Fix

Signed-off-by: Guyue Huang <guyueh@nvidia.com>

Author: guyueh1

nemo_rl/models/generation/vllm/vllm_worker.py

@@ -452,8 +452,11 @@ def _patch_vllm_hermes_tool_parser_thread_safety():
 
             # Use Ray for distributed execution in parallel mode
             vllm_kwargs["distributed_executor_backend"] = "ray"
+        elif self.expert_parallel_size > 1:
+            vllm_kwargs["distributed_executor_backend"] = "mp"
         else:
-            # For non-parallel mode, explicitly set executor to mp; Otherwise, it will fall back to
+            # For non-parallel mode, explicitly set executor to None to avoid Ray issues
+            vllm_kwargs["distributed_executor_backend"] = None
             # ray, which causes hang when we vllm_dp > 1 (i.e., ep_size > tp_size)
             vllm_kwargs["distributed_executor_backend"] = "mp"