perf(glm52): launch-ahead decode — hide the ~0.7ms cuGraphLaunch under the previous step's execution#570

Merged

xiaguan merged 5 commits into

mainfrom

feat/glm52-device-fed-decode

Jul 5, 2026

Commits on Jul 5, 2026

perf(glm52): launch-ahead decode — speculative next-step replay hides the ~0.7ms cuGraphLaunch

xiaguan
and
claude
committed
fix(glm52): size the pinned argmax landing buffers per bucket

xiaguan
and
claude
committed
fix(glm52): make launch-ahead speculation all-ranks-or-none

xiaguan
and
claude
committed
docs(glm52): launch-ahead decode execution record — boundary idle 641-804us -> 43us, solo 19.5 ms/step

xiaguan
and
claude
committed
fix(glm52): toxic-review fixes — active_rows in the step shape, pre-capture shutdown broadcast, testable flag decision

xiaguan
and
claude
committed