IT Home, October 20: On October 19 local time, Amazon announced the launch of the open-source tool TorchNeuron, which enables developers to run PyTorch code natively on Amazon’s custom Trainium chips with just a single configuration change, without requiring code rewrites.
IT Home learns that TorchNeuron will be officially released under an open-source license on October 20 local time at the North American PyTorch Conference in San Jose, California.
PyTorch is an open-source framework led by Meta and adopted by the majority of AI developers. NVIDIA has spent over a decade developing CUDA tools that enable PyTorch code to run efficiently on NVIDIA chips.
According to the report, Amazon engineers ran the same AI training script on both traditional chips and Trainium chips. The only difference was the configuration parameter specifying which chip to use. The results showed that the AI model learned identical content and at the same speed in both environments. This means developers can use their existing tools for testing, debugging, and scaling AI models on Trainium without any modifications.
Amazon says TorchNeuron brings a long-missing key element to the AI industry: genuine choice. When developers can freely switch between different chips without incurring additional costs, chip manufacturers must compete by offering more compelling products to win developer support; in turn, developers benefit from reduced costs and faster iteration.
In addition, Amazon is actively contributing to upstream PyTorch development, collaborating with the PyTorch community to remove code that ties the framework to specific chips. Their goal is to make PyTorch more “hardware-agnostic” for the entire developer community, not just Trainium users.