Fix: force adapter safetensors load onto CPU (device_map=cpu) to avoid ZeroGPU emulation gap

app.py

@@ -49,7 +49,17 @@ SYSTEM_PROMPT = SYSTEM_PROMPT_TEMPLATE.format(
 
 tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
 base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype="auto")
-model = PeftModel.from_pretrained(base_model, ADAPTER, subfolder=ADAPTER_SUBFOLDER)
+# device_map="cpu" is required here: PeftModel.from_pretrained's internal
+# safetensors load otherwise infers a CUDA target device from
+# torch.cuda.is_available() (which ZeroGPU's emulation layer reports as
+# True even at module-load time, before a real GPU is allocated), and that
+# specific direct-to-device safetensors load path is not covered by the
+# emulation, causing "RuntimeError: No CUDA GPUs are available". Loading
+# on CPU here and moving the whole composed model to CUDA afterward avoids
+# that code path entirely.
+model = PeftModel.from_pretrained(
+    base_model, ADAPTER, subfolder=ADAPTER_SUBFOLDER, device_map="cpu"
+)
 model.to("cuda")
 model.eval()