File size: 18,984 Bytes
d5147ba | 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 | # MODIFIED FILE NOTICE (2026-10-17)
# This file contains project modifications to upstream LOOM commit
# a2ee5a7ecd38b083d9c049d97271ece38f4bce4c for the training/runtime implementation.
# Those modifications are preserved in the original trained-model artifact.
# This distribution adds this notice only; executable Python syntax is
# unchanged from that artifact.
from typing import Tuple
import os
import torch
from torch import nn
from torch import Tensor
import torch.distributed as dist
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
from pydantic import BaseModel
from models.layers import LinearInit, ScaledEmbeddingInit, Carry
from models.common import IGNORE_LABEL_ID, packing_sequence_sum
class LMHeadConfig(BaseModel):
vocab_size: int
class LMHead(nn.Module):
def __init__(self, model: nn.Module, config_dict: dict) -> None:
super().__init__()
self.model = model
# Create cache function
self.create_cache = self.model.create_cache
# Train extra args function
self.compute_train_extra_args = self.model.compute_train_extra_args
config = LMHeadConfig(**config_dict)
head_hint: dict = self.model.head_hint # pyright: ignore[reportAssignmentType]
# LMHead input and output
self.embed_tokens = ScaledEmbeddingInit(config.vocab_size, head_hint["in"]["dim"], init_std=head_hint["in"]["init_std"]) # pyright: ignore[reportArgumentType]
self.lm_head = LinearInit(head_hint["out"]["dim"], config.vocab_size, bias=False, init_std=head_hint["out"]["init_std"]) # pyright: ignore[reportArgumentType]
def _linear_ce_chunk(self, hidden: Tensor, labels: Tensor) -> tuple[Tensor, Tensor]:
logits = self.lm_head(hidden)
loss = F.cross_entropy(
logits.to(torch.float32), labels.to(torch.long),
ignore_index=IGNORE_LABEL_ID, reduction="sum",
)
return loss, logits.detach().argmax(dim=-1)
def _chunked_linear_ce(self, hidden: Tensor, labels: Tensor, chunk_size: int) -> tuple[Tensor, Tensor]:
"""Sum checkpointed packed-token CE without retaining full-vocabulary logits."""
if hidden.shape[0] == 0:
return self._linear_ce_chunk(hidden, labels)
losses = []
predictions = []
for start in range(0, hidden.shape[0], chunk_size):
chunk_hidden = hidden[start:start + chunk_size]
chunk_labels = labels[start:start + chunk_size]
if self.training and torch.is_grad_enabled():
loss, predicted = checkpoint(
self._linear_ce_chunk, chunk_hidden, chunk_labels,
use_reentrant=False,
)
else:
loss, predicted = self._linear_ce_chunk(chunk_hidden, chunk_labels)
losses.append(loss)
predictions.append(predicted)
return torch.stack(losses).sum(), torch.cat(predictions)
def ce_from_logits(self, logits: Tensor, batch: dict[str, Tensor]) -> tuple[Tensor, dict[str, tuple[Tensor, Tensor]]]:
"""CE + metrics from hidden-state logits (used by full forward and segmented BP)."""
labels = batch["labels"]
masks = labels != IGNORE_LABEL_ID
loss = F.cross_entropy(logits.to(torch.float32), labels.to(torch.long), ignore_index=IGNORE_LABEL_ID, reduction="sum")
loss_divisor = masks.sum().to(torch.float32)
dist.all_reduce(loss_divisor, op=dist.ReduceOp.AVG)
normalized_loss = loss / loss_divisor
with torch.no_grad():
is_correct = torch.argmax(logits, dim=-1) == labels
local_valid_counts = masks.sum()
seq_num_tokens_correct = packing_sequence_sum(is_correct, batch["cu_seqlens"])
seq_num_valid_tokens = packing_sequence_sum(masks, batch["cu_seqlens"])
seq_is_valid = seq_num_valid_tokens > 0
metrics = {
"loss": (normalized_loss.detach() * local_valid_counts, local_valid_counts),
"accuracy": (is_correct.sum(), local_valid_counts),
"exact_accuracy": (((seq_num_tokens_correct == seq_num_valid_tokens) & seq_is_valid).sum(), seq_is_valid.sum()),
}
return normalized_loss, metrics
def forward(self, carry: Carry, batch: dict[str, Tensor], **kwargs) -> Tuple[Carry, Tensor] | Tuple[Carry, Tensor, dict[str, Tuple[Tensor, Tensor]]]:
# Token embedding
input_embedding = self.embed_tokens(batch["inputs"])
# Model forward
new_carry, hidden = self.model(carry,
input_embedding,
**{k: v for k, v in batch.items() if k not in ("inputs", "labels")},
**kwargs)
chunk_env = "LM_HEAD_CE_CHUNK_SIZE" if self.training else "LM_HEAD_EVAL_CE_CHUNK_SIZE"
chunk_size = int(os.environ.get(chunk_env, "0"))
chunked_ce = "labels" in batch and chunk_size > 0
if chunked_ce:
loss, predicted = self._chunked_linear_ce(hidden, batch["labels"], chunk_size)
else:
logits = self.lm_head(hidden)
# Loss & Metrics
if "labels" in batch:
# Masks & labels
labels = batch["labels"]
masks = labels != IGNORE_LABEL_ID
# Loss (CE in F32)
if not chunked_ce:
loss = F.cross_entropy(logits.to(torch.float32), labels.to(torch.long), ignore_index=IGNORE_LABEL_ID, reduction="sum")
# AllReduce loss divisor. Divide by mean of valid tokens across all processes, as gradient will be averaged.
loss_divisor = masks.sum().to(torch.float32)
dist.all_reduce(loss_divisor, op=dist.ReduceOp.AVG)
router_z_coef = getattr(self.model, "router_z_loss_coef", 0.0)
z_loss = None
if router_z_coef > 0 and hasattr(self.model, "pop_router_z_loss"):
z_loss = self.model.pop_router_z_loss()
normalized_loss = loss / loss_divisor
if z_loss is not None:
normalized_loss = normalized_loss + router_z_coef * z_loss
# Deep supervision: mean CE over intermediate H-cycle states.
deep_sup_coef = float(getattr(self.model, "deep_sup_coef", 0.0) or 0.0)
deep_sup_loss = None
if (
deep_sup_coef > 0.0
and self.training
and hasattr(self.model, "pop_deep_sup_states")
):
inter = self.model.pop_deep_sup_states()
if inter:
aux_sum = None
for h in inter:
if chunked_ce:
aux_i, _ = self._chunked_linear_ce(h, labels, chunk_size)
else:
aux_logits = self.lm_head(h)
aux_i = F.cross_entropy(
aux_logits.to(torch.float32),
labels.to(torch.long),
ignore_index=IGNORE_LABEL_ID,
reduction="sum",
)
aux_sum = aux_i if aux_sum is None else (aux_sum + aux_i)
deep_sup_loss = (aux_sum / float(len(inter))) / loss_divisor
normalized_loss = normalized_loss + deep_sup_coef * deep_sup_loss
elif hasattr(self.model, "pop_deep_sup_states"):
# Eval / coef=0: drop any stashed states so they don't leak.
self.model.pop_deep_sup_states()
# Scale-visible norm penalty (arXiv:2606.24898): λ · mean_k E[‖Hk‖²_rms].
# ‖v‖²_rms = mean_d(v²); keeps RMSNorm readout but makes scale visible to loss.
norm_pen_coef = float(getattr(self.model, "state_norm_penalty_coef", 0.0) or 0.0)
norm_pen_loss = None
if (
norm_pen_coef > 0.0
and self.training
and hasattr(self.model, "pop_norm_pen_states")
):
states = self.model.pop_norm_pen_states()
if states:
rms2_sum = None
for h in states:
# E[‖h‖²_rms] over tokens; h is [N, D] packed.
rms2 = h.to(torch.float32).pow(2).mean()
rms2_sum = rms2 if rms2_sum is None else (rms2_sum + rms2)
norm_pen_loss = rms2_sum / float(len(states))
normalized_loss = normalized_loss + norm_pen_coef * norm_pen_loss
elif hasattr(self.model, "pop_norm_pen_states"):
self.model.pop_norm_pen_states()
# Loop-Δz RMS match (CoD / residual-scaling): unitless update-norm balance.
# Final-z RMS is LN-flat (rms2 saw match=0); use RMS(z_t−z_{t−1}) instead.
# Full-H: L = (rms_ΔH / rms_Δ2 − 1)². Plus stream-gain:
# λ* = clamp(λ · ratio, λ_min, λ_max); L += (λ − stopgrad(λ*))²
# Plus depth prior (fixed-λ BEST / transfer point ≈1.6), gated by |ratio−1|
# so it only pulls when the stream is imbalanced — not pack-specific LR.
rms_match = bool(getattr(self.model, "cycle_lambda_rms_match", False))
rms_match_loss = None
stream_gain_loss = None
depth_prior_loss = None
if rms_match and self.training:
ref, delta_last = (None, None)
if hasattr(self.model, "pop_lambda_rms_full_h"):
ref, delta_last = self.model.pop_lambda_rms_full_h()
if ref is not None and delta_last is not None:
rms_last = (
delta_last.to(torch.float32).pow(2).mean().clamp_min(1e-12).sqrt()
)
ratio = rms_last / ref
rms_match_loss = (ratio - 1.0).pow(2)
normalized_loss = normalized_loss + rms_match_loss
get_lam = getattr(self.model, "_cycle_lambda_res_live", None)
if not callable(get_lam):
get_lam = getattr(self.model, "_cycle_lambda_res", None)
if callable(get_lam):
lam = get_lam()
if isinstance(lam, torch.Tensor):
lam_f = lam.to(torch.float32).reshape(())
lam_min = float(
getattr(self.model, "cycle_scale_lambda_min", 1.5) or 1.5
)
lam_max = float(
getattr(self.model, "cycle_scale_lambda_max", 0.0) or 0.0
)
if lam_max <= 0.0:
lam_max = lam_min + 2.0
# Operating band (residual-scaling BEST ≈1.6–1.7).
lo = float(
getattr(self.model, "cycle_lambda_depth_prior_lo", 1.58)
or 1.58
)
hi = float(
getattr(self.model, "cycle_lambda_depth_prior_hi", 1.75)
or 1.75
)
# Match residual H-aware hi so prior/stream gates align with detach.
H = float(
max(int(getattr(self.model, "num_loops", 1) or 1), 1)
)
hi = min(hi, 1.65 + 0.10 * (9.0 / H) ** 0.5)
mid = 0.5 * (lo + hi) # ≈1.65
# Climb zone [lo, mid): keep stream-gain so init15 can rise
# (rms7 froze at lo→λ≈1.58→eval 7.2).
# rms7b: gate stream in [mid,hi]; rms8: also gate when λ>hi
# (CE+stream yank caused soft15→1.80 / soft20→2.17 rebound).
mode = str(
getattr(self.model, "cycle_lambda_detach_mode", "rms7b")
or "rms7b"
)
if mode == "rms8":
stream_off = lam_f >= mid
elif mode == "rms9":
latched = getattr(
self.model, "_cycle_lambda_settled_latched", None
)
latched_b = bool(
latched.detach().item()
if isinstance(latched, torch.Tensor)
else bool(latched)
)
stream_off = latched_b or ((lam_f >= mid) & (lam_f <= hi))
elif mode == "rms10":
# Stream-gain off at/above hi (ceiling); keep below mid live.
stream_off = lam_f >= hi
elif mode == "rms11":
# Settled band OR above hi (ceiling zone): no stream yank.
stream_off = (lam_f >= mid) | (lam_f >= hi)
elif mode == "rms12":
# STE ceiling: keep stream/prior ON when λ>hi (pull down);
# gate only in settled [mid,hi].
stream_off = (lam_f >= mid) & (lam_f <= hi)
else:
stream_off = (lam_f >= mid) & (lam_f <= hi)
ratio_d = ratio.detach()
lam_star = (lam_f.detach() * ratio_d).clamp(lam_min, lam_max)
stream_raw = (lam_f - lam_star).pow(2)
stream_gain_loss = torch.where(
stream_off, lam_f.new_zeros(()), stream_raw
)
normalized_loss = normalized_loss + stream_gain_loss
lam_ref = float(
getattr(self.model, "cycle_lambda_depth_prior", 1.6) or 0.0
)
if lam_ref > 0.0:
# Asymmetric operating band:
# λ > hi → (λ−ref)²
# λ < mid → (λ−mid)² (climb through [lo,mid) too)
# mid≤λ≤hi → 0 (settled; stream gated above)
coef = float(
getattr(self.model, "cycle_lambda_depth_prior_coef", 1.0)
or 1.0
)
high_pen = torch.where(
lam_f > hi, (lam_f - lam_ref).pow(2), lam_f.new_zeros(())
)
low_pen = torch.where(
lam_f < mid, (lam_f - mid).pow(2), lam_f.new_zeros(())
)
depth_prior_loss = coef * (high_pen + low_pen)
normalized_loss = normalized_loss + depth_prior_loss
elif hasattr(self.model, "pop_lambda_rms_states"):
# Legacy fallback (non-segmented / short H).
states = self.model.pop_lambda_rms_states()
if len(states) >= 2:
rms_list = [
h.to(torch.float32).pow(2).mean().clamp_min(1e-12).sqrt()
for h in states
]
ref0 = rms_list[0].detach()
acc = None
for r in rms_list[1:]:
term = (r / ref0 - 1.0).pow(2)
acc = term if acc is None else (acc + term)
rms_match_loss = acc / float(len(rms_list) - 1)
normalized_loss = normalized_loss + rms_match_loss
elif hasattr(self.model, "pop_lambda_rms_states"):
self.model.pop_lambda_rms_states()
# Accuracy
with torch.no_grad():
is_correct = (predicted if chunked_ce else torch.argmax(logits, dim=-1)) == labels
local_valid_counts = masks.sum()
# Sequence-level statistics
seq_num_tokens_correct = packing_sequence_sum(is_correct, batch["cu_seqlens"])
seq_num_valid_tokens = packing_sequence_sum(masks, batch["cu_seqlens"])
seq_is_valid = seq_num_valid_tokens > 0
# Metrics
metrics = {
"loss": (normalized_loss.detach() * local_valid_counts, local_valid_counts),
"accuracy": (is_correct.sum(), local_valid_counts),
"exact_accuracy": (((seq_num_tokens_correct == seq_num_valid_tokens) & seq_is_valid).sum(), seq_is_valid.sum()),
}
if z_loss is not None:
metrics["router_z_loss"] = (z_loss.detach(), torch.ones((), device=z_loss.device, dtype=torch.float32))
if deep_sup_loss is not None:
metrics["deep_sup_loss"] = (
deep_sup_loss.detach(),
torch.ones((), device=deep_sup_loss.device, dtype=torch.float32),
)
if norm_pen_loss is not None:
metrics["state_norm_penalty"] = (
norm_pen_loss.detach(),
torch.ones((), device=norm_pen_loss.device, dtype=torch.float32),
)
if rms_match_loss is not None:
metrics["lambda_rms_match"] = (
rms_match_loss.detach(),
torch.ones((), device=rms_match_loss.device, dtype=torch.float32),
)
if stream_gain_loss is not None:
metrics["lambda_stream_gain"] = (
stream_gain_loss.detach(),
torch.ones((), device=stream_gain_loss.device, dtype=torch.float32),
)
if depth_prior_loss is not None:
metrics["lambda_depth_prior"] = (
depth_prior_loss.detach(),
torch.ones((), device=depth_prior_loss.device, dtype=torch.float32),
)
return new_carry, normalized_loss, metrics
return new_carry, logits
|