All pull requests

Pull requests

• [benchmark] Make fp8_gemm compatible with TritonBench's scale API

  CLA Signed
  #3659·etaf opened Sep 24, 2026·

  ·

• [metal] Support strided MPP matmul operands

  CLA Signed
  #3658·aditvenk opened Sep 23, 2026Contributor·

  ·

• [benchmark] Enable XPU benchmarks

  CLA Signed
  #3657·etaf opened Sep 23, 2026·

  ·

• [autotuner] Back off the reference config when the backend compiler rejects it

  CLA Signed
  #3656·etaf opened Sep 23, 2026·

  ·
  3

• [megakernel] add gpt oss moe, deespeek v3 nvfp4 moe, deepseek v3 mla, and standalone comparisons

  CLA Signed
  #3655·ethche opened Sep 22, 2026Contributor·4/4·

  ·

• [triton] Enable host-constructed tensor descriptors

  CLA Signed
  #3654·ethche opened Sep 22, 2026Contributor·3/4·

  ·

• [megakernel] prove completion-covered conditional payload dependencies

  CLA Signed
  #3652·ethche opened Sep 22, 2026Contributor·2/4·

  ·

• [benchmark] Support old and new TritonBench rope input APIs

  CLA Signed
  #3643·etaf opened Sep 17, 2026·

  ·
  3

• [triton] Add limited torch.chunk and torch.unbind support

  CLA Signed
  #3612·jjjxia opened Sep 12, 2026·

  ·
  3

• [CuTe] Improve cluster reductions and vector memory operations

  CLA Signed
  #3609·yushangdi opened Sep 11, 2026Contributor·

  ·
  1

• [xpu][feat] Avoid implicit CUDA device fallbacks for autodiff

  CLA Signedmodule: xpu
  #3605·Stonepia opened Sep 10, 2026Contributor·

  ·
  1

• [xpu][tests][1/N] Unskip XPU tests

  CLA Signedmodule: xpu
  #3604·Stonepia opened Sep 10, 2026Contributor·

  ·
  2

• [metal] honor begin and step for scalar root grids

  CLA Signed
  #3596·aditvenk opened Sep 9, 2026Contributor·

  ·
  2

• Remove redundant barriers for inter-loop RAW hazards

  #3582·NaOHCC opened Sep 7, 2026·

  ·
  1

• [noland] Hillclimb plan, prompt, and per-step records for CuTe attention on GB300

  CLA Signed
  #3579·yushangdi opened Sep 7, 2026Contributor·

  ·

• [cutedsl] Retune the sm_103 flash seeds onto the one-CTA and resident lowerings

  CLA Signed
  #3578·yushangdi opened Sep 7, 2026Contributor·

  ·

• [cutedsl] Gate the causal resident softmax on its schedule, and close two KV-width holes

  CLA Signed
  #3577·yushangdi opened Sep 7, 2026Contributor·

  ·

• Warm the GPU to its steady clock before the autotuner's final verification

  CLA Signed
  #3572·yushangdi opened Sep 7, 2026Contributor·

  ·

• Add an fa4-tuned attention baseline to the backend comparison

  CLA Signed
  #3571·yushangdi opened Sep 7, 2026Contributor·

  ·

• Ellipsis indexing for StackTensorType

  CLA Signed
  #3563·parsshar-RH opened Sep 4, 2026Contributor·

  ·
  12

• [metal] Honor begin/step for 1-D root grids

  CLA Signed
  #3562·rk9595 opened Sep 4, 2026·

  ·
  3

• [benchmark] Add --timing-protocol steady to the grouped GEMM campaign, match the attention kernel campaign

  CLA Signed
  #3552·yushangdi opened Sep 3, 2026Contributor·

  ·
  1

• [autotuner] Allow scalar overrides across autotune_multi shapes

  CLA Signed
  #3549·ethche opened Sep 3, 2026Contributor·2/2·

  ·

• [cutedsl] Enable SM103 causal resident softmax and 2-CTA at head_dim=128

  CLA Signed
  #3526·yushangdi opened Sep 1, 2026Contributor·

  ·

• [cute] tcgen05 fragment aux prefetch

  CLA Signed
  #3476·ethche opened Aug 28, 2026Contributor·

  ·