G-series — four arms, lr 2e-6

10-19 21:36 UTC · auto-refresh 30s · TensorBoard →
armstepprogresslrstatusnexteval
G0
control (E3 recipe, lr 2e-6)
1,501
/24,704 · 48,032 samples · saved 32m ago
6.1% · warmup
1.215e-06 (61%)Running 4n 4.2hwarmup in 2.4h-
G1
rim fix + un-gated highlights + ci + prompt
1,501
/24,704 · 48,032 samples · saved 28m ago
6.1% · warmup
1.215e-06 (61%)Running 4n 4.2hwarmup in 2.4h-
G2
reference lit differently from target
1,001
/19,763 · 40,040 samples · saved 49m ago
5.1% · warmup
1.013e-06 (51%)Running 5n 3.4hwarmup in 2.5h-
G3
o-mask from 0917 + bg 0/0.5/1
1,501
/24,704 · 48,032 samples · saved 18m ago
6.1% · warmup
1.215e-06 (61%)Running 4n 4.1hwarmup in 2.4h-
lr is read from scheduler.bin, not assumed — every run before this series used 1e-5, so a stale config is the easiest mistake to make. Red lr = above 3e-6, i.e. almost certainly 1e-5. G2 runs 5 nodes / 40 ranks, so its steps are NOT comparable to the others': compare it sample-matched — use the samples figure, which is the one number that means the same thing in all four arms (G2 moves 40 per step, the rest 32). A red node count means volc disagrees with the configured nodes, which would make this arm's steps/epoch and every % here wrong.