Skip to content
dormouse

jepa3 — усиление JEPA (2026-10-03)

Lane: jepa3. Worktree wt/jepa3 off 10b526d. No GPU arms were runnable: night2e held the single GPU (PID 2511393, 8.4–9.9 GB used, b8 s512, ~245 ms/step, step ~14 500 of 80 000 during the lane window) and the brief forbids evicting it (pgrep -x train empty + nvidia-smi <1 GB required; neither held in any polled window). Arms are scripted, not run.

Done

Section titled “Done”

  1. Mask-спан стал флагом — dd27bb3. jepa_mask_span already existed in the schema (default 8, reachable only via --set); it is now the typed flag --jepa-span, off-default = preset value (8). Smoke: 1 CPU step, nano, b2 s64, --jepa-span 4 → ce=5.566 aux=0.0257, checkpoint written; cfg snapshot test green (16 passed). Deliverable 3 ✓.

  2. Precompute-качество (главный) — NOT RUN (night2e). Ready to launch when the card frees: /tmp/opencode/jepa3/lane.sh — 2k steps × {live EMA teacher vs --jepa-precompute 2000 + --jepa-targets} × seeds {1,2}, b8 s512 fp8-auto, held-out window 20·8·512 = 81920 B, then a 200-step DM_LAUNCH_ATLAS=1 pair for the launch delta. Expectation from the atlas lane (docs/reviews/launch-atlas-2026-10-02.md): the JEPA teacher+aux arm owns 22.3% of step launches, so cutting the per-step teacher forward should show up as roughly that share of launches and a comparable share of the fixed ~119 ms teacher-forward ceiling. Verdict rule per brief: precompute within the 2-seed spread → банк.

  3. Весовой sweep 0.05 / 0.15 / 0.3 — NOT RUN (night2e). Filter-level 2k × 1 seed per value once the card frees; winner vs control goes to the 3-seed spread ladder.

Commits

Section titled “Commits”
  • dd27bb3 feat(train): –jepa-span flag for the JEPA mask span

Follow-ups (owner decides)

Section titled “Follow-ups (owner decides)”
  • Run /tmp/opencode/jepa3/lane.sh after night2e exits 0 (check ~/logs/night_20261003.log tail for done), then fill the verdict table here and in the lane report.
  • history.tsv untouched — no GPU measurement taken this lane.