All pull requests

Pull requests

• Avoid SSD kernel respecialization when sequence lengths change

  #1056·ziqing-pplx opened Oct 6, 2026·

  ·
  1

• Skip missing gradients in sequence parallel reduction

  #1055·Chessing234 opened Oct 5, 2026Contributor·

  ·

• Reset the batch offset when reusing inference parameters

  #1054·Chessing234 opened Oct 5, 2026Contributor·

  ·

• Allow parallel embeddings without a process group

  #1053·Chessing234 opened Oct 4, 2026Contributor·

  ·

• Stop decoding when the prompt reaches max_length

  #1052·Chessing234 opened Oct 4, 2026Contributor·

  ·

• Keep at least one token during top-p filtering

  #1050·Chessing234 opened Oct 1, 2026Contributor·

  ·

• Match grouped RMSNorm in the split scan reference

  #1049·Chessing234 opened Oct 1, 2026Contributor·

  ·
  1

• [Bug] Use per-rank ngroups for the Mamba2 RMSNormGated group size

  #1048·gss10282023 opened Sep 26, 2026·

  ·
  1

  1

• Fix OOB writes in Mamba3 step() when headdim % tile_D != 0

  #1046·imao81 opened Sep 18, 2026·

  ·
  1

• docs: add Simplified Chinese translation (README.zh.md)

  #1045·JasonYeYuhe opened Sep 13, 2026·

  ·
  1

• Deterministic mode: pin tile sizes per kernel from TRITON_AUTOTUNE_BLOCK* variables

  #1043·NolenLiang opened Sep 11, 2026·

  ·
  1

• Migrate selective scan to the PyTorch stable ABI

  #1042·stmcgovern opened Sep 10, 2026·

  ·
  1

  31

• fix: treat non-positive sample temperature as greedy

  #1041·Chessing234 opened Sep 4, 2026Contributor·1 task done·

  ·

• fix: honor list/tensor eos_token_id in decode stop

  #1040·Chessing234 opened Sep 4, 2026Contributor·1 task done·

  ·

• fix: accept (B, 1, D) in Mamba3.step like Mamba2

  #1039·Chessing234 opened Sep 4, 2026Contributor·1 of 2 tasks·

  ·
  1

• [1 of 5] Sequences longer than ~61K tokens silently corrupt Mamba2's chunk state

  #1038·chipsncheese-ie opened Sep 3, 2026·

  ·
  4

• fix(triton): pass False not None for delta_softplus after manual softplus

  #1029·Chessing234 opened Aug 31, 2026Contributor·2 tasks·

  ·

• fix(triton): handle None dt_bias/D in selective_state_update

  #1028·Chessing234 opened Aug 31, 2026Contributor·2 tasks·

  ·

• fix(mha): store pre-conv qkv in conv_state during prefill

  #1027·Chessing234 opened Aug 31, 2026Contributor·2 tasks·

  ·

• fix: only create CUDA timing events when enable_timing is set

  #1026·Chessing234 opened Aug 31, 2026Contributor·2 tasks·

  ·

• fix(mamba3): support GH200 256x256 SISO backward

  #1025·klokedm opened Aug 29, 2026·

  ·

• Fix layer_norm backward to reuse the forward dropout mask and scale dx1 by rowscale

  #1023·cyboTiger opened Aug 25, 2026·

  ·

• mamba3: access ctx.saved_tensors exactly once in backward

  #1021·jgcb00 opened Aug 18, 2026·

  ·
  2

• Add an opt-in Helion kernel backend for the Mamba-2 SSD path

  #1012·shcho1118 opened Aug 3, 2026·

  ·

• fix: apply min-p filtering on softmax probabilities

  #1010·Chessing234 opened Jul 31, 2026Contributor·2 tasks·

  ·