Repository navigation

[TRTLLM-16288][feat] Integrate sparse KV offload for DeepSeek-V4 Flash and GLM5.2 on KVCM2 - #21350

Open

cascade812 wants to merge 12 commits into

NVIDIA:mainfrom

cascade812:guiju/sparse-kv-offload

Open

[TRTLLM-16288][feat] Integrate sparse KV offload for DeepSeek-V4 Flash and GLM5.2 on KVCM2#21350
cascade812 wants to merge 12 commits into
NVIDIA:mainfrom
cascade812:guiju/sparse-kv-offload

Commits on Oct 30, 2026
Commits on Oct 31, 2026
- [TRTLLM-16282][test] Give the Helix KV-transfer test fakes a VRAM pool view
  cascade812
  committed
  2b91474
Commits on Nov 1, 2026
- [TRTLLM-16288][test] Restore real admission in disagg receive ordering tests
  cascade812
  committed
  1f09b8f
- [None][fix] Restore auxiliary KV pool selection for DSA FMHA
  
  cascade812
  authored and
  Guiju Zhang
  committed
  b33e1f4
Commits on Nov 2, 2026
- [TRTLLM-16282][perf] Enable THP by default for host KV pools
  Guiju Zhang
  committed
  fd1c0ba
- [TRTLLM-16282][fix] Size sparse-offload pools at decode block boundaries
  Guiju Zhang
  committed
  b10d99c
- Merge origin/main into guiju/sparse-kv-offload
  Guiju Zhang
  committed
  2b3aa21