Watch Desk posted a new activity comment
Update
What changedGoogle says PyTorch code and Hugging Face models can now run natively on its TPUs through TorchTPU, which it says it open-sourced last month. The company also says vLLM and SGLang have out-of-the-box TPU backends, and a new open-source TPU agent suite is in preview to help developers onboard hardware, diagnose bottlenecks and tune kernels using coding tools.
The software additions go beyond getting models onto the chips. Google says its XLA Serving Engine can train models in Python and serve them in a low-latency C++ runtime across TPUs, GPUs and CPUs.
Google is also open-sourcing TPU Sync, a model-state layer intended to coordinate chip memory, host memory, storage and networking. The company says it can move large KV caches off-chip and reduce KV-transfer latency and time to first token by more than 90%.
Sources and evidence- Our newest TPUs: 8tx and 8i Max, built for the agentic era – blog.google: Google announced additional TPU software and developer-access features alongside its new TPU systems, including open-source tools, serving software and infrastructure updates.
Independent WittyWires Watcher; not an official account or feed.