noonghunnaandClaude Opus 4.8 6cfcbc6bc0 consolidate(gemma-31b): single bf16 dual slug on v0.24.0, retire v0.22.0 composes ...

vllm/gemma-31b-dual now serves cyankiwi QAT-AWQ-int4 + bf16 KV @224K on stock
vLLM v0.24.0 (base.yml), OVERLAY-FREE — folds onto vllm-stable, retiring the
31b's vllm-gemma-stable dependence. Promoted to ⚠️ Production w/ caveats
(validated 2026-07-02: verify-full 9/9, verify-stress→210K @ 1162MB VRAM
margin, bench ~59 TPS, soak PASS).

bf16 not int8-PTH: on v0.24.0 int8-PTH allocates 262K but silently craters
recall past ~32K (needs PR #40391, open/unmerged upstream — the same cyankiwi
weights recall clean to 112K+ on v0.22.0+#40391). int8-PTH 262K returns free
when #40391 merges. MTP disabled (Gemma-4 MTP x tools broken on v0.24.0,
vLLM #39043 / #42006 closed-unmerged).

- remove the broken #537 int8.yml (recall cliff); base.yml replaces it
- DEFAULTS (gemma-4-31b,vllm,dual) -> vllm/gemma-31b-dual
- deprecate gemma-int8-mtp / gemma-bf16-mtp / qat-w4a16 (registry + header)
- fix 2 false "gemma int8-PTH native @262K" claims in vllm-stable.yml
- repoint launch/preflight/setup/switch hints off the deprecated slugs
- update resolver + setup-picker fixtures; add BENCHMARKS row

Suite: 59 pass / 1 pre-existing (test-submit-bench worktree-fixture-absent).

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EfF565T9eSLaqGzidyJ1Pm

2026-07-02 07:58:13 +00:00

..

fixtures

…

lib

consolidate(gemma-31b): single bf16 dual slug on v0.24.0, retire v0.22.0 composes

2026-07-02 07:58:13 +00:00

maintenance

…

systemd

Add nvidia-power-cap.service for boot-time GPU power cap (#483) (#485)

2026-06-26 21:08:26 +05:00

tests

consolidate(gemma-31b): single bf16 dual slug on v0.24.0, retire v0.22.0 composes

2026-07-02 07:58:13 +00:00

beellama-pin-bump.sh

…

bench-agentic.sh

feat(bench-agentic): decouple context ramp from tool-call success (#255) (#396)

2026-06-13 18:53:54 +05:00

bench.sh

scripts: auto-detect served model from /v1/models in verify/bench (#372) (#388)

2026-06-12 14:54:51 +05:00

c3t

Add c3t — the club-3090 test-console TUI (built by Qwen Max from the spec) (#428)

2026-06-18 07:31:02 +05:00

detect_nvlink.sh

detect_nvlink: auto-enable PCIe P2P when nvidia-smi reports it OK

2026-06-24 14:07:06 +00:00

diagnose-estate.sh

…

diagnose-profile.sh

…

generate-compose.sh

chore(254): deprecate the now-unused Genesis nightly engines + patches [Phase 3+4]

2026-06-05 14:44:21 +00:00

gpu-mode.sh

fix(studio): pin COMFYUI_ROOT to .env so ComfyUI mounts the models tree (#510, #530) (#531)

2026-06-30 18:29:51 +05:00

health.sh

Phase 2b: add --json/CLI contracts to 7 stack scripts (additive)

2026-06-18 08:43:00 +00:00

launch.sh

consolidate(gemma-31b): single bf16 dual slug on v0.24.0, retire v0.22.0 composes

2026-07-02 07:58:13 +00:00

power-cap-sweep.sh

…

preflight.sh

consolidate(gemma-31b): single bf16 dual slug on v0.24.0, retire v0.22.0 composes

2026-07-02 07:58:13 +00:00

pull.sh

Phase 2b: add --json/CLI contracts to 7 stack scripts (additive)

2026-06-18 08:43:00 +00:00

quality-baseline.sh

Add curated quality-baseline corpus + auto-diff (#252 Phase 1) (#397)

2026-06-13 19:11:35 +05:00

quality-test.sh

docs: correct --no-thinking help — four packs default thinking-on

2026-06-30 01:48:37 +00:00

rebench-full.sh

quality-test / rebench-full: forward --max-tokens to benchlocal-cli (#426)

2026-06-18 05:57:19 +05:00

rebench-report.py

…

rebench-runtime.sh

…

report.sh

report.sh: add --studio flag (AI Studio / ComfyUI container log tails) (#526)

2026-06-30 00:04:19 +00:00

setup-ai-studio.sh

setup-ai-studio.sh: make the "sign up then install the pipe" step unmissable (#510) (#527)

2026-06-30 00:08:52 +00:00

setup-image-studio.sh

studio: one-command setup-ai-studio.sh + auto-install the OWUI pipe

2026-06-24 03:47:18 +00:00

setup-video-studio.sh

studio: one-command setup-ai-studio.sh + auto-install the OWUI pipe

2026-06-24 03:47:18 +00:00

setup.sh

consolidate(gemma-31b): single bf16 dual slug on v0.24.0, retire v0.22.0 composes

2026-07-02 07:58:13 +00:00

soak-helper.py

soak-helper: count silent-empty by completion_tokens, not decode_tps

2026-06-11 12:41:17 +05:00

soak-test.sh

Fix soak-test container auto-detect: match by engine port, not model allowlist (#405) (#414)

2026-06-15 07:07:26 +05:00

stream-toolcall-probe.py

verify-full: add streaming tool-call check + a reusable streaming probe (#404)

2026-06-14 04:17:04 +05:00

submit-bench.sh

…

switch.sh

consolidate(gemma-31b): single bf16 dual slug on v0.24.0, retire v0.22.0 composes

2026-07-02 07:58:13 +00:00

update.sh

…

verify-full.sh

verify-full: add streaming tool-call check + a reusable streaming probe (#404)

2026-06-14 04:17:04 +05:00

verify-stress.sh

scripts: auto-detect served model from /v1/models in verify/bench (#372) (#388)

2026-06-12 14:54:51 +05:00

verify.sh

scripts: auto-detect served model from /v1/models in verify/bench (#372) (#388)

2026-06-12 14:54:51 +05:00