fix: 同时修复 config.json 和 generation_config.json 的 eos_token_id

train_mother.py

@@ -1,73 +1,75 @@
 #!/usr/bin/env python3
+"""全参数SFT训练 - 光湖母模型
+Qwen2.5-7B -> sft.jsonl (11,470条纯净对话)
+每条消息独立分词，只对assistant回复计算loss
 """
-Qwen2.5-7B SFT 训练脚本
 
-用法:
+import os, json, torch, sys
-  python3 train_mother.py
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
-依赖:
-  pip3 install transformers accelerate datasets
-"""
-import os, sys, json, torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
-from datasets import load_dataset
+from datasets import Dataset
+from tqdm import tqdm
 
-BS = int(os.environ.get("BATCH_SIZE", "4"))
+# ========== Config ==========
-GA = int(os.environ.get("GRAD_ACCUM", "8"))
+MODEL_PATH = "/root/autodl-tmp/cache/Qwen/Qwen2___5-7B"
-LR = float(os.environ.get("LEARNING_RATE", "1e-5"))
+DATA = "/root/autodl-tmp/data/sft.jsonl"
-EPOCHS = int(os.environ.get("EPOCHS", "3"))
+OUT = "/root/autodl-tmp/output/qwen25-7b-sft"
-DS  = os.environ.get("DATASET", "autodl-tmp/data/sft.jsonl")
+EPOCHS = 3
-OUT = os.environ.get("OUTPUT_DIR", "autodl-tmp/output/qwen25-7b-sft")
+BS = 1
+GA = 8
+LR = 2e-5
+MAX_LEN = 2048
 
-print("="*50)
+os.makedirs(OUT, exist_ok=True)
-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 ==========
+# ========== 1. Load data ==========
-print("[1/5] Loading model...")
+print("[1/5] Loading data...")
-model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B", torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto")
+with open(DATA) as f:
-tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B", trust_remote_code=True)
+    raw = [json.loads(line) for line in f]
+raw = [{"messages": [m for m in obj["messages"] if m["role"] != "system"]} for obj in raw if any(m["role"] != "system" for m in obj["messages"])]
+print(f"  {len(raw)} examples (system filtered)")
 
+# ========== 2. Load model ==========
+print(f"[2/5] Loading Qwen/Qwen2.5-7B...")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
 tokenizer.pad_token = tokenizer.eos_token
+tokenizer.padding_side = "right"
 
-total = sum(p.numel() for p in model.parameters())
+model = AutoModelForCausalLM.from_pretrained(
-print(f"  Parameters: {total/1e9:.2f}B")
+    MODEL_PATH, trust_remote_code=True,
-print(f"  Device: {model.device}")
+    torch_dtype=torch.bfloat16, attn_implementation="sdpa",
-
+).cuda()
-# ========== 2. Load data ==========
+model.config.use_cache = False
-print("[2/5] Loading data...")
+model.gradient_checkpointing_enable()
-dataset = load_dataset("json", data_files=DS, split="train")
+print(f"  Params: {sum(p.numel() for p in model.parameters())/1e9:.2f}B (full FT)")
-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...")
+print("[3/5] Tokenizing...")
+processed = []
+for d in tqdm(raw, desc="Tokenize"):
+    ids, labs = [], []
+    for msg in d["messages"]:
+        c = msg["content"]
+        if not c.strip():
+            continue
+        t = f"<|im_start|>{msg['role']}\n{c}<|im_end|>\n"
+        tok = tokenizer.encode(t, add_special_tokens=False)
+        ids.extend(tok)
+        labs.extend(tok if msg["role"] == "assistant" else [-100] * len(tok))
+    if len(ids) > MAX_LEN:
+        ids, labs = ids[:MAX_LEN], labs[:MAX_LEN]
+    processed.append({"input_ids": ids, "labels": labs, "attention_mask": [1]*len(ids)})
 
-def tokenize(example):
+ds = Dataset.from_list(processed)
-    # Build full chat text using Qwen chat template
+total_tok = sum(len(d["input_ids"]) for d in processed)
-    texts = tokenizer.apply_chat_template(example["messages"], tokenize=False)
+loss_tok = sum(sum(1 for l in d["labels"] if l != -100) for d in processed)
-    enc = tokenizer(texts, truncation=True, max_length=8192, add_special_tokens=False)
+print(f"  Dataset: {len(ds)} ex, {total_tok:,} tokens, {loss_tok:,} loss ({loss_tok/max(total_tok,1)*100:.1f}%)")
-    return {"input_ids": enc["input_ids"], "labels": enc["input_ids"].copy()}
+sys.stdout.flush()
-
-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:,}")
 
+# ========== 4. Train ==========
+print("[4/5] Training config...")
 def collate(features):
     max_len = max(len(f["input_ids"]) for f in features)
     batch = {}
@@ -76,11 +78,6 @@ def collate(features):
         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,
@@ -115,6 +112,13 @@ tokenizer.save_pretrained(final)
 model.config.eos_token_id = 151645
 model.config.save_pretrained(final)
 
+# ⚠️ generation_config.json 也必须修复！
+# HuggingFace 的 model.generate() 读取 generation_config.json，不是 config.json
+# 不修这个 → 部署后仍然无限生成 → 乱码
+model.generation_config.eos_token_id = 151645
+model.generation_config.pad_token_id = 151645
+model.generation_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:
@@ -122,8 +126,7 @@ with open(tok_cfg_path) as 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!")
+print("DONE!")