guanghulab/train_mother.py

#!/usr/bin/env python3
"""
Qwen2.5-7B SFT 训练脚本

用法:
  python3 train_mother.py

依赖:
  pip3 install transformers accelerate datasets
"""
import os, sys, json, torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from datasets import load_dataset

BS = int(os.environ.get("BATCH_SIZE", "4"))
GA = int(os.environ.get("GRAD_ACCUM", "8"))
LR = float(os.environ.get("LEARNING_RATE", "1e-5"))
EPOCHS = int(os.environ.get("EPOCHS", "3"))
DS  = os.environ.get("DATASET", "autodl-tmp/data/sft.jsonl")
OUT = os.environ.get("OUTPUT_DIR", "autodl-tmp/output/qwen25-7b-sft")

print("="*50)
print("Qwen2.5-7B SFT Training")
print(f"  Batch: {BS}, GradAccum: {GA}, Eff: {BS*GA}")
print(f"  LR: {LR}, Epochs: {EPOCHS}")
print(f"  Data: {DS}")
print(f"  Out:  {OUT}")
print("="*50)

# ========== 1. Load model ==========
print("[1/5] Loading model...")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B", torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B", trust_remote_code=True)

tokenizer.pad_token = tokenizer.eos_token

total = sum(p.numel() for p in model.parameters())
print(f"  Parameters: {total/1e9:.2f}B")
print(f"  Device: {model.device}")

# ========== 2. Load data ==========
print("[2/5] Loading data...")
dataset = load_dataset("json", data_files=DS, split="train")
print(f"  Samples: {len(dataset)}")

print(f"  Format:")
for k in dataset[0]:
    v = dataset[0][k]
    if isinstance(v, list):
        print(f"    {k}: [{len(v)} msgs]")
        for m in v[:2]:
            print(f"      {m['role']}: {m['content'][:50]}...")
    else:
        print(f"    {k}: {v}")

# ========== 3. Tokenize ==========
print("[3/5] Tokenizing data...")

def tokenize(example):
    # Build full chat text using Qwen chat template
    texts = tokenizer.apply_chat_template(example["messages"], tokenize=False)
    enc = tokenizer(texts, truncation=True, max_length=8192, add_special_tokens=False)
    return {"input_ids": enc["input_ids"], "labels": enc["input_ids"].copy()}

dataset = dataset.map(tokenize, remove_columns=["messages"], num_proc=8)

total_tokens = sum(len(x["input_ids"]) for x in dataset)
post_pad = sum(x["input_ids"].count(tokenizer.pad_token_id) for x in dataset) if hasattr(tokenizer, "pad_token_id") else 0
print(f"  Total tokens: {total_tokens:,}")

def collate(features):
    max_len = max(len(f["input_ids"]) for f in features)
    batch = {}
    for k in ["input_ids", "labels", "attention_mask"]:
        pad = tokenizer.pad_token_id if k != "labels" else -100
        batch[k] = torch.tensor([f[k] + [pad]*(max_len-len(f[k])) for f in features])
    return batch

model.config.use_cache = False

# ========== 4. Training args ==========
print("[4/5] Training config...")

args = TrainingArguments(
    output_dir=OUT, num_train_epochs=EPOCHS,
    per_device_train_batch_size=BS, gradient_accumulation_steps=GA,
    learning_rate=LR, warmup_ratio=0.05, lr_scheduler_type="cosine",
    bf16=True, tf32=True, logging_steps=10,
    save_strategy="epoch", save_total_limit=3,
    remove_unused_columns=False, dataloader_num_workers=4,
    gradient_checkpointing=True, optim="adamw_torch",
    report_to="none", ddp_find_unused_parameters=False,
)

trainer = Trainer(model=model, args=args, train_dataset=ds, data_collator=collate)

# ========== 5. Go ==========
print("[5/5] Starting training!")
gpu = torch.cuda.get_device_name(0)
mem = torch.cuda.get_device_properties(0).total_memory / 1e9
print(f"  GPU: {gpu} ({mem:.1f}GB) | Epochs: {EPOCHS} | Eff batch: {BS*GA} | LR: {LR}")
sys.stdout.flush()

trainer.train()

# ========== 6. Save ==========
print("Saving model...")
final = os.path.join(OUT, "final")
trainer.save_model(final)
tokenizer.save_pretrained(final)

# ⚠️ 关键修复：Qwen chat template 使用 <|im_end|> (151645) 作为对话EOS
# 但 config.json 中默认 eos_token_id=151643 (<|endoftext|>)
# 不修复会导致部署时模型无限生成 → 死循环乱码
model.config.eos_token_id = 151645
model.config.save_pretrained(final)

# 修复 tokenizer 默认system prompt
tok_cfg_path = os.path.join(final, "tokenizer_config.json")
with open(tok_cfg_path) as f:
    tok_cfg = json.load(f)
tok_cfg["default_system"] = ""
with open(tok_cfg_path, "w") as f:
    json.dump(tok_cfg, f, indent=2, ensure_ascii=False)

peak = torch.cuda.max_memory_allocated() / 1e9
print(f"  Model: {final}")
print(f"  Peak VRAM: {peak:.2f}GB / {mem:.1f}GB")
print(f"  DONE!")