[Perf] Eliminate padding and slicing op for GPT-OSS with Flashinfer MXFP4 MXFP8 MoE (#30647)

Signed-off-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
This commit is contained in:
elvischenv
2026-03-18 15:01:26 +00:00
committed by GitHub
parent c373b5c00d
commit 296839a1b0
6 changed files with 40 additions and 3 deletions
@@ -415,7 +415,10 @@ class DefaultMoERunner(MoERunner):
# This is the dimension after transform (for routed expert output slicing)
transformed_hidden_dim = hidden_states.shape[-1]
if self.moe_config.hidden_dim != transformed_hidden_dim:
if (
not self.quant_method.skip_forward_padding
and self.moe_config.hidden_dim != transformed_hidden_dim
):
hidden_states = F.pad(
hidden_states,
(0, self.moe_config.hidden_dim - transformed_hidden_dim),