From f670a036dff0a6b522a1b146e390d90c744481a5 Mon Sep 17 00:00:00 2001 From: bhsueh_NV <11360707+byshiue@users.noreply.github.com> Date: Wed, 7 May 2025 11:06:37 +0800 Subject: [PATCH] [Qwen3] chore: fix bug of fused_moe on tp > 1 (#4093) * fix bug of fused_moe on tp > 1 Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com> * refine codes Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com> --------- Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com> --- tensorrt_llm/_torch/modules/fused_moe.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/tensorrt_llm/_torch/modules/fused_moe.py b/tensorrt_llm/_torch/modules/fused_moe.py index fe69d899b0..f2728fbf98 100755 --- a/tensorrt_llm/_torch/modules/fused_moe.py +++ b/tensorrt_llm/_torch/modules/fused_moe.py @@ -821,10 +821,7 @@ class FusedMoE(nn.Module): final_hidden_states = final_hidden_states[0] if not self.enable_alltoall: - if self.reduce_results and self.parallel_size > 1: - return self.all_reduce(final_hidden_states) - else: - return final_hidden_states + return final_hidden_states else: return self.alltoall_combine(final_hidden_states, alltoall_info, token_count)