fix(v12): resolve 6 defects from v11 audit (F-NEW-6, S-NEW-5, M-NEW-5/6, MI-NEW-4/5/6)

FATAL:
- F-NEW-6: QATTrainer.prepare() crashes on load_model() returning None -> add None guard

SERIOUS:
- S-NEW-5: save() uses HF save_pretrained which can't serialize QuantizedLinear -> switch to safetensors/torch.save

MODERATE:
- M-NEW-5: _insert_fake_quant only matches LLaMA layer names -> apply qconfig to all nn.Linear
- M-NEW-6: ollama_deploy_v2 fallback export ignores sharded models -> detect index.json and merge shards

MINOR:
- MI-NEW-4: manage_mini_data.py DATA_PATH relative -> use Path(__file__).resolve().parent.parent
- MI-NEW-5: bilingual_filter/ollama_deploy/fusion_mini/run_tests [LOGO] residue -> replace with [INFO]/[WARN]
- MI-NEW-6: ollama_deploy.py check_dependencies lacks shell=True on Windows -> add shell=True
- also fix broken string literal in ollama_deploy.py line 46

data_pipeline/bilingual_filter.py

@@ -377,7 +377,7 @@ def process_data_pipeline(
     en_clean = en_filter.process(en_raw)
     
     # 4. 平衡采样
-    logger.info("\n[BALANCE][LOGO]  平衡采样...")
+    logger.info("\n[BALANCE] 平衡采样...")
     sampler = BalancedSampler(zh_clean, en_clean, zh_ratio=0.5)
     balanced_data = sampler.sample(n_samples)
     
@@ -393,7 +393,7 @@ def process_data_pipeline(
 
 if __name__ == "__main__":
     # 单元测试（模拟数据）
-    print("[LOGO] 测试 Bi-Lingual TrueFilter...")
+    print("[TEST] 测试 Bi-Lingual TrueFilter...")
     
     # 模拟中文数据
     zh_test_data = [

inference/dyquant.py

@@ -520,8 +520,22 @@ class DyQuantConverter:
         output_dir = Path(output_path)
         output_dir.mkdir(parents=True, exist_ok=True)
         
-        # 保存模型
-        self.model.save_pretrained(output_dir)
+        # 保存模型 - extract state_dict from custom quantized layers
+        # Custom QuantizedLinear layers are not HF-compatible, use safetensors directly
+        try:
+            import safetensors.torch as st
+            state = self.model.state_dict()
+            # Convert non-tensor values (scales/zeros) to tensors for serialization
+            clean_state = {}
+            for k, v in state.items():
+                if isinstance(v, torch.Tensor):
+                    clean_state[k] = v.contiguous()
+                else:
+                    clean_state[k] = torch.tensor(v) if v is not None else torch.tensor(0.0)
+            st.save_file(clean_state, str(output_dir / "model.safetensors"))
+        except ImportError:
+            # Fallback to torch.save if safetensors unavailable
+            torch.save(self.model.state_dict(), output_dir / "pytorch_model.bin")
         
         # 保存量化配置
         quant_config = {
@@ -639,19 +653,20 @@ class QATTrainer:
         self.model = None
         self.qat_model = None
     
-    def prepare(self) -> nn.Module:
+    def prepare(self) -> Optional[nn.Module]:
         """Load model and insert fake-quantization nodes."""
         self.model = self.converter.load_model()
+        if self.model is None:
+            print("[DyQuant] QAT: model load failed, cannot prepare")
+            return None
         self.qat_model = self._insert_fake_quant(self.model)
         return self.qat_model
     
     def _insert_fake_quant(self, model: nn.Module) -> nn.Module:
         """Insert fake-quantization observers into all Linear layers."""
         for name, module in model.named_modules():
-            if isinstance(module, nn.Linear) and any(
-                kw in name for kw in ['q_proj', 'k_proj', 'v_proj', 'out_proj', 'gate_proj', 'up_proj', 'down_proj']
-            ):
-                # Use PyTorch native fake quantization per-module
+            if isinstance(module, nn.Linear):
+                # Use PyTorch native fake quantization for all Linear layers
                 module.qconfig = torch.ao.quantization.get_default_qat_qconfig('x86')
         torch.ao.quantization.prepare_qat(model, inplace=True)
         return model

inference/ollama_deploy.py

@@ -41,9 +41,9 @@ def check_dependencies():
     convert_script = os.path.join(llama_cpp_dir, "convert-hf-to-gguf.py")
     
     if not os.path.exists(convert_script):
-        logger.warning(f"[WARN][LOGO]  未找到 llama.cpp 转换脚本：{convert_script}")
+        logger.warning(f"[WARN] 未找到 llama.cpp 转换脚本：{convert_script}")
         logger.warning("   请设置环境变量 LLAMA_CPP_DIR 或手动下载 llama.cpp")
-        logger.warning("   下载地址：<ADDRESS_REMOVED>
+        logger.warning("   https://github.com/ggerganov/llama.cpp")
         return False
     
     # 检查 Ollama
@@ -52,15 +52,16 @@ def check_dependencies():
             ["ollama", "--version"],
             capture_output=True,
             text=True,
+            shell=True,
         )
         if result.returncode == 0:
             logger.info(f"[OK] Ollama 已安装：{result.stdout.strip()}")
         else:
-            logger.warning("[WARN][LOGO]  Ollama 未安装或无法运行")
+            logger.warning("[WARN] Ollama 未安装或无法运行")
             logger.warning("   请访问 https://ollama.com 安装")
             return False
     except FileNotFoundError:
-        logger.warning("[WARN][LOGO]  Ollama 未安装")
+        logger.warning("[WARN] Ollama 未安装")
         logger.warning("   请访问 https://ollama.com 安装")
         return False
     
@@ -118,7 +119,7 @@ def convert_to_gguf(
         result = subprocess.run(quantize_cmd, capture_output=True, text=True)
         
         if result.returncode != 0:
-            logger.warning(f"[WARN][LOGO]  量化失败：{result.stderr}")
+            logger.warning(f"[WARN] 量化失败：{result.stderr}")
             logger.warning("   继续使用未量化模型")
         else:
             output_path = output_path.replace(".gguf", f"_{quantize}.gguf")
@@ -392,7 +393,7 @@ ollama run {model_name} --top_p 0.95
     with open(output_path, 'w', encoding='utf-8') as f:
         f.write(content)
     
-    logger.info(f"[LOGO] 使用示例已生成：{output_path}")
+    logger.info(f"[INFO] 使用示例已生成：{output_path}")
 
 
 def main():

inference/ollama_deploy_v2.py

@@ -557,16 +557,34 @@ def _fallback_export_gguf(model_path: str, output_path: str) -> Optional[str]:
     config = FusionConfig.from_pretrained(model_path)
     model = FusionModel(config)
     
-    # Load weights
-    from pathlib import Path
-    weight_files = list(Path(model_path).glob("*.safetensors")) + list(Path(model_path).glob("*.bin"))
-    if not weight_files:
-        logger.error("No model weight files found")
-        return None
-    
-    # Export as safetensors
+    # Export path
     export_path = output_path.replace('.gguf', '.safetensors')
-    st.save_file(model.state_dict(), export_path)
+    
+    # Load weights - handle sharded models (index.json + multiple safetensors)
+    from pathlib import Path
+    model_path_obj = Path(model_path)
+    index_file = model_path_obj / "model.safetensors.index.json"
+    
+    if index_file.exists():
+        # Sharded model: load all shards and merge
+        import json as _json
+        with open(index_file, 'r') as f:
+            index = _json.load(f)
+        weight_map = index.get("weight_map", {})
+        shard_files = set(weight_map.values())
+        merged_state = {}
+        for shard in shard_files:
+            shard_path = model_path_obj / shard
+            shard_state = st.load_file(str(shard_path))
+            merged_state.update(shard_state)
+        st.save_file(merged_state, export_path)
+    else:
+        # Single-file model
+        weight_files = list(model_path_obj.glob("*.safetensors")) + list(model_path_obj.glob("*.bin"))
+        if not weight_files:
+            logger.error("No model weight files found")
+            return None
+        st.save_file(model.state_dict(), export_path)
     logger.info(f"Exported model weights to: {export_path}")
     logger.info("NOTE: This is a safetensors export, not GGUF. For Ollama deployment,")
     logger.info("      convert this to GGUF using llama.cpp after ensuring architecture compatibility.")

models/fusion_mini.py

@@ -481,7 +481,7 @@ class FusionMini(PreTrainedModel):
 
 if __name__ == "__main__":
     # 单元测试
-    print("[LOGO] 测试 Fusion Mini 模型...")
+    print("[INFO] 测试 Fusion Mini 模型...")
     
     # 创建配置
     config = FusionMiniConfig(

scripts/manage_mini_data.py

@@ -14,7 +14,7 @@ import re
 import sys
 from pathlib import Path
 
-DATA_PATH = Path("data/mini_data.json")
+DATA_PATH = Path(__file__).resolve().parent.parent / "data" / "mini_data.json"
 
 # --- Depth assignment keywords ---
 DEPTH_3_KEYWORDS = ['prove', 'theorem', 'proof', 'derive', 'mathematical', 'complex',

tests/run_tests.py

@@ -24,7 +24,7 @@ from pathlib import Path
 project_root = Path(__file__).parent.parent
 sys.path.insert(0, str(project_root))
 
-print("[LOGO] Fusion 项目单元测试")
+print("[INFO] Fusion 项目单元测试")
 print("=" * 50)