File size: 11,919 Bytes
e2bfccc | 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 | """Small SFT diagnostics for checkpoint quality and trainability.
This script intentionally bypasses the full trainer so it can answer one narrow
question quickly: can the checkpoint reduce response-only SFT loss on a tiny,
fixed batch?
"""
from __future__ import annotations
import argparse
import json
import math
from pathlib import Path
from typing import Any
import torch
from taoTrain.checkpointing.checkpoint import CheckpointManager
from taoTrain.config import TrainingModeEnum, load_config
from taoTrain.core import create_model
from taoTrain.data.sft_utils import build_sft_sequence_tokens, parse_sft_record
try:
from taoTrain.data.sft_utils import build_response_only_next_token_labels
except ImportError:
def build_response_only_next_token_labels(input_ids: list[int], mask: list[int]) -> list[int]:
labels = [token_id if mask_value else -100 for token_id, mask_value in zip(input_ids, mask)]
return labels[1:] + [-100]
from taoTrain.data.tokenizer import SentencePieceTokenizerWrapper
from taoTrain.utils import set_seed
def load_tokenizer(tokenizer_path: str):
path = Path(tokenizer_path)
if path.suffix == ".model":
import sentencepiece as spm
sp = spm.SentencePieceProcessor()
sp.Load(str(path))
return SentencePieceTokenizerWrapper(sp)
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
if getattr(tokenizer, "pad_token", None) is None and getattr(tokenizer, "eos_token", None):
tokenizer.pad_token = tokenizer.eos_token
return tokenizer
def read_jsonl_records(path: str, limit: int) -> list[dict[str, Any]]:
records = []
with open(path, "r", encoding="utf-8") as handle:
for line in handle:
line = line.strip()
if not line:
continue
records.append(json.loads(line))
if len(records) >= limit:
break
return records
def build_batch(config, tokenizer, records: list[dict[str, Any]], device: torch.device) -> dict[str, torch.Tensor]:
input_rows = []
attention_rows = []
label_rows = []
train_tokens = []
for record in records:
turns, _ = parse_sft_record(record, config)
if not turns:
continue
input_ids, attention_mask, mask = build_sft_sequence_tokens(
turns=turns,
tokenizer=tokenizer,
user_token=getattr(config, "user_token", "<user>"),
assistant_token=getattr(config, "assistant_token", "<assistant>"),
max_seq_length=config.model.max_seq_length,
)
labels = build_response_only_next_token_labels(input_ids, mask)
input_rows.append(input_ids)
attention_rows.append(attention_mask)
label_rows.append(labels)
train_tokens.append(sum(1 for value in labels if value != -100))
if not input_rows:
raise ValueError("No valid SFT records found for the diagnostic batch")
return {
"input_ids": torch.tensor(input_rows, dtype=torch.long, device=device),
"attention_mask": torch.tensor(attention_rows, dtype=torch.long, device=device),
"labels": torch.tensor(label_rows, dtype=torch.long, device=device),
"train_tokens": torch.tensor(train_tokens, dtype=torch.long),
}
@torch.no_grad()
def score_batch(model, batch: dict[str, torch.Tensor], dtype: torch.dtype) -> float:
model.eval()
device_type = "cuda" if batch["input_ids"].is_cuda else "cpu"
enabled = device_type == "cuda" and dtype in (torch.float16, torch.bfloat16)
with torch.autocast(device_type=device_type, dtype=dtype, enabled=enabled):
outputs = model(
input_ids=batch["input_ids"],
attention_mask=batch["attention_mask"],
labels=batch["labels"],
)
return float(outputs["loss"].detach().cpu())
def grad_l2_norm(parameters) -> float:
total = 0.0
for parameter in parameters:
if parameter.grad is None:
continue
grad = parameter.grad.detach()
total += float(torch.sum(grad.float() * grad.float()).cpu())
return math.sqrt(total)
def grad_summary(named_parameters, max_items: int = 12) -> dict[str, Any]:
groups: dict[str, dict[str, Any]] = {}
worst = []
nonfinite = []
for name, parameter in named_parameters:
if parameter.grad is None:
continue
grad = parameter.grad.detach().float()
finite = torch.isfinite(grad)
finite_count = int(finite.sum().cpu())
numel = grad.numel()
finite_abs_max = float(grad[finite].abs().max().cpu()) if finite_count else float("inf")
has_nonfinite = finite_count != numel
if has_nonfinite:
nonfinite.append(name)
if ".layers." in name:
parts = name.split(".")
try:
idx = parts.index("layers")
group = "layer_" + parts[idx + 1]
except (ValueError, IndexError):
group = "layers"
else:
group = name.split(".", 1)[0]
entry = groups.setdefault(group, {
"numel": 0,
"finite": 0,
"nonfinite_tensors": 0,
"max_abs_grad": 0.0,
})
entry["numel"] += numel
entry["finite"] += finite_count
entry["nonfinite_tensors"] += int(has_nonfinite)
entry["max_abs_grad"] = max(entry["max_abs_grad"], finite_abs_max)
worst.append((finite_abs_max, name))
worst.sort(reverse=True, key=lambda item: item[0])
return {
"groups": groups,
"worst_tensors": [{"name": name, "max_abs_grad": value} for value, name in worst[:max_items]],
"nonfinite_tensors": nonfinite[:max_items],
"nonfinite_tensor_count": len(nonfinite),
}
def freeze_ssm_core_parameters(model) -> int:
frozen = 0
markers = (
".ssm_lanes.",
".ssm.",
)
for name, parameter in model.named_parameters():
if any(marker in name for marker in markers):
parameter.requires_grad_(False)
frozen += parameter.numel()
return frozen
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--config", required=True)
parser.add_argument("--checkpoint", required=True)
parser.add_argument("--output", required=True)
parser.add_argument("--samples", type=int, default=2)
parser.add_argument("--steps", type=int, default=80)
parser.add_argument("--lr", type=float, default=3e-4)
parser.add_argument("--log-every", type=int, default=10)
parser.add_argument("--device", default="cuda")
parser.add_argument("--dtype", choices=["config", "float32", "float16", "bfloat16"], default="config")
parser.add_argument("--no-clip", action="store_true")
parser.add_argument("--freeze-ssm-core", action="store_true")
parser.add_argument("--ssm-branch-rms-norm", action="store_true")
parser.add_argument("--ssm-branch-clip-value", type=float, default=None)
parser.add_argument("--block-residual-rms-norm", action="store_true")
parser.add_argument("--block-residual-rms-target", type=float, default=None)
parser.add_argument("--seed", type=int, default=123)
args = parser.parse_args()
set_seed(args.seed)
config = load_config(args.config, TrainingModeEnum.SFT)
if args.ssm_branch_rms_norm:
config.model.ssm_branch_rms_norm = True
if args.ssm_branch_clip_value is not None:
config.model.ssm_branch_clip_value = args.ssm_branch_clip_value
if args.block_residual_rms_norm:
config.model.block_residual_rms_norm = True
if args.block_residual_rms_target is not None:
config.model.block_residual_rms_target = args.block_residual_rms_target
device = torch.device(args.device if args.device == "cpu" or torch.cuda.is_available() else "cpu")
if args.dtype == "float32":
dtype = torch.float32
elif args.dtype == "float16":
dtype = torch.float16
elif args.dtype == "bfloat16":
dtype = torch.bfloat16
else:
dtype = torch.bfloat16 if str(config.dtype) == "DataTypeEnum.BFLOAT16" or str(config.dtype) == "bfloat16" else torch.float32
tokenizer = load_tokenizer(config.dataset.tokenizer_path)
records = read_jsonl_records(config.dataset.jsonl_path, args.samples)
batch = build_batch(config, tokenizer, records, device)
model = create_model(config, device)
checkpoint = CheckpointManager(config.checkpoint_dir).load(args.checkpoint, device=device)
model.load_state_dict(checkpoint["model_state"], strict=False)
frozen_params = freeze_ssm_core_parameters(model) if args.freeze_ssm_core else 0
initial_loss = score_batch(model, batch, dtype)
trainable_params = [parameter for parameter in model.parameters() if parameter.requires_grad]
optimizer = torch.optim.AdamW(trainable_params, lr=args.lr, weight_decay=0.0)
history = []
device_type = "cuda" if device.type == "cuda" else "cpu"
autocast_enabled = device_type == "cuda" and dtype in (torch.float16, torch.bfloat16)
model.train()
for step in range(1, args.steps + 1):
optimizer.zero_grad(set_to_none=True)
with torch.autocast(device_type=device_type, dtype=dtype, enabled=autocast_enabled):
outputs = model(
input_ids=batch["input_ids"],
attention_mask=batch["attention_mask"],
labels=batch["labels"],
)
loss = outputs["loss"]
loss.backward()
grad_norm = grad_l2_norm(trainable_params)
stats = None
if step == 1 or step % args.log_every == 0 or step == args.steps:
stats = grad_summary(model.named_parameters())
if not args.no_clip:
torch.nn.utils.clip_grad_norm_(trainable_params, 1.0)
optimizer.step()
if step == 1 or step % args.log_every == 0 or step == args.steps:
item = {
"step": step,
"loss": float(loss.detach().cpu()),
"grad_l2_norm": grad_norm,
}
if stats is not None:
item["grad_summary"] = stats
history.append(item)
final_loss = score_batch(model, batch, dtype)
result = {
"checkpoint": str(Path(args.checkpoint)),
"config": str(Path(args.config)),
"dataset": config.dataset.jsonl_path,
"samples": len(records),
"sequence_length": config.model.max_seq_length,
"train_tokens_per_sample": batch["train_tokens"].tolist(),
"lr": args.lr,
"steps": args.steps,
"clip_grad_norm": not args.no_clip,
"freeze_ssm_core": args.freeze_ssm_core,
"ssm_branch_rms_norm": config.model.ssm_branch_rms_norm,
"ssm_branch_clip_value": config.model.ssm_branch_clip_value,
"block_residual_rms_norm": config.model.block_residual_rms_norm,
"block_residual_rms_target": config.model.block_residual_rms_target,
"frozen_params": frozen_params,
"trainable_params": sum(parameter.numel() for parameter in trainable_params),
"initial_loss": initial_loss,
"final_loss": final_loss,
"loss_delta": final_loss - initial_loss,
"history": history,
"device": str(device),
"dtype": str(dtype),
}
output = Path(args.output)
output.parent.mkdir(parents=True, exist_ok=True)
output.write_text(json.dumps(result, indent=2), encoding="utf-8")
print(json.dumps(result, indent=2))
if __name__ == "__main__":
main()
|