InfoQ 中文T2·Infra
78 pts
Google and Meta have open-sourced TorchTPU, a native PyTorch backend for TPUs that replaces the intrusive PyTorch/XLA approach. Developers can migrate existing workloads by swapping the cuda device for tpu and the nccl backend for tpu_dist, gaining eager execution, compiled execution and distributed computing support. The project is still early-stage, with a roadmap covering dynamic shapes, operator coverage, reinforcement learning and quantization.