Pull requests · meta-pytorch/autoparallel

Add agent-driven AutoParallel optimization workflow

#559fmassa opened

Oct 13, 2026

Contributor

Handle per-channel QAT observer mutations in AutoParallel

#554fmassa opened

Oct 6, 2026

Contributor

Pass AutoParallel runtime estimates to the overlap scheduler in ms

#536AlbedoWang opened

Sep 27, 2026

Add calibrated H100 communication cost profile

#535aditvenk opened

Sep 24, 2026

Contributor

Preserve physical shard order for 3-D CP and HSDP

#533AlbedoWang opened

Sep 22, 2026

Generalize ordered FSDP shard layouts to N-D

#529AlbedoWang opened

Sep 1, 2026

2

Fix bandwidth modeling for sparse two-node communicators

#527AlbedoWang opened

Aug 11, 2026

Estimate flex_local_map alternative costs automatically

#525AlbedoWang opened

Aug 10, 2026

Add corrected split-dimension seed search

#523AlbedoWang opened

Jul 27, 2026

7

Build TRW-S costs lazily without PuLP

#522AlbedoWang opened

Jul 27, 2026

3

Add LP relaxation and eager TRW-S solvers

#521AlbedoWang opened

Jul 27, 2026

5

Speed up sharding optimizer construction

#520AlbedoWang opened

Jul 27, 2026

4

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

#519AlbedoWang opened

Jul 23, 2026

Add flex_local_map: multiple sharding alternatives for the AutoParallel solver

#518AlbedoWang opened

Jul 14, 2026

1

Avoid logical/physical placement divergence for Partial targets

#517yanboliang opened

Jul 13, 2026

Support context-parallel attention local_map

#516AlbedoWang opened

Jul 9, 2026

Support 3D+ MoE local_map mesh variants

#514AlbedoWang opened

Jul 7, 2026

Bypass aliases for consumers that would redistribute to the producer

#479fmassa opened

May 27, 2026

Contributor

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

#472fmassa opened

May 21, 2026

Contributor

Fix double recomputation in torch.compile round-trip

#467fmassa opened

May 20, 2026

Contributor

Skip _check_forward_args validation during make_fx tracing

#454IvanKobzarev opened

May 11, 2026

Contributor

2

Run TorchTitan GraphTrainer AutoParallel CI

#452sanketpurandare opened

May 8, 2026

Contributor

Add TorchTitan DSv3 equivalence coverage

#444sanketpurandare opened

May 4, 2026

Contributor

3

Trim non-pipeline input grads before caching in bwd_cache

#431aditvenk opened

Apr 22, 2026

Contributor

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

#429weifengpy opened

Apr 21, 2026

Contributor

Navigation Menu

Repository navigation

All pull requests

Add agent-driven AutoParallel optimization workflow

Handle per-channel QAT observer mutations in AutoParallel

Pass AutoParallel runtime estimates to the overlap scheduler in ms

Add calibrated H100 communication cost profile

Preserve physical shard order for 3-D CP and HSDP

Generalize ordered FSDP shard layouts to N-D

Fix bandwidth modeling for sparse two-node communicators

Estimate flex_local_map alternative costs automatically

Add corrected split-dimension seed search

Build TRW-S costs lazily without PuLP

Add LP relaxation and eager TRW-S solvers

Speed up sharding optimizer construction

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

Add flex_local_map: multiple sharding alternatives for the AutoParallel solver

Avoid logical/physical placement divergence for Partial targets

Support context-parallel attention local_map

Support 3D+ MoE local_map mesh variants

Bypass aliases for consumers that would redistribute to the producer

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

Fix double recomputation in torch.compile round-trip

Skip _check_forward_args validation during make_fx tracing

Run TorchTitan GraphTrainer AutoParallel CI

Add TorchTitan DSv3 equivalence coverage

Trim non-pipeline input grads before caching in bwd_cache

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