Files
MeltMind/docs/track-a-shared-postfit-2026-11-19.md
T

2.4 KiB
Raw Blame History

Track A shared-slot dispatch delta

Claude 1795024014.778499 replaces the whole-GPU2 idle condition with one shared slot on GPU0–2. This entry remains a serial four-split prediction and CPU scoring job. It uses the existing c/b registry, upstream admission and original guard transfer and release operations. Earlier pending folds keep dispatch priority, including the already running b queue. GPU3 stays outside this entry. Reservations hold 4 GiB each of VRAM and host RAM, enforced by the b guard's own-process/recorded-peak and process-tree monitoring. A lease transfers to one live guard and is released only after its own guard and worker exit. No launch occurs as part of preparing the binding.

The scientific review is MeltAssist's a5fa4a6e frozen at 958abd8c. The original metric, A predictions, H2 checkpoints and normalization, four selected test lists, split 3 exclusion and paired bootstrap still come from that binding. The scoring function is AST-identical. The prediction worker adds lease and GPU assignment identity before using the same H2 inference and write path. Both the original scientific review and a new independent PASS tied to the exact shared dispatch binding are required. The complete frozen prediction manifest remains mandatory before any test target access.

The timing proxy uses c's recorded same-card epochs, scaled by the largest relative part, node or edge workload across the four selected Track A test lists and the c learning workloads. Each split uses the largest such proxy, an additional 300 seconds for the previous per-fold CPU allowance, a 25% margin on those two terms, and the fixed 60-second setup and 120-second cleanup allowances. The postfit budget reserves one largest split envelope for aggregate queue waiting, plus coordinator setup/cleanup. This is a conservative planning proxy, not a shared measurement of V4 predictions or compression. Real waiting is charged to the postfit total, outside each guard's prediction cap. The former 7,200 GPU-second and separate 162,000 GPU-second limits are upper bounds, and the 3,600 CPU-second limit remains in force. Original training cost 78,579.79901044304 GPU seconds is counted once.

The entry root is /home/frank14f/MeltMind-runs/track-a-postfit-four-split-shared-2026-11-19. The former unused four-split GPU2 binding remains historical evidence. Live c, seed 1 and b use their own frozen roots.