All pull requests

Pull requests

• [CPU] Support serial cumsum and cummax lowering

  bugcpuenhancementUX
  #3897·penguin-wwy opened Nov 9, 2026Collaborator·

  ·
  4

• [Ascend][Example] Support dynamic sequence lengths in MHA

  ascendenhancementpython
  #3896·dangyw18 opened Nov 9, 2026Contributor·

  ·
  2

• [BugFix] Add loop-carried sync for uniform while loops

  bug
  #3895·edragain2nd opened Nov 9, 2026Contributor·

  ·
  1

  8

• [Ascend][PTO] Fix VMI descriptors and remove redundant logical coverage checks

  arithascendbug
  #3894·KurrinQu opened Nov 9, 2026Contributor·

  ·
  13

• [Ascend] Dequantize an int32 L0C accumulator in the L0C->GM copy

  ascendbug
  #3893·Rachmanino opened Nov 9, 2026Collaborator·2 tasks·

  ·
  2

• [Ascend][Graph] Add torch.compile support for AscendC and PTO

  #3892·erhsh opened Nov 9, 2026Contributor·

  ·
  2

• [Ascend] Use aclrtDeviceL2CacheFlush for benchmark L2 flushing

  ascendpython
  #3891·LJC00118 opened Nov 9, 2026Collaborator·

  ·
  7

• [BugFix][Metal] Emit float constants as hex floats so they round-trip

  bugmetal
  #3889·Chr0mium24 opened Nov 9, 2026·

  ·
  1

  2

• [BugFix][Language] Preserve ceildiv semantics for signed divisors

  bugpython
  #3888·lexiaox opened Nov 9, 2026·

  ·
  2

• [Release] Bump version to 0.1.16

  #3887·LeiWang1999 opened Nov 9, 2026Member·

  ·
  2

• [BugFix][ROCm] Handle nonconstant int8 broadcasts

  bug
  #3886·lexiaox opened Nov 9, 2026·

  ·
  1

  2

• [Metal] Use half2 arithmetic for packed FP16 vectors

  bugmetal
  #3884·LiRunGuo opened Nov 8, 2026·

  ·
  1

  14

• [BugFix][WS] Decline warp specialization on loop-carried post-pipeline updates

  bugwarp specialization
  #3883·VikingDeng opened Nov 8, 2026·

  ·
  1

  14

• [BugFix][JIT] Raise a cython kernel's dynamic shared memory limit on every device it launches on

  bugpython
  #3882·lucifer1004 opened Nov 8, 2026Collaborator·

  ·
  1

  5

• [BugFix][NVRTC] Launch kernels with a single argument

  bugpython
  #3880·lucifer1004 opened Nov 8, 2026Collaborator·

  ·
  1

  9

• [Enhancement][NVRTC] Allow compile_cuda without -lineinfo

  enhancementpython
  #3879·stealth-rklopper opened Nov 8, 2026Contributor·

  ·
  1

  2

• [BugFix][CuTeDSL] Compare unsigned atomics unsigned

  bugenhancementpython
  #3876·lucifer1004 opened Nov 8, 2026Collaborator·

  ·
  1

  8

• [BugFix][CuTeDSL] Keep values read from unsigned buffers unsigned

  bugpython
  #3875·lucifer1004 opened Nov 8, 2026Collaborator·

  ·
  1

  5

• [BugFix][Layout] Run the atomics and read-modify-writes of a replicated loop once

  arithbugUX
  #3873·lucifer1004 opened Nov 7, 2026Collaborator·

  ·
  1

  44

• [BugFix][JIT] Restore the caller's current device after a tvm_ffi call

  bugpython
  #3872·lucifer1004 opened Nov 7, 2026Collaborator·

  ·
  1

  3

• [BugFix][WarpSpecialize] Split the partitions outside an if around the pipeline

  bug
  #3869·lucifer1004 opened Nov 7, 2026Collaborator·

  ·
  1

  16

• [BugFix][JIT] Launch on the inputs' device when it is not the current device

  bugpython
  #3867·lucifer1004 opened Nov 7, 2026Collaborator·

  ·
  1

  26

• [Fix][Eager] Pass type_params to _eval_type for Python 3.15

  bugpython
  #3865·YidaWeng opened Nov 7, 2026·

  ·
  1

  2

• [Fix][TVM] Preserve reverse FlattenBuffer indices via non-positive IterMap divisors

  arithbugdependencies
  #3864·YidaWeng opened Nov 7, 2026·

  ·
  1

  2

• [Ascend] Reject FP4 VMI lane counts that pack to an invalid size

  ascendbugpythonUX
  #3863·cn-sys opened Nov 7, 2026·

  ·
  1

  4