All pull requests

Pull requests

• [Ascend] Improve PTO codegen for DeepGEMM kernels

  #3330·erhsh opened Sep 30, 2026Contributor·

  ·
  3

• [Ascend] Add TileLang agents for operator generation and tuning

  #3329·yuqiha opened Sep 30, 2026·

  ·
  6

• feat(ascend): support L0C fp32->f16/bf16 dual_copy on-path cast

  ascend
  #3328·CeleNewYear opened Sep 30, 2026Collaborator·

  ·
  6

• fix(pto): isolate compiler imports from Triton

  #3327·erhsh opened Sep 30, 2026Contributor·

  ·
  2

• refactor(ascend): compile aibin in single stage without ld.lld

  ascend
  #3326·CeleNewYear opened Sep 30, 2026Collaborator·

  ·
  2

• [Ascend] Add T.PersistentKernel and AutoPersistent pass

  ascend
  #3325·CeleNewYear opened Sep 30, 2026Collaborator·

  ·
  2

• feat(ascend): Add torch_npu task queue support

  ascend
  #3324·CeleNewYear opened Sep 30, 2026Collaborator·

  ·
  4

• [Ascend][Gemm] Support gemm_l1 in the Python sugar via a NormalizeGemm pass

  ascend
  #3323·CeleNewYear opened Sep 30, 2026Collaborator·

  ·
  2

• feat(ascend): add causal sparse attention example and tests

  ascend
  #3322·yangsichan opened Sep 30, 2026Collaborator·

  ·
  2

• [BugFix][Ascend] correlate host operators with device kernels in profiling

  ascend
  #3321·yangsichan opened Sep 30, 2026Collaborator·

  ·
  4

• feat(ascend): Add AutoSimtVF pass and register it in Ascend pipeline

  ascend
  #3320·yangsichan opened Sep 30, 2026Collaborator·

  ·
  8

• [BugFix][Language] Preserve wide predicate truth values

  #3319·ZenAlexa opened Sep 30, 2026Contributor·

  ·
  4

• [BugFix][JIT] Preserve caller-owned output indices

  #3318·ZenAlexa opened Sep 30, 2026Contributor·

  ·
  3

• [Ascend] Route PTO execution through PTO codegen

  ascend
  #3316·erhsh opened Sep 30, 2026Contributor·

  ·
  2

• [Test] Improve testing infrastructure

  #3313·penguin-wwy opened Sep 30, 2026Contributor·

  ·
  2

• [BugFix][CUDA] Fix int64 atomic max/min crash and int64/bf16/fp16 crash in AtomicStore

  #3307·KareemMusleh opened Sep 29, 2026Contributor·

  ·
  3

  

• [Example] Add DeepSelect Top-K

  #3304·sepcnt opened Sep 29, 2026Contributor·

  ·
  5

• [BugFix][CUDA] Support popcount for signed and narrow integers

  #3300·KareemMusleh opened Sep 28, 2026Contributor·

  ·
  1

  3

  

• [CUDA] Fix bulk TMA alignment and cache hints

  #3296·sepcnt opened Sep 28, 2026Contributor·

  ·
  1

  4

• [BugFix][Transform] using z3 to infer all the possible threads that pass the constraint

  #3291·y1lan opened Sep 27, 2026Contributor·

  ·
  9

• [CUDA] Add an SM120 W4A16 NVFP4 GEMM example for decode-sized batches

  #3290·ghostrider0470 opened Sep 27, 2026·

  ·
  2

• [Transform] Ensure import C source ends with a newline

  #3289·jjppp opened Sep 27, 2026Contributor·

  ·
  1

  9

• [BugFix][CUDA] Keep warp-specialized stages until moved non-TMA copies are consumed

  #3288·ghostrider0470 opened Sep 27, 2026·

  ·
  1

  2

• [CUDA] SM120 block-scaled register-A GEMM: unroll K atoms and load scales once per atom

  #3286·ghostrider0470 opened Sep 27, 2026·

  ·
  11

• feat(cuda): support NVRTC device compilation with TVM-FFI

  #3285·ZenAlexa opened Sep 26, 2026Contributor·

  ·
  9