All pull requests

Pull requests

• [CUDA] Synchronize TMEM users before automatic deallocation

  bug
  #3975·huahuaquq opened Nov 12, 2026·

  ·
  2

• [Ascend] Serialize the device source map with the module

  #3974·cn-sys opened Nov 12, 2026·

  ·
  1

  2

• [Ascend] Map hf32 mode 1 to NEAREST_ZERO on the CCE backend

  #3973·cn-sys opened Nov 12, 2026·

  ·
  1

  3

• [Ascend] Align the GEMM input dtype table with both codegen backends

  #3972·cn-sys opened Nov 12, 2026·

  ·
  1

  3

• [Ascend] Accept the .dyn spelling in the scope equality checks

  ascendbug
  #3971·cn-sys opened Nov 12, 2026·

  ·
  1

  4

• [Ascend] Reject batched reductions instead of emitting a scalar call

  ascendbug
  #3970·cn-sys opened Nov 12, 2026·

  ·
  1

  2

• [Ascend] Reject pad_value for non-UB copy destinations

  ascendbugpythonUX
  #3969·cn-sys opened Nov 12, 2026·

  ·
  1

  3

• [Ascend] Advance the ND-to-NZ scatter mask count so the loop terminates

  ascendbug
  #3968·cn-sys opened Nov 12, 2026·

  ·
  1

  4

• [Ascend] Reject the bitwise reductions instead of emitting an undefined type

  ascendbugUX
  #3967·cn-sys opened Nov 12, 2026·

  ·
  1

  4

• [Ascend] Reject PK4_B32 stores on 16-bit vectors

  ascendbugUX
  #3966·cn-sys opened Nov 12, 2026·

  ·
  1

  5

• [Ascend] Skip the signal deferral when msprof runs off the main thread

  ascendbugpython
  #3965·cn-sys opened Nov 12, 2026·

  ·
  1

  5

• [Ascend] Fix the annotation docstring examples for the Ascend dialect

  ascenddocumentation
  #3964·cn-sys opened Nov 12, 2026·

  ·
  1

  3

• [Ascend] Use a runtime tail mask for ND-to-NZ partial passes

  ascendbug
  #3963·cn-sys opened Nov 12, 2026·

  ·
  1

  3

• [Language] Add T.reducer_scale for sum reducers rescaled each step

  enhancementpython
  #3961·ghostrider0470 opened Nov 11, 2026Contributor·

  ·
  5

• [ROCm] Keep fp16 fragments held across a loop in registers

  enhancement
  #3960·ghostrider0470 opened Nov 11, 2026Contributor·

  ·
  5

• [ROCm] Skip the reduce workspace when all reducing threads share a wavefront

  enhancement
  #3959·ghostrider0470 opened Nov 11, 2026Contributor·

  ·
  4

• [ROCm] Add occupancy_report to hip_resource_info

  enhancementpythontoolsUX
  #3958·ghostrider0470 opened Nov 11, 2026Contributor·

  ·
  5

• [ROCm] Emit waves-per-EU launch bounds from T.annotate_min_blocks_per_sm

  bugdocumentationpython
  #3957·ghostrider0470 opened Nov 11, 2026Contributor·

  ·
  4

• [ROCm] Explain T.gemm limits on tiles with few rows per wavefront

  UX
  #3956·ghostrider0470 opened Nov 11, 2026Contributor·

  ·
  6

• [BugFix][ROCm] Broadcast 8-bit integers with 8 and 16 lanes

  bug
  #3955·ghostrider0470 opened Nov 11, 2026Contributor·

  ·
  2

• [BugFix][ROCm] Pass compile_flags and tl.device_compile_flags to hipcc

  bugpython
  #3954·ghostrider0470 opened Nov 11, 2026Contributor·

  ·
  2

• [BugFix][Transform] Convert packed shared-memory byte offsets in elements

  arithascendbug
  #3953·YidaWeng opened Nov 11, 2026·

  ·
  22

• [Example] Fit BHSD attention autotuning in the sm_12x shared-memory budget

  bugpython
  #3952·losewayy opened Nov 11, 2026·

  ·
  5

• [Ascend] Reject non-positive T.Kernel blocks and negative Pipelined num_stages

  ascendbugpythonUX
  #3926·JINITAIMEI121 opened Nov 11, 2026·

  ·
  2

  5

• [CI](ascend) Add nightly Ascend wheel pipeline with NPU smoke test and rolling release

  ascenddependenciesenhancementgithub_actionspython
  #3924·xuedinge233 opened Nov 11, 2026·

  ·
  5