[Sparse & HICache]: Enables hierarchical sparse KV cache management and scheduling for DeepSeek V32.#14619

Open

hzh0425 wants to merge 41 commits intomainfrom

Commits on Nov 20, 2025

init framework
hzh0425
committed

Commits on Nov 25, 2025

[nsa sparse]: Optimize Code style
hzh0425
committed

Commits on Nov 29, 2025

tmp optimized
hzh0425
committed

Commits on Dec 3, 2025

[nsa sparse]: Using unified kernel to process all spares requests to reduce cpu launch overhead

LingYeAI
authored and
hzh0425
committed

Commits on Dec 4, 2025

[Sparse]: Support cuda graph
hzh0425
committed

Commits on Dec 5, 2025

Commits on Dec 8, 2025

[Sparse]: Fix triton accuracy bug
hzh0425
committed
[io kernel]:optimize transfer kernel temporarily

LingYeAI
authored and
hzh0425
committed
Optimize Code Style
hzh0425
committed

Commits on Dec 10, 2025

Commits on Dec 11, 2025

[Sparse]: Refactor algorithm layer
hzh0425
committed

Commits on Dec 13, 2025

Commits on Dec 16, 2025

tmp optimization for io_kernel block_quota

huangtingwei9988
authored and
hzh0425
committed

Commits on Dec 17, 2025

Commits on Dec 22, 2025

Refactor mem alloc, only seperate alloc index_k on decode mode
hzh0425
committed

Commits on Dec 23, 2025

Remove NSAReqToTokenPool
hzh0425
committed

Commits on Jan 3, 2026

[1/N][Sparse With Hicache]: Add Sparse Interface (#14741)
hzh0425
committed
Refactoring sparse diff triton kernel

huangtingwei9988
authored and
hzh0425
committed
support lru evict for sparse kv cache

huangtingwei9988
authored and
hzh0425
committed