Repository navigation

fix(kimi-k3): reduce attention residual peak memory - #1

Draft

qijiajin wants to merge 6 commits into

maoxx241:codex/kimi-k3-model-adapters-mainfrom

qijiajin:codex/fix-kimi3-attn-res-memory-only

Draft

fix(kimi-k3): reduce attention residual peak memory#1
qijiajin wants to merge 6 commits into
maoxx241:codex/kimi-k3-model-adapters-mainfrom
qijiajin:codex/fix-kimi3-attn-res-memory-only

Commits on Aug 21, 2026
- feat(model): register Kimi K3 model adapters
  maoxx241
  committed
  ae0693c
- fix(model): load K3 mixed and DSpark weights
  maoxx241
  committed
  18ef992
- chore(model): annotate stacked K3 imports
  maoxx241
  committed
  7bd46ca
- test(kimi-k3): annotate captured DSpark weight names
  maoxx241
  committed
  0d5cd2e
- test(kimi-k3): add storage-light nightly parity guard
  maoxx241
  committed
  7420dc3
- fix(kimi-k3): reduce attention residual peak memory
  qijiajin
  committed
  ef8425f