All pull requests

Pull requests

• Fix multiple issues in Op Builder and refactor implementation

  skip-gpu-test
  #9085·hwchen2017 opened Nov 14, 2026Contributor·

  ·
  1

• [ZeRO-3] Optimize GatheredTensor.contiguous group interval lookup with bisect (#9062)

  skip-gpu-test
  #9084·davrens101 opened Nov 14, 2026·

  ·
  3

• Gate merge queue GPU CI by PR label

  skip-gpu-test
  #9083·tohtana opened Nov 13, 2026Collaborator·

  ·
  10

• Load universal checkpoints across ZeRO stages

  skip-gpu-test
  #9080·jinyouzhi opened Nov 13, 2026Contributor·

  ·

• Run the whole static rollout decode step inside the CUDA graph

  #9079·jinyouzhi opened Nov 13, 2026Contributor·

  ·

• Fix DeepCompile NCCL header discovery for PyTorch wheels

  skip-gpu-test
  #9078·hwchen2017 opened Nov 13, 2026Contributor·

  ·
  1

• Reuse the static rollout decode graph across calls

  #9077·jinyouzhi opened Nov 13, 2026Contributor·

  ·

• [zero] Run ZeRO-3's rank-agreement checks and barriers over the partition group

  skip-gpu-test
  #9076·LiRunGuo opened Nov 13, 2026Contributor·

  ·
  7

• [rollout] In-memory relayout between sharded layouts (two-stage relay, pipelined execution) and ZeRO-3 training-state move

  #9074·LiRunGuo opened Nov 13, 2026Contributor·

  ·
  4

• write the latest checkpoint tag only after every rank has saved

  skip-gpu-test
  #9071·dundysm opened Nov 12, 2026Contributor·

  ·
  1

  5

• [AffineIR] Plan flat shards that keep the parameter's own strides

  skip-gpu-test
  #9070·0z5a opened Nov 12, 2026Contributor·

  ·
  4

• ZeRO-3: take the AutoTP shard from universal checkpoints converted without TP

  skip-gpu-test
  #9069·Achyuthan-S opened Nov 12, 2026Contributor·

  ·
  3

• [Triton] Add Ascend Triton variants and a per-kernel probe policy

  #9065·VenusTZZ opened Nov 12, 2026Contributor·5 tasks done·

  ·

• Fix replicated K/V all-reduce to run once per accumulation window

  skip-gpu-test
  #9064·VenusTZZ opened Nov 12, 2026Contributor·4 tasks done·

  ·

• Fix flops profiler MAC count for conv with padding='same' and even kernel

  skip-gpu-test
  #9063·davrens101 opened Nov 12, 2026·

  ·
  1

• Broadcast the mask in Triton inference masked softmax

  #9061·cervoliu opened Nov 11, 2026·

  ·
  1

  8

• fix(module-inject): preserve HF precision boundaries in segKI kernels

  #9057·nathon-lee opened Nov 11, 2026Contributor·

  ·

• [XPU] Open the Triton gate on Intel XPU

  skip-gpu-test
  #9055·jinyouzhi opened Nov 11, 2026Contributor·

  ·

• Fix checkpoint and broadcast tests to use the accelerator backend

  #9054·VenusTZZ opened Nov 11, 2026Contributor·4 tasks done·

  ·
  6

• Describe a parameter held in several layouts as one affine map

  skip-gpu-test
  #9052·Achyuthan-S opened Nov 11, 2026Contributor·

  ·
  7

• Fix rank split when a full optimizer state is missing

  skip-gpu-test
  #9049·Shr3yash opened Nov 10, 2026·

  ·
  4

• Fix OneCycle overwriting beta2, starting at the wrong momentum, and get_last_lr before step

  skip-gpu-test
  #9048·davrens101 opened Nov 10, 2026·

  ·
  3

  5

• Avoid quadratic copies in curriculum cluster selection

  skip-gpu-test
  #9046·Hang-get opened Nov 10, 2026·

  ·
  1

• [rollout] Affine weight-sync building block: ZeRO storage into a tensor-parallel inference copy, shard to shard

  #9045·LiRunGuo opened Nov 10, 2026Contributor·

  ·
  5

• Fix oversized zero_to_fp32 shards for ZeRO-1/2 checkpoints

  skip-gpu-test
  #9040·Hang-get opened Nov 9, 2026·

  ·
  1