May 2026: TorchRec bug fix and codebase cleanup delivering more reliable batch output scaling and easier rollout. The central change unconditionalizes batch_outputs scaling by input_length, removing the JustKnobs gating after production validation, which improves accuracy and reduces gating-related risks in production. Accompanied by targeted cleanup removing JustKnobs dependencies and gating code across multiple modules (estimator/types.py, fb/distributed/planner/shard_estimators.py, OSS shard_estimators.py). Implemented in commit a330956aa556a6408b876e6f63b8bd29f734fea5 and reflected in PR 4217 (D99850647).

April 2026 across torchrec and FBGEMM focused on improving resource planning accuracy, benchmarking realism, and batch-processing reliability. Key work includes scaling planner performance estimation to account for sparse features, stabilizing planner context hashes across machines, enabling per-feature pooling factors in TBE benchmarking, and ensuring external batch sizes are correctly forwarded in VBE request building. These changes improve resource allocation, reduce mispredictions in comms and compute, increase reproducibility and caching efficiency, and enhance end-to-end determinism in tests.

November 2025 monthly summary for pytorch/torchrec focused on delivering robust planner usability enhancements and establishing reliable caching and planning foundations for distributed workloads. Key deliverable: TorchRec Planner now supports explicit hashing and equality for DeviceHardware and BasicCommsBandwidths, enabling efficient, deterministic lookups in planner collections and preventing subtle bugs from identity-based hashing. Implemented via explicit __hash__ and __eq__ methods; DeviceHardware hashes on rank, storage, and perf fields with equality verifying all three; BasicCommsBandwidths hashes on inter_host_bw and intra_host_bw with equality verifying both values. The change is tracked in PR 3522, including commit a34ff1e3c7ee1e467da404713e36b2817c9482b8.

April 2025 monthly summary for pytorch/torchrec focused on strengthening distributed systems performance estimation via bandwidth modeling. Implemented GeneralizedCommsBandwidth to let users derive bandwidth estimates based on hardware and network configurations, improving planning and resource allocation in large-scale distributed training/inference.

March 2025 performance summary: Focused on stabilizing distributed training workflows by addressing accuracy in performance estimates and reliability of tooling. No new features delivered; two high-impact bug fixes completed across torchrec and param.