update custom_all_reduce

vllm/config.py

@@ -1629,11 +1629,11 @@ class ParallelConfig:
             from vllm.executor import ray_utils
             ray_utils.assert_ray_available()
 
-        if not current_platform.use_custom_allreduce():
-            self.disable_custom_all_reduce = True
-            logger.info(
-                "Disabled the custom all-reduce kernel because it is not "
-                "supported on current platform.")
+        # if not current_platform.use_custom_allreduce():
+        #     self.disable_custom_all_reduce = True
+        #     logger.info(
+        #         "Disabled the custom all-reduce kernel because it is not "
+        #         "supported on current platform.")
         if self.ray_workers_use_nsight and not self.use_ray:
             raise ValueError("Unable to use nsight profiling unless workers "
                              "run with Ray.")

vllm/distributed/device_communicators/custom_all_reduce.py

@@ -132,7 +132,8 @@ class CustomAllreduce:
         assert current_platform.is_cuda_alike()
         fully_connected = current_platform.is_fully_connected(
             physical_device_ids)
-        if world_size > 2 and not fully_connected:
+        # if world_size > 2 and not fully_connected:
+        if not fully_connected:
             logger.warning(
                 "Custom allreduce is disabled because it's not supported on"
                 " more than two PCIe-only GPUs. To silence this warning, "