Preload LoRA tensors for ZeroGPU

app.py

@@ -32,6 +32,12 @@ def load_runtime():
                 trust_remote_code=True,
                 dtype=torch.bfloat16,
             ).to("cuda").eval()
+            # Register both adapters before ZeroGPU packs the model tensors.
+            # Adding CUDA parameters later inside a request is less reliable.
+            for adapter_repo in (repo for repo in ADAPTERS.values() if repo):
+                _model.load_lora_pretrained(adapter_repo, activate=False)
+                _loaded_adapters.add(adapter_repo)
+            _model.activate_single_lora(None)
         return _model, _tokenizer