From e8eb0490ce098b1add05877363b185f3a7b570c5 Mon Sep 17 00:00:00 2001 From: Netanel Haber <58652339+netanel-haber@users.noreply.github.com> Date: Fri, 24 Apr 2026 14:53:23 +0300 Subject: [PATCH] [Bugfix][MoE] Unpad routed output before shared expert add [Fixes #35949] (#40794) Signed-off-by: Netanel Haber --- vllm/model_executor/layers/fused_moe/runner/moe_runner.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/vllm/model_executor/layers/fused_moe/runner/moe_runner.py b/vllm/model_executor/layers/fused_moe/runner/moe_runner.py index d6a6c0502c..d37e29938f 100644 --- a/vllm/model_executor/layers/fused_moe/runner/moe_runner.py +++ b/vllm/model_executor/layers/fused_moe/runner/moe_runner.py @@ -550,10 +550,14 @@ class MoERunner(MoERunnerInterface): hidden_states ) + # Record before `_maybe_pad_hidden_states` pads activations to match + # `moe_config.hidden_dim`, e.g. after `align_trtllm_fp4_moe_hidden_dim_for_fi` + routed_hidden_dim = hidden_states.shape[-1] hidden_states, og_hidden_dim = self._maybe_pad_hidden_states( shared_experts_input, hidden_states, ) + hidden_dim_was_padded = hidden_states.shape[-1] > routed_hidden_dim result = self._forward_entry( hidden_states, @@ -573,6 +577,8 @@ class MoERunner(MoERunnerInterface): # Extract outputs from result shared_output, fused_output = _unpack(result) + if hidden_dim_was_padded: + fused_output = fused_output[..., :routed_hidden_dim] # If combine kernel already reduced fused, reduce shared to match. # See note above re: the two all-reduce points.