Repository navigation

[None][feat] Kimi K3 NVFP4 on Rubin (SM107): Modeling V2 target and the runtime it needs - #21996

Draft

xguannv wants to merge 14 commits into

NVIDIA:mainfrom

xguannv:user/xguan/k3-rubin-modeling-v2

Draft

[None][feat] Kimi K3 NVFP4 on Rubin (SM107): Modeling V2 target and the runtime it needs#21996
xguannv wants to merge 14 commits into
NVIDIA:mainfrom
xguannv:user/xguan/k3-rubin-modeling-v2

Commits on Nov 10, 2026
- [None][perf] Parallelize long KV cache block-offset copies
  
  dc3671
  authored and
  xguannv
  committed
  76637bd
- [None][fix] Reserve device memory for allocations outside PyTorch
  
  dc3671
  authored and
  xguannv
  committed
  a1b8638
- [None][perf] Accept caller-quantized FP8 Q / K / V in MLA context attention
  
  dc3671
  authored and
  xguannv
  committed
  6b20017
- [None][feat] Share hybrid KV cache pools with speculative drafters
  
  dc3671
  authored and
  xguannv
  committed
  1921268
- [None][fix] Check a one-sided disaggregated config for its own side of shared draft KV
  xguannv
  committed
  f0cbfb8
- [None][fix] Align DWDP worker discovery in disaggregated tests
  xguannv
  committed
  1fc5fa9
- [None][test] Perf-sanity log scan: allow a refused clock-offset forward at startup
  xguannv
  committed
  fa495ad
- [None][feat] Kimi K3 SM107 target: features, routing, core layers and collectives
  xguannv
  committed
  b069eb4
- [None][feat] Kimi K3 SM107 target: FP4 / FP8 MLA attention with Helix
  xguannv
  committed
  2b42cfd
- [None][feat] Kimi K3 SM107 target: KDA linear attention
  xguannv
  committed
  95afa1d
- [None][feat] Kimi K3 SM107 target: MoE and FP8 / MXFP8 GEMMs
  xguannv
  committed
  29a0f55
- [None][feat] Kimi K3 SM107 target: DSpark speculative decoding
  xguannv
  committed
  50fddfd
- [None][feat] Kimi K3 SM107 target: the model
  xguannv
  committed
  1f764b6
- [None][test] Kimi K3 SM107 target: accuracy test, deployments and CI lists
  xguannv
  committed
  ac90a78

Commits

Commits on Nov 10, 2026