fix: audit v2 P0/P1 fixes (D1/D2/D3/D7/D9/D11/D14)

- D1: inference/dyquant.py load_model() 优先 FusionMini/FusionModel，回退 AutoModelForCausalLM 时警告
- D2: train/train_grpo.py fallback 到 AutoModelForCausalLM 时显式警告 Thinking Dial 不可用
- D3: requirements.txt 添加 gradio>=4.0.0（inference/dashboard.py 依赖）
- D7: 创建 data_pipeline/__init__.py（setup.py 声明为包）
- D9: evaluation/model_interpretability.py hash() 替换为 hashlib.md5()（确定性跨会话）
- D11: train/lora_finetune.py main() 添加 CUDA 可用性检查
- D14: evaluation/visualization.py save_visualization_report() stdout 劫持改用 try/finally 保护

25/25 测试通过

data_pipeline/__init__.py

@@ -0,0 +1,2 @@
+# data_pipeline package
+# 数据管道模块：T-KD 蒸馏训练、双语过滤等

evaluation/model_interpretability.py

@@ -9,6 +9,7 @@ SHAP: SHapley Additive exPlanations
 """
 import sys
 import math
+import hashlib
 import torch
 import numpy as np
 from pathlib import Path
@@ -86,7 +87,7 @@ class SimplifiedLIME:
         input_ids = []
         for t in tokens:
             # 简单 hash 到 vocab 范围
-            token_id = hash(t) % vocab_size
+            token_id = int(hashlib.md5(t.encode()).hexdigest(), 16) % vocab_size
             input_ids.append(token_id)
         
         input_ids = torch.tensor([input_ids])
@@ -188,7 +189,7 @@ class SimplifiedSHAP:
         
         input_ids = []
         for t in tokens:
-            token_id = hash(t) % vocab_size
+            token_id = int(hashlib.md5(t.encode()).hexdigest(), 16) % vocab_size
             input_ids.append(token_id)
         
         input_ids = torch.tensor([input_ids])

evaluation/visualization.py

@@ -162,8 +162,8 @@ def save_visualization_report(model, attention_weights, losses, output_path):
         import sys
         original_stdout = sys.stdout
         sys.stdout = f
-        
-        print("=" * 60)
+        try:
+            print("=" * 60)
         print("Fusion-LLM 可视化报告")
         print("=" * 60)
         print()
@@ -184,8 +184,9 @@ def save_visualization_report(model, attention_weights, losses, output_path):
         print("报告结束")
         print("=" * 60)
         
-        # 恢复 stdout
-        sys.stdout = original_stdout
+        finally:
+            # 恢复 stdout
+            sys.stdout = original_stdout
     
     print(f"   报告已保存到: {output_path}")
     print()

inference/dyquant.py

@@ -93,24 +93,46 @@ class DyQuantConverter:
         print(f"   按头量化：{config.per_head}")
     
     def load_model(self):
-        """加载模型"""
+        """加载模型（优先 FusionModel/FusionMini，回退 AutoModelForCausalLM）"""
         print(f"\n[DyQuant] 加载模型：{self.config.model_path}")
         
+        self.model = None
+        
+        # Try FusionMini first
+        try:
+            from models.fusion_mini import FusionMini
+            self.model = FusionMini._load_from_safetensors(self.config.model_path)
+            self.model.eval()
+            print(f"[DyQuant] FusionMini 加载成功")
+            return self.model
+        except Exception as e1:
+            pass  # fallback
+        
+        # Try FusionModel
+        try:
+            from models.fusion_model import FusionModel
+            self.model = FusionModel.from_pretrained(self.config.model_path)
+            self.model.eval()
+            print(f"[DyQuant] FusionModel 加载成功")
+            return self.model
+        except Exception as e2:
+            pass  # fallback
+        
+        # Fallback to AutoModelForCausalLM
         try:
             from transformers import AutoModelForCausalLM
-            
-            # 加载真实模型
             self.model = AutoModelForCausalLM.from_pretrained(
                 self.config.model_path,
                 torch_dtype=torch.bfloat16,
-                device_map="cpu",  # 量化在 CPU 上进行
+                device_map="cpu",
                 trust_remote_code=True,
             )
             self.model.eval()
-            print(f"[DyQuant] 模型加载成功")
+            print(f"[DyQuant] AutoModelForCausalLM 加载成功（回退模式）")
+            print(f"[DyQuant] 警告：非 Fusion 模型，SBLA/ThinkingDial 量化路径未验证")
             return self.model
-        except Exception as e:
-            print(f"[DyQuant] 模型加载失败：{e}")
+        except Exception as e3:
+            print(f"[DyQuant] 模型加载失败：{e3}")
             import traceback; traceback.print_exc()
             self.model = None
             return None

requirements.txt

@@ -40,6 +40,7 @@ rouge-score>=0.1.2
 # 可视化（可选）
 tensorboard>=2.15.0
 wandb>=0.16.0
+gradio>=4.0.0  # inference/dashboard.py 依赖
 
 
 # ONNX 导出

train/lora_finetune.py

@@ -1,4 +1,4 @@
-"""
+"""
 Fusion 模型 LoRA/QLoRA 微调脚本
 
 支持：
@@ -367,6 +367,14 @@ def train(args):
 
 
 def main():
+    """Main entry point with CUDA availability check."""
+    # Check CUDA availability
+    if not torch.cuda.is_available():
+        print("[ERROR] CUDA is not available. This training script requires a GPU.")
+        print("[INFO] For CPU training, use train_mini.py (FusionMini).")
+        sys.exit(1)
+    print(f"[INFO] CUDA available: {torch.cuda.get_device_name(0)}")
+
     parser = argparse.ArgumentParser(description="Fusion 模型 LoRA/QLoRA 微调")
     
     # 模型参数

train/train_grpo.py

@@ -85,6 +85,14 @@ def load_fusion_model(model_path: str, device: str = "auto"):
 
     # Fallback to HF AutoModel
     if model is None:
+        import warnings
+        warnings.warn(
+            "无法加载 FusionModel/FusionMini，回退到 AutoModelForCausalLM。\n"
+            "Thinking Dial（推理深度控制）在此模式下不可用。\n"
+            "建议确认模型路径指向有效的 Fusion 模型。",
+            UserWarning
+        )
+        logger.warning("[WARNING] 回退到 AutoModelForCausalLM，Thinking Dial 将不可用！")
         from transformers import AutoModelForCausalLM
         model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
         config = model.config