Amazon opens TorchNeuron to run PyTorch on Trainium chips
Amazon says developers need to change only one setting to move PyTorch code to Trainium, and the tool goes open source on October 20.
Amazon said on its official blog on the 19th (local time) that it is releasing TorchNeuron, a tool that lets PyTorch code run directly on its Trainium AI chips without modification, according to digitaltoday.co.kr.
With TorchNeuron, developers need to change only one setting to move their PyTorch code to the new chip, and they can keep using their existing testing, debugging and scaling tools. The tool supports the Trainium2 and Trainium3 chips.
Amazon will release TorchNeuron as open source at PyTorch Conference North America on October 20.
The company pointed to early users: voice AI company Cartesia is running inference workloads on Trainium with almost no changes to its PyTorch code, and River Neuroscience Institute, which researches treatments for neuropsychiatric disorders such as schizophrenia, is using TorchNeuron in drug development. Hugging Face offers default support for Trainium in its AI development tools.
Background
Until now, developers using PyTorch had to rewrite and verify code over several months to move it to a new chip. Amazon said it is also working with the PyTorch community on code improvements to keep PyTorch from being tied to specific chips.
Quick answers
When is TorchNeuron released?
Amazon said it will release TorchNeuron as open source at PyTorch Conference North America on October 20.
Which chips does TorchNeuron support?
TorchNeuron supports Amazon's Trainium2 and Trainium3 chips.
Do developers have to rewrite their PyTorch code to use TorchNeuron?
No. Amazon says only one setting needs to change, and existing testing, debugging and scaling tools can still be used.