In the aftermath of Amazon's TorchNeuron, Google makes it possible to switch a PyTorch application to its TPU chips by a simple configuration change, without rewriting the code.
Google has released TorchTPU, its open source version of PyTorch, the framework used by a large majority of AI applications. Since the latter was designed at a time when Nvidia's GPUs were already the norm, many of its functions assume this hardware. TorchTPU allows these applications to run on Google's Tensor Processing Units, and therefore on Google Cloud, by only modifying the configuration.
The announcement comes a day after TorchNeuron, Amazon's equivalent for its Trainium chips. Google is highlighting a collaboration with Meta, Hugging Face, Inferact (vLLM) and RadixArk (SGLang) to ensure compatibility with the most popular tools. The group is also launching a suite of agents to support the migration to TPUs and fix errors, and is updating LiteRT, its runtime engine on phones and embedded devices, to better support PyTorch models.
Read the full article on the original source
Go to the source →