Repository navigation

GLM paged attention and multi-sequence serving with prefix-aware scheduling - #642

Open

ClintMoody wants to merge 6 commits into

ClintMoody:paged-attn-main

Open

GLM paged attention and multi-sequence serving with prefix-aware scheduling#642
ClintMoody wants to merge 6 commits into
mw00:mainfrom
ClintMoody:paged-attn-main

Commits on Nov 12, 2026
- GLM: a paged latent cache - the page-table indirection, the prefix store and the gather kernel
  
  clintm
  authored and
  ClintMoody
  committed
  3e20fd6
- GLM: a dynamic multi-sequence scheduler, MTP speculation and the paged path's wiring
  
  clintm
  authored and
  ClintMoody
  committed
  c8fbf31
- GLM: seq_layout must mirror the load's mode pick, or the seq guard's recompute diverges in the slow legs (STRATA_GLM_SLOW=1: the load carves SlowF32 with the plain ring and full-ctx latents, the B4…
  
  clintm
  authored and
  ClintMoody
  committed
  8abc270
- glm: read the paged knobs under their MAYA_ names
  clintm
  committed
  e57d159
- test: the paged parity tests set the knobs under their MAYA_ names
  clintm
  committed
  23d5441
- glm: read every paged knob under its MAYA_ name
  ClintMoody
  committed
  beacc14