All pull requests

Pull requests

• Guard the _chunk_scan_fwd kernel launch with the input's device

  #1073·toddwbucy opened Oct 31, 2026·

  ·
  1

• Mamba-3 SISO: take the angle-gradient carry across chunks from the st…

  #1071·triphopp opened Oct 27, 2026·

  ·

• Mamba-3 SISO: keep the rotary angle rates in fp32 (follow-up to #1059)

  #1070·Johnsonms opened Oct 27, 2026·

  ·

• Bump tilelang to 0.1.15, apache-tvm-ffi to 0.1.12, quack-kernels to 0.6.5

  #1069·KyleMylonakisProtopia opened Oct 26, 2026Contributor·

  ·
  1

• Keep mamba_ssm importable when Mamba-3 kernel deps fail to load

  #1068·konstde00 opened Oct 25, 2026·

  ·

• feat(mamba3-siso): Ascend NPU A2 (Triton-Ascend/CANN) support for SISO fwd/bwd/decode

  #1066·Kanraaaaa opened Oct 22, 2026·2 of 4 tasks·

  ·

• [Ascend NPU] 新增 Mamba-1 selective_scan fwd/bwd AscendC 算子（npu/ 包）

  #1064·yyoean opened Oct 20, 2026·

  ·

• Ascend NPU (910B) support for Mamba-2 Triton kernels

  #1063·Ymr419 opened Oct 20, 2026·

  ·

• [Bugfix] Fix NaN output of _chunk_scan_fwd_kernel on sm90 (port of vllm#61718)

  #1062·maxli0182 opened Oct 17, 2026·

  ·
  1

• Mamba-3 MIMO: fix backward for seqlen not divisible by chunk_size

  #1060·Johnsonms opened Oct 14, 2026·

  ·

• Mamba-3: keep cumulative rotary angles in fp32 in angle_dt_fwd (halves fwd/bwd error vs reference)

  #1059·Johnsonms opened Oct 13, 2026·

  ·
  2

• tests: fix tensor star-unpack in test_mamba3_siso_fwd_blackwell_regression

  #1058·Johnsonms opened Oct 13, 2026·

  ·

• Mamba-3 SISO bwd: drop the Q @ States^T MMA so num_warps >= 4 compiles on SM100 (2.4-2.9x faster dqkv on B200)

  #1057·Johnsonms opened Oct 13, 2026·

  ·

• 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