These patch files were generated from the active local sources that reproduced the 89-class MiniMax result on 2026-05-20.
vllm-active-promoted-minimax-89tps-20260520.patch.gz.b64
vllm-project/vllm checkout at commit c51df43005726a09c6eb7348e8c1b00501c70a8e.llm-scaler-active-promoted-minimax-89tps-20260520.patch.gz.b64
vllm/custom-esimd-kernels-vllm subdirectory of the llm-scaler checkout at commit 4bfc0070090cc54afdb2d46b8e57882359141568.The patches are intentionally broad active snapshots rather than a polished upstream patch series. They are meant to let another machine reproduce this benchmark path first. Upstreamable cleanup should be done separately after reproducing quality and speed.
Decode manually with:
base64 -d vllm-active-promoted-minimax-89tps-20260520.patch.gz.b64 | gzip -dc > /tmp/vllm.patch
base64 -d llm-scaler-active-promoted-minimax-89tps-20260520.patch.gz.b64 | gzip -dc > /tmp/llm-scaler.patch