AWS introduces TorchNeuron for native PyTorch support on AWS Trainium
AWS has unveiled TorchNeuron, a native PyTorch backend for AWS Trainium, now available in public beta. This tool aids developers in efficiently running PyTorch models on Trainium hardware.
Amazon Web Services (AWS) has announced the public beta release of TorchNeuron, an open-source native PyTorch backend specifically designed for AWS Trainium, a suite of AI chips aimed at enhancing performance and cost efficiency for deep learning AI training and inference tasks. TorchNeuron empowers AI researchers and developers to utilize PyTorch natively on Trainium, facilitating the experimentation and construction of models.
With the introduction of TorchNeuron, developers can easily adapt their PyTorch code for Trainium by using the command .to('neuron'), while retaining full compatibility with PyTorch’s eager execution, torch.compile, and standard APIs. Furthermore, TorchNeuron integrates seamlessly with the Neuron Kernel Interface (NKI), offering support for custom operators and performance optimization tailored to Trainium hardware.
The TorchNeuron backend is accessible as an open-source resource via the TorchNeuron GitHub repository, applicable for both Amazon EC2 Trainium2 and Trainium3 instances. Developers and researchers interested in further exploring TorchNeuron and AWS Trainium accelerators are encouraged to visit the AWS Neuron page for additional information.