Amazon has launched TorchNeuron in public beta, an open-source backend under the Apache 2.0 license that enables PyTorch code to run on Trainium2 and Trainium3 chips. Developers can migrate with a one-line change, replacing .to(‘cuda’) with .to(‘neuron’), without rewriting or revalidating their codebase. More than 30 companies and organizations participated in the testing phases, including Cartesia, Deepgram, and the Lieber Institute for Brain Development, which is using the tool for pharmaceutical research. Amazon says its custom-chip division has reached an annual revenue run rate of $25 billion and projects that figure could reach $50 billion, but the current version skips approximately 22,000 PyTorch tests and does not yet support advanced features such as dynamic input shapes.
Source: Read the original article
