Update app.py

app.py

@@ -47,8 +47,7 @@ def fixed_pad(self, *args, **kwargs):
 ChatGLMTokenizer._pad = fixed_pad
 
 # ---------------------------
-# Load Kolors components
-# NOTE: dtype is fp16 on CUDA, fp32 on CPU to avoid NaNs on CPU
+# Load Kolors components (dtype fp16 on CUDA, fp32 on CPU)
 # ---------------------------
 text_encoder = ChatGLMModel.from_pretrained(
     f"{ckpt_dir}/text_encoder",
@@ -75,7 +74,6 @@ clip_image_encoder = CLIPVisionModelWithProjection.from_pretrained(
     torch_dtype=DTYPE,
     use_safetensors=True
 )
-# Prefer from_pretrained for config parity
 clip_image_processor = CLIPImageProcessor.from_pretrained(
     "openai/clip-vit-large-patch14-336"
 )
@@ -95,19 +93,11 @@ pipe = StableDiffusionXLPipeline(
 print("Models loaded successfully!")
 
 # ---------------------------
-# InsightFace helper (CPU by default; GPU if available)
+# InsightFace helper (force CPU provider to avoid CUDA init errors)
 # ---------------------------
 class FaceInfoGenerator:
     def __init__(self, root_dir: str = "./.insightface/"):
-        providers = ["CPUExecutionProvider"]
-        # Try to prefer CUDA provider if available in runtime
-        try:
-            import onnxruntime as ort
-            if "CUDAExecutionProvider" in ort.get_available_providers():
-                providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
-        except Exception:
-            pass
-
+        providers = ["CPUExecutionProvider"]  # GPU 없는 환경에서 안전
         self.app = FaceAnalysis(
             name="antelopev2",
             root=root_dir,
@@ -142,10 +132,9 @@ face_info_generator = FaceInfoGenerator()
 
 # ---------------------------
 # Inference function
-# - Uses fp16 autocast only on CUDA
-# - Ensures dtype/device consistency to avoid NaNs
+# - No @spaces.GPU decorator (GPU 없을 때 충돌 방지)
+# - Autocast only on CUDA
 # ---------------------------
-@spaces.GPU(duration=120)
 def infer(
     prompt,
     image=None,
@@ -159,7 +148,7 @@ def infer(
         gr.Warning("Please upload an image with a face.")
         return None, 0
 
-    # Detect face (InsightFace)
+    # Detect face (InsightFace on CPU)
     face_info = face_info_generator.get_faceinfo_one_img(image)
     if face_info is None:
         raise gr.Error("No face detected. Please upload an image with a clear face.")
@@ -179,7 +168,6 @@ def infer(
     pipe.text_encoder = pipe.text_encoder.to(device, dtype=DTYPE)
     pipe.unet = pipe.unet.to(device, dtype=DTYPE)
     pipe.face_clip_encoder = pipe.face_clip_encoder.to(device, dtype=DTYPE)
-
     face_embeds = face_embeds.to(device, dtype=DTYPE)
 
     # Load IP-Adapter weights (FaceID Plus)
@@ -235,6 +223,10 @@ def infer(
 
     return result, seed
 
+# If CUDA is available, optionally wrap with spaces.GPU for scheduling
+if torch.cuda.is_available():
+    infer = spaces.GPU(duration=120)(infer)
+
 # ---------------------------
 # Gradio UI
 # ---------------------------