Fix GPU detection and CPU fallback: remove ssr_mode, add runtime GPU check, support CPU mode for FaceInference

app.py

@@ -17,13 +17,22 @@ from PIL import Image
 from huggingface_hub import snapshot_download
 
 # 检查 GPU 可用性（参考 Meigen-MultiTalk）
-is_gpu_available = torch.cuda.is_available()
-if is_gpu_available:
-    # 初始化 CUDA，确保设备可用
+# 参考: https://huggingface.co/spaces/fffiloni/Meigen-MultiTalk/blob/main/app.py
+is_shared_ui = True if os.environ.get('SPACE_ID', '').startswith('C4G-HKUST/AnyTalker') else False
+is_gpu_associated = torch.cuda.is_available()
+
+if is_gpu_associated:
     try:
-        _ = torch.cuda.current_device()
-    except RuntimeError:
-        is_gpu_available = False
+        num_gpus = torch.cuda.device_count()
+        print(f"GPU AVAILABLE: {num_gpus} GPU(s)")
+        if num_gpus > 0:
+            gpu_name = torch.cuda.get_device_name(0)
+            print(f"GPU Name: {gpu_name}")
+    except Exception as e:
+        print(f"GPU detection error: {e}")
+        is_gpu_associated = False
+else:
+    print("No CUDA-compatible GPU found. Will use CPU (slower).")
 
 # 导入 AnyTalker 相关的模块
 import wan
@@ -374,6 +383,25 @@ def run_graio_demo(args):
 
     os.makedirs(args.audio_save_dir, exist_ok=True)
 
+    # 运行时动态检测 GPU 可用性（参考 Meigen-MultiTalk）
+    if torch.cuda.is_available():
+        try:
+            num_gpus = torch.cuda.device_count()
+            if num_gpus > 0:
+                gpu_name = torch.cuda.get_device_name(0)
+                logging.info(f"GPU AVAILABLE: {num_gpus} GPU(s), Name: {gpu_name}")
+                # 使用 GPU
+                device = local_rank if world_size > 1 else 0
+            else:
+                logging.warning("CUDA is available but no GPU devices found. Using CPU.")
+                device = -1  # 使用 CPU
+        except Exception as e:
+            logging.warning(f"GPU detection error: {e}. Using CPU.")
+            device = -1  # 使用 CPU
+    else:
+        logging.warning("No CUDA-compatible GPU found. Using CPU (slower).")
+        device = -1  # 使用 CPU
+
     logging.info("Creating AnyTalker pipeline.")
     # 加载模型
     wan_a2v = wan.WanAF2V(
@@ -390,7 +418,9 @@ def run_graio_demo(args):
     )
     
     # 创建 InsightFace 人脸检测器
-    face_processor = FaceInference(det_thresh=args.det_thresh, ctx_id=local_rank)
+    # ctx_id=-1 表示使用 CPU，否则使用 GPU
+    face_processor_ctx_id = -1 if device == -1 else local_rank
+    face_processor = FaceInference(det_thresh=args.det_thresh, ctx_id=face_processor_ctx_id)
     logging.info("Model and face processor loaded successfully.")
 
     def generate_video(img2vid_image, img2vid_prompt, n_prompt, img2vid_audio_1, img2vid_audio_2, img2vid_audio_3,
@@ -672,7 +702,8 @@ def run_graio_demo(args):
         )
     # 参考 Meigen-MultiTalk 的成功配置
     # 在 Hugging Face Spaces 上，Gradio 会自动处理端口和服务器配置
-    demo.queue(max_size=4).launch(ssr_mode=False, show_error=True, show_api=False)
+    # 参考 Meigen-MultiTalk 的启动方式
+    demo.queue(max_size=4).launch(show_error=True, show_api=False)
 
         
 

utils/get_face_bbox.py

@@ -17,12 +17,21 @@ class FaceInference:
         Args:
             det_thresh: 检测阈值
             det_size: 检测图像尺寸
-            ctx_id: GPU设备ID，如果为None则自动检测当前进程的local_rank
+            ctx_id: GPU设备ID，如果为-1则使用CPU，否则使用GPU
         """
+        # 如果 ctx_id 为 -1，使用 CPU；否则使用 GPU
+        if ctx_id == -1:
+            providers = ['CPUExecutionProvider']
+            provider_options = [{}]
+            ctx_id = -1  # InsightFace 使用 -1 表示 CPU
+        else:
+            providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
+            provider_options = [{"device_id": str(ctx_id)}, {}]
+        
         self.face_analysis = FaceAnalysis(
             allowed_modules=['detection'], 
-            providers=['CUDAExecutionProvider'], 
-            provider_options=[{"device_id": str(ctx_id)}], # make sure use provider_options to specify GPU ran
+            providers=providers, 
+            provider_options=provider_options,
         )
         
         self.face_analysis.prepare(ctx_id=ctx_id, det_thresh=det_thresh, det_size=det_size)