|
|
@@ -36,13 +36,15 @@
|
|
|
# - Engram on NFS: 158/176 (90), 401 pp, 33.7 tg, 5,061 ms TTFT — prefill
|
|
|
# collapse (io_uring row reads over NFS).
|
|
|
# - Engram on local NVMe (this recipe, composed_model.local_files + the
|
|
|
-# eugr lane's per-rank sparse copies): 149/176 (85), 3,654 pp (2.2x
|
|
|
-# prod), 43.9 tg, 826 ms TTFT (2.1x better than prod), KV 3.34M @131K.
|
|
|
-# Storage suspect CONFIRMED (prefill 9.1x faster than NFS run). Quality
|
|
|
-# 149 = run variance vs the NFS run (11 tool-discipline wobbles); the
|
|
|
-# r36-vs-prod gap is the engine stack (K7 vs K5). Status: CANDIDATE lane
|
|
|
-# — fastest prefill/TTFT we have, -18 quality vs prod. Full write-up:
|
|
|
-# history/night-shifts/2026-09-12.md ("Day follow-up 2 09-13").
|
|
|
+# eugr lane's per-rank sparse copies), 131K: 149/176 (85), 3,654 pp,
|
|
|
+# 43.9 tg, 826 ms TTFT. At 512K (our floor): **154/176 (88), 3,774 pp,
|
|
|
+# 45.7 tg, 822 ms TTFT, KV 9.62M tokens (18.3x @512K)** — prefill 2.3x
|
|
|
+# prod, TTFT 2.1x better, decode −3%. Storage suspect CONFIRMED (9.1x
|
|
|
+# prefill vs the NFS run). Quality −13 vs prod = the K7 stack's
|
|
|
+# long-horizon cluster (TC-82/85 fail-or-partial in every r36 run,
|
|
|
+# TC-60/74 flapping). Status: CANDIDATE lane — best prefill/TTFT in
|
|
|
+# the fleet, −13 quality. Full write-up: history/night-shifts/2026-09-12.md
|
|
|
+# ("Day follow-up 2" / "3 09-13").
|
|
|
#
|
|
|
# A/B target: the current dsv41 prod (eugr image, DSpark K5 probabilistic,
|
|
|
# 167/176, 1,642 pp / 47.0 tg / 1,738 ms TTFT @d0 c1, 512K, dev65). Same
|
