noonghunnaandClaude Opus 4.8
3a5ece7bb5
gemma-26ba4b-single: promote INT8-PTH single → ⚠️ Production-w/-caveats (gate PASS)
rebench-full (gemma-26ba4b-int8r, 1x 3090 @370W) PASSED every gate:
verify-full ✓ · bench 168 narr / 217 code TPS (MTP AL 3.06-3.79) ·
verify-stress NIAH-clean → 161K (91% of 176K) ·
quality 109/150 think-ON (98/150 off) — on par with gemma-4-31B's 107/150 ·
soak 20x5 PASS, 0 MiB growth.
Flip Status 🧪 Experimental → ⚠️ Production-w/-caveats in the int8.yml header
(+ Quality line) and compose_registry (status experimental→caveats + gate-result
note). Add the BENCHMARKS.md row. Caveats: needs the #40391 overlay; 176K @
mem_util 0.94 (262K only without the MTP drafter — drafter weights + cudagraph
capture cost ~86K tok, 0.96 OOMs the capture tail); think-OFF agentic/extraction
softer (cli-40 30% / DataExtract 60%, recover to 52%/73% with thinking). Gate 42/42.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>