Upload app.py with huggingface_hub

app.py

@@ -41,7 +41,9 @@ processor = AudioFlamingo3TemporalProcessor.from_pretrained(BASE_MODEL)
 model = AudioFlamingo3ForTemporalConditionalGeneration.from_pretrained(
     BASE_MODEL, dtype=torch.bfloat16, attn_implementation="sdpa"
 )
-model = PeftModel.from_pretrained(model, ADAPTER).merge_and_unload()
+# `torch_device="cpu"`: peft otherwise infers "cuda" and safetensors would try to
+# materialise the adapter on a GPU that does not exist yet under ZeroGPU.
+model = PeftModel.from_pretrained(model, ADAPTER, torch_device="cpu").merge_and_unload()
 model = model.eval().to("cuda")
 
 # ------------------------------------------------------------- utilities -----