noonghunna
8e7f5a87a5
feat(gemma): gemma duals → vLLM v0.22.0 (rebase #40391 lean + re-instate #42006) (#287)
Both gemma duals on immutable v0.22.0. int8-mtp = #40391 (rebased, lean diff-apply) + #42006; bf16-mtp = #42006. #40391 rebased onto v0.22.0 (old full-module copies ImportError'd) + re-delivered as install_script diff (−13K lines). #42006 streaming-multi-tool fix re-instated on both (live-repro'd: streamed multi-tool dropped non-last args). Dropped #41800/#41991 (in stock v0.22.0); fixed stale #41800 patches.yml entry. int8 default 98K→262K. Engine-profile-injection gotcha documented in CLAUDE.md. Validated on real v0.22.0 (docker-inspect): int8 pool 447K@262K bench 95.7/125.8, bf16 pool 195K@131K, streaming multi-tool keeps all args on both. Suite 41/41.