yy-fighting
|
97c208cfdb
|
b300eq: TP2PP4 MRR64 retest (pass-1 + ordered v2 + MRR48 control) — c64 gains +27%/+19%, TTFT collapse, -14% 1K c32 config cost attributed
- deploy_glm53_pp4_mrr64.sh: byte-diff vs original = MRR 48->64 only; decode graph
stack-default bs<=256 (no graph change needed, arm never fell off graph)
- pass-1 + v2 (fresh instance, retraction-prone 16K c64 ordered last): all 6 points
aligned <=1.8% -> post-retraction contamination hypothesis disproved; anchor
regression is reproducible MRR64 behavior
- MRR48 fresh-instance control (4.1 c32 = 266.42): decomposes -17% anchor into
-3.3% instance freshness + -14.4% MRR64 config cost (per-req TPOT p50 87->102ms,
three instances, mechanism unidentified)
- 16K c64: pool-capped ~59 active, 3 retractions both runs (deterministic);
out -5% vs MRR48 queue-mode but TTFT P95 135->89s
- report/README/provenance/CURRENT.md overwritten in place per user instruction;
feishu wiki revision 13
|
2026-09-10 21:29:05 +08:00 |
|