fix: load LoRA adapter weights to CPU for ZeroGPU startup compat

model_inference.py

@@ -88,7 +88,11 @@ class TransformersLLM:
 
         if adapter:
             print(f"🔗 Applying LoRA adapter: {adapter}")
-            model = PeftModel.from_pretrained(model, adapter)
+            # torch_device="cpu": load adapter weights to CPU first. On
+            # ZeroGPU, safetensors loading straight to cuda fails at startup
+            # ("No CUDA GPUs are available") — copying CPU tensors into the
+            # model's (emulated) CUDA params works fine.
+            model = PeftModel.from_pretrained(model, adapter, torch_device="cpu")
 
         model.eval()
         self.model = model