use flashattn as default

vllm-project · Sep 6, 2024 · e144cb6 · e144cb6
1 parent 3e2cb99
commit e144cb6
Showing 1 changed file with 3 additions and 2 deletions.
diff --git a/vllm/executor/ray_gpu_executor.py b/vllm/executor/ray_gpu_executor.py
@@ -242,8 +242,9 @@ def sort_by_driver_then_worker_ip(worker):
             VLLM_INSTANCE_ID,
             "VLLM_TRACE_FUNCTION":
             str(envs.VLLM_TRACE_FUNCTION),
-            "VLLM_ATTENTION_BACKEND": (str(envs.VLLM_ATTENTION_BACKEND) if
-                                       envs.VLLM_ATTENTION_BACKEND else None),
+            "VLLM_ATTENTION_BACKEND":
+            (envs.VLLM_ATTENTION_BACKEND
+             if envs.VLLM_ATTENTION_BACKEND else "FLASH_ATTN"),
         }, ) for (node_id, _) in worker_node_and_gpu_ids]
 
         self._env_vars_for_all_workers = (