All pull requests

Pull requests

• DMP: harden _init_parameters against FakeTensorMode leaks

  CLA Signed
  #4895·warrendeng opened Oct 16, 2026·

  ·

• Add forward-only DMP construction (#4894)

  CLA Signedmeta-exported
  #4894·xywang9334 opened Oct 16, 2026·

  ·
  1

• Add batch attribution decorators that keep the original exception

  CLA Signedmeta-exported
  #4893·kaanbaloglu opened Oct 16, 2026Contributor·

  ·
  1

• Preserve sharding metadata in keyed checkpoints

  CLA Signedmeta-exported
  #4891·tangyuhao opened Oct 15, 2026·

  ·
  1

• Add forward-only DMP construction

  CLA Signedmeta-exported
  #4888·amirafzali opened Oct 15, 2026Member·

  ·
  1

• Defer sharded VLE backward to the next forward pass (#4881)

  CLA Signedmeta-exported
  #4881·axeisghost opened Oct 13, 2026Contributor·

  ·
  1

• Enroll the managed-collision and feature-processed wrappers in the checkpoint golden test (#4877)

  CLA Signedmeta-exported
  #4877·kaanbaloglu opened Oct 12, 2026Contributor·

  ·
  1

• Cache recat tensors in KJT all-to-all

  CLA Signedmeta-exported
  #4873·che-sh opened Oct 12, 2026Contributor·

  ·
  1

• Add table-aware UVM prefetch resolution (#4872)

  CLA Signedmeta-exported
  #4872·bharathvemula opened Oct 11, 2026·

  ·
  1

• Expose per-shard UVM cache weight bytes (#4871)

  CLA Signedmeta-exported
  #4871·bharathvemula opened Oct 11, 2026·

  ·
  1

• Allow callers to select the backend for the 2D sparse device mesh

  CLA Signedmeta-exported
  #4868·darshanvs-m opened Oct 9, 2026Contributor·

  ·
  1

• Suppress type errors for Pyre upgrade (landing group 13)

  CLA Signedmeta-exported
  #4866·maggiemoss opened Oct 9, 2026Contributor·

  ·
  1

• Replace per-tensor batch record_stream with one memcpy fence

  CLA Signedmeta-exported
  #4857·aporialiao opened Oct 7, 2026Contributor·

  ·
  1

• Plumb module_fqn through the ManagedCollision sharders (#4851)

  CLA Signedmeta-exported
  #4851·mengtian-x opened Oct 6, 2026Contributor·

  ·
  1

• Stage safe-list CUDA readbacks through pinned memory (#4846)

  CLA Signedmeta-exported
  #4846·dsaluja opened Oct 5, 2026·

  ·
  1

• Avoid recompiling capped Triton TBE for variable batch sizes

  CLA Signedmeta-exported
  #4843·xywang9334 opened Oct 5, 2026·

  ·
  1

• Recombine the per-step loss under micro-batching (#4841)

  CLA Signedmeta-exported
  #4841·amitaga opened Oct 5, 2026Contributor·

  ·
  1

• Count a threshold tie as a true positive in recall

  CLA Signed
  #4839·shaneraphel opened Oct 4, 2026·

  ·
  1

  2

• Speed up MCH remapping on CUDA with a fused Triton kernel

  CLA Signed
  #4832·CharlesXu-HQ opened Oct 3, 2026·

  ·

• Fix ManagedCollision shared-feature eviction with RowWiseAdagrad

  CLA Signed
  #4831·CharlesXu-HQ opened Oct 3, 2026·

  ·
  1

  2

• Emit SDD pipelining failure reasons into pipeline_module_summary (#4822)

  CLA Signedmeta-exported
  #4822·mengtian-x opened Oct 1, 2026Contributor·

  ·
  1

• Register process-group cleanup at init in metric and optimizer tests (#4804)

  CLA Signedmeta-exported
  #4804·kaanbaloglu opened Sep 29, 2026Contributor·

  ·
  1

• Cover batch index select validation

  CLA Signedmeta-exported
  #4802·jeffkbkim opened Sep 28, 2026Contributor·

  ·
  1

• Migrate current_stream calls to torch.accelerator.current_stream

  CLA Signed
  #4799·yilingyuan opened Sep 28, 2026·

  ·
  2

• Log ZORM QPS event-time preservation

  CLA Signedmeta-exported
  #4782·jeffkbkim opened Sep 23, 2026Contributor·

  ·
  1