Fix use_ep

Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
2025-10-20 14:53:52 +08:00 · 2025-04-07 19:56:41 +00:00
1 changed files with 1 additions and 1 deletions
--- a/vllm/model_executor/layers/fused_moe/layer.py
+++ b/vllm/model_executor/layers/fused_moe/layer.py
@ -437,7 +437,7 @@ class FusedMoE(torch.nn.Module):
        # Use expert parallelism instead of tensor parallelism?
        vllm_config = get_current_vllm_config()
        use_ep = (vllm_config.parallel_config.enable_expert_parallel
-                  and self.tp_size > 1)
+                  and (self.tp_size != 1 or self.dp_size != 1))

        # For smuggling this layer into the fused moe custom op
        self.use_direct_call = self.dp_size == 1