File size: 34,190 Bytes
d303b8b | 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 | """
v11 model — same trunk as v8 so we can warm-start from v8's final checkpoint.
The architecture differences vs v8 are the prediction heads:
v8: reg_head = Linear(d_model, 2) # mean, log_var
v8: cls_head = Linear(d_model, max_classes)
v11: reg_head = BarDistributionHead(d_model, n_bins=1024)
v11: cls_head = BinClassificationHead(d_model, max_classes=10)
Everything else (feature_weights, y_embed, class_embed, type_embed,
shared_layers, reg_layers, cls_layers, *_norm) keeps the same module
names and parameter shapes, so:
v11_model.load_state_dict(v8_ckpt, strict=False)
will load the trunk and leave only the heads as randomly-initialized.
The v11 trainer's head-warmup phase trains only the heads + reg_norm /
cls_norm for the first 5k steps, exactly as v10 did.
Tokenization is identical to v8: 2D grid [B, n_rows, n_cols, d_model]
with one token per cell. Each layer alternates feature-attention (within
a row) and datapoint-attention (within a column with the
context-vs-query mask).
For now, v11 SKIPS v8's metadata conditioning (the column-statistics
encoder). The v11 plan defers architectural cleanups to v13; the goal
here is data-prior work, not arch work. Once warm-started, the
metadata-related parameters in the v8 ckpt are simply ignored.
"""
from __future__ import annotations
import math
from dataclasses import dataclass
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint as grad_checkpoint
from .heads import (
BarDistributionHead,
BinClassificationHead,
bar_distribution_loss,
cls_masked_loss,
standardize_y_per_task,
decode_bar_distribution,
cls_predict,
)
# ─── config ──────────────────────────────────────────────────────────────────
@dataclass
class V11Config:
d_model: int = 256
n_layers: int = 12 # 8 shared + 4 task-specific per branch
n_heads: int = 8
d_ffn: int = 1024
dropout: float = 0.0
max_features: int = 128 # warm-start slices v8's feature_weights[500] → [128] in warm_start_from_v8
max_classes: int = 10
max_context: int = 1024
max_query: int = 256
n_periodic_freqs: int = 8
n_bins: int = 1024
cls_label_smoothing: float = 0.05
# v11.0.6-tiny architecture toggles. Defaults preserve v11.0 behavior so
# existing ckpts load unchanged via warm_start_from_v8 / strict=False.
mlp_variant: str = "gelu" # "gelu" (legacy) or "swiglu"
norm_variant: str = "layernorm" # "layernorm" (legacy) or "rmsnorm"
# ALBERT-style cross-layer parameter sharing. share_factor>1 means the
# `n_layers`-deep stack uses only `n_layers // share_factor` UNIQUE
# modules; each unique block is applied `share_factor` times via index
# cycling. share_factor=1 = legacy (no sharing).
share_factor: int = 1
def v11_default_config() -> V11Config:
return V11Config()
# ─── v11.0.6-tiny blocks (drop-in upgrades behind config flag) ──────────────
class SwiGLUFFN(nn.Module):
"""SwiGLU MLP (Shazeer 2020, arXiv 2002.05202). Default in PaLM/LLaMA.
Pattern: Linear(d, 8d/3) gate + Linear(d, 8d/3) value, silu*gate, Linear(8d/3, d).
Hidden dim scaled to (8/3)d_ffn/4 = (2/3)d_ffn to hold param count constant
vs the legacy GELU FFN (Linear(d, d_ffn), GELU, Linear(d_ffn, d)).
"""
def __init__(self, d_model: int, d_ffn: int):
super().__init__()
# Match legacy FFN's parameter count: legacy is 2 * d_model * d_ffn.
# SwiGLU is 3 linears (gate, value, out), each d_model * d_hidden.
# So set d_hidden = (2/3) * d_ffn for parity.
d_hidden = int(round(d_ffn * 2 / 3))
self.w_gate = nn.Linear(d_model, d_hidden, bias=False)
self.w_value = nn.Linear(d_model, d_hidden, bias=False)
self.w_out = nn.Linear(d_hidden, d_model, bias=False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.w_out(F.silu(self.w_gate(x)) * self.w_value(x))
class RMSNorm(nn.Module):
"""Root Mean Square Layer Norm (Zhang & Sennrich 2019). LLaMA default.
No mean subtraction, no learned bias. Cheaper than LayerNorm; works as
a drop-in for transformer pre-norm.
"""
def __init__(self, d_model: int, eps: float = 1e-6):
super().__init__()
self.weight = nn.Parameter(torch.ones(d_model))
self.eps = eps
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.weight * (x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps))
def _build_ffn(d_model: int, d_ffn: int, variant: str = "gelu") -> nn.Module:
"""Factory: return GELU MLP (legacy) or SwiGLU MLP based on variant."""
if variant == "swiglu":
return SwiGLUFFN(d_model, d_ffn)
return nn.Sequential(
nn.Linear(d_model, d_ffn),
nn.GELU(),
nn.Linear(d_ffn, d_model),
)
def _build_norm(d_model: int, variant: str = "layernorm") -> nn.Module:
"""Factory: return LayerNorm (legacy) or RMSNorm based on variant."""
if variant == "rmsnorm":
return RMSNorm(d_model)
return nn.LayerNorm(d_model)
# ─── blocks (verbatim from v8 so state_dict keys match) ───────────────────────
class FlashPreLNAttention(nn.Module):
"""Pre-LN attention + FFN using F.scaled_dot_product_attention (Flash)."""
def __init__(self, d_model: int, n_heads: int, d_ffn: int, dropout: float = 0.0,
mlp_variant: str = "gelu", norm_variant: str = "layernorm"):
super().__init__()
self.n_heads = n_heads
self.head_dim = d_model // n_heads
self.d_model = d_model
self.norm1 = _build_norm(d_model, norm_variant)
self.q_proj = nn.Linear(d_model, d_model)
self.k_proj = nn.Linear(d_model, d_model)
self.v_proj = nn.Linear(d_model, d_model)
self.o_proj = nn.Linear(d_model, d_model)
self.norm2 = _build_norm(d_model, norm_variant)
self.ffn = _build_ffn(d_model, d_ffn, mlp_variant)
def _heads(self, x: torch.Tensor) -> torch.Tensor:
B, S, _ = x.shape
return x.view(B, S, self.n_heads, self.head_dim).transpose(1, 2)
def forward(
self,
x: torch.Tensor,
key_padding_mask: Optional[torch.Tensor] = None,
attn_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
residual = x
x = self.norm1(x)
q = self._heads(self.q_proj(x))
k = self._heads(self.k_proj(x))
v = self._heads(self.v_proj(x))
sdpa_mask = None
if attn_mask is not None:
# attn_mask may be 2D [seq, seq] (shared across batch) or 3D [B, seq, seq]
if attn_mask.dim() == 2:
amask = torch.zeros_like(attn_mask, dtype=q.dtype)
amask.masked_fill_(attn_mask, float("-inf"))
sdpa_mask = amask.unsqueeze(0).unsqueeze(0) # [1,1,seq,seq]
else:
amask = torch.zeros_like(attn_mask, dtype=q.dtype)
amask.masked_fill_(attn_mask, float("-inf"))
sdpa_mask = amask.unsqueeze(1) # [B,1,seq,seq]
if key_padding_mask is not None:
pad_mask = torch.zeros(
key_padding_mask.shape[0], 1, 1, key_padding_mask.shape[1],
dtype=q.dtype, device=q.device,
)
pad_mask.masked_fill_(key_padding_mask.unsqueeze(1).unsqueeze(2), float("-inf"))
sdpa_mask = pad_mask if sdpa_mask is None else sdpa_mask + pad_mask
attn_out = F.scaled_dot_product_attention(q, k, v, attn_mask=sdpa_mask, dropout_p=0.0)
attn_out = attn_out.transpose(1, 2).contiguous().view(x.shape[0], x.shape[1], self.d_model)
x = self.o_proj(attn_out) + residual
residual = x
x = self.norm2(x)
x = self.ffn(x) + residual
return x
class AlternatingLayerV8(nn.Module):
"""Feature attention (within rows) → Datapoint attention (within cols).
Name matches v8 verbatim so state_dict keys align for warm-start.
"""
def __init__(self, d_model: int, n_heads: int, d_ffn: int, dropout: float = 0.0,
mlp_variant: str = "gelu", norm_variant: str = "layernorm"):
super().__init__()
self.feature_attn = FlashPreLNAttention(d_model, n_heads, d_ffn, dropout,
mlp_variant=mlp_variant, norm_variant=norm_variant)
self.datapoint_attn = FlashPreLNAttention(d_model, n_heads, d_ffn, dropout,
mlp_variant=mlp_variant, norm_variant=norm_variant)
def forward(
self,
x: torch.Tensor, # [B, n_rows, n_cols, d_model]
feature_pad_mask: torch.Tensor,
datapoint_mask: torch.Tensor, # [n_rows, n_rows] OR [B, n_rows, n_rows]
) -> torch.Tensor:
B, n_rows, n_cols, d_model = x.shape
# within-row feature attn
x_feat = x.reshape(B * n_rows, n_cols, d_model)
feat_pad = feature_pad_mask.unsqueeze(1).expand(B, n_rows, n_cols).reshape(B * n_rows, n_cols)
x_feat = self.feature_attn(x_feat, key_padding_mask=feat_pad)
x = x_feat.reshape(B, n_rows, n_cols, d_model)
# within-col datapoint attn — expand per-batch mask along n_cols if needed
x_data = x.permute(0, 2, 1, 3).reshape(B * n_cols, n_rows, d_model)
if datapoint_mask.dim() == 3:
# [B, n_rows, n_rows] → [B*n_cols, n_rows, n_rows]
dp_mask = (
datapoint_mask.unsqueeze(1)
.expand(B, n_cols, n_rows, n_rows)
.reshape(B * n_cols, n_rows, n_rows)
)
else:
dp_mask = datapoint_mask
x_data = self.datapoint_attn(x_data, attn_mask=dp_mask)
x = x_data.reshape(B, n_cols, n_rows, d_model).permute(0, 2, 1, 3)
return x
# ─── numerical-value embedding (matches v8's NumericalFeatureEmbedding) ──────
class NumericalFeatureEmbedding(nn.Module):
"""Embed a scalar numerical value into a d_model vector via Fourier features."""
def __init__(self, d_model: int = 256, n_freqs: int = 8):
super().__init__()
self.d_model = d_model
self.n_freqs = n_freqs
freqs = 2.0 ** torch.arange(n_freqs, dtype=torch.float32)
self.register_buffer("freqs", freqs)
in_dim = 1 + 1 + 2 * n_freqs # sign + log_mag + sin/cos at each freq
self.mlp = nn.Sequential(
nn.Linear(in_dim, d_model),
nn.GELU(),
nn.Linear(d_model, d_model),
)
self.missing_token = nn.Parameter(torch.randn(d_model) * 0.02)
def forward(self, values: torch.Tensor, mask: Optional[torch.Tensor] = None) -> torch.Tensor:
sign = torch.sign(values)
log_mag = torch.log1p(torch.abs(values))
# Sinusoidal features at multiple frequencies
f = self.freqs.to(values.device).view(*([1] * (values.dim() - 1)), self.n_freqs)
scaled = values.unsqueeze(-1) * f
sins = torch.sin(scaled)
coss = torch.cos(scaled)
feats = torch.cat([sign.unsqueeze(-1), log_mag.unsqueeze(-1), sins, coss], dim=-1)
emb = self.mlp(feats)
if mask is not None:
emb = torch.where(mask.unsqueeze(-1), self.missing_token.expand_as(emb), emb)
return emb
# ─── main v11 model ──────────────────────────────────────────────────────────
@dataclass
class V11Output:
"""Single forward pass output."""
reg_logits: Optional[torch.Tensor] = None # [B, n_query, n_bins] for reg
cls_logits: Optional[torch.Tensor] = None # [B, n_query, max_classes] for cls
y_mean: Optional[torch.Tensor] = None # [B] context y mean (reg only)
y_std: Optional[torch.Tensor] = None # [B] context y std (reg only)
class PredictLMv11(nn.Module):
"""
v11 model: same trunk as v8, new heads.
Forward returns either reg_logits (for regression) or cls_logits (for
classification). For mixed-batch joint training, the trainer should
call the model twice — once with task_type='regression' and once with
task_type='classification' — sharing the trunk pass via gradient
accumulation. (Per-batch-element task_type would require padding to
a max-class shape and we keep it simple.)
State-dict keys match v8's PredictLMv8 exactly EXCEPT:
- reg_head (Linear → BarDistributionHead.mlp)
- cls_head (Linear → BinClassificationHead.mlp)
All other keys load via load_state_dict(strict=False).
"""
def __init__(self, cfg: V11Config = None):
super().__init__()
cfg = cfg or v11_default_config()
self.cfg = cfg
# Toggle gradient checkpointing. Default True (memory-conservative,
# for H100/T4 sized batches). On A100 80GB we can disable for ~2-3×
# throughput when memory permits. Set via `model.use_grad_checkpoint = False`.
self.use_grad_checkpoint = True
# Per-feature projection (same as v8)
self.feature_weights = nn.Parameter(torch.randn(cfg.max_features, cfg.d_model) * 0.02)
self.feature_biases = nn.Parameter(torch.zeros(cfg.max_features, cfg.d_model))
# y embeddings
self.y_embed = NumericalFeatureEmbedding(cfg.d_model, n_freqs=cfg.n_periodic_freqs)
self.class_embed = nn.Embedding(cfg.max_classes, cfg.d_model)
nn.init.normal_(self.class_embed.weight, std=0.02)
# tokens
self.query_token = nn.Parameter(torch.randn(cfg.d_model) * 0.02)
self.type_embed = nn.Embedding(2, cfg.d_model)
nn.init.normal_(self.type_embed.weight, std=0.02)
self.col_type_embed = nn.Embedding(2, cfg.d_model)
nn.init.normal_(self.col_type_embed.weight, std=0.02)
# trunk: 8 shared + 4 reg + 4 cls
# v11.0.6-tiny: variant flags flow through to FFN/norm choice; defaults
# preserve v11.0 layout for backward-compat with existing ckpts.
mv = getattr(cfg, "mlp_variant", "gelu")
nv = getattr(cfg, "norm_variant", "layernorm")
share = max(1, int(getattr(cfg, "share_factor", 1)))
_layer = lambda: AlternatingLayerV8(
cfg.d_model, cfg.n_heads, cfg.d_ffn, cfg.dropout,
mlp_variant=mv, norm_variant=nv,
)
n_shared = cfg.n_layers - 4
# Under share_factor>1, build only n//share unique blocks; the
# forward pass cycles through them. n_shared and n_branch (=4) must
# both be divisible by share_factor.
if n_shared % share != 0 or 4 % share != 0:
raise ValueError(
f"share_factor={share} must divide both n_shared={n_shared} and 4 (branch layers)"
)
n_shared_unique = n_shared // share
n_branch_unique = 4 // share
self.shared_layers = nn.ModuleList([_layer() for _ in range(n_shared_unique)])
self.reg_layers = nn.ModuleList([_layer() for _ in range(n_branch_unique)])
self.cls_layers = nn.ModuleList([_layer() for _ in range(n_branch_unique)])
self.shared_norm = _build_norm(cfg.d_model, nv)
self.reg_norm = _build_norm(cfg.d_model, nv)
self.cls_norm = _build_norm(cfg.d_model, nv)
# Stored for forward to know how many depth-passes to do.
self.effective_n_shared = n_shared
self.effective_n_branch = 4
# v11 heads
self.reg_head = BarDistributionHead(
d_model=cfg.d_model, n_bins=cfg.n_bins, dropout=cfg.dropout,
)
self.cls_head = BinClassificationHead(
d_model=cfg.d_model, max_classes=cfg.max_classes, dropout=cfg.dropout,
)
# NOTE: v8's `log_var_reg` / `log_var_cls` Kendall-style task weights
# are intentionally NOT instantiated here. They were declared but
# never read in the v11 trainer, and ratio-balancing reg/cls via
# alternation + curriculum bias is sufficient at this scale per
# Expert 4. If they appear in a v8 checkpoint, `warm_start_from_v8`
# filters them out via `strict=False` (they land in `unexpected_keys`).
self._init_weights()
def _init_weights(self):
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight)
if m.bias is not None:
nn.init.zeros_(m.bias)
# ──────────────────────────────────────────────────────────────
# Internal: build the [B, n_rows, n_cols, d_model] grid
# ──────────────────────────────────────────────────────────────
def _build_grid(
self,
X_ctx: torch.Tensor, # [B, n_ctx, n_features]
y_ctx: torch.Tensor, # [B, n_ctx]
X_query: torch.Tensor, # [B, n_query, n_features]
feature_mask: torch.Tensor, # [B, n_features] bool, True=padded
task_type: str,
ctx_row_mask: Optional[torch.Tensor] = None, # [B, n_ctx] bool, True=padded
query_row_mask: Optional[torch.Tensor] = None, # [B, n_query] bool, True=padded
):
B, n_ctx, n_features = X_ctx.shape
n_query = X_query.shape[1]
n_rows = n_ctx + n_query
max_f = self.cfg.max_features
device = X_ctx.device
# Effective feature count
if feature_mask.any():
real_per_item = (~feature_mask).sum(dim=1)
n_real = min(int(real_per_item.max().item()), max_f)
else:
n_real = min(n_features, max_f)
n_real = max(n_real, 2)
n_cols = n_real + 1
X_all = torch.cat([X_ctx, X_query], dim=1) # [B, n_rows, n_features]
X_real = X_all[:, :, :n_real] # [B, n_rows, n_real]
# Per-feature projection
feat_grid = (
X_real.unsqueeze(-1) * self.feature_weights[:n_real]
+ self.feature_biases[:n_real]
) # [B, n_rows, n_real, d_model]
# Target column embedding
if task_type == "classification":
y_clamped = y_ctx.long().clamp(0, self.cfg.max_classes - 1)
y_emb_ctx = self.class_embed(y_clamped) # [B, n_ctx, d_model]
else:
y_emb_ctx = self.y_embed(y_ctx.float()) # [B, n_ctx, d_model]
y_emb_q = self.query_token.unsqueeze(0).unsqueeze(0).expand(B, n_query, -1)
y_emb = torch.cat([y_emb_ctx, y_emb_q], dim=1).unsqueeze(2) # [B, n_rows, 1, d_model]
grid = torch.cat([feat_grid, y_emb], dim=2) # [B, n_rows, n_cols, d_model]
# Type (ctx vs query) and column-type (feature vs target) embeds
type_ids = torch.zeros(B, n_rows, dtype=torch.long, device=device)
type_ids[:, n_ctx:] = 1
grid = grid + self.type_embed(type_ids).unsqueeze(2)
col_types = torch.zeros(n_cols, dtype=torch.long, device=device)
col_types[-1] = 1
grid = grid + self.col_type_embed(col_types).unsqueeze(0).unsqueeze(0)
# Feature-pad mask
feature_pad_mask = torch.zeros(B, n_cols, dtype=torch.bool, device=device)
if feature_mask.shape[1] >= n_real:
feature_pad_mask[:, :n_real] = feature_mask[:, :n_real]
# Datapoint mask: query rows can't attend to other query rows (they each
# predict independently). If ctx_row_mask / query_row_mask are provided,
# padded rows are also blocked from being keys (per-batch [B, n_rows, n_rows]).
# Without row-pad masks, build the simple [n_rows, n_rows] shared mask.
if ctx_row_mask is None and query_row_mask is None:
datapoint_mask = torch.zeros(n_rows, n_rows, dtype=torch.bool, device=device)
datapoint_mask[n_ctx:, n_ctx:] = True
for i in range(n_query):
datapoint_mask[n_ctx + i, n_ctx + i] = False
else:
row_pad = torch.zeros(B, n_rows, dtype=torch.bool, device=device)
if ctx_row_mask is not None:
row_pad[:, :n_ctx] = ctx_row_mask
if query_row_mask is not None:
row_pad[:, n_ctx:] = query_row_mask
# base [n_rows, n_rows] block-mask: query↔query disallowed except diag
base = torch.zeros(n_rows, n_rows, dtype=torch.bool, device=device)
base[n_ctx:, n_ctx:] = True
for i in range(n_query):
base[n_ctx + i, n_ctx + i] = False
base = base.unsqueeze(0).expand(B, n_rows, n_rows).clone()
# block any KEY row that is padded (broadcast over queries)
base = base | row_pad.unsqueeze(1).expand(B, n_rows, n_rows)
datapoint_mask = base
return grid, feature_pad_mask, datapoint_mask, n_ctx
# ──────────────────────────────────────────────────────────────
# Forward
# ──────────────────────────────────────────────────────────────
def forward(
self,
X_ctx: torch.Tensor,
y_ctx: torch.Tensor,
X_query: torch.Tensor,
feature_mask: torch.Tensor,
task_type: str = "regression",
ctx_row_mask: Optional[torch.Tensor] = None,
query_row_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Returns logits over bins (reg) or classes (cls).
For regression, the trainer is responsible for calling
`standardize_y_per_task(y_ctx_orig)` BEFORE this forward to obtain
the standardized y_ctx (and stash mean/std for un-standardization).
Optional ctx_row_mask / query_row_mask (bool, True=padded row)
block padded rows from attention as keys, preventing
zero-padded fake-context contamination.
"""
grid, feat_pad, dp_mask, n_ctx = self._build_grid(
X_ctx, y_ctx, X_query, feature_mask, task_type,
ctx_row_mask=ctx_row_mask, query_row_mask=query_row_mask,
)
# Shared trunk. Under share_factor>1, len(self.shared_layers) may be
# < effective_n_shared; cycle via modulo index (ALBERT pattern).
n_uniq_shared = len(self.shared_layers)
for i in range(self.effective_n_shared):
layer = self.shared_layers[i % n_uniq_shared]
if self.training and torch.is_grad_enabled() and self.use_grad_checkpoint:
grid = grad_checkpoint(layer, grid, feat_pad, dp_mask, use_reentrant=False)
else:
grid = layer(grid, feat_pad, dp_mask)
grid = self.shared_norm(grid)
# Task-specific layers
if task_type == "regression":
h = grid
n_uniq_branch = len(self.reg_layers)
for i in range(self.effective_n_branch):
layer = self.reg_layers[i % n_uniq_branch]
if self.training and torch.is_grad_enabled() and self.use_grad_checkpoint:
h = grad_checkpoint(layer, h, feat_pad, dp_mask, use_reentrant=False)
else:
h = layer(h, feat_pad, dp_mask)
h = self.reg_norm(h)
query_target = h[:, n_ctx:, -1, :] # [B, n_query, d_model]
return self.reg_head(query_target) # [B, n_query, n_bins]
# classification — symmetric grad flow with reg path. Earlier
# versions had `h = 0.5*grid + 0.5*grid.detach()` here, which
# halved the cls branch's gradient into the shared trunk while
# the reg branch passed full gradient. Combined with bar-dist
# reg loss being ~3× larger by magnitude than cls (ln(1024) vs
# ln(10)) and 50/50 step alternation, the trunk was receiving
# ~6× more reg signal than cls signal per step. Removed.
h = grid
n_uniq_branch = len(self.cls_layers)
for i in range(self.effective_n_branch):
layer = self.cls_layers[i % n_uniq_branch]
if self.training and torch.is_grad_enabled() and self.use_grad_checkpoint:
h = grad_checkpoint(layer, h, feat_pad, dp_mask, use_reentrant=False)
else:
h = layer(h, feat_pad, dp_mask)
h = self.cls_norm(h)
query_target = h[:, n_ctx:, -1, :]
return self.cls_head(query_target) # [B, n_query, max_classes]
# ──────────────────────────────────────────────────────────────
# Convenience: warm-start from v8 checkpoint
# ──────────────────────────────────────────────────────────────
@torch.no_grad()
def warm_start_from_v8(self, v8_state_dict: dict, verbose: bool = True) -> dict:
"""Load v8 trunk weights, leave heads at random init.
Args:
v8_state_dict: a v8 checkpoint's state_dict
Returns:
dict with `loaded`, `missing`, `unexpected` key counts
"""
# Filter out v8's old reg_head / cls_head (shape-incompatible) and
# the dead log_var weights (removed in v11).
skip_prefixes = ("reg_head.", "cls_head.", "log_var_reg", "log_var_cls")
filtered = {
k: v for k, v in v8_state_dict.items()
if not k.startswith(skip_prefixes)
}
# Slice feature_weights / feature_biases if v8 ckpt has more features
# than v11's max_features (v8 used 500, v11 default 128 for VRAM).
# Keep the first N rows (v8 trained on tasks that primarily used the
# earliest column slots).
target_max = self.cfg.max_features
for k in ("feature_weights", "feature_biases"):
if k in filtered and filtered[k].shape[0] > target_max:
filtered[k] = filtered[k][:target_max]
result = self.load_state_dict(filtered, strict=False)
if verbose:
print(f"[v11.warm_start_from_v8] loaded {len(filtered)} keys")
if result.missing_keys:
print(f" missing ({len(result.missing_keys)}): {result.missing_keys[:5]}…")
if result.unexpected_keys:
print(f" unexpected ({len(result.unexpected_keys)}): {result.unexpected_keys[:5]}…")
return {
"loaded": len(filtered),
"missing": len(result.missing_keys),
"unexpected": len(result.unexpected_keys),
}
@torch.no_grad()
def warm_start_slice_from_v11(self, v11_state_dict: dict, verbose: bool = True) -> dict:
"""Initialize this (smaller) model from a v11.0 ckpt by SLICING layers.
Used when this model has `share_factor > 1`: the v11.0 trunk has
`n_layers` unique blocks, but this model has only `n_layers /
share_factor` unique blocks (each used `share_factor` times via
cycling). We copy every-`share_factor`-th v11.0 block into the
student's unique-blocks list.
Non-layer modules (feature_weights, y_embed, class_embed, query_token,
col_type_embed, shared_norm/reg_norm/cls_norm, reg_head, cls_head)
copy verbatim — they're share-factor-independent.
Requires this model use legacy (gelu + layernorm) MLP/norm variants
for the layer slicing to be shape-compatible.
"""
if self.cfg.mlp_variant != "gelu" or self.cfg.norm_variant != "layernorm":
raise ValueError(
"warm_start_slice_from_v11 requires mlp_variant=gelu, "
"norm_variant=layernorm for shape compatibility with v11.0 ckpt. "
f"Got mlp_variant={self.cfg.mlp_variant}, norm_variant={self.cfg.norm_variant}."
)
share = max(1, int(self.cfg.share_factor))
# Build the source→target index map for layer slicing.
# v11.0 trunk: 8 shared + 4 reg + 4 cls
v11_n_shared = self.cfg.n_layers - 4 # 8 typically
v11_n_branch = 4
# Student unique counts
s_n_shared = v11_n_shared // share
s_n_branch = v11_n_branch // share
# Pick every share-th index from v11.0
shared_src = list(range(0, v11_n_shared, share))[:s_n_shared]
branch_src = list(range(0, v11_n_branch, share))[:s_n_branch]
new_state = {}
layer_keys_copied = 0
non_layer_keys_copied = 0
for k, v in v11_state_dict.items():
# Layer-keyed weights: rewrite the layer index per the slicing map.
if k.startswith("shared_layers."):
# k = "shared_layers.<idx>.<rest>"
parts = k.split(".", 2)
src_idx = int(parts[1])
if src_idx in shared_src:
tgt_idx = shared_src.index(src_idx)
new_state[f"shared_layers.{tgt_idx}.{parts[2]}"] = v
layer_keys_copied += 1
elif k.startswith("reg_layers."):
parts = k.split(".", 2)
src_idx = int(parts[1])
if src_idx in branch_src:
tgt_idx = branch_src.index(src_idx)
new_state[f"reg_layers.{tgt_idx}.{parts[2]}"] = v
layer_keys_copied += 1
elif k.startswith("cls_layers."):
parts = k.split(".", 2)
src_idx = int(parts[1])
if src_idx in branch_src:
tgt_idx = branch_src.index(src_idx)
new_state[f"cls_layers.{tgt_idx}.{parts[2]}"] = v
layer_keys_copied += 1
else:
# Non-layer weights copy verbatim.
new_state[k] = v
non_layer_keys_copied += 1
result = self.load_state_dict(new_state, strict=False)
param_names = {n for n, _ in self.named_parameters()}
missing_params = [k for k in result.missing_keys if k in param_names]
if verbose:
print(f"[v11.warm_start_slice] share_factor={share}, slice indices: "
f"shared={shared_src}, branch={branch_src}")
print(f" copied {layer_keys_copied} layer-keys + {non_layer_keys_copied} non-layer keys")
if missing_params:
print(f" WARN: {len(missing_params)} trainable params unmatched: "
f"{missing_params[:5]}{'...' if len(missing_params) > 5 else ''}")
if result.unexpected_keys:
print(f" ignored {len(result.unexpected_keys)} unexpected keys (e.g., v11.0 layers we didn't slice)")
return {
"share_factor": share,
"layer_keys_copied": layer_keys_copied,
"non_layer_keys_copied": non_layer_keys_copied,
"missing_params": len(missing_params),
"unexpected": len(result.unexpected_keys),
}
def count_params(model: nn.Module) -> int:
return sum(p.numel() for p in model.parameters() if p.requires_grad)
# ─── self-test: forward pass shapes + warm-start sanity ───────────────────────
if __name__ == "__main__":
torch.manual_seed(0)
cfg = V11Config()
model = PredictLMv11(cfg)
print(f"v11 model: {count_params(model)/1e6:.1f}M params (cfg={cfg})")
B, n_ctx, n_q, n_f = 2, 64, 16, 8
X_ctx = torch.randn(B, n_ctx, n_f)
y_ctx = torch.randn(B, n_ctx)
X_q = torch.randn(B, n_q, n_f)
feat_mask = torch.zeros(B, n_f, dtype=torch.bool)
# Regression path
reg_logits = model(X_ctx, y_ctx, X_q, feat_mask, task_type="regression")
print(f"[reg] logits shape: {tuple(reg_logits.shape)} (expected (2,16,1024))")
assert reg_logits.shape == (B, n_q, cfg.n_bins)
loss = bar_distribution_loss(reg_logits, y_ctx[:, :n_q], model.reg_head)
print(f"[reg] uniform-prior loss: {loss.item():.3f} (≈ ln(1024) = 6.93)")
# Classification path
y_ctx_cls = torch.randint(0, 5, (B, n_ctx))
cls_logits = model(X_ctx, y_ctx_cls, X_q, feat_mask, task_type="classification")
print(f"[cls] logits shape: {tuple(cls_logits.shape)} (expected (2,16,10))")
assert cls_logits.shape == (B, n_q, cfg.max_classes)
n_classes_per_task = torch.tensor([3, 5])
y_q_cls = torch.stack([
torch.randint(0, 3, (n_q,)),
torch.randint(0, 5, (n_q,)),
])
loss_c = cls_masked_loss(cls_logits, y_q_cls, n_classes_per_task)
print(f"[cls] masked loss: {loss_c.item():.3f}")
# Warm-start dry run: simulate a v8 ckpt with wrong-shape heads
fake_v8_ckpt = {k: v.clone() for k, v in model.state_dict().items()
if not k.startswith("reg_head.") and not k.startswith("cls_head.")}
fake_v8_ckpt["reg_head.weight"] = torch.zeros(2, cfg.d_model) # v8 shape
fake_v8_ckpt["reg_head.bias"] = torch.zeros(2)
fake_v8_ckpt["cls_head.weight"] = torch.zeros(cfg.max_classes, cfg.d_model)
fake_v8_ckpt["cls_head.bias"] = torch.zeros(cfg.max_classes)
fresh = PredictLMv11(cfg)
info = fresh.warm_start_from_v8(fake_v8_ckpt)
print(f"[warm-start] loaded={info['loaded']}, missing={info['missing']}, unexpected={info['unexpected']}")
assert info['unexpected'] == 0, "v8 reg/cls heads should be filtered, got unexpected"
print("[OK] v11 model self-test passed")
|