During 2025-10, the team delivered major enhancements to memory reporting in pytorch/FBGEMM, focusing on richer memory visibility and planning alignment. Key accomplishments include detailed tensor memory logging and a new memory calculation helper, plus targeted refactors that improve maintainability and observability. The work enables a granular breakdown of memory usage (static vs ephemeral) and per-component HBM/UVM categorization, supported by new Scuba metrics. This instrumentation enables validation of planner memory estimates against actual runtime usage, helping reduce memory-related risk and informing optimization opportunities in production. Note: No major bugs fixed were reported for this repo this month based on the available data.

September 2025 monthly summary for pytorch/torchrec. Focus: Shard Management Size Calculation Utilities refactor to modularize size calculation logic (tensors, optimizers, cache state) and cache checks, enabling more accurate sizing and more efficient shard management in distributed systems. Impact: improved performance, scalability, and maintainability; easier testing and future enhancements.