File size: 18,453 Bytes
9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c bc7101b 9477b5c | 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 | """BLT-Reasoner training loop.
Usage:
python -m experiments.blt_reasoner.train --config configs/pilot_qwen15b_gsm8k.json
Pre-registered success criterion (do not evaluate raw GSM8K accuracy first):
Δ(normal-z − random-z) ≥ 15pp AND Δ(normal-z − zero-z) ≥ 25pp
on the held-out GSM8K mini-eval. If both hold, H1 (z carries information)
is supported and we proceed to scale.
"""
from __future__ import annotations
import argparse
import json
import math
import os
import random
import sys
import time
from dataclasses import asdict
from pathlib import Path
from typing import Optional
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from .data import GSM8KDataset, MATHDataset, collate_batch
def _build_dataset(name, split, max_examples):
"""Dispatch to either GSM8K or MATH based on cfg['dataset'] (default gsm8k)."""
name = (name or "gsm8k").lower()
if name == "math":
return MATHDataset(split=split, max_examples=max_examples)
return GSM8KDataset(split=split, max_examples=max_examples)
from .losses import (
InfoNCEHead, LossWeights, encode_answer_for_infonce, infonce_loss,
kl_to_gaussian, lm_loss_on_y,
)
from .model import BLTConfig, LatentProjector, build_base, forward_with_latent
def set_seed(s: int):
random.seed(s)
torch.manual_seed(s)
torch.cuda.manual_seed_all(s)
def make_optimizer(model, projector, head, cfg):
"""Three param groups: LoRA / projector / InfoNCE head."""
try:
from bitsandbytes.optim import PagedAdamW8bit as AdamW
use_8bit = True
except Exception:
from torch.optim import AdamW
use_8bit = False
lora_params = [p for p in model.parameters() if p.requires_grad]
groups = [
{"params": lora_params, "lr": cfg["lr_lora"]},
{"params": projector.parameters(), "lr": cfg["lr_proj"]},
{"params": head.parameters(), "lr": cfg["lr_head"]},
]
opt = AdamW(groups, weight_decay=cfg["weight_decay"])
return opt, use_8bit
def get_K_for_step(step: int, curriculum) -> int:
"""curriculum = [[step_threshold, K], ...] (ascending)."""
K = curriculum[0][1]
for thr, k in curriculum:
if step >= thr:
K = k
return K
def cosine_lr(step, warmup, total, base_lr):
if step < warmup:
return base_lr * step / max(1, warmup)
progress = (step - warmup) / max(1, total - warmup)
return base_lr * 0.5 * (1.0 + math.cos(math.pi * min(progress, 1.0)))
def evaluate_quick(model, projector, tokenizer, val_loader, device, K, block_z_to_x: bool = False) -> dict:
"""Quick eval: per-token LM perplexity + InfoNCE accuracy on the val set."""
model.eval()
total_tok, total_loss, total_correct, total_total = 0, 0.0, 0, 0
with torch.no_grad():
for batch in val_loader:
x_ids = batch.x_ids.to(device)
x_attn = batch.x_attn.to(device)
y_ids = batch.y_ids.to(device)
y_attn = batch.y_attn.to(device)
logits_y, z, _ = forward_with_latent(
model, x_ids, x_attn, y_ids, projector, K,
block_y_to_x=True, block_z_to_x=block_z_to_x,
)
B, L_y, V = logits_y.shape
ce = torch.nn.functional.cross_entropy(
logits_y.reshape(-1, V), y_ids.reshape(-1), reduction="none"
).reshape(B, L_y)
mask = y_attn.float()
total_loss += (ce * mask).sum().item()
total_tok += mask.sum().item()
preds = logits_y.argmax(dim=-1)
total_correct += ((preds == y_ids) * mask).sum().item()
total_total += mask.sum().item()
model.train()
return {
"val_lm_ppl": math.exp(total_loss / max(total_tok, 1)),
"val_tok_acc": total_correct / max(total_total, 1),
}
def _load_state_from_ckpt(model, projector, head, ckpt_dir, device):
"""Restore LoRA adapter, projector, and InfoNCE head from a ckpt dir.
Optimizer state is NOT saved by save_every (see train loop), so resuming
re-initializes Adam moments. Loss curves will spike for a few hundred
steps but the latent geometry survives — adequate for instance-failure
recovery.
ckpt_dir may be a local path or a `<hf-namespace>/<repo>[:subfolder]` ref.
NB: peft serializes LoRA weights with keys like
`...lora_A.weight` but the wrapped PeftModel's state_dict uses
`...lora_A.default.weight` (the `.default` is the adapter name). Naive
`model.load_state_dict(sd, strict=False)` silently discards all 224 LoRA
matrices. We rewrite the keys here and verify nothing is "unexpected".
"""
from pathlib import Path
if "/" in ckpt_dir and not Path(ckpt_dir).exists():
# HF Hub form: optional :subfolder suffix selects a directory inside the repo.
from huggingface_hub import snapshot_download
repo, _, sub = ckpt_dir.partition(":")
local = snapshot_download(repo_id=repo, allow_patterns=(f"{sub}/*" if sub else None))
ckpt_dir = str(Path(local) / sub) if sub else local
ckpt = Path(ckpt_dir)
from safetensors.torch import load_file
adapter_file = ckpt / "model" / "adapter_model.safetensors"
if adapter_file.exists():
sd_raw = load_file(str(adapter_file))
# Rewrite peft v0.19 serialization keys -> adapter-name-aware keys.
sd = {}
for k, v in sd_raw.items():
# Insert ".default" before ".weight" inside the LoRA tensor path.
nk = k
for tag in (".lora_A.weight", ".lora_B.weight",
".lora_A.bias", ".lora_B.bias"):
if nk.endswith(tag):
nk = nk[: -len(tag)] + tag.replace(".weight", ".default.weight") \
.replace(".bias", ".default.bias")
break
sd[nk] = v
missing, unexpected = model.load_state_dict(sd, strict=False)
# All "unexpected" would indicate the rewrite is wrong; we expect 0.
# "missing" is huge because the base-model weights aren't in this file.
lora_missing = [m for m in missing if "lora" in m.lower()]
print(f"[resume] adapter: total_missing={len(missing)} unexpected={len(unexpected)} "
f"lora_missing={len(lora_missing)} (lora_missing>0 ⇒ LoRA load failed)", flush=True)
if lora_missing:
print(f"[resume] WARNING: first 3 lora_missing keys: {lora_missing[:3]}", flush=True)
if unexpected:
print(f"[resume] WARNING: first 3 unexpected keys: {unexpected[:3]}", flush=True)
proj_file = ckpt / "projector.pt"
if proj_file.exists():
projector.load_state_dict(torch.load(proj_file, map_location=device))
print(f"[resume] projector loaded", flush=True)
head_file = ckpt / "head.pt"
if head_file.exists():
head.load_state_dict(torch.load(head_file, map_location=device))
print(f"[resume] head loaded", flush=True)
# Try to recover step number from ckpt dir name like "ckpt-step6000".
m = ckpt.name
import re
mm = re.match(r"ckpt-step(\d+)", m)
return int(mm.group(1)) if mm else 0
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--config", required=True)
parser.add_argument("--resume_from", default=None,
help="Local ckpt dir OR 'hf-namespace/repo[:subfolder]'. "
"Restores LoRA + projector + InfoNCE head; opt state is re-initialized.")
args = parser.parse_args()
with open(args.config) as f:
cfg = json.load(f)
set_seed(cfg["seed"])
out_dir = Path(cfg["output_dir"])
out_dir.mkdir(parents=True, exist_ok=True)
log_path = out_dir / "train.log"
metrics_path = out_dir / "metrics.jsonl"
log_f = open(log_path, "a", buffering=1)
met_f = open(metrics_path, "a", buffering=1)
def log(msg):
line = f"[{time.strftime('%H:%M:%S')}] {msg}"
print(line, flush=True)
log_f.write(line + "\n")
log(f"config={json.dumps(cfg)}")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
log(f"device={device}")
# Build model + tokenizer + projector + head
bcfg = BLTConfig(
base_model=cfg["base_model"],
use_lora=cfg["use_lora"],
lora_r=cfg["lora_r"], lora_alpha=cfg["lora_alpha"],
lora_dropout=cfg["lora_dropout"],
lora_target_modules=tuple(cfg["lora_target_modules"]),
K_latents=cfg["K_latents"], block_y_to_x=cfg["block_y_to_x"],
block_z_to_x=cfg.get("block_z_to_x", False),
proj_init_scale=cfg["proj_init_scale"],
dtype=cfg["dtype"], attn_impl=cfg["attn_impl"],
gradient_checkpointing=cfg.get("gradient_checkpointing", False),
)
model, tokenizer = build_base(bcfg)
model.to(device)
inner = model.get_base_model() if hasattr(model, "get_base_model") else model
d_model = inner.config.hidden_size
model_dtype = {"bfloat16": torch.bfloat16, "float16": torch.float16, "float32": torch.float32}[cfg["dtype"]]
projector = LatentProjector(
d_model, init_scale=cfg["proj_init_scale"],
use_mlp=cfg.get("proj_mlp", False),
hidden_mult=cfg.get("proj_hidden_mult", 4),
).to(device).to(model_dtype)
# InfoNCE head stays fp32 for contrastive numerical stability (z_pool and f_y are .float()'d at use).
head = InfoNCEHead(d_z=d_model, d_y=d_model, d_out=256).to(device)
n_params_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad) \
+ sum(p.numel() for p in projector.parameters()) \
+ sum(p.numel() for p in head.parameters())
log(f"trainable params (LoRA+proj+head) = {n_params_trainable/1e6:.2f}M")
# Data — dataset selected by cfg['dataset'] (default gsm8k)
ds_name = cfg.get("dataset", "gsm8k")
train_ds = _build_dataset(ds_name, "train", cfg.get("data_train_size"))
val_ds = _build_dataset(ds_name, "test", cfg.get("data_eval_size") or 200)
log(f"dataset={ds_name} train={len(train_ds)} val={len(val_ds)}")
def make_loader(ds, bs, shuffle):
return DataLoader(ds, batch_size=bs, shuffle=shuffle, drop_last=shuffle,
collate_fn=lambda b: collate_batch(
b, tokenizer,
max_prompt_len=cfg["max_prompt_len"],
max_answer_len=cfg["max_answer_len"],
))
train_loader = make_loader(train_ds, cfg["batch_size"], True)
val_loader = make_loader(val_ds, cfg["batch_size"], False)
opt, used_8bit = make_optimizer(model, projector, head, cfg)
log(f"optimizer 8bit={used_8bit}")
weights = LossWeights(
lambda_lm=cfg["lambda_lm"], lambda_id=cfg["lambda_id"],
lambda_kl=cfg["lambda_kl"], tau_infonce=cfg["tau_infonce"],
)
step = 0
if args.resume_from:
step = _load_state_from_ckpt(model, projector, head, args.resume_from, device)
log(f"[resume] restored from {args.resume_from} at step={step}")
accum_idx = 0
t0 = time.time()
running = {"loss": 0.0, "lm": 0.0, "id": 0.0, "kl": 0.0, "z_norm": 0.0, "n": 0}
model.train()
while step < cfg["max_steps"]:
for batch in train_loader:
if step >= cfg["max_steps"]:
break
K = get_K_for_step(step, cfg["K_curriculum"])
x_ids = batch.x_ids.to(device)
x_attn = batch.x_attn.to(device)
y_ids = batch.y_ids.to(device)
y_attn = batch.y_attn.to(device)
# InfoNCE positives. Three modes:
# - default ("answer-only"): encode "#### N" → [B, d]; pool z over K → [B, d];
# single contrastive loss. Used in the original recipe.
# - infonce_full_answer: same shape, but target encodes full reasoning chain.
# - infonce_per_slot: encode each of K y-chunks separately → [B, K, d]; do
# PER-SLOT contrastive (each slot k tries to identify chunk_k of its
# problem from a B*K negative pool). Forces slot specialization.
use_per_slot = cfg.get("infonce_per_slot", False)
if use_per_slot:
from .data import split_y_into_chunks
from .losses import encode_chunks_per_slot, infonce_per_slot_loss
src = batch.full_answer_strs if batch.full_answer_strs is not None else batch.final_strs
chunks_per_problem = [split_y_into_chunks(s, K) for s in src]
_chunk_max_len = cfg.get("infonce_chunk_max_len", 32)
f_y_chunks = encode_chunks_per_slot(
model, tokenizer, chunks_per_problem, device=device, max_len=_chunk_max_len,
) # [B, K, d]
else:
if cfg.get("infonce_full_answer", False):
_target_text = batch.full_answer_strs if batch.full_answer_strs is not None \
else batch.final_strs
_target_max_len = cfg.get("infonce_target_max_len", 128)
else:
_target_text = batch.final_strs
_target_max_len = cfg.get("infonce_target_max_len", 16)
f_y = encode_answer_for_infonce(
model, tokenizer, _target_text, device=device, max_len=_target_max_len,
) # [B, d]
logits_y, z, _ = forward_with_latent(
model, x_ids, x_attn, y_ids, projector, K,
block_y_to_x=cfg["block_y_to_x"],
block_z_to_x=cfg.get("block_z_to_x", False),
)
L_lm = lm_loss_on_y(logits_y, y_ids, y_attn)
if use_per_slot:
_ps_info = infonce_per_slot_loss(z, f_y_chunks, head, tau=weights.tau_infonce)
L_id = _ps_info["loss"]
# Stash diagnostic acc values for logging
running.setdefault("ps_acc_full", 0.0)
running.setdefault("ps_acc_within", 0.0)
running["ps_acc_full"] += float(_ps_info["acc_z2y"].item())
running["ps_acc_within"] += float(_ps_info["acc_within_problem"].item())
else:
# InfoNCE: pool z over K (single target per problem)
z_pool = z.mean(dim=1) # [B, d]
z_emb, y_emb = head(z_pool.float(), f_y.float())
L_id = infonce_loss(z_emb, y_emb, tau=weights.tau_infonce)
L_kl = kl_to_gaussian(z.float())
# Optional slot-decorrelation regularizer.
lambda_decorr = cfg.get("lambda_decorr", 0.0)
if lambda_decorr > 0:
from .losses import slot_decorrelation_loss
L_decorr = slot_decorrelation_loss(z)
else:
L_decorr = torch.zeros((), device=device)
loss = (weights.lambda_lm * L_lm
+ weights.lambda_id * L_id
+ weights.lambda_kl * L_kl
+ lambda_decorr * L_decorr)
(loss / cfg["grad_accum"]).backward()
running["loss"] += loss.item()
running["lm"] += L_lm.item()
running["id"] += L_id.item()
running["kl"] += L_kl.item()
running.setdefault("decorr", 0.0)
running["decorr"] += float(L_decorr.item())
running["z_norm"] += z.float().pow(2).sum(dim=-1).mean().sqrt().item()
running["n"] += 1
accum_idx += 1
if accum_idx % cfg["grad_accum"] == 0:
lr_now = cosine_lr(step, cfg["warmup_steps"], cfg["max_steps"], cfg["lr_lora"])
for pg, base in zip(opt.param_groups,
[cfg["lr_lora"], cfg["lr_proj"], cfg["lr_head"]]):
pg["lr"] = base * lr_now / cfg["lr_lora"]
torch.nn.utils.clip_grad_norm_(
[p for p in model.parameters() if p.requires_grad]
+ list(projector.parameters()) + list(head.parameters()),
cfg["max_grad_norm"],
)
opt.step()
opt.zero_grad(set_to_none=True)
step += 1
if step % cfg["log_every"] == 0:
n = max(running["n"], 1)
log(f"step={step} K={K} loss={running['loss']/n:.4f} "
f"lm={running['lm']/n:.4f} id={running['id']/n:.4f} "
f"kl={running['kl']/n:.4f} z_norm={running['z_norm']/n:.3f} "
f"elapsed={time.time()-t0:.0f}s")
met_f.write(json.dumps({
"step": step, "K": K,
"loss": running['loss']/n,
"lm": running['lm']/n,
"id": running['id']/n,
"kl": running['kl']/n,
"z_norm": running['z_norm']/n,
"elapsed_s": time.time()-t0,
}) + "\n")
running = {k: (0.0 if k != "n" else 0) for k in running}
if cfg["eval_every"] > 0 and step % cfg["eval_every"] == 0:
quick = evaluate_quick(model, projector, tokenizer, val_loader, device, K,
block_z_to_x=cfg.get("block_z_to_x", False))
log(f"[eval] step={step} {quick}")
met_f.write(json.dumps({"step": step, "eval": quick}) + "\n")
if cfg["save_every"] > 0 and step % cfg["save_every"] == 0:
save_dir = out_dir / f"ckpt-step{step}"
save_dir.mkdir(exist_ok=True)
model.save_pretrained(save_dir / "model")
torch.save(projector.state_dict(), save_dir / "projector.pt")
torch.save(head.state_dict(), save_dir / "head.pt")
log(f"[save] {save_dir}")
# Final save
save_dir = out_dir / "final"
save_dir.mkdir(exist_ok=True)
model.save_pretrained(save_dir / "model")
torch.save(projector.state_dict(), save_dir / "projector.pt")
torch.save(head.state_dict(), save_dir / "head.pt")
log(f"[done] final saved at {save_dir}")
log_f.close()
met_f.close()
if __name__ == "__main__":
main()
|