Account for grouped checkpoint, head, and enclosing forward buffers - #898

Merged

bradhilton merged 42 commits into

mainfrom

schulman/848-checkpoint-head-demand-20260915

Sep 24, 2026

Commits on Sep 16, 2026

Account for grouped checkpoint and shared-expert memory demand
bradhilton
committed
b430f5e
Narrow optional model owners before metadata inspection
bradhilton
committed
688de07
Accept ART-owned GDN norm wrappers in pending memory floor
bradhilton
committed
23ab6ae
Price gated shared saves in checkpoint gradient workspace
bradhilton
committed
c00a459
Price qualified expert conversion and checkpoint return stages
bradhilton
committed
5a61685
Run converted-memory regressions in Megatron CI environment
bradhilton
committed
0c708ab

Commits on Sep 17, 2026

Commits on Sep 19, 2026

Account for checkpoint input gradients in static peak admission
bradhilton
committed
32c874f
Preserve split forward ordering with checkpoint gradient peak allowance
bradhilton
committed
be3fe54
style: format checkpoint gradient memory tests
bradhilton
committed
f570804
Route checkpoint gradient memory tests to Megatron environment
bradhilton
committed
bb39499
test(trainer-rank): include checkpoint gradients in head peak expectations
bradhilton
committed
8b3e1df
fix(trainer-rank): retain generic admission for declined MoE layouts
bradhilton
committed
935e0e3
Private integration: merge main 5b7acf06d209 into PR #898
Codex Halley
committed
7e60bce
Fix cold-signature planner fixture slot references
Codex Halley
committed
05e8ebf

Commits on Sep 21, 2026

Compose reviewed ART #898 with merged #929
bradhilton
committed
f26eda2

Commits on Sep 22, 2026

Commits on Sep 23, 2026

Commits on Sep 24, 2026

Merge landed cache-recovery dependency into admission branch
bradhilton
committed
25bff2e