = compute_adaptive_rank_for_linear_diffs(w_diff_out_list, per_task=True)
global_.moduleName_2_adaRank[module_name] = [ranks_in, ranks_out]
else:
r_info = global_.moduleName_2_adaRank[module_name]
if FORCE_SAME_RANK_ACROSS_TASKS: rank_in, rank_out = r_info
else: ranks_in, ranks_out = r_info
if 1:
# fallback decision: (1) tiny feature dims
min_dim_in = min(self.in_features, self.mid_features)
min_dim_out = min(self.mid_features, self.out_features)
if (min_dim_in < DONT_lora_if_dim_lt) or (min_dim_out < DONT_lora_if_dim_lt):
# print(f"[LoRA fallback][FFN] {module_name} {min_dim_in=} {min_dim_out=} {DONT_lora_if_dim_lt=}")
self.dont_lora = True; return
# per-task LoRA adapters
_l_in = []
_l_out = []
for idx, f in enumerate(l_ffn):
if self.is_glu:
tin: nn.Linear = f.net[0].proj
else:
tin: nn.Linear = f.net[0][0]
tout: nn.Linear = f.net[2]
if not FORCE_SAME_RANK_ACROSS_TASKS:
rank_in = ranks_in[idx]
rank_out = ranks_out[idx]
frac_in = float(rank_in) / min(self.in_features, self.mid_features)
frac_out = float(rank_out) / min(self.mid_features, self.out_features)
frac_avg = 0.5 * (frac_in + frac_out)
if frac_avg > DONT_lora_if_rankFrac_gt:
lora_in = ResidualAdapterLinearOnly(self.in_features, self.mid_features, scaling=1.0, use_bias_delta=True)
lora_out = ResidualAdapterLinearOnly(tout.in_features, tout.out_features, scaling=1.0, use_bias_delta=True)
else:
lora_in = LoRAAdapterLinearOnly(self.in_features, self.mid_features, rank=rank_in, dropout=0.0, scaling=1.0)
lora_out = LoRAAdapterLinearOnly(tout.in_features, tout.out_features, rank=rank_out, dropout=0.0, scaling=1.0)
# init from diffs
if FOR_upcycle_ckpt_GEN_or_USE:
with torch.no_grad():
w_diff_in = tin.weight.data - in_linear.weight.data
b_diff_in = (tin.bias.data - in_linear.bias.data) if tin.bias is not None else None
lora_in.init_from_diff(w_diff_in, b_diff_in)
w_diff_out = tout.weight.data - out_linear.weight.data
b_diff_out = (tout.bias.data - out_linear.bias.data) if tout.bias is not None else None
lora_out.init_from_diff(w_diff_out, b_diff_out)
_l_in.append(lora_in)
_l_out.append(lora_out)
self.task_lora_in = ModuleDict_W(_l_in, self.tasks)
self.task_lora_out = ModuleDict_W(_l_out, self.tasks)
# reuse dropout and activation behavior
self.dropout_p = self.shared_ffn.net[1].p if isinstance(self.shared_ffn.net[1], nn.Dropout) else 0.0
self.dropout = nn.Dropout(self.dropout_p) if self.dropout_p > 0 else nn.Identity()
# Sparse/Dense MoE experts (small inner dim) + gate
if EXTRA_MoE_enable:
small_inner = self.mid_features // EXTRA_MoE_inner_divisor
self.num_moe_expert = EXTRA_MoE_num_ep
gate_in_dim = self.in_features + self.in_features + len(self.tasks) + 2 + 2*NUM_lmk_pick
hidden = gate_in_dim // 8
self.moe_gate_mlp = nn.Sequential(
nn.Linear(gate_in_dim, hidden),
nn.SiLU(),
nn.Linear(hidden, self.num_moe_expert),
)
if EXTRA_MoE_routing_mode == 'dense':
self.moe_experts_batched = BatchedFeedForward(
dim=self.in_features, dim_out=self.out_features,
glu=self.is_glu, dropout=self.dropout_p,
inner_dim=small_inner, num_expert=self.num_moe_expert,
)
else:
mult = small_inner / self.in_features
experts = []
for _ in range(self.num_moe_expert):
expert = FeedForward(self.in_features, dim_out=self.out_features, mult=mult, glu=self.is_glu, dropout=self.dropout_p)
experts.append(expert)
self.moe_experts_list = nn.ModuleList(experts)
if 0: # log MoE gate and expert architecture to file only (no terminal output)
log_dir = Path("4debug/moe_ffn_struc"); log_dir.mkdir(exist_ok=True)
mod_name = self.module_name; log_path = log_dir / f"{mod_name}.txt"
gate_desc = f"GateMLP: Linear({gate_in_dim},{hidden})->SiLU->Linear({hidden},{self.num_moe_expert})"
if EXTRA_MoE_routing_mode == 'dense':
ep_desc = f"BatchedFeedForward(glu={self.is_glu}, num={self.num_moe_expert}, inner={small_inner}, in={self.in_features}, out={self.out_features})"
else:
ep_desc = f"FeedForwardList(glu={self.is_glu}, num={self.num_moe_expert}, inner≈{self.in_features*mult}, in={self.in_features}, out={self.out_features})"
with open(log_path, 'a') as f:
f.write(f"{mod_name} | routing={EXTRA_MoE_routing_mode} | {gate_desc} | {ep_desc}\n")
print(f"{log_path}")
if FOR_upcycle_ckpt_GEN_or_USE:
self.verify_approximation(orig_ffn_list=l_ffn)
def forward(self, x: torch.Tensor, token_pos_grid__cur=None):
t = global_.task
# in linear + LoRA
if self.is_glu:
base = self.shared_ffn.net[0].proj(x)
delta = self.task_lora_in[t](x)
z = base + delta
v, gate = z.chunk(2, dim=-1)
h = v * F.gelu(gate)
else:
base = self.shared_ffn.net[0][0](x)
delta = self.task_lora_in[t](x)
h = F.gelu(base + delta)
h = self.dropout(h)
# out linear + LoRA
y_base = self.shared_ffn.net[2](h)
y_delta = self.task_lora_out[t](h)
y = y_base + y_delta
if EXTRA_MoE_enable:
# gate input
gate_in, _ = build_ffn_gate_input_common(x, token_pos_grid__cur, self.tasks)
scores = self.moe_gate_mlp(gate_in).to(dtype=x.dtype) # b,n,k
if EXTRA_MoE_add_noise and self.training:
scores = scores + torch.randn_like(scores) * EXTRA_MoE_noise_std
v_topk, idx_topk = scores.topk(k=EXTRA_MoE_topK, dim=-1)
if EXTRA_MoE_routing_mode == 'dense':
raise Exception('not carefully checked yet')
else: # sparse: forward only the selected experts and aggregate by top-k weights
if 1: weights_topk = torch.softmax(v_topk, dim=-1) # b,n,topk
else: weights_topk = v_topk # b,n,topk. use top-k expert scores directly as weights
b, n, d = x.shape
dim_out = self.out_features
y_moe_flat = x.new_zeros(b*n, dim_out) # flattened tensor accumulating outputs from all experts (bs*N, D_out)
x_flat = x.reshape(b*n, d) # flatten input tensor (bs*N, D_in)
unique_experts = torch.unique(idx_topk) # set of expert IDs actually selected in this batch
for j in unique_experts.tolist(): # iterate only over active experts
mask_j = (idx_topk == j) # b,n,topk boolean mask indicating which tokens picked expert j
sel_token_mask = mask_j.any(dim=-1) # b,n boolean mask for tokens that selected expert j
if not sel_token_mask.any(): # skip if expert j was not selected by any token
continue
flat_pos = sel_token_mask.view(-1).nonzero(as_tuple=False).squeeze(1) # T_j flattened indices of tokens assigned to expert j
x_sel = x_flat.index_select(0, flat_pos) # T_j,d select those tokens from flattened input
# run expert only on selected tokens (n = T_j)
y_sel = self.moe_experts_list[j](x_sel.view(1, x_sel.shape[0], d)).squeeze(0) # T_j,dim_out expert j handles only its tokens
w_tok = (weights_topk * mask_j).sum(dim=-1).view(-1).index_select(0, flat_pos).unsqueeze(-1) # T_j,1 weights for each token assigned to expert j
y_moe_flat.index_add_(0, flat_pos, w_tok * y_sel) # add weighted expert output back into flattened tensor (in-place)
y = y + y_moe_flat.view(b, n, dim_out) # reshape aggregated MoE output and add back to backbone output
if EXTRA_MoE_en_auxLoss and self.training:
raise Exception('not carefully checked yet')
importance = torch.zeros(self.num_moe_expert, device=scores.device, dtype=weights_topk.dtype)
importance = importance.scatter_add(0, idx_topk.reshape(-1), weights_topk.reshape(-1))
load = torch.zeros(self.num_moe_expert, device=scores.device, dtype=weights_topk.dtype)
load = load.scatter_add(0, idx_topk.reshape(-1), torch.ones_like(weights_topk.reshape(-1)))
k = importance.shape[0]
target_imp = torch.full_like(importance, fill_value=importance.sum() / k)
target_load = torch.full_like(load, fill_value=load.sum() / k)
aux_imp = F.mse_loss(importance, target_imp)
aux_load = F.mse_loss(load, target_load)
aux = 0.5 * (aux_imp + aux_load) * EXTRA_MoE_aux_coef
global_.moe_aux_loss = aux # expose aux loss to the training loop for aggregation
return y
@torch.no_grad()
def verify_approximation(self, num_tokens: int = 16, batch_size: int = 2, orig_ffn_list: list = None):
if EXTRA_MoE_enable: return
device = next(self.shared_ffn.parameters()).device
dtype = next(self.shared_ffn.parameters()).dtype
x = torch.randn(batch_size, num_tokens, self.in_features, device=device, dtype=dtype)
old_task = getattr(global_, 'task', None)
for i,t in enumerate(self.tasks):
_log2(orig_ffn_list[i], [self.task_lora_in[t], self.task_lora_out[t]])
global_.task = t
y_lora = self.forward(x)
y_avg = self.shared_ffn(x)
assert orig_ffn_list is not None, "orig_ffn_list must be provided for verification"
y_orig = orig_ffn_list[i](x)
d_avg = torch.norm((y_avg - y_orig).float()).item()
d_lora = torch.norm((y_lora - y_orig).float()).item()
_log1(f"[FFN verify] task={t} rank_in={self.task_lora_in[t].rank} rank_out={self.task_lora_out[t].rank} L2(avg,orig)={d_avg:.6f} L2(lora,orig)={d_lora:.6f}")
global_.task = old_task
class Linear_Shared_Plus_TaskLoRA(nn.Module):
def __init__(self, l_proj: list, l_task: list, module_name: str = None):
super().__init__()
# _log1(f"-------- {module_name} --------")
assert len(l_proj) >= 1
self.dont_lora = False
p0: nn.Linear = l_proj[0]
# build shared from p0 and load avg
self.shared: nn.Linear = copy.deepcopy(p0)
if FOR_upcycle_ckpt_GEN_or_USE:
avg_sd = _average_state_dict(l_proj)
self.shared.load_state_dict(avg_sd)
for p in self.shared.parameters():
p.requires_grad = False
self.in_features = self.shared.in_features
self.out_features = self.shared.out_features
self.tasks = l_task
# cal/read adaptive rank across tasks
if 1:
if FOR_upcycle_ckpt_GEN_or_USE:
w_diff_list = []
for lin in l_proj:
w_diff_list.append(lin.weight.data - self.shared.weight.data)
if FORCE_SAME_RANK_ACROSS_TASKS:
rank_lin = compute_adaptive_rank_for_linear_diffs(w_diff_list)
global_.moduleName_2_adaRank[module_name] = rank_lin
else:
ranks_lin = compute_adaptive_rank_for_linear_diffs(w_diff_list, per_task=True)
global_.moduleName_2_adaRank[module_name] = ranks_lin
else:
r_info = global_.moduleName_2_adaRank[module_name]
if FORCE_SAME_RANK_ACROSS_TASKS: rank_lin = r_info
else: ranks_lin = r_info
if 1: # fallback decision for Linear
min_dim = min(self.in_features, self.out_features)
if min_dim < DONT_lora_if_dim_lt:
# print(f"[LoRA fallback][Linear] {module_name} {min_dim=} < {DONT_lora_if_dim_lt}")
self.dont_lora = True; return
_l = [] # per-task LoRA adapters
for idx, lin in enumerate(l_proj):
if not FORCE_SAME_RANK_ACROSS_TASKS:
rank_lin = ranks_lin[idx]
frac = float(rank_lin) / min(self.in_features, self.out_features)
if frac > DONT_lora_if_rankFrac_gt:
lora = ResidualAdapterLinearOnly(self.in_features, self.out_features, scaling=1.0, use_bias_delta=True)
else:
lora = LoRAAdapterLinearOnly(self.in_features, self.out_features, rank=rank_lin, dropout=0.0, scaling=1.0)
if FOR_upcycle_ckpt_GEN_or_USE:
with torch.no_grad():
w_diff = lin.weight.data - self.shared.weight.data
b_diff = (lin.bias.data - self.shared.bias.data) if (lin.bias is not None and self.shared.bias is not None) else None
lora.init_from_diff(w_diff, b_diff)
_l.append(lora)
self.task_lora = ModuleDict_W(_l, self.tasks)
if FOR_upcycle_ckpt_GEN_or_USE:
self.verify_approximation(orig_linear_list=l_proj)
def forward(self, x: torch.Tensor) -> torch.Tensor:
y = self.shared(x)
y = y + self.task_lora[global_.task](x)
return y
@torch.no_grad()
def verify_approximation(self, batch_size: int = 2, in_dim_override: int = None, orig_linear_list: list = None):
device = next(self.shared.parameters()).device
dtype = next(self.shared.parameters()).dtype
d_in = self.in_features if in_dim_override is None else in_dim_override
x = torch.randn(batch_size, d_in, device=device, dtype=dtype)
old_task = getattr(global_, 'task', None)
for i,t in enumerate(self.tasks):
_log2(orig_linear_list[i], self.task_lora[t])
global_.task = t
y_lora = self.forward(x)
y_avg = self.shared(x)
assert orig_linear_list is not None, "orig_linear_list must be provided for verification"
y_orig = orig_linear_list[i](x)
d_avg = torch.norm((y_avg - y_orig).float()).item()
d_lora = torch.norm((y_lora - y_orig).float()).item()
_log1(f"[Linear verify] task={t} rank={self.task_lora[t].rank} L2(avg,orig)={d_avg:.6f} L2(lora,orig)={d_lora:.6f}")
global_.task = old_task
class Conv_Shared_Plus_TaskLoRA(nn.Module):
def __init__(self, l_conv: list, l_task: list, module_name: str = None):
super().__init__()
# _log1(f"-------- {module_name} --------")
assert len(l_conv) >= 1
self.dont_lora = False
c0: nn.Conv2d = l_conv[0]
# build shared conv
self.shared = nn.Conv2d(
c0.in_channels, c0.out_channels,
kernel_size=c0.kernel_size, stride=c0.stride,
padding=c0.padding, dilation=c0.dilation,
groups=c0.groups, bias=(c0.bias is not None),
padding_mode=c0.padding_mode,
)
if FOR_upcycle_ckpt_GEN_or_USE:
avg_sd = _average_state_dict(l_conv)
self.shared.load_state_dict(avg_sd)
for p in self.shared.parameters():
p.requires_grad = False
# per-task LoRA
self.tasks = l_task
_l = []
# cal/read adaptive rank across tasks
if 1:
if FOR_upcycle_ckpt_GEN_or_USE:
w_diff_list = []
for c in l_conv:
w_diff_list.append(c.weight.data - self.shared.weight.data)
if FORCE_SAME_RANK_ACROSS_TASKS:
rank_conv = compute_adaptive_rank_for_conv_diffs(w_diff_list)
global_.moduleName_2_adaRank[module_name] = rank_conv
else:
ranks_conv = compute_adaptive_rank_for_conv_diffs(w_diff_list, per_task=True)
global_.moduleName_2_adaRank[module_name] = ranks_conv
else:
r_info = global_.moduleName_2_adaRank[module_name]
if FORCE_SAME_RANK_ACROSS_TASKS: rank_conv = r_info
else: ranks_conv = r_info
if 1: # fallback decision for Conv
kH, kW = self.shared.kernel_size
min_dim = min(self.shared.out_channels, self.shared.in_channels * kH * kW )
if min_dim < DONT_lora_if_dim_lt:
# print(f"[LoRA fallback][Conv] {module_name} {min_dim=} {DONT_lora_if_dim_lt=} (in={self.shared.in_channels}, out={self.shared.out_channels}, k=({kH},{kW}))")
self.dont_lora = True; return
for idx, c in enumerate(l_conv):
if not FORCE_SAME_RANK_ACROSS_TASKS:
rank_conv = ranks_conv[idx]
frac = float(rank_conv) / min(self.shared.out_channels, self.shared.in_channels * kH * kW)
if frac > DONT_lora_if_rankFrac_gt:
lora = ResidualAdapterConv2dOnly(
in_channels=c.in_channels, out_channels=c.out_channels,
kernel_size=c.kernel_size, stride=c.stride,
padding=c.padding, dilation=c.dilation, groups=c.groups,
scaling=1.0, use_bias_delta=True,
)
else:
lora = LoRAAdapterConv2dOnly(
in_channels=c.in_channels, out_channels=c.out_channels,
kernel_size=c.kernel_size, stride=c.stride,
padding=c.padding, dilation=c.dilation, groups=c.groups,
rank=rank_conv, dropout=0.0, scaling=1.0,
)
if FOR_upcycle_ckpt_GEN_or_USE:
with torch.no_grad():
w_diff = c.weight.data - self.shared.weight.data
b_diff = (c.bias.data - self.shared.bias.data) if c.bias is not None and self.shared.bias is not None else None
lora.init_from_diff(w_diff, b_diff)
_l.append(lora)
self.task_lora = ModuleDict_W(_l, self.tasks)
if FOR_upcycle_ckpt_GEN_or_USE:
self.verify_approximation(orig_conv_list=l_conv)
def forward(self, x: torch.Tensor) -> torch.Tensor:
y = self.shared(x)
y = y + self.task_lora[global_.task](x)
return y
@torch.no_grad()
def verify_approximation(self, spatial_hw=(32, 32), batch_size: int = 2, orig_conv_list: list = None):
device = next(self.shared.parameters()).device
dtype = next(self.shared.parameters()).dtype
H, W = spatial_hw
x = torch.randn(batch_size, self.shared.in_channels, H, W, device=device, dtype=dtype)
old_task = getattr(global_, 'task', None)
for i,t in enumerate(self.tasks):
_log2(orig_conv_list[i], self.task_lora[t])
global_.task = t
y_lora = self.forward(x)
y_avg = self.shared(x)
assert orig_conv_list is not None, "orig_conv_list must be provided for verification"
y_orig = orig_conv_list[i](x)
d_avg = torch.norm((y_avg - y_orig).float()).item()
d_lora = torch.norm((y_lora - y_orig).float()).item()
_log1(f"[Conv2d verify] task={t} rank={self.task_lora[t].rank} L2(avg,orig)={d_avg:.6f} L2(lora,orig)={d_lora:.6f}")
global_.task = old_task
|