AWS will run ordinary PyTorch on Trainium. It is a public beta, and the chip is addressed as “neuron”

AWS said on 21 October that Neuron 2.33 is out, and that TorchNeuron, a native PyTorch backend for its Trainium chips, is in public beta. A model written with the ordinary PyTorch tools can run on EC2 Trn2 and Trn3 instances, either immediately or through torch.compile. The change AWS describes is a device name: code is sent to “neuron” instead of to another chip. The same release updates the vLLM plugin to vLLM 0.29 and includes a plugin for Wan 2.2 video generation, both able to compile through TorchNeuron. A new machine image and container ship with PyTorch 2.14. The code is open source. A day earlier AWS had already posted the beta itself. The 21 October note is the release that puts it in the numbered Neuron package.

There is a second tool in the same note, and it is about failure rather than speed. Neuron Observer, also in public beta, is supposed to notice a sick Trainium node or device in under a second. That is AWS’s claim for the detector. It is not a figure for how often a training job now finishes. The note does not give a tokens-per-second result, a cost per million tokens, or a customer who has moved a named model off another chip.

The same week, the other cloud did this too

Google’s TorchTPU, filed on this desk yesterday, is the same idea on a different processor: make the chip look like a normal PyTorch device so a lab does not rewrite the model to try it. AWS is now offering the Trainium version of that door. Two betas in the same week do not make either chip a substitute. They mean a developer can point the same code at both and see what breaks. The beta label is the part to keep. Public beta is an invitation to try the device name. It is not a statement that the operator coverage, or the cost, matches the chip a team already runs.