All pull requests

Pull requests

• Replace fcntl.flock with filelock for cross-platform compatibility

  #216·woct0rdho opened Oct 9, 2026·

  ·

• feat: add per-token weights to linear_cross_entropy

  #215·m4xig1 opened Oct 1, 2026·

  ·

• [Fix] Fall back when the async compile executor is unavailable

  #214·adenzhou1350 opened Sep 30, 2026·

  ·
  2

• Include 32-thread RMSNorm configs in forward autotuning

  #213·aryanputta opened Sep 29, 2026·

  ·
  1

• Full-width outputs in acc_pair epilogues; restore_value fixes in the autotuner

  #212·HanGuo97 opened Sep 26, 2026Contributor·

  ·

• Fix unbatched blocked scale unpacking

  #210·adenzhou1350 opened Sep 21, 2026·

  ·
  1

• Fix retained backward in chunked linear cross entropy

  #209·adenzhou1350 opened Sep 20, 2026·

  ·

• [Epi] Size reduce-sink partials from the device, not the config tag

  #208·GarlGuo opened Sep 17, 2026Member·

  ·

• [SM100] Wait for PDL dependencies before grouped tile decoding

  #207·mcwitt opened Sep 16, 2026·

  ·

• Fix autotuning under selective activation checkpointing

  #204·S1ro1 opened Sep 12, 2026·

  ·
  1

• fix illegal memory access error

  #201·ptxu78 opened Sep 8, 2026·

  ·

• [SM100][GEMM]: Support fallback cluster shapes by default and add larger preferred cluster shapes to default autotune list

  #200·thakkarV opened Sep 3, 2026Collaborator·

  ·

• [Rounding] Widen the SR stream offset from 16 to 28 bits; rename tid to offset

  #199·thakkarV opened Sep 3, 2026Collaborator·

  ·

• [Blockscaled] Add square 2D scale-block MXFP8 formats for quant-free transpose in bwd

  #198·thakkarV opened Sep 3, 2026Collaborator·

  ·

• [Act] Add optional clamp limit to SwiGLU

  #195·gupta-abhay opened Aug 30, 2026·

  ·
  3

• Skip SMEM for small N in softmax forward (#62)

  #194·XFDG opened Aug 28, 2026·

  ·
  1

  2

• fix: compat with cuda-bindings 12.9 and CUTLASS DSL signature change (#187, #188)

  #192·XFDG opened Aug 26, 2026·

  ·
  1

  4

• feat(rmsnorm): add a FlyDSL forward backend for ROCm gfx950

  #190·jhinpan opened Aug 14, 2026·10 tasks done·

  ·
  4

• [Gemm, SM90] Add optional decoupled seqused_k (real per-batch K-length) for varlen_k GEMM

  #189·jaimec00 opened Aug 12, 2026·

  ·

• Support CUDA 12.9 stream capture info

  #186·JiaxuanBai opened Aug 12, 2026·

  ·
  1

• [top-k] leader-only bitonic top-k warp merge

  #185·ajuneja23 opened Aug 8, 2026·

  ·

• [Gemm] Add Gemm + ReduceScatter/AllReduce

  #173·Nicyzk opened Jul 16, 2026Contributor·

  ·
  11

• Migrate deprecated cutlass.utils spellings to canonical namespaces

  #171·yiwangchunyu opened Jul 13, 2026Contributor·

  ·
  1

• [Gemm,Sm100] MXFP8 grouped GEMM (offs dispatcher)

  #161·santoshmo opened Jun 24, 2026Contributor·

  ·

• [Autotuner] Fix BrokenPipeError root cause

  #149·akshattripathi-art opened Jun 10, 2026·

  ·
  1