Model retained activations and allow selective recompute in TrainerRank - #915

Merged

bradhilton merged 15 commits into

mainfrom

fix/913-model-recompute-memory

Sep 17, 2026

Commits on Sep 16, 2026

Model retained activations for selective recompute admission
bradhilton
committed
9411a65
Calibrate recompute memory estimates on H200
bradhilton
committed
8dece8e
Normalize calibration CSV line endings
bradhilton
committed
be21a25
Account for TP, sequence parallelism and hybrid layer counts
bradhilton
committed
e1d00a3
Record expert topology and gradient health during calibration
bradhilton
committed
ad4e86b
Measure the minimum-memory admission layout before refusing
bradhilton
committed
d31f942
Cover eager MLP retention and calibrate GDN storage separately
bradhilton
committed
aef12a9
Record sharded GPU calibration and declare legacy test recompute mode
bradhilton
committed
57f9de2

Commits on Sep 17, 2026