111

app.py

@@ -33,57 +33,92 @@ def load_cosyvoice_model():
     print("="*60)
     
     try:
-        # 使用transformers加载（推荐用于Hugging Face Space）
-        print("\n尝试使用transformers加载...")
-        from transformers import AutoModel
+        # 方法1: 尝试使用huggingface_hub下载模型文件
+        print("\n尝试从Hugging Face Hub下载模型...")
+        from huggingface_hub import snapshot_download
         import torch
         
         model_name = "FunAudioLLM/CosyVoice-300M"
-        print(f"从 {model_name} 加载...")
+        print(f"从 {model_name} 下载模型文件...")
         
-        # CosyVoice需要trust_remote_code=True来加载自定义模型代码
-        model = AutoModel.from_pretrained(
-            model_name, 
-            trust_remote_code=True,
-            torch_dtype=torch.float32,
-            low_cpu_mem_usage=True
+        # 下载模型到本地缓存
+        model_dir = snapshot_download(
+            repo_id=model_name,
+            allow_patterns=["*.pt", "*.pth", "*.bin", "*.json", "*.yaml", "*.txt"],
+            ignore_patterns=["*.md", "*.gitattributes"]
         )
         
-        # 设置为评估模式
-        model.eval()
+        print(f"✓ 模型文件已下载到: {model_dir}")
         
-        # 检查模型类型和方法
-        print(f"模型类型: {type(model).__name__}")
-        
-        # 检查关键推理方法
-        has_inference_sft = hasattr(model, 'inference_sft')
-        has_inference = hasattr(model, 'inference')
-        has_generate = hasattr(model, 'generate')
-        
-        print(f"推理方法检查:")
-        print(f"  - inference_sft: {has_inference_sft}")
-        print(f"  - inference: {has_inference}")
-        print(f"  - generate: {has_generate}")
-        
-        # 如果模型有这些方法，说明加载成功
-        if has_inference_sft or has_inference or has_generate:
+        # 方法1.1: 尝试导入CosyVoice官方包
+        try:
+            print("\n尝试使用CosyVoice官方包加载...")
+            from cosyvoice.cli.cosyvoice import CosyVoice
+            model = CosyVoice(model_dir)
+            print("✓ 使用CosyVoice官方包加载成功")
+            
             cosyvoice_model = {
                 'model': model,
-                'type': 'transformers',
+                'type': 'official',
                 'has_inference': True
             }
             model_loaded = True
-            print("✓ 成功加载CosyVoice模型")
+            print("✓ 成功加载CosyVoice模型 (官方包)")
+            print("="*60 + "\n")
+            return cosyvoice_model
+            
+        except ImportError:
+            print("⚠ CosyVoice官方包未安装")
+            print("提示: pip install cosyvoice")
+        
+        # 方法1.2: 尝试直接加载PyTorch模型
+        print("\n尝试直接加载PyTorch模型文件...")
+        import glob
+        
+        # 查找模型文件
+        model_files = glob.glob(os.path.join(model_dir, "**/*.pt"), recursive=True)
+        model_files += glob.glob(os.path.join(model_dir, "**/*.pth"), recursive=True)
+        model_files += glob.glob(os.path.join(model_dir, "**/*.bin"), recursive=True)
+        
+        if model_files:
+            print(f"找到模型文件: {model_files[0]}")
+            model = torch.load(model_files[0], map_location='cpu')
+            
+            # 检查模型是否可用
+            if isinstance(model, dict):
+                print("✓ PyTorch模型字典加载成功")
+                cosyvoice_model = {
+                    'model': model,
+                    'type': 'pytorch_dict',
+                    'model_dir': model_dir,
+                    'has_inference': False
+                }
+            else:
+                if hasattr(model, 'eval'):
+                    model.eval()
+                print(f"✓ PyTorch模型加载成功: {type(model).__name__}")
+                cosyvoice_model = {
+                    'model': model,
+                    'type': 'pytorch',
+                    'model_dir': model_dir,
+                    'has_inference': hasattr(model, 'inference_sft') or hasattr(model, 'inference')
+                }
+            
+            model_loaded = True
+            print("✓ 模型文件已下载并缓存")
             print("="*60 + "\n")
             return cosyvoice_model
         else:
-            print("⚠ 模型加载成功但缺少推理方法")
+            print("⚠ 未找到模型权重文件")
+            # 模型文件已下载，但没有找到权重文件
             cosyvoice_model = {
-                'model': model,
-                'type': 'transformers',
+                'model': None,
+                'type': 'downloaded',
+                'model_dir': model_dir,
                 'has_inference': False
             }
             model_loaded = True
+            print("✓ 模型文件已下载（但未找到权重文件）")
             print("="*60 + "\n")
             return cosyvoice_model
             
@@ -95,9 +130,10 @@ def load_cosyvoice_model():
         # 演示模式（加载失败）
         print("\n⚠ 使用演示模式")
         print("提示: 要使用完整功能，请确保:")
-        print("  1. 网络连接正常，可以访问Hugging Face")
+        print("  1. 网络连接正常，可以访问Hugging Face Hub")
         print("  2. 有足够的磁盘空间（约2GB）")
-        print("  3. transformers包版本 >= 4.35.0")
+        print("  3. 已安装 huggingface_hub: pip install huggingface_hub")
+        print("  4. (可选) 安装CosyVoice官方包: pip install cosyvoice")
         print("="*60 + "\n")
         
         cosyvoice_model = None