Release notes from torchrec
Articles about Release notes from torchrec
V2026.10.12.00: Shared memory-stashing config logging (#4869)
Summary: Pull Request resolved: #4869 Four pipelines drive MemoryStashingManager, and each built its own config event. TrainPipelineSparseDistEmbStash logged ems_config with its own table scan. TrainPipelineSparseDistOptStash logged nothing. The APS and MVAI pipelines (next diff) were adding two more copies, each with different keys, its own tagging rule, and no protection against a logging error. This adds one utility next to MemoryStashingManager and moves the two TorchRec pipelines onto it.
V2026.10.05.00: Support DP lookup deferral in Fused SDD (#4837)
Summary: Pull Request resolved: #4837 Context With embedding_lookup_before_optimizer sets to True, FusedSDD starts batch i+1 embedding lookup after batch i backward but before optimizer.step(), allowing embedding output_dist to overlap dense optimizer work. This is correct for most sparse tables updated by the backward-fused optimizers. DP sharded embedding tables are updated by the dense optimizer, however, so reading them at the same early point would use stale weights.
Release v1.9.1 · meta-pytorch/torchrec
TorchRec 1.9.1 Release Notes Announcement TorchRec 1.9.1 is a patch release based on TorchRec 1.9.0. It adds the experimental TPU SparseCore embedding package that was not included in 1.9.0. Existi
V2026.09.28.00: Configure lazy DDP bucket allocation (#4798)
Summary: Pull Request resolved: #4798 TL;DR Expose opt-in lazy DDP bucket allocation through ShardingConfig and enable it for the heavy-dense benchmark. Reshape the dense stack into twenty wide bottleneck pairs so the benchmark creates twenty 1 GiB activations while retaining approximately 9.87 GiB of dense parameters. Make model copying opt-in so benchmark initialization does not retain a redundant model copy by default. Motivation D122067773 added model-initialization profiling and DDP bucket-cap configuration.
V2026.09.21.00: Move UVM TBE variants into a dedicated module (#4762)
Summary: Pull Request resolved: #4762 Context The main Triton TBE module mixed the general implementation with UVM-specific subclasses, making the file harder to navigate and making the caching follow-up difficult to review. This diff is a behavior-preserving module split. It moves the existing direct-UVM and capped-UVM classes into a dedicated UVM module and introduces a cache-agnostic forward-launch hook used by the caching follow-up.
V2026.09.14.00: Add structured activation layouts (#4717)
Summary: Pull Request resolved: #4717 Context Maglev deliberately transports inter-layer state as Activations, a flat tuple containing only tensors. That wire format is efficient for pipeline send and receive operations, but production models reason about named boundaries containing tensors, containers, and static configuration. Without a shared codec, every model must manually assign tuple positions and duplicate the same reconstruction logic on both sides of each boundary.