Disable custom all reduce by default (#2808)

vllm/config.py

@@ -388,6 +388,7 @@ class ParallelConfig:
         if self.pipeline_parallel_size > 1:
             raise NotImplementedError(
                 "Pipeline parallelism is not supported yet.")
+        if not self.disable_custom_all_reduce and self.world_size > 1:
             if is_hip():
                 self.disable_custom_all_reduce = True
                 logger.info(
@@ -399,6 +400,15 @@ class ParallelConfig:
                     "Disabled the custom all-reduce kernel because it is not "
                     "supported with pipeline parallelism.")
 
+        # FIXME(woosuk): Fix the stability issues and re-enable the custom
+        # all-reduce kernel.
+        if not self.disable_custom_all_reduce and self.world_size > 1:
+            self.disable_custom_all_reduce = True
+            logger.info(
+                "Custom all-reduce kernels are temporarily disabled due to "
+                "stability issues. We will re-enable them once the issues are "
+                "resolved.")
+
 
 class SchedulerConfig:
     """Scheduler configuration.