File size: 20,563 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 | #!/usr/bin/env python3
"""Model and layer helpers for fuse_layers."""
import os
from typing import Dict, List, Optional, Tuple
import torch
try:
from tqdm import tqdm
except Exception: # pragma: no cover - optional dependency
tqdm = None
def _tqdm_enabled() -> bool:
value = os.environ.get("DISABLE_TQDM", os.environ.get("TQDM_DISABLE", "0"))
return value.strip().lower() not in {"1", "true", "yes", "on"}
def get_dtype(dtype: str):
if dtype == "auto":
return None
if dtype == "float16":
return torch.float16
if dtype == "bfloat16":
return torch.bfloat16
return torch.float32
def resolve_attr(root: object, path: str) -> Optional[object]:
cur = root
for part in path.split("."):
if not hasattr(cur, part):
return None
cur = getattr(cur, part)
return cur
def resolve_attr_with_parent(root: object, path: str) -> Tuple[object, str, object]:
parts = path.split(".")
cur = root
for part in parts[:-1]:
if not hasattr(cur, part):
raise ValueError(f"'{path}' not found on model")
cur = getattr(cur, part)
name = parts[-1]
if not hasattr(cur, name):
raise ValueError(f"'{path}' not found on model")
return cur, name, getattr(cur, name)
def find_layer_container(model, layer_path: Optional[str]) -> Tuple[object, str, object]:
if layer_path:
parent, name, container = resolve_attr_with_parent(model, layer_path)
return parent, name, container
candidate_paths = [
"model.layers", # LLaMA, Mistral, Qwen2, Gemma
"model.decoder.layers", # OPT
"transformer.h", # GPT-2, GPT-J, Bloom, Falcon
"transformer.blocks", # MPT
"gpt_neox.layers", # GPT-NeoX
"layers", # fallback
]
for path in candidate_paths:
candidate = resolve_attr(model, path)
if candidate is None:
continue
try:
list(candidate)
except TypeError:
continue
parent, name, container = resolve_attr_with_parent(model, path)
return parent, name, container
raise ValueError(
"Could not locate transformer layers. Pass --layer_path explicitly."
)
def find_attention_module(layer: torch.nn.Module) -> torch.nn.Module:
if hasattr(layer, "self_attn"):
return getattr(layer, "self_attn")
if hasattr(layer, "attn"):
return getattr(layer, "attn")
if hasattr(layer, "attention"):
return getattr(layer, "attention")
for _, module in layer.named_modules():
if all(
hasattr(module, attr) for attr in ("q_proj", "k_proj", "v_proj", "o_proj")
):
return module
raise ValueError("Could not find attention module with q_proj/k_proj/v_proj/o_proj")
def find_mlp_module(layer: torch.nn.Module) -> torch.nn.Module:
if hasattr(layer, "mlp"):
return getattr(layer, "mlp")
for attr in ("feed_forward", "feedforward", "ffn", "ff"):
if hasattr(layer, attr):
return getattr(layer, attr)
for _, module in layer.named_modules():
if all(hasattr(module, attr) for attr in ("gate_proj", "up_proj", "down_proj")):
return module
if all(hasattr(module, attr) for attr in ("fc1", "fc2")):
return module
if all(
hasattr(module, attr)
for attr in ("dense_h_to_4h", "dense_4h_to_h")
):
return module
if all(hasattr(module, attr) for attr in ("w1", "w2")):
return module
raise ValueError("Could not find MLP/FFN module on layer")
def get_head_info(
attn: torch.nn.Module, hidden_size: int, config
) -> Tuple[int, int, int]:
num_heads = getattr(attn, "num_heads", None)
if num_heads is None:
num_heads = getattr(attn, "num_attention_heads", None)
if num_heads is None and config is not None:
num_heads = getattr(
config,
"num_attention_heads",
getattr(config, "num_heads", getattr(config, "n_head", None)),
)
num_key_value_heads = getattr(attn, "num_key_value_heads", None)
if num_key_value_heads is None:
num_key_value_heads = getattr(attn, "num_kv_heads", None)
if num_key_value_heads is None and config is not None:
num_key_value_heads = getattr(
config,
"num_key_value_heads",
getattr(config, "num_kv_heads", getattr(config, "n_head_kv", None)),
)
head_dim = getattr(attn, "head_dim", None)
if head_dim is None and config is not None:
head_dim = getattr(config, "head_dim", None)
if num_heads is None:
if hasattr(attn, "q_proj"):
q_out = attn.q_proj.weight.shape[0]
if head_dim is not None:
num_heads = q_out // head_dim
elif num_key_value_heads is not None and hasattr(attn, "k_proj"):
k_out = attn.k_proj.weight.shape[0]
head_dim = k_out // max(int(num_key_value_heads), 1)
num_heads = q_out // head_dim
if num_heads is None:
raise ValueError(
"Attention module missing num_heads/num_attention_heads; "
"pass --layer_path or add config overrides."
)
if num_key_value_heads is None:
num_key_value_heads = num_heads
if head_dim is None:
head_dim = hidden_size // int(num_heads)
if num_key_value_heads is None and hasattr(attn, "k_proj"):
k_out = attn.k_proj.weight.shape[0]
num_key_value_heads = k_out // int(head_dim)
return int(num_heads), int(num_key_value_heads), int(head_dim)
def cosine_cost_matrix(
a: torch.Tensor, b: torch.Tensor, eps: float = 1e-8
) -> torch.Tensor:
a_norm = a / (a.norm(dim=1, keepdim=True) + eps)
b_norm = b / (b.norm(dim=1, keepdim=True) + eps)
sim = a_norm @ b_norm.t()
return 1.0 - sim
def hungarian(cost: torch.Tensor) -> List[int]:
# Kuhn-Munkres for square cost matrix (minimization).
n = cost.size(0)
u = [0.0] * (n + 1)
v = [0.0] * (n + 1)
p = [0] * (n + 1)
way = [0] * (n + 1)
for i in range(1, n + 1):
p[0] = i
j0 = 0
minv = [float("inf")] * (n + 1)
used = [False] * (n + 1)
while True:
used[j0] = True
i0 = p[j0]
delta = float("inf")
j1 = 0
for j in range(1, n + 1):
if used[j]:
continue
cur = cost[i0 - 1, j - 1].item() - u[i0] - v[j]
if cur < minv[j]:
minv[j] = cur
way[j] = j0
if minv[j] < delta:
delta = minv[j]
j1 = j
for j in range(0, n + 1):
if used[j]:
u[p[j]] += delta
v[j] -= delta
else:
minv[j] -= delta
j0 = j1
if p[j0] == 0:
break
while True:
j1 = way[j0]
p[j0] = p[j1]
j0 = j1
if j0 == 0:
break
assignment = [-1] * n
for j in range(1, n + 1):
if p[j] > 0:
assignment[p[j] - 1] = j - 1
return assignment
def compute_head_means(
model,
attn_i: torch.nn.Module,
attn_j: torch.nn.Module,
dataloader,
device: str,
hidden_size: int,
) -> Tuple[torch.Tensor, torch.Tensor, int, int, int]:
num_heads_i, num_kv_i, head_dim_i = get_head_info(attn_i, hidden_size, model.config)
num_heads_j, num_kv_j, head_dim_j = get_head_info(attn_j, hidden_size, model.config)
if num_heads_i != num_heads_j or head_dim_i != head_dim_j:
raise ValueError("Head counts or head_dim differ between layers; cannot align")
sums_i = torch.zeros(num_heads_i, head_dim_i, device="cpu")
sums_j = torch.zeros(num_heads_j, head_dim_j, device="cpu")
count_i = [0]
count_j = [0]
def make_hook(
sums: torch.Tensor, count_ref: List[int], num_heads: int, head_dim: int
):
def hook(_module, inputs, _output):
hidden = inputs[0].detach()
if hidden.dim() != 3:
return
batch, seq, width = hidden.shape
if width != num_heads * head_dim:
return
reshaped = hidden.view(batch, seq, num_heads, head_dim)
sums.add_(reshaped.sum(dim=(0, 1)).float().cpu())
count_ref[0] += batch * seq
return hook
hook_i = attn_i.o_proj.register_forward_hook(
make_hook(sums_i, count_i, num_heads_i, head_dim_i)
)
hook_j = attn_j.o_proj.register_forward_hook(
make_hook(sums_j, count_j, num_heads_j, head_dim_j)
)
model.eval()
iterator = dataloader
if tqdm is not None and _tqdm_enabled():
iterator = tqdm(dataloader, desc="Head stats", unit="batch")
with torch.no_grad():
for batch in iterator:
input_ids = batch[0].to(device)
_ = model(input_ids=input_ids)
hook_i.remove()
hook_j.remove()
if count_i[0] == 0 or count_j[0] == 0:
raise RuntimeError("Failed to capture head outputs; check attention modules.")
mean_i = sums_i / count_i[0]
mean_j = sums_j / count_j[0]
return mean_i, mean_j, num_heads_i, num_kv_i, head_dim_i
def build_head_permutation(
mean_i: torch.Tensor,
mean_j: torch.Tensor,
num_heads: int,
num_kv_heads: int,
eps: float,
) -> List[int]:
group_size = num_heads // num_kv_heads
if group_size * num_kv_heads != num_heads:
raise ValueError("num_heads must be divisible by num_key_value_heads")
perm = list(range(num_heads))
for g in range(num_kv_heads):
start = g * group_size
end = start + group_size
cost = cosine_cost_matrix(mean_i[start:end], mean_j[start:end], eps=eps)
assignment = hungarian(cost)
for local_idx, match in enumerate(assignment):
perm[start + local_idx] = start + match
return perm
def permute_attention_heads(
attn: torch.nn.Module,
perm: List[int],
num_heads: int,
num_kv_heads: int,
head_dim: int,
) -> None:
hidden_size = num_heads * head_dim
def permute_out_proj_weight(weight: torch.Tensor) -> torch.Tensor:
out_features, in_features = weight.shape
if in_features != hidden_size:
raise ValueError(
"o_proj in_features ({} ) != num_heads*head_dim ({})".format(
in_features, hidden_size
)
)
reshaped = weight.view(out_features, num_heads, head_dim)
reshaped = reshaped[:, perm, :]
return reshaped.reshape(out_features, in_features)
def permute_proj_weight(weight: torch.Tensor) -> torch.Tensor:
out_features, in_features = weight.shape
if out_features != hidden_size:
raise ValueError(
"proj out_features ({}) != num_heads*head_dim ({})".format(
out_features, hidden_size
)
)
reshaped = weight.view(num_heads, head_dim, in_features)
reshaped = reshaped[perm, :, :]
return reshaped.reshape(out_features, in_features)
def permute_proj_bias(bias: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
if bias is None:
return None
reshaped = bias.view(num_heads, head_dim)
reshaped = reshaped[perm, :]
return reshaped.reshape(num_heads * head_dim)
with torch.no_grad():
attn.q_proj.weight.copy_(permute_proj_weight(attn.q_proj.weight))
if attn.q_proj.bias is not None:
attn.q_proj.bias.copy_(permute_proj_bias(attn.q_proj.bias))
if num_kv_heads == num_heads:
attn.k_proj.weight.copy_(permute_proj_weight(attn.k_proj.weight))
if attn.k_proj.bias is not None:
attn.k_proj.bias.copy_(permute_proj_bias(attn.k_proj.bias))
attn.v_proj.weight.copy_(permute_proj_weight(attn.v_proj.weight))
if attn.v_proj.bias is not None:
attn.v_proj.bias.copy_(permute_proj_bias(attn.v_proj.bias))
attn.o_proj.weight.copy_(permute_out_proj_weight(attn.o_proj.weight))
def compute_fisher(
model,
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
dataloader,
fisher_mode: str,
device: str,
) -> Tuple[List[Dict[str, object]], int, List[Dict[str, int]]]:
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: 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] = 0.0
layer_numels[name] = param.numel()
fisher_sums.append(layer_sums)
param_numels.append(layer_numels)
num_batches = 0
model.eval()
iterator = dataloader
if tqdm is not None and _tqdm_enabled():
iterator = tqdm(dataloader, desc="Fisher", unit="batch")
for batch in iterator:
input_ids = batch[0].to(device)
outputs = model(input_ids=input_ids, labels=input_ids)
loss = outputs.loss
loss.backward()
for layer_idx, layer in enumerate((layer_a, layer_b)):
layer_sums = fisher_sums[layer_idx]
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] += float(grad_sq.sum().item())
model.zero_grad(set_to_none=True)
num_batches += 1
if num_batches == 0:
raise RuntimeError("No batches processed; check dataset or text inputs.")
return fisher_sums, num_batches, param_numels
def merge_layers(
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
fisher_a: Dict[str, object],
fisher_b: Dict[str, object],
num_batches: int,
numels_a: Dict[str, int],
numels_b: Dict[str, int],
fisher_mode: str,
eps: float,
) -> int:
merged = 0
params_b = {name: param for name, param in layer_b.named_parameters()}
with torch.no_grad():
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_a[name] / num_batches
fb = fisher_b[name] / num_batches
# Fisher tensors are accumulated on CPU to save VRAM; move to the
# parameter device for the actual merge.
if isinstance(fa, torch.Tensor) and fa.device != param_a.device:
fa = fa.to(param_a.device)
if isinstance(fb, torch.Tensor) and fb.device != param_a.device:
fb = fb.to(param_a.device)
denom = fa + fb
denom_mean = float(denom.mean().item())
if denom_mean <= eps:
merged_param = 0.5 * (param_a.float() + param_b.float())
else:
merged_param = (fa * param_a.float() + fb * param_b.float()) / (
denom + eps
)
else:
fa = fisher_a[name] / (num_batches * numels_a[name])
fb = fisher_b[name] / (num_batches * numels_b[name])
denom = fa + fb
if denom <= eps:
merged_param = 0.5 * (param_a.float() + param_b.float())
else:
merged_param = (
fa * param_a.float() + fb * param_b.float()
) / (denom + eps)
param_a.copy_(merged_param.to(dtype=param_a.dtype))
merged += 1
return merged
def merge_layers_with_gates(
layer_a: torch.nn.Module,
layer_b: torch.nn.Module,
gates: Dict[str, torch.Tensor],
) -> int:
"""Merge layer_b into layer_a using precomputed gates.
Each gate is a lambda in [0, 1] that mixes parameters as:
W = lambda * W_a + (1 - lambda) * W_b
Gate tensors may be scalars (per-tensor gating) or full tensors matching the
parameter shape (per-parameter gating).
"""
merged = 0
params_b = {name: param for name, param in layer_b.named_parameters()}
with torch.no_grad():
for name, param_a in layer_a.named_parameters():
gate = gates.get(name)
if gate is None:
continue
param_b = params_b.get(name)
if param_b is None or param_b.shape != param_a.shape:
continue
lam = gate
if not isinstance(lam, torch.Tensor):
lam = torch.tensor(lam)
if lam.device != param_a.device:
lam = lam.to(param_a.device)
merged_param = lam * param_a.float() + (1.0 - lam) * param_b.float()
param_a.copy_(merged_param.to(dtype=param_a.dtype))
merged += 1
return merged
def drop_layer(container: object, index: int) -> object:
if isinstance(container, torch.nn.ModuleList):
return torch.nn.ModuleList(
[layer for idx, layer in enumerate(container) if idx != index]
)
if isinstance(container, list):
del container[index]
return container
raise TypeError("Layer container must be ModuleList or list")
def decrement_config(config) -> None:
for attr in ("num_hidden_layers", "n_layer", "num_layers"):
if hasattr(config, attr):
value = getattr(config, attr)
if isinstance(value, int) and value > 0:
setattr(config, attr, value - 1)
normalize_config(config)
def normalize_config(config) -> None:
num_hidden_layers = getattr(config, "num_hidden_layers", None)
layer_types = getattr(config, "layer_types", None)
if (
isinstance(num_hidden_layers, int)
and num_hidden_layers >= 0
and isinstance(layer_types, (list, tuple))
and len(layer_types) != num_hidden_layers
):
config.layer_types = list(layer_types[:num_hidden_layers])
def find_colon_modules(module: torch.nn.Module) -> List[str]:
found: List[str] = []
for name, child in module._modules.items():
if ":" in name:
found.append(name)
if isinstance(child, torch.nn.Module):
for sub in find_colon_modules(child):
found.append(f"{name}.{sub}")
return found
def get_norm_pair(
layer: torch.nn.Module,
) -> Tuple[
Optional[torch.nn.Module],
Optional[torch.nn.Module],
Tuple[Optional[str], Optional[str]],
]:
candidates = [
("input_layernorm", "post_attention_layernorm"),
("ln_1", "ln_2"),
("norm1", "norm2"),
("norm_1", "norm_2"),
("layer_norm_1", "layer_norm_2"),
("self_attn_layer_norm", "final_layer_norm"),
]
for n1, n2 in candidates:
if hasattr(layer, n1) and hasattr(layer, n2):
return getattr(layer, n1), getattr(layer, n2), (n1, n2)
return None, None, (None, None)
def clone_state_dict(module: torch.nn.Module) -> Dict[str, torch.Tensor]:
return {k: v.detach().clone() for k, v in module.state_dict().items()}
def apply_norm_policy(
layer: torch.nn.Module,
norm_policy: str,
norm1_state: Optional[Dict[str, torch.Tensor]],
norm2_state: Optional[Dict[str, torch.Tensor]],
norm_names: Tuple[Optional[str], Optional[str]],
) -> None:
norm1, norm2, _ = get_norm_pair(layer)
if norm_policy in {"copy_n1", "hybrid"} and norm1_state is not None and norm1 is not None:
norm1.load_state_dict(norm1_state)
if norm_policy == "copy_n1_n2" and norm2_state is not None and norm2 is not None:
norm2.load_state_dict(norm2_state)
|