[Build] Monkey patch RayWorkerWrapper to make CI work on 0.7.3-dev (#140)

examples/offline_distributed_inference_npu.py

@@ -32,7 +32,7 @@
 llm = LLM(
     model="Qwen/Qwen2.5-0.5B-Instruct",
     tensor_parallel_size=2,
-    distributed_executor_backend="mp",
+    distributed_executor_backend="ray",
     trust_remote_code=True,
 )
 

vllm_ascend/patch/ray_patch.py

@@ -0,0 +1,22 @@
+import torch_npu  # noqa: F401
+import vllm
+from vllm.executor.ray_utils import RayWorkerWrapper
+
+if RayWorkerWrapper is not None:
+
+    class NPURayWorkerWrapper(RayWorkerWrapper):
+        """Importing torch_npu in other Ray processes through an empty class and 
+        a monkey patch.
+
+        When Ray performs a remote call, it serializes the Task or Actor and passes 
+        it to the Worker process, where it is deserialized and executed. 
+
+        If no patch is applied, the default code of the RayWorkerWrapper provided 
+        by vLLM is used, which does not import torch_npu, causing an error in the 
+        Worker process.
+        See https://github.com/vllm-project/vllm-ascend/pull/92.
+        """
+
+        pass
+
+    vllm.executor.ray_utils.RayWorkerWrapper = NPURayWorkerWrapper

vllm_ascend/platform.py

@@ -103,6 +103,9 @@ def mem_get_info(cls) -> Tuple[int, int]:
 
     @classmethod
     def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
+        # RayWorkerWrapper monkey patch when setup
+        from vllm_ascend.patch import ray_patch  # noqa: F401
+
         parallel_config = vllm_config.parallel_config
         if parallel_config.worker_cls == "auto":
             parallel_config.worker_cls = "vllm_ascend.worker.NPUWorker"