Repository navigation

[NPU] support kimi k3 on A5 and improve performance - #39589

Merged

sglang-npu-bot merged 40 commits into

sgl-project:mainfrom

zhaozx-cn:a5-k3-0915

Sep 18, 2026

Commits on Sep 15, 2026
- support k3 on a5
  
  zhaozx-cn
  authored and
  root
  committed
  bada6df
- refractor
  
  zhaozx-cn
  authored and
  root
  committed
  59b2075
- Run the shared-expert MLP on the alternate stream while keeping attention-TP all-gather and reduce-scatter on the current stream.
  
  hanwlax
  authored and
  root
  committed
  0eacad5
- run fused_qkvg_proj with Kimi-K3-W4A8 quanted with modelslim
  
  hanwlax
  authored and
  root
  committed
  98ba765
- fix(npu): skip KDA autotune during first request (#2)
  
  McZyWu
  authored and
  root
  committed
  2324c84
- skip qkv_proj (#4)
  
  hanwlax
  authored and
  root
  committed
  75c0cb4
- fix spec acc
  
  zhaozx-cn
  authored and
  root
  committed
  ac83e8b
- k3 use fias v2 (#6)
  
  hanwlax
  authored and
  root
  committed
  955bf4b
- feat: support Kimi-K3 DSPARK PD decode radix cache (#7)
  
  qybnb
  authored and
  root
  committed
  e83efb3
- fix(npu): use DeepEP MXFP8 dispatch for W4A8 MXFP MoE (#11)
  
  hanwlax
  authored and
  root
  committed
  10cdafa
- A5 k3 fp8 fused (#13)
  
  authored and
  root
  committed
  af7152c
- perf(npu): reuse graph update worker
  
  McZyWu
  authored and
  root
  committed
  eec5cb7
- fix(npu): use NPU chain rejection sampling for DSPARK
  
  hanwlax
  authored and
  root
  committed
  21009b2
- feat(npu): add fine-grained Kimi-K3 shared expert overlap
  
  Hexq0210
  authored and
  root
  committed
  f1179a6
- [NPU] Skip DSpark Mamba tracking when no boundary can be crossed
  
  Hexq0210
  authored and
  root
  committed
  4782966
- feat(npu): switch Ascend KDA extend to fused chunk_kda_fwd (#30)
  
  sherdavincl9
  authored and
  root
  committed
  21a4a01
- nz
  
  zhaozx-cn
  authored and
  root
  committed
  7ac8292
- fix(npu): restore logical MLA prefix pages from NZ cache
  
  hanwlax
  authored and
  root
  committed
  4cd9ed4
- [Kimi-K3] Allow shared-expert TP subgroups below attention TP
  
  Hexq0210
  authored and
  root
  committed
  58ae5ec
- revert kda triton kernel modification
  zhaozx-cn
  committed
  91cb6a0
- Revert "feat: support Kimi-K3 DSPARK PD decode radix cache (#7)"
  zhaozx-cn
  committed
  23ba155
- fix
  zhaozx-cn
  committed
  83a12a4
- fix lint
  zhaozx-cn
  committed
  da481f4
- fix shares expert tp
  zhaozx-cn
  committed
  eb64d9c
- fix lint
  zhaozx-cn
  committed
  f9b16ad
- fix dispatch
  zhaozx-cn
  committed
  ce651b1
- fix lint
  zhaozx-cn
  committed
  db368a2
Commits on Sep 16, 2026
- remove unused code
  zhaozx-cn
  committed
  929a91b
- fia v2 for mtp
  zhaozx-cn
  committed
  478f45f
- fia v2
  zhaozx-cn
  committed
  361de6a
- Merge branch 'main' into a5-k3-0915
  zhaozx-cn
  authored
  957f4b7
- fix kimi k3 pd acc on npu
  zhaozx-cn
  committed
  a122d31
- fix pd kimi k3 acc on npu
  zhaozx-cn
  committed
  c1b396b
- fix ci
  zhaozx-cn
  committed
  bd0c6a5
- fix
  zhaozx-cn
  committed
  9742bb3
Commits on Sep 17, 2026
- fix ci
  zhaozx-cn
  committed
  09d6d8a
- Merge branch 'main' into a5-k3-0915
  zhaozx-cn
  authored
  acafebd
- fix
  zhaozx-cn
  committed
  775bcc4
- Merge branch 'main' into a5-k3-0915
  sglang-npu-bot
  authored
  0a236c4
Commits on Sep 18, 2026
- Merge branch 'main' into a5-k3-0915
  zhaozx-cn
  authored
  4ab4643

Commits

Commits on Sep 15, 2026

Commits on Sep 16, 2026

Commits on Sep 17, 2026

Commits on Sep 18, 2026