Repository navigation

GLM-5.3-Flash on two CUDA ranks: DFlash2 drafts (--drafter) - #979

Merged

ashhart merged 48 commits into

ashhart:mainfrom

BobClawblaw:pr/glm-dflash2

Oct 18, 2026

Merged

GLM-5.3-Flash on two CUDA ranks: DFlash2 drafts (--drafter)#979
ashhart merged 48 commits into
ashhart:mainfrom
BobClawblaw:pr/glm-dflash2

Commits on Oct 14, 2026
Commits on Oct 15, 2026
Commits on Oct 16, 2026
Commits on Oct 17, 2026
- glm cuda: DFlash2 drafts sized by their chances (the selector's, chained, cut at glm_cuda's 0.7) against measured window costs; TF_GLM_DFLASH2_LOG and _PROF
  BobClawblaw
  committed
  9f638dc
- glm cuda: DFlash2's linears at 4 bits by default (--drafter-bits 4|0): quantize4 and q4 tiles, inputs' group sums by df_gsum; down and the selector projection stay bf16
  BobClawblaw
  committed
  70faca3