Repository navigation

EXL3 Flash Next: draft lane guard for offloaded experts (8/8) - #1446

Draft

BlivionIaG wants to merge 8 commits into

ashhart:mainfrom

BlivionIaG:pr/exl3-s6-draft-guard

Draft

EXL3 Flash Next: draft lane guard for offloaded experts (8/8)#1446
BlivionIaG wants to merge 8 commits into
ashhart:mainfrom
BlivionIaG:pr/exl3-s6-draft-guard

Commits on Oct 29, 2026
- EXL3 Flash Next on CUDA: one rank, experts resident or offloaded to pinned host RAM, sm_120 gate
  BlivionIaG
  committed
  725041a
- EXL3 Flash Next: expert-parallel TP and MTP drafts
  BlivionIaG
  committed
  bd23ac2
- EXL3 Flash Next: deep prefill
  BlivionIaG
  committed
  505a649
- EXL3 Flash Next: KV pages from the core pool
  BlivionIaG
  committed
  5980c09
- KV pool: host tier
  BlivionIaG
  committed
  1fa5296
- EXL3 Flash Next: host-resident K/V on the tier
  BlivionIaG
  committed
  5809006
- EXL3 Flash Next: CUDA graphs and the device planner
  BlivionIaG
  committed
  3d310e6
- EXL3 Flash Next: draft lane guard for offloaded experts
  BlivionIaG
  committed
  8743eaa