Repository navigation

fix(dspark): reduce V2 working VRAM - #27

Closed

alexbi29 wants to merge 8 commits into

jasl:codex/ds4-sm120-min-enablefrom

alexbi29:codex/ds4-vram-kv-cleanups

Commits on Jul 27, 2026
- fix(dspark): reduce V2 working VRAM
  alexbi29
  committed
  64ed3f6
- fix(dspark): reduce MHC working buffers
  alexbi29
  committed
  65d769b
- fix(dspark): reuse current draft logits for rejection
  alexbi29
  committed
  56523cc
- fix(dspark): keep rejection logits graph-stable
  alexbi29
  committed
  e64d680
- fix(dspark): reuse dead MHC input buffer
  alexbi29
  committed
  6f386f8
- test(dspark): update config fixture after upstream sync
  alexbi29
  committed
  47efda7
- chore(dspark): remove dead inv-rope-layout and greedy-argmax kernels
  alexbi29
  committed
  f056ff7
- fix(config): keep sequence-parallel MoE enabled at DP=1
  alexbi29
  committed
  f02be76