Repository navigation

GLM-5.3-Flash on two CUDA ranks: 16 streams at a 40960-token window - #984

Merged

ashhart merged 47 commits into

ashhart:mainfrom

BobClawblaw:pr/glm-16-streams

Oct 18, 2026

Merged

GLM-5.3-Flash on two CUDA ranks: 16 streams at a 40960-token window#984
ashhart merged 47 commits into
ashhart:mainfrom
BobClawblaw:pr/glm-16-streams

Commits on Oct 14, 2026
Commits on Oct 15, 2026
Commits on Oct 16, 2026
Commits on Oct 17, 2026
- glm cuda: index keys and gates in 4096-row rings (pool_keys reads only the current pool on): 0.674 GiB a stream at 40960, so --parallel 16 fits; rings rebase before a pass that would pass them, kep…
  BobClawblaw
  committed
  da126f2
- glm cuda: 16 streams a shared verify (2 rows each under the 32-row cap); needs a shared page pool to serve 16 at a long window
  BobClawblaw
  committed
  3fa37f7