From 67e30b0ccee54eb1be39217267895422bda77783 Mon Sep 17 00:00:00 2001
From: Matthias Reso <13337103+mreso@users.noreply.github.com>
Date: Wed, 11 Sep 2024 22:59:00 +0000
Subject: [PATCH] Use spawn instead of fork method for vllm

---
 ts/torch_handler/vllm_handler.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/ts/torch_handler/vllm_handler.py b/ts/torch_handler/vllm_handler.py
index 927efe93e2b..02f7ee77b0a 100644
--- a/ts/torch_handler/vllm_handler.py
+++ b/ts/torch_handler/vllm_handler.py
@@ -1,5 +1,6 @@
 import asyncio
 import logging
+import os
 import pathlib
 import time
 from unittest.mock import MagicMock
@@ -40,6 +41,8 @@ def initialize(self, ctx):
         vllm_engine_config = self._get_vllm_engine_config(
             ctx.model_yaml_config.get("handler", {})
         )
+        
+        os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 
         self.vllm_engine = AsyncLLMEngine.from_engine_args(vllm_engine_config)