[None][fix] Fix CppMambaHybridCacheManager functional and perf issues - #14003

Merged

VALLIS-NERIA merged 16 commits into

NVIDIA:mainfrom

VALLIS-NERIA:user/xiweny/mamba-recurrent-cuda-graph-snapshots

May 22, 2026

Commits on May 11, 2026

[None][fix] CppMambaHybridCacheManager: handle ranks with zero local mamba layers (By Agent)
VALLIS-NERIA
committed
ac30ded
[None][fix] KVCacheManager: reserve recurrent-state slots for CUDA graph padding and draft-len sentinels (By Agent)
VALLIS-NERIA
committed
992ec1b
[None][fix] KVCacheManager: fix max_draft_tokens typo and add slot reservation tests (By Agent)
VALLIS-NERIA
committed
3310de0

Commits on May 12, 2026

Commits on May 13, 2026

fix missing is_draft
VALLIS-NERIA
committed
3f0e36c
fix replay check on sm120
VALLIS-NERIA
committed
69899c5
Merge branch 'main' into user/xiweny/mamba-recurrent-cuda-graph-snapshots
VALLIS-NERIA
committed
e00c230

Commits on May 15, 2026

[None][fix] Address recurrent-state pool sizing and layer-first transfer bugs (By Agent)
VALLIS-NERIA
committed
1813212

Commits on May 16, 2026

fix schedule issue, improve runtime perf, allow change manager
VALLIS-NERIA
committed
717edeb

Commits on May 17, 2026

[None][fix] Cap CppMamba warmup tokens by recurrent-state pool capacity (By Agent)
VALLIS-NERIA
committed
7a6e90f

Commits on May 18, 2026

[None][fix] Scale recurrent-state pool by pp_size for hybrid Mamba models (By Agent)
VALLIS-NERIA
committed
c9933a8

Commits on May 21, 2026