Cherry-pick Lora Fix (#2960)

Author: Lokiiiiii

engines/python/setup/djl_python/lmi_vllm/vllm_async_service.py

@@ -270,6 +270,8 @@ async def add_lora(self, lora_name: str, lora_alias: str, lora_path: str):
         logging.info(f"Adding LoRA {lora_name} from {lora_path}")
         lora_id = self.lora_id_counter.inc(1)
         lora_request = create_lora_request(lora_name, lora_id, lora_path, None)
+        # Register the LoRA request with the model registry so vLLM can find it
+        self.model_registry.lora_requests[lora_name] = lora_request
         self.lora_requests[lora_request.lora_name] = lora_request
         result = await self.vllm_engine.add_lora(lora_request)
         logging.info(f"LoRA {lora_name} added to engine: {result}")
@@ -282,6 +284,7 @@ async def remove_lora(self, lora_name: str, lora_alias: str):
         lora_request = get_lora_request(lora_name, self.lora_requests)
         result = await self.vllm_engine.remove_lora(lora_request.lora_int_id)
         del self.lora_requests[lora_name]
+        del self.model_registry.lora_requests[lora_name]
         return result
 
     async def pin_lora(self, lora_name: str, lora_alias: str):