Repository navigation

[None][feat] Rubin merge back: MLA, DSpark, DWDP, Helix - #21067

Open

mingyangHao wants to merge 21 commits into

NVIDIA:mainfrom

mingyangHao:port/rubin-advance-main-20261023

Open

[None][feat] Rubin merge back: MLA, DSpark, DWDP, Helix#21067
mingyangHao wants to merge 21 commits into
NVIDIA:mainfrom
mingyangHao:port/rubin-advance-main-20261023

Commits on Oct 28, 2026
- [None][perf] Fuse the Kimi K3 routed RMSNorm into the MNNVL all-reduce
  
  dc3671
  authored and
  mingyangHao
  committed
  52b715e
- [None][fix] Join x_ready_event in FC12/B12x CuteDSL MoE backends
  
  sunnyqgg
  authored and
  mingyangHao
  committed
  c4051d7
- [None][fix] DFlash/DSpark: size slot-indexed rejection-sampling buffers by the seq-slot pool and share them with CUDA-graph metadata
  
  xguannv
  authored and
  mingyangHao
  committed
  87359f8
- [None][feat] Transfer the standalone DSpark draft KV from CTX to GEN on rubin-advance (ports !10774 + !10802)
  
  dc3671
  authored and
  mingyangHao
  committed
  272f7bf
- [None][perf] Split PARALLEL autotuning across Helix CP ranks
  
  reasonsolo
  authored and
  mingyangHao
  committed
  a5329b6
- [None][fix] Helix all-to-all: fence shared memory before the async-proxy bulk copies
  
  xguannv
  authored and
  mingyangHao
  committed
  b91aec3
- [None][perf] FP4 MLA packed decode: shorter tile critical path and producer pipelining
  mingyangHao
  committed
  80ecd72
- [None][perf] FP4 MLA packed decode: simulate the waves to pick the task length
  
  lancelly
  authored and
  mingyangHao
  committed
  358b1cc
- [None][fix] Adapt imported runtime and kernel changes to current main
  mingyangHao
  committed
  f189f03
- [None][fix] Merge main and align Helix V weight binding
  mingyangHao
  committed
  3df8d42
Commits on Oct 29, 2026
- Merge main and repair Rubin mergeback integration regressions
  mingyangHao
  committed
  b386e74
- Merge latest main through 5be3334257
  mingyangHao
  committed
  e7ea5f8
- Fix DSpark unsupported-architecture fallback fixture
  mingyangHao
  committed
  6f765ba
- Declare absent draft layers in disaggregation test manager
  mingyangHao
  committed
  646c479
- Fix unified-cache warmup and restore DSpark SM107 routing
  mingyangHao
  committed
  f4587da
- Remove legacy unified DSpark draft-cache transfer
  mingyangHao
  committed
  36ae25e
- Merge current main and resolve B12x input admission conflict
  mingyangHao
  committed
  b60aa1f
Commits on Oct 30, 2026
- Merge main and combine DSpark import additions
  mingyangHao
  committed
  a1002ef
- Fix DSpark test fixtures after mergeback
  mingyangHao
  committed
  1d08c8c
- Preserve FP8 cache precision in windowed DFlash attention
  mingyangHao
  committed
  0671031
- Fix FC12 input-ready event joins
  
  longlee0622
  authored and
  mingyangHao
  committed
  72e6cc7

Commits

Commits on Oct 28, 2026

Commits on Oct 29, 2026

Commits on Oct 30, 2026