feat: v1.3-lite-flash 训练代码 + 新词表(vocab_size=47725)

- 用 CharTokenizer.build_vocab 从 977K 数据生成词表 (vocab_size=47725)
- train_cnb.py: batch=224 + accum=2 (等效448) + AMP + torch.compile
- build_vocab.py: 词表生成脚本
- vocab_full.json: 含特殊token完整词表
- 模型 41.85M参数, L40 GPU训练

scripts/train_cnb.py

@@ -76,15 +76,26 @@ def main():
     vocab_path = os.path.join(MODEL_DIR, 'vocab.json')
     with open(vocab_path, 'r', encoding='utf-8') as f:
         vocab = json.load(f)
-    print(f"词表: {len(vocab)} (BPE)")
+    # vocab.json 不含特殊token(0-6)，需合并进来，否则：
+    # 1) tokenizer 编码 <SOS>/<SEP>/<EOS> 等标记时会逐字符错码
+    # 2) vocab 的 id 范围是 7..max，超过 len(vocab) 会造成 embedding 索引越界
+    special_tokens = {'<PAD>': 0, '<UNK>': 1, '<SOS>': 2, '<EOS>': 3,
+                      '<START>': 4, '<END>': 5, '<SEP>': 6}
+    for tok, tid in special_tokens.items():
+        if tok not in vocab:
+            vocab[tok] = tid
+    real_vocab_size = max(vocab.values()) + 1
+    print(f"词表: {len(vocab)}项, vocab_size={real_vocab_size} (字符+BPE, 来自 vocab.json)")
     tokenizer_instance.char_to_id = vocab.copy()
     tokenizer_instance.id_to_char = {v: k for k, v in vocab.items()}
+    tokenizer_instance.special_tokens = special_tokens.copy()
+    tokenizer_instance._sorted_vocab_cache = None
     for attr in ['_multi_trie', '_multi_keys', '_single_map']:
         if hasattr(tokenizer_instance, attr):
             delattr(tokenizer_instance, attr)
 
     # ====== 模型 ======
-    model = LLMModel(vocab_size=max(vocab.values()) + 1, d_model=320, n_heads=8,
+    model = LLMModel(vocab_size=real_vocab_size, d_model=320, n_heads=8,
                      n_layers=6, d_ff=1536, max_len=256, dropout=0.1,
                      use_memory=False, use_thinking=False, use_moe=False)
     total = sum(p.numel() for p in model.parameters())
@@ -96,7 +107,7 @@ def main():
         if ckpt_path:
             ckpt = torch.load(ckpt_path, map_location=device)
             old_cfg = ckpt.get('model_config', {})
-            if (old_cfg.get('vocab_size', 0) == len(vocab)
+            if (old_cfg.get('vocab_size', 0) == real_vocab_size
                 and old_cfg.get('d_model', 0) == 320
                 and old_cfg.get('n_layers', 0) == 6):
                 model.load_state_dict(ckpt['model_state_dict'])
@@ -109,12 +120,25 @@ def main():
     
     model = model.to(device)
 
+    # ====== 火力全开：torch.compile (inductor融合, 无CUDA graphs) + 梯度累积 ======
+    # 物理batch=192(不OOM), accum=2 → 等效batch=384
+    fwd_model = model
+    if device == 'cuda':
+        print("🔥 启用 torch.compile (mode=default, inductor融合)...")
+        fwd_model = torch.compile(model, mode='default')
+        print("✓ torch.compile 已就绪（首步触发编译约30-60s）")
+
     # ====== DataLoader ======
+    n_workers = 16  # CPU火力全开，16核全用
     train_ds = ConversationDataset(train_set, vocab, max_length=256)
-    train_loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=True)
+    train_loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=True,
+                              num_workers=n_workers, pin_memory=(device == 'cuda'),
+                              persistent_workers=True, prefetch_factor=4, drop_last=True)
     val_ds = ConversationDataset(val_set, vocab, max_length=256)
-    val_loader = DataLoader(val_ds, batch_size=args.batch_size, shuffle=False)
-    print(f"批次: 训练{len(train_loader)} | 验证{len(val_loader)}")
+    val_loader = DataLoader(val_ds, batch_size=args.batch_size, shuffle=False,
+                            num_workers=n_workers, pin_memory=(device == 'cuda'),
+                            persistent_workers=True, prefetch_factor=4, drop_last=True)
+    print(f"批次: 训练{len(train_loader)} | 验证{len(val_loader)} | num_workers={n_workers} | prefetch=4 | drop_last=True")
 
     # ====== Trainer ======
     config = {
@@ -125,7 +149,7 @@ def main():
         'd_ff': 1536, 'max_length': 256, 'dropout': 0.1,
         'use_amp': device == 'cuda',
     }
-    trainer = LLMTrainer(model, device, lr=args.lr, use_amp=config['use_amp'])
+    trainer = LLMTrainer(fwd_model, device, lr=args.lr, use_amp=config['use_amp'])
 
     # ====== 训练 ======
     print(f"\n开始训练 epoch {start_epoch+1} → {args.epochs}")