Repository navigation

[None][feat] Add native KV cache retention to KVCM V2 - #20086

Merged

yizhang-nv merged 23 commits into

NVIDIA:mainfrom

yizhang-nv:codex/kvcm-v2-native-retention

Oct 20, 2026

Merged

[None][feat] Add native KV cache retention to KVCM V2#20086
yizhang-nv merged 23 commits into
NVIDIA:mainfrom
yizhang-nv:codex/kvcm-v2-native-retention

Commits on Oct 19, 2026
- feat: implement native V2 KV cache retention
  yizhang-nv
  committed
  e3b4392
- fix: preserve retention across partial reuse and lifecycle boundaries
  yizhang-nv
  committed
  eccc995
- fix: retain V1 policy across prefetch and cache lifecycle transitions
  yizhang-nv
  committed
  5240193
- test: cover retention lifecycle and rank-local policy boundaries
  yizhang-nv
  committed
  173f457
- fix: keep prefetch a no-op for GPU-only caches
  yizhang-nv
  committed
  f77f698
- test: cover GPU-only retention prefetch
  yizhang-nv
  committed
  b081744
- test: explicitly configure GPU-only prefetch coverage
  yizhang-nv
  committed
  f925150
- fix: align retention type declarations and test fixtures
  yizhang-nv
  committed
  6f911b7
- test: provide native manager in disaggregated ordering fixture
  yizhang-nv
  committed
  ea9c3f8
- fix: convert prefetch cache levels at the native binding
  yizhang-nv
  committed
  edfe5d5
- test: assert the manager used by default retention controls
  yizhang-nv
  committed
  00635f2
- Rename the cache offload threshold for cold tiers
  yizhang-nv
  committed
  a396fea
- fix: address KV cache retention review feedback
  yizhang-nv
  committed
  abc7d4a
- test: enable V2 in admission fixtures and tidy docs
  yizhang-nv
  committed
  2d21198
- fix: validate layer groups in V2 priority queries
  yizhang-nv
  committed
  b119b99
- fix: remove obsolete V2 retention warning references
  yizhang-nv
  committed
  b207266
- fix: preserve grouped connectors during retention validation
  yizhang-nv
  committed
  6f6a51a
- refactor: consolidate executor transfer mode import
  yizhang-nv
  committed
  30f6a53
- perf: skip connector priority queries without new blocks
  yizhang-nv
  committed
  1949ee4
- fix: select parameterized connector cancellation cases on B200
  yizhang-nv
  committed
  9bfb115
- fix: supply retention config on KV cache stats test requests
  yizhang-nv
  committed
  5603573
- fix: include SHA256 headers for KV cache typed index tests
  yizhang-nv
  committed
  68b9c2f
- fix: regenerate serve schemas for cold tier offload priority
  yizhang-nv
  committed
  9315100

Commits

Commits on Oct 19, 2026