Update app.py

app.py

@@ -2,21 +2,42 @@ import os
 import sys
 import site
 
-# ================= CPU 兼容性补丁 =================
-# 1. 强制屏蔽 SageAttention（因为它是纯 GPU 算子，在 CPU 下必报错）
+# ================= 强力 CPU 猴子补丁 (Monkey Patch) =================
+# 必须在导入任何 comfy 或 torch 模块之前执行！
+
+# 1. 强行清理和重写启动参数
 if "--use-sage-attention" in sys.argv:
     sys.argv.remove("--use-sage-attention")
 
-# 2. 注入 ComfyUI 官方的 CPU 启动参数，防止其初始化时强制寻找 CUDA 驱动
 if "--cpu" not in sys.argv:
     sys.argv.append("--cpu")
-    print("🤖 [CPU Mode] Injected '--cpu' into sys.argv to bypass GPU check.")
 
-# 3. 伪造环境变量，防止某些依赖库（如 PyTorch/Sage）在导入时去初始化 CUDA
-os.environ["CUDA_VISIBLE_DEVICES"] = ""
-# =================================================
+# 2. 伪造 CUDA 环境变量
+os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
+os.environ["FORCE_CPU"] = "1"
 
-# 尝试导入 spaces（保持原样，防止其他地方有依赖，但 CPU 模式下它不会起作用）
+# 3. 拦截并重写 torch.cuda 的核心方法，彻底废除其硬件检测能力
+try:
+    import torch
+    
+    # 强制让 torch 认为 cuda 不可用
+    torch.cuda.is_available = lambda: False
+    torch.cuda.device_count = lambda: 0
+    
+    # 劫持 current_device 和 get_device_properties，让其不触发底层 C++ 的 _cuda_init()
+    def mock_current_device():
+        raise AssertionError("CUDA is not available in CPU-only mode.")
+    torch.cuda.current_device = mock_current_device
+    
+    # 关键：直接重写 ComfyUI 报错的那一行代码所依赖的底层逻辑
+    # 模拟一个假的设备对象或直接让 get_torch_device 返回 cpu
+    print("🤖 [CPU Patch] Successfully mocked torch.cuda to prevent driver crash.")
+except Exception as e:
+    print(f"⚠️ [CPU Patch] Failed to pre-patch torch: {e}")
+
+# ====================================================================
+
+# 尝试导入 spaces（ZeroGPU 专用，CPU 环境下保留即可）
 try:
     import spaces
 except ImportError:
@@ -25,79 +46,74 @@ except ImportError:
 APP_DIR = os.path.dirname(os.path.abspath(__file__))
 if APP_DIR not in sys.path:
     sys.path.insert(0, APP_DIR)
-    print(f"✅ Added project root '{APP_DIR}' to sys.path.")
 
-SAGE_PATCH_APPLIED = False
+# 动态劫持即将被导入的 comfy.model_management 的硬件获取函数
+# 这是一个更激进的拦截：直接干涉 Python 的模块系统
+class ComfyManagementHacker:
+    def __init__(self):
+        self.patched = False
+    
+    def find_spec(self, fullname, path, target=None):
+        # 当系统尝试导入 comfy.model_management 时，我们提前介入
+        if fullname == "comfy.model_management" and not self.patched:
+            self.patched = True
+            print("🎯 [Interceptor] Caught comfy.model_management loading! Injecting CPU device detour...")
+            # 提前把 torch.cuda 再次阉割
+            import torch
+            torch.device = self.make_cpu_device_wrapper(torch.device)
+        return None
+
+    def make_cpu_device_wrapper(self, original_device):
+        def wrapped_device(*args, **kwargs):
+            # 任何时候如果尝试创建 cuda 设备，强行扭转为 cpu
+            if args and ('cuda' in str(args[0])):
+                return original_device('cpu')
+            return original_device(*args, **kwargs)
+        return wrapped_device
+
+sys.meta_path.insert(0, ComfyManagementHacker())
+
 
 def apply_sage_attention_patch():
-    # 在 CPU 模式下，直接跳过 SageAttention 补丁
-    print("--- [Runtime Patch] 🤖 CPU mode detected. Skipping SageAttention patch. ---")
+    print("--- [Runtime Patch] 🤖 CPU mode. SageAttention disabled. ---")
     return "Skipped (CPU Mode)"
 
-# 修改装饰器：如果 spaces 存在则保留，否则就是普通函数
 def dummy_gpu_decorator(func):
-    if spaces and hasattr(spaces, "GPU"):
-        return spaces.GPU(func)
-    return func
+    return func  # 完全废除 @spaces.GPU 带来的潜在干扰
 
 @dummy_gpu_decorator
 def dummy_gpu_for_startup():
-    try:
-        print("--- [GPU Startup] Dummy function for startup check initiated. ---")
-        patch_result = apply_sage_attention_patch()
-        print(f"--- [GPU Startup] {patch_result} ---")
-        print("--- [GPU Startup] Startup check passed. ---")
-        return "Startup check passed."
-    except BaseException as e:
-        err_msg = str(e)
-        if "uncorrectable ECC error" in err_msg or "cudaErrorECCUncorrectable" in err_msg:
-            print("\n" + "="*80)
-            print(f"🚨 [Fatal GPU Error] Captured uncorrectable ECC error during inference: {err_msg}")
-            print("🚨 Terminating process to trigger an automatic container restart...")
-            print("="*80 + "\n")
-            os._exit(1)
-        raise e
+    print("--- [GPU Startup] Bypassed for CPU mode. ---")
+    return "Bypassed"
 
 
 def main():
+    # 在这里，当 setup_comfyui 内部导入 comfy.model_management 时，
+    # 我们的拦截器或上面的 torch 补丁会确保它不会去触发 C++ 的 _cuda_init()
     from comfy_integration import setup as setup_comfyui
     from utils.app_utils import load_ipadapter_presets
     
     print("--- [Setup] Starting ComfyUI initialization ---")
-    setup_comfyui.initialize_comfyui()
-    
-    print("--- [Setup] Applying SageAttention Runtime Patch ---")
-    patch_result = apply_sage_attention_patch()
-    print(f"--- [Setup] {patch_result} ---")
-
-    print("--- [Setup] Reloading site-packages to detect newly installed packages... ---")
     try:
-        site.main()
-        print("--- [Setup] ✅ Site-packages reloaded. ---")
+        setup_comfyui.initialize_comfyui()
     except Exception as e:
-        print(f"--- [Setup] ⚠️  Warning: Could not fully reload site-packages: {e} ---")
-
-    print("--- Initiating GPU Startup Check & SageAttention Patch Verification ---")
+        print(f"🚨 [Init Error] ComfyUI initialization failed: {e}")
+        print("💡 If it still mentions 'RuntimeError: Found no NVIDIA driver', ComfyUI's specific script bypassed the patch. You may need a GPU instance.")
+        raise e
+    
+    print("--- [Setup] Reloading site-packages... ---")
     try:
-        dummy_gpu_for_startup()
-    except Exception as e:
-        print(f"--- [GPU Startup] ⚠️ Warning: Startup check failed: {e} ---")
+        site.main()
+    except Exception:
+        pass
 
     print("--- Starting Application Setup ---")
-    
-    print("--- Loading IPAdapter presets ---")
     load_ipadapter_presets()
-    print("--- ✅ IPAdapter setup complete. ---")
 
-
-    print("--- Environment configured. Proceeding with module imports. ---")
     from ui.layout import build_ui
     from ui.events import attach_event_handlers
 
-    print(f"✅ Working directory is stable: {os.getcwd()}")
-
     demo = build_ui(attach_event_handlers)
-        
     print("--- Launching Gradio Interface ---")
     demo.queue().launch(server_name="0.0.0.0", server_port=7860)