fix(space): lazy-load model to avoid import-time errors; pin huggingface-hub 0.14.1 (cached_download)

requirements.txt

@@ -8,7 +8,7 @@ opencv-python
 pillow
 numpy
 scipy
-huggingface-hub==0.17.3
+huggingface-hub==0.14.1
 controlnet-aux==0.0.10
 safetensors
 einops

server.py

@@ -12,7 +12,7 @@ import torch
 import numpy as np
 from PIL import Image
 
-from infer_full import StableHair
+# Lazy import performed in get_model() to avoid import-time failures on Space
 
 
 LOGGER = logging.getLogger("hair_server")
@@ -53,15 +53,17 @@ class HairSwapRequest(BaseModel):
 
 
 # Initialize model lazily on first request
-_model: Optional[StableHair] = None
+_model = None  # type: ignore[assignment]
 
 
-def get_model() -> StableHair:
+def get_model():
     global _model
     if _model is None:
         LOGGER.info("Loading StableHair model ...")
         device = "cuda" if torch.cuda.is_available() else "cpu"
         dtype = torch.float16 if device == "cuda" else torch.float32
+        # Import here to defer importing diffusers/transformers until needed
+        from infer_full import StableHair  # noqa: WPS433
         _model = StableHair(config="./configs/hair_transfer.yaml", device=device, weight_dtype=dtype)
         LOGGER.info("Model loaded")
     return _model