[Refactor] Switch gradient checkpointing to the non-reentrant implementation (1/3) - #1979

Open

HAOCHENYE wants to merge 10 commits into

mainfrom

feat/sac-nonlegacy-base

Commits on Aug 19, 2026

[Refactor] Switch gradient checkpointing to the non-reentrant implementation

HAOCHENYE
authored and
jayhenry
committed
03f91e0
[Fix] Keep MTP gradients and the DSA top-k lifecycle correct under the new checkpointing

HAOCHENYE
authored and
jayhenry
committed
c8310bb
[Refactor] Drop the legacy reentrant checkpoint path

HAOCHENYE
authored and
jayhenry
committed
6716186
[Refactor] Remove the now-unused DSA lifecycle predicate

HAOCHENYE
authored and
jayhenry
committed
e35cd4d
[Fix] Let a checkpointed layer's input tensors reach the caller's saved-tensor hooks

HAOCHENYE
authored and
jayhenry
committed
d3063a8
[Fix] Preserve binding for checkpointed compiled forwards
jayhenry
committed
3dc5fd6
[Fix] Consume keyed decoder layer outputs
jayhenry
committed
864cbac
Fix FP8 MoE checkpoint replay with dynamic token shapes
jayhenry
committed
36f56fb
Default gradient checkpointing to reentrant
jayhenry
committed
f861fe5
[Fix] Keep FSDP outside activation checkpoint replay
jayhenry
committed
dca9485