Upload app.py with huggingface_hub

app.py

@@ -96,7 +96,12 @@ def load_model():
         hf_hub_download(ADAPTER_REPO, f"{ADAPTER_SUBFOLDER}/{fname}")
     adapter_dir = Path(hf_hub_download(
         ADAPTER_REPO, f"{ADAPTER_SUBFOLDER}/adapter_config.json")).parent
-    model = PeftModel.from_pretrained(base, str(adapter_dir))
+    # torch_device="cpu": peft otherwise infers "cuda" and safetensors tries to
+    # open the file straight onto the (ZeroGPU-virtual) GPU at import time,
+    # which dies with "No CUDA GPUs are available" in the main process.
+    # Loading on CPU here is free — .to("cuda") below moves everything.
+    model = PeftModel.from_pretrained(base, str(adapter_dir),
+                                      torch_device="cpu")
     model.eval()
     return model