Repository navigation

[Feature][Model] Add Kimi K3 support for vLLM 0.27 - #14454

Merged

linfeng-yuan merged 50 commits into

vllm-project:mainfrom

maoxx241:codex/kimi-k3-model-features-main

Aug 28, 2026

Commits on Aug 27, 2026
- feat(ops): port Kimi K3 AscendC operators
  maoxx241
  committed
  c4f6030
- refactor(ops): move KDA torch adapters into operator directories
  maoxx241
  committed
  0cc801a
- fix(mamba): align hybrid state copy on Ascend
  maoxx241
  committed
  80282f8
- feat(attention): add Kimi K3 KDA execution
  maoxx241
  committed
  814c2fd
- feat(mla): support Kimi K3 attention on Ascend
  maoxx241
  committed
  37cf61c
- perf(ops): add Kimi K3 attention residual fusion
  maoxx241
  committed
  3db03ec
- docs(ops): document K3 attention residual
  maoxx241
  committed
  88252bc
- fix(ops): enforce attention residual stream capacity
  maoxx241
  committed
  7d21d38
- feat(moe): support Kimi K3 SiTU execution
  maoxx241
  committed
  bdec553
- refactor(moe): integrate SiTU quantization into common flow
  maoxx241
  committed
  8fa59cd
- feat(model): register Kimi K3 adapters
  maoxx241
  committed
  dae7c13
- feat(spec-decode): add Kimi K3 DSpark execution
  maoxx241
  committed
  c2c3af5
- feat(kv-cache): group Kimi K3 hybrid caches
  maoxx241
  committed
  b2aa0ec
- refactor(kv-cache): address K3 grouping review
  maoxx241
  committed
  8f2cb31
- fix(pd): harden Kimi K3 disaggregated serving
  maoxx241
  committed
  0e83f71
- refactor(proxy): inline multimodal retry handling
  maoxx241
  committed
  bf40dad
- docs(kimi-k3): add deployment and validation guide
  maoxx241
  committed
  60c6cfc
- fix(ops): use relative path for shared KDA adapter header
  maoxx241
  committed
  3f43cb8
- fix(model): load QuaRot DSpark boundaries in modeling
  maoxx241
  committed
  d533d5d
- refactor(spec_decode): remove QuaRot model hooks
  maoxx241
  committed
  5b5512f
- feat(ops): update fused chunk KDA forward
  
  weinachuan
  authored and
  maoxx241
  committed
  6c039a1
- [Feature] Add mla_prolog_v3 and optional RoPE for MLA prolog (#13355)
  
  authored and
  maoxx241
  committed
  e888390
- [Ops][Feature] Support first-axis non-contiguous kv/kr cache for mla_prolog_v3 (#13422)
  
  yolic66
  authored and
  maoxx241
  committed
  20bd5a7
- [Feature][MLA] Support Kimi K3 no-RoPE MLAPO on A5 (#13507)
  
  Dawn952
  authored and
  maoxx241
  committed
  cc00487
- chore(ops): normalize MLA prolog source line endings
  
  Dawn952
  authored and
  maoxx241
  committed
  1dc3630
- fix(quantization): use forward config for MX scale selection
  maoxx241
  committed
  adf50d7
- fix(quantization): carry MX scale algorithm into forward context
  maoxx241
  committed
  da6e071
- fix(ci): cover Kimi KDA and format MX context helper
  maoxx241
  committed
  9e8b8f7
- refactor(kimi-k3): remove operator-integration detours
  maoxx241
  committed
  8a3fdcd
- fix(quantization): preserve ModelSlim optional metadata
  maoxx241
  committed
  ac63c66
- fix(ci): restore 310P unified-core output path
  maoxx241
  committed
  eb11817
- fix(kimi-k3): align CI coverage with vLLM versions
  maoxx241
  committed
  530884a
- fix(ci): execute SiTU custom op eagerly in CPU tests
  maoxx241
  committed
  fb5c390
- fix(mla): preserve defaults for existing attention models
  maoxx241
  committed
  45804ed
- fix(mla): default existing models to RoPE
  maoxx241
  committed
  d7a93b2
- fix(kv-cache): scope RoPE mode to MLA backends
  maoxx241
  committed
  342ee0c
- refactor(mla): remove no-rope identity metadata
  maoxx241
  committed
  2c83f01
- fix(kv-cache): preserve grouped Mamba block capacity
  maoxx241
  committed
  11bc0aa
- test(kimi-k3): remove dummy execution parity nightly
  maoxx241
  committed
  548c0dc
- fix(moe): compute SiTU without a runtime CustomOp
  maoxx241
  committed
  4c6924c
- fix(pd): preserve decode graphs for stateful handoffs
  maoxx241
  committed
  6da8cfa
- fix(mamba): preserve accepted snapshots and per-rank table capacity
  maoxx241
  committed
  5e327cb
- fix(moe): normalize optional SiTU beta before activation
  maoxx241
  committed
  4a96af9
- test(moe): isolate SiTU reference configuration
  maoxx241
  committed
  17d2491
- fix(mla): preserve absent RoPE metadata during prefill
  maoxx241
  committed
  b106ea1
- perf(kda): restore Ascend fused RMSNorm gate dispatch
  maoxx241
  committed
  5dc5cb3
- fix(ci): register fused norm gate test duration
  maoxx241
  committed
  3b6c845
- test(kimi-k3): cover A3 dummy deployment variants
  maoxx241
  committed
  8c8deed
- test(kimi-k3): streamline A3 smoke coverage
  maoxx241
  committed
  b6dce8c
- test(kimi-k3): prune redundant unit test scaffolding
  maoxx241
  committed
  ba14b85

Commits

Commits on Aug 27, 2026