All pull requests

Pull requests

• Handle per-channel QAT observer mutations in AutoParallel

  CLA Signed
  #554·fmassa opened Oct 6, 2026Contributor·

  ·

• Pass AutoParallel runtime estimates to the overlap scheduler in ms

  CLA Signed
  #536·AlbedoWang opened Sep 27, 2026·

  ·

• Add calibrated H100 communication cost profile

  CLA Signed
  #535·aditvenk opened Sep 24, 2026Contributor·

  ·

• Preserve physical shard order for 3-D CP and HSDP

  CLA Signed
  #533·AlbedoWang opened Sep 22, 2026·

  ·

• Generalize ordered FSDP shard layouts to N-D

  CLA Signed
  #529·AlbedoWang opened Sep 1, 2026·

  ·
  2

• Fix bandwidth modeling for sparse two-node communicators

  CLA Signed
  #527·AlbedoWang opened Aug 11, 2026·

  ·

• Estimate flex_local_map alternative costs automatically

  CLA Signed
  #525·AlbedoWang opened Aug 10, 2026·2/2·

  ·

• Add corrected split-dimension seed search

  CLA Signed
  #523·AlbedoWang opened Jul 27, 2026·5/5·

  ·
  7

• Build TRW-S costs lazily without PuLP

  CLA Signed
  #522·AlbedoWang opened Jul 27, 2026·4/5·

  ·
  3

• Add LP relaxation and eager TRW-S solvers

  CLA Signed
  #521·AlbedoWang opened Jul 27, 2026·3/5·

  ·
  5

• Speed up sharding optimizer construction

  CLA Signed
  #520·AlbedoWang opened Jul 27, 2026·2/5·

  ·
  4

• Aggregate: optimizer build, TRW-S, lazy costs, and split-dim seed

  CLA Signed
  #519·AlbedoWang opened Jul 23, 2026·

  ·

• Add flex_local_map: multiple sharding alternatives for the AutoParallel solver

  CLA Signed
  #518·AlbedoWang opened Jul 14, 2026·1/2·

  ·
  1

• Avoid logical/physical placement divergence for Partial targets

  CLA Signed
  #517·yanboliang opened Jul 13, 2026·

  ·

• Support context-parallel attention local_map

  CLA Signed
  #516·AlbedoWang opened Jul 9, 2026·

  ·

• Support 3D+ MoE local_map mesh variants

  CLA Signed
  #514·AlbedoWang opened Jul 7, 2026·1/5·

  ·

• Bypass aliases for consumers that would redistribute to the producer

  CLA Signed
  #479·fmassa opened May 27, 2026Contributor·

  ·

• Fuse chained allgathers on different subgroups into single full-mesh allgather

  CLA Signed
  #472·fmassa opened May 21, 2026Contributor·

  ·

• Fix double recomputation in torch.compile round-trip

  CLA Signed
  #467·fmassa opened May 20, 2026Contributor·

  ·

• Skip _check_forward_args validation during make_fx tracing

  CLA Signed
  #454·IvanKobzarev opened May 11, 2026Contributor·

  ·
  2

• Run TorchTitan GraphTrainer AutoParallel CI

  CLA Signed
  #452·sanketpurandare opened May 8, 2026Contributor·

  ·

• Add TorchTitan DSv3 equivalence coverage

  CLA Signed
  #444·sanketpurandare opened May 4, 2026Contributor·

  ·
  3

• Trim non-pipeline input grads before caching in bwd_cache

  CLA Signed
  #431·aditvenk opened Apr 22, 2026Contributor·

  ·

• ENABLE_SINGLE_DIM_MM_FAMILY: opt-in DTensor strided sharding for mm-family ops

  CLA Signed
  #429·weifengpy opened Apr 21, 2026Contributor·

  ·

• Pre-overlap collective bucketing pass for FSDP/DDP

  CLA Signed
  #412·fmassa opened Apr 3, 2026Contributor·

  ·