Upload folder using huggingface_hub

app.py

@@ -68,18 +68,15 @@ def load_pipeline() -> PixArtAlphaPipeline:
     pipe.vae.to(dtype=torch.float32)
 
     # Load LoRA adapters for transformer and text encoder.
-    # PeftModel.from_pretrained calls safe_load_file(device="cuda") because
-    # spaces patches torch.cuda.is_available() to True at module scope.
-    # Patch load_peft_weights to force CPU during loading, then .to("cuda").
+    # ZeroGPU patches torch.cuda.is_available() to True at module scope, so
+    # peft's infer_device() returns "cuda" and safe_load_file tries to use
+    # CUDA — which fails because there's no GPU at startup.  Temporarily
+    # make CUDA "unavailable" so peft loads weights on CPU; the subsequent
+    # pipe.to("cuda") is intercepted by ZeroGPU's hijack as expected.
     from peft import PeftModel
-    from peft.utils.save_and_load import load_peft_weights as _orig_load_peft
-    from peft.utils import save_and_load as _peft_sl
-
-    def _cpu_load_peft_weights(peft_model_local_path, device="cpu", **kw):
-        return _orig_load_peft(peft_model_local_path, device="cpu", **kw)
-
-    _peft_sl.load_peft_weights = _cpu_load_peft_weights
 
+    _orig_is_available = torch.cuda.is_available
+    torch.cuda.is_available = lambda: False
     try:
         pipe.transformer = PeftModel.from_pretrained(
             pipe.transformer,
@@ -94,7 +91,7 @@ def load_pipeline() -> PixArtAlphaPipeline:
             torch_dtype=DTYPE,
         )
     finally:
-        _peft_sl.load_peft_weights = _orig_load_peft
+        torch.cuda.is_available = _orig_is_available
 
     pipe.to("cuda")
     return pipe