[Fix] Avoid pickling entire LLMEngine for Ray workers (#3207)

Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>

vllm/engine/llm_engine.py

@@ -158,6 +158,11 @@ class LLMEngine:
         if USE_RAY_COMPILED_DAG:
             self.forward_dag = self._compiled_ray_dag()
 
+    def __reduce__(self):
+        # This is to ensure that the LLMEngine is not referenced in
+        # the closure used to initialize Ray worker actors
+        raise RuntimeError("LLMEngine should not be pickled!")
+
     def get_tokenizer_for_seq(self, sequence: Sequence):
         return self.tokenizer.get_lora_tokenizer(sequence.lora_request)
 
@@ -280,6 +285,8 @@ class LLMEngine:
         parallel_config = copy.deepcopy(self.parallel_config)
         scheduler_config = copy.deepcopy(self.scheduler_config)
         device_config = copy.deepcopy(self.device_config)
+        lora_config = copy.deepcopy(self.lora_config)
+        kv_cache_dtype = self.cache_config.cache_dtype
 
         for rank, (worker, (node_id,
                             _)) in enumerate(zip(self.workers,
@@ -295,22 +302,22 @@ class LLMEngine:
                     local_rank,
                     rank,
                     distributed_init_method,
-                    lora_config=self.lora_config,
-                    kv_cache_dtype=self.cache_config.cache_dtype,
+                    lora_config=lora_config,
+                    kv_cache_dtype=kv_cache_dtype,
                 ))
 
         driver_rank = 0
         driver_local_rank = node_workers[driver_node_id].index(driver_rank)
         self.driver_worker = Worker(
-            model_config,
-            parallel_config,
-            scheduler_config,
-            device_config,
+            self.model_config,
+            self.parallel_config,
+            self.scheduler_config,
+            self.device_config,
             driver_local_rank,
             driver_rank,
             distributed_init_method,
             lora_config=self.lora_config,
-            kv_cache_dtype=self.cache_config.cache_dtype,
+            kv_cache_dtype=kv_cache_dtype,
             is_driver_worker=True,
         )