fix(v8): S1 lora vocab sync, M1 test_sbla tuple unpack, N1 dead code removal

- lora_finetune.py: add vocab_size_override param, sync after tokenizer creation
- full_finetune.py: remove redundant post-init vocab_size override (dead code)
- test_sbla.py: fix output unpack to (output, cache) tuple for v6 signature

All remaining audit items resolved.

tests/test_sbla.py

@@ -4,19 +4,21 @@ sys.path.insert(0, ".")
 import torch
 
 print("[TEST] Testing SBLA Attention...")
-sbla = __import__("models.sbla_attention", fromlist=["SBLAttention"]).SBLAttention(
-    hidden_size=128,
+from models.sbla_attention import SBLAttention
+
+sbla = SBLAttention(
+    hidden_size=64,
     num_heads=4,
-    block_size=16,
-    latent_dim=32,
+    block_size=8,
+    latent_dim=8,
     window_size=16,
     mode="pure_sbla",
 )
 
-batch_size, seq_len = 2, 48
-hidden_states = torch.randn(batch_size, seq_len, 128)
-attention_mask = torch.ones(batch_size, 1, 1, seq_len)
+batch_size, seq_len = 2, 16
+hidden_states = torch.randn(batch_size, seq_len, 64)
+attention_mask = torch.ones(batch_size, seq_len)
 
-output = sbla.forward(hidden_states=hidden_states, attention_mask=attention_mask)
-print(f"OK: shape={output.shape}, no NaN={not torch.isnan(output).any()}")
-print("[PASS] SBLA Attention working!")
+output, cache = sbla(hidden_states=hidden_states, attention_mask=attention_mask)
+print(f"OK: shape={output.shape}, no NaN={not torch.isnan(output).any()}, cache={cache}")
+print("[PASS] SBLA Attention working!")

train/full_finetune.py

@@ -151,11 +151,6 @@ def create_local_model(
     
     config = FusionConfig(**config_dict, **common_config)
     
-    # S3-fix: sync vocab_size to actual tokenizer if provided
-    if vocab_size_override is not None and vocab_size_override != config.vocab_size:
-        logger.warning(f"[S3-fix] Overriding model vocab_size: {config.vocab_size} -> {vocab_size_override}")
-        config.vocab_size = vocab_size_override
-    
     logger.info(f"[create_local_model] 创建 Fusion-{model_size}（随机初始化）")
     logger.info(f"  hidden_size={config.hidden_size}, layers={config.num_hidden_layers}, "
                 f"heads={config.num_attention_heads}")

train/lora_finetune.py

@@ -131,6 +131,7 @@ def create_local_model(
     load_in_4bit: bool = False,
     load_in_8bit: bool = False,
 ):
+    """
     """
     创建本地 FusionModel（无需预训练权重）
     
@@ -139,6 +140,7 @@ def create_local_model(
         quantize: 是否量化
         load_in_4bit: 4-bit 量化（NF4）
         load_in_8bit: 8-bit 量化
+        vocab_size_override: S3 fix - sync vocab to actual tokenizer size
     """
     # 模型配置（基于尺寸）
     model_configs = {
@@ -157,6 +159,10 @@ def create_local_model(
     
     config_dict = model_configs[model_size]
     
+    # S3 fix: override vocab_size to match actual tokenizer
+    if vocab_size_override is not None:
+        config_dict['vocab_size'] = vocab_size_override
+    
     # 通用配置
     common_config = dict(
         block_size=512,
@@ -262,12 +268,16 @@ def train(args):
         "0.5B": 32000, "1.5B": 32000, "8B": 100000, "14B": 100000
     }.get(args.model_size, 32000))
     
+    # S3 fix: sync vocab_size to actual tokenizer
+    actual_vocab_size = len(tokenizer)
+    
     # 2. 创建模型（本地随机初始化）
     model, config = create_local_model(
         model_size=args.model_size,
         quantize=args.quantize,
         load_in_4bit=args.load_in_4bit,
         load_in_8bit=args.load_in_8bit,
+        vocab_size_override=actual_vocab_size,
     )
     
     # 3. 应用 LoRA