Upload app.py with huggingface_hub

app.py

@@ -17,7 +17,7 @@ import cv2
 import numpy as np
 import torch
 import torch._dynamo
-from huggingface_hub import snapshot_download, login as hf_login
+from huggingface_hub import snapshot_download
 import huggingface_hub.constants as _hf_const
 if not hasattr(_hf_const, "HF_HUB_ENABLE_HF_TRANSFER"):
     _hf_const.HF_HUB_ENABLE_HF_TRANSFER = False
@@ -69,8 +69,11 @@ except Exception as _e:
 # Space restarts without re-downloading.
 LORA_DIR  = Path("/data/loras")
 HF_TOKEN  = os.environ.get("HF_TOKEN", "")
+# Set both env-var spellings so all HF libraries pick up the token
+# without needing to call login() (which hits whoami and gets rate-limited).
 if HF_TOKEN:
-    hf_login(token=HF_TOKEN)
+    os.environ["HF_TOKEN"] = HF_TOKEN
+    os.environ["HUGGING_FACE_HUB_TOKEN"] = HF_TOKEN
 
 # Delete the old lkzd7 LoRA directory if it still exists — it consumed 15-20 GB
 # of storage and contained no identity-preserving LoRAs.
@@ -341,12 +344,10 @@ _aoti_saved: list[tuple] = []
 def _init_pipeline():
     global pipe, original_scheduler, _aoti_saved
 
-    # Re-authenticate inside the GPU worker — the login from startup may not
-    # carry into this process context.
+    # Ensure token env-vars are set in this worker context.
     if HF_TOKEN:
         os.environ["HF_TOKEN"] = HF_TOKEN
         os.environ["HUGGING_FACE_HUB_TOKEN"] = HF_TOKEN
-        hf_login(token=HF_TOKEN)
 
     print(f"Loading pipeline: {MODEL_ID}")
     pipe = WanImageToVideoPipeline.from_pretrained(