Force adapter load onto CPU (PEFT infer_device picks cuda on ZeroGPU)

app.py

@@ -38,14 +38,17 @@ DEFAULT_INSTRUCTION = (
 
 processor = AutoProcessor.from_pretrained(BASE_MODEL_ID)
 
-# Build and merge entirely on CPU. With device_map="cuda" here, PEFT loads the adapter
-# safetensors straight onto CUDA, which doesn't exist yet on ZeroGPU at startup
-# ("No CUDA GPUs are available"). The single .to("cuda") below is the pattern `spaces`
-# intercepts and replays once a GPU is attached.
+# Build and merge entirely on CPU, then move once.
+#
+# torch_device="cpu" is load-bearing on ZeroGPU: PEFT picks the adapter's load device via
+# infer_device(), which returns "cuda" because torch.cuda.is_available() reports True at
+# startup even though no GPU is attached yet. safetensors then materialises straight onto
+# CUDA and dies with "No CUDA GPUs are available". Module .to("cuda") is patched by
+# `spaces` and replayed when a GPU attaches; a direct safetensors CUDA load is not.
 model = Gemma3ForConditionalGeneration.from_pretrained(
     BASE_MODEL_ID, dtype=torch.bfloat16, device_map="cpu"
 )
-model = PeftModel.from_pretrained(model, ADAPTER_ID)
+model = PeftModel.from_pretrained(model, ADAPTER_ID, torch_device="cpu")
 model = model.merge_and_unload().eval().to("cuda")
 
 _eot = processor.tokenizer.convert_tokens_to_ids("<end_of_turn>")