March 2026 monthly summary for pytorch/pytorch: Focused on stabilizing the TorchTitan XPU path. Delivered a bug fix that generalizes the Split_Group API calls via the accelerator API for the TorchComms backend, enabling TP>1 on XPU and preventing execution hangs. Merged PR 178236 with commit e41371ce3a045f4306e0816921d38060e666b697, expanding XPU compatibility beyond CUDA and improving reliability for large-scale TorchTitan workloads. Impact: reduced downtime, improved scalability, and stronger business value for customers deploying TorchTitan on XPU.

December 2025 focused on strengthening deterministic behavior and test reliability for DTensor on XPU accelerator devices within PyTorch. Delivered a key feature that extends RNG state management to XPU devices in DTensor tests, enabling per-rank RNG state collection and setting to ensure deterministic results across ranks during op dispatch. This work completes the RNG-state handling extension from CPU/CUDA to accelerator devices and mitigates unit-test failures related to RNG state management on XPU devices.

July 2025 ROCm/pytorch monthly summary focusing on delivering a Unified Device Management API for DistributedDataParallel (DDP) and integrating essential XCCL changes to support scalable multi-GPU training. This work reduces setup complexity, improves training usability, and strengthens multi-node accelerator support.