[PyTorch] Pair delayed-scaling FP8 recompute metadata per module - #3394

Open

nvegesna-netizen wants to merge 7 commits into

NVIDIA:mainfrom

nvegesna-netizen:fix/fp8-recompute-stash-pairing

Commits on Aug 21, 2026

Pair FP8 recompute restores with their stashes
nvegesna-netizen
committed
f075cb1
Test minimal FP8 eval recompute stash fix
nvegesna-netizen
committed
a182f75
Bypass FP8 recompute bookkeeping without autograd
nvegesna-netizen
committed
31b535c
Test explicit grad inside no-grad checkpoint
nvegesna-netizen
committed
45dadab
Address FP8 recompute review feedback
nvegesna-netizen
committed
a5800cc
Guard FP8 recompute state by autograd eligibility
nvegesna-netizen
committed
e9a8c2a

Commits on Aug 22, 2026

Merge branch 'main' into fix/fp8-recompute-stash-pairing
nvegesna-netizen
authored
71cb312