From 12a3f6454b973d7cd8806d398ba287a7e1d22c63 Mon Sep 17 00:00:00 2001 From: Netanel Haber <58652339+netanel-haber@users.noreply.github.com> Date: Sat, 25 Apr 2026 23:50:12 +0300 Subject: [PATCH] [Bugfix][MoE] Only unpad routed output before shared expert add or routed output transform (#40865) Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com> Co-authored-by: Jee Jee Li --- vllm/model_executor/layers/fused_moe/runner/moe_runner.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) 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 d37e29938f..6b13c0c363 100644 --- a/vllm/model_executor/layers/fused_moe/runner/moe_runner.py +++ b/vllm/model_executor/layers/fused_moe/runner/moe_runner.py @@ -552,6 +552,8 @@ class MoERunner(MoERunnerInterface): # 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` + # so routed output can be trimmed before + # shared+routed add / latent up proj if needed. routed_hidden_dim = hidden_states.shape[-1] hidden_states, og_hidden_dim = self._maybe_pad_hidden_states( shared_experts_input, @@ -577,7 +579,9 @@ class MoERunner(MoERunnerInterface): # Extract outputs from result shared_output, fused_output = _unpack(result) - if hidden_dim_was_padded: + if ( + shared_output is not None or self.routed_output_transform is not None + ) and hidden_dim_was_padded: fused_output = fused_output[..., :routed_hidden_dim] # If combine kernel already reduced fused, reduce shared to match.