Voxtral STT: unblock concurrent batch and ship tuned L40S config - #1

Open

AzimovS wants to merge 6 commits into

Ramshreyas:mainfrom

AzimovS:voxtral-tuning

Commits on May 14, 2026

fix(sweep): write VLLM_MXFP4_USE_MARLIN=0 instead of empty string
AzimovS
committed
d30531e
feat(voxtral): add patched image with PR #39229 for concurrent batch
AzimovS
committed
9f59a69
bench: add four L40S-tuned voxtral patched variants for sweep
AzimovS
committed
905f628
bench: drop FP8 KV from broken voxtral variants, add 0.92 safety variant
AzimovS
committed
3fd1583
voxtral: consolidate to single production config, 128 concurrent users on L40S
AzimovS
committed
f006aad
cleanup: stale comment, restore baseline level, Dockerfile retirement note
AzimovS
committed
ea769de