perf(dflash): split-chain fast rollback with F32 SSM checkpoints (+29% decode) - #506

Merged

davide221 merged 7 commits into

Luce-Org:mainfrom

Amidwestnoob:perf/dflash-split-chain-fast-rollback-main

Jul 30, 2026

Commits on Jul 29, 2026

perf(dflash): split-chain fast rollback with F32 SSM checkpoints (+29% decode)

Drew Schuyler
authored and
peterabutle
committed
60577cd
fix(dflash): address split rollback review findings

Drew Schuyler
authored and
peterabutle
committed
5b7629f
fix(dflash): gate split capture diagnostics

Drew Schuyler
authored and
peterabutle
committed
7fe92ee
perf(qwen35): batch recurrent snapshot synchronization

authored and
peterabutle
committed
abf2f24
fix(qwen35): gate split rollback checkpoints and support HIP
davide221
committed
8585b6c
Merge main into perf/dflash-split-chain-fast-rollback-main
davide221
committed
b22570e

Commits on Jul 30, 2026

feat(server): expose split rollback option
davide221
committed
19622d7