atom GOOD ckpt 430080000 ids (ema_tf 1.130)

Atom trainer: spike guard, rollback request, good checkpoints, cosine LR; docs on the two divergences

Repair checkpoints/atom: restore healthy 409.6M-id checkpoint (replaces diverged 829.4M push)

atom ckpt periodic 2026-10-13 11:52:48

atom ckpt first 2026-10-13 01:49:36

Archive healthy AdamW checkpoint 409.6M ids (before divergence at ~710M)

atom ckpt periodic 2026-10-12 16:01:57

atom ckpt first 2026-10-12 05:59:16

Fix NaN gradient in elu_plus_one (exp overflow in discarded torch.where branch); docs

atom ckpt first 2026-10-12 03:24:26

ATOM training fixes (load, ss_prob, LR warmup, optimizer state) + Oct 2026 training log (synced with GitHub 78199e5)

SGD atom manifest 61.44M

SGD atom ckpt 61.44M ids (archive)

atom ckpt first 2026-10-12 02:08:29

atom ckpt first 2026-10-11 20:36:19

rm test file

test _t3.bin

test _t2.pt

test small pt

atom ckpt first (pt)

atom ckpt first 2026-10-11 20:33:58

README: fix Status contradiction.

README: 5896 ids/s at B=16.

convert_atom.py skips existing.

Atom corpus partial (110 files, 8.3GB). Resume-ready.

Compile the pure core. B=16 hits 5896 ids/s.

Capacity MoE.

Grouped MoE.

Per-head carry.

Remove item syncs.

Measured numbers.

Factor path.

Triton MoE first cut.

GPU benchmark numbers.

Compare TF rates.

Lanes and per-row BOS reset.

README: BOS reset and carry metric.

Reset at BOS.

Status 2026-10-09. Exact CPU rates.

Speech measured on smoke.

GPU benchmark protocol.

Calibrate phase_proj to c/sqrt(d).

Learned phase encode. Expert collapse fixed.

Document real expert hits and theta-bar vs spacing.

Per-token MoE routing. Causal leak closed.

Document last-theta chunk routing and the inert leak.

Equal-T Kuramoto check: one step matches four, gates unchanged.

Document single RK4 step and the r sweep.

Planning table: days = 2.4B ids / measured ids per second.

fast4gpu_atom trainer, guards passed, CPU smoke 1378 ids/s.