noonghunna 4d47d77fce Prune dual vLLM composes: qwen-27b -> one config; gemma-31b default -> gemma-int8 (#279) ...

Settle the per-model dual vLLM set. qwen3.6-27b dual -> ONE (vllm/dual fp8 262K vision MTP); deprecate dual-dflash/dflash-noviz/tq3-nomtp/bf16/int8. gemma-4-31b dual -> keep TWO: DEFAULTS moved gemma-mtp -> gemma-int8 (full 262K + vision + 4 streams; rides v0.21.0+#40391 overlay), gemma-mtp kept as the stable v0.22.0 32K fallback. Registry entries kept (deprecated, not deleted) so patches.yml + hardware-gating tests stay valid. Out of scope: carnice/qwopus fine-tunes, gemma-26b-a4b, multi4. Suite 38/38.

2026-05-31 07:45:07 +05:00

..

profiles

Prune dual vLLM composes: qwen-27b -> one config; gemma-31b default -> gemma-int8 (#279)

2026-05-31 07:45:07 +05:00

bench-row-formatter.sh

refactor(launch): derive launcher tables from the registry + <engine>/default resolver

2026-05-26 18:35:35 +00:00

compose-meta.sh

Prune Gemma vLLM variants (9->3) and pin survivors to v0.21.0

2026-05-29 07:59:42 +05:00

generate_compose.py

feat(pull): v0.8.2 STEP V3 — arch-registry expansion + chat-template attribution/drift_guard

2026-05-18 18:29:58 +00:00

registry-emit.sh

feat: model-default resolver + user-pinnable defaults (PR-B) (#266)

2026-05-30 22:00:36 +05:00

report_calib.sh

fix(#168): scope report.sh kv-calc calibration to the running model

2026-05-23 22:19:37 +00:00