vLLM reverses yesterday's DeepSeek-V4.1-Flash MegaMoE padding and proves zero output error on GB200

GitHub·high signalXBlueskyLinkedInCopy link

PR #57204, merged 2026-09-16T19:09Z, removes the MegaMoE intermediate-width padding and the shared-weight padding workaround from #56568 that had widened DeepSeek-V4.1-Flash's 2304 checkpoint width to 2560. The bundled DeepGEMM uses `layout::Data(..., false)` for activation-scale rows, so the old 16-byte TMA row alignment rationale no longer applies; removing it strips 10% of the intermediate dimension from the padded expert GEMMs while keeping native shared-expert fusion enabled. Validation on four GB200 GPUs against upstream `42919b49c5` across token counts 1 through 2048 reported 48 of 48 comparisons with exactly zero output error, and prior GSM8K runs at TP4/EP4 held at 96.36% for the native-width path.