All pull requests

Pull requests

• [BugFix][JIT] Avoid stale lazy kernels when mutable defaults change

  #3417·luicarus opened Oct 3, 2026·

  ·
  2

• [BugFix] Complete bfloat16 scalar input binding and Cython marshalling

  #3416·dajiaohuang opened Oct 3, 2026Contributor·

  ·
  5

• [BugFix][Autotuner] Restore compatible input tensor reuse

  #3415·cyanseek opened Oct 2, 2026·

  ·
  1

  2

• [BugFix] Preserve parameter identity in NVRTC and CuTeDSL wrappers

  #3414·adenzhou1350 opened Oct 2, 2026·

  ·
  13

• [BugFix][CUDA] Respect runtime alignment in 256-bit global copies

  #3405·adenzhou1350 opened Oct 2, 2026·

  ·
  1

  2

• [BugFix][CuTeDSL] Elect one grid-barrier leader per block

  #3400·adenzhou1350 opened Oct 2, 2026·

  ·
  2

• [Feature][Autotune] Add opt-in structured trial reporting

  #3393·WenzheWang opened Oct 2, 2026Contributor·

  ·
  2

• [BugFix] Route shared-to-fragment scans through the fragment path

  #3392·Eliaaazzz opened Oct 2, 2026·

  ·
  1

  2

• [BugFix][CUDA] Preserve enclosing predicates on nested LDG

  #3390·adenzhou1350 opened Oct 2, 2026·

  ·
  1

  2

• fix(autodd): reject non-positive --jobs instead of hanging

  #3385·li-lizhe opened Oct 2, 2026·

  ·
  1

  2

• [BugFix][Language] Avoid eager JIT generated-name collisions

  #3384·adenzhou1350 opened Oct 2, 2026·

  ·
  1

  2

• [BugFix] Preserve scalar bit counts in T.view

  #3382·adenzhou1350 opened Oct 2, 2026·

  ·
  1

  3

• fix(language): coerce coalesced_width to IntImm in T.Parallel

  #3374·Adarsh-mk7 opened Oct 2, 2026·

  ·
  1

  3

• Warn when Hopper WGMMA layout constraints force fallback

  #3373·dajiaohuang opened Oct 2, 2026Contributor·

  ·
  1

  6

• [BugFix][Language] Accept integer per-axis shift flags

  #3372·Metis-dot opened Oct 2, 2026·

  ·
  2

• Report unsupported 16-bit integer atomic adds clearly

  #3370·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  1

  4

• Reject unsupported atomic_addx4 destination dtypes

  #3369·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  1

  2

• Validate mbarrier arrive-count upper bound

  #3368·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  2

• Reject sparse MMA K tiles with unhandled tails

  #3367·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  1

  2

• Preserve non-restrict buffers with handle suffixes

  #3366·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  1

  2

• Fix constant row indices for TMA gather and scatter

  #3365·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  1

  2

• fix: validate T.copy region bounds

  #3364·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  2

• Fix warp specialization for while-wrapped pipelines

  #3363·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  1

  5

• Document alignment requirements for shared PTX stores

  #3362·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  1

  2

• Honor current CUDA stream in NVRTC and CuTeDSL adapters

  #3359·dajiaohuang opened Oct 1, 2026Contributor·

  ·
  8