Amazon's Trainium chips have long been a cheaper place to train models on paper and an awkward one in practice, because PyTorch code had to be adapted to run on them. TorchNeuron, a native PyTorch backend for Trainium, aims to remove that friction. It entered public beta on 20 October 2026, installs with pip install torch-neuronx, and Hugging Face says Transformers, Diffusers, Accelerate, PEFT and TRL now run on it directly.
Until now, the Hugging Face route to Trainium and Inferentia, AWS's training and inference chips, went through Optimum Neuron, built on a PyTorch/XLA package. That meant thinking like a compiler: lazy tensors instead of eager execution, static shapes only, and a recompilation whenever the graph changed. TorchNeuron exposes the chip as a normal neuron device with eager execution, torch.compile, and PyTorch's own distributed tools such as FSDP2 and DTensor. Hugging Face now recommends using its libraries directly rather than through Optimum Neuron.
The coverage figures are the useful part. With TorchNeuron, 89% of the 100 most-used text-generation architectures run out of the box, against 46% with Optimum Neuron v0.4.5. Fine-tuning uses a standard TRL script with a few Neuron-specific settings, and the resulting checkpoint is an ordinary Transformers file that can be pushed to the Hub and loaded on any hardware. In the worked example, 32 processes on one instance split each model across four cores with tensor parallelism and shard the rest with FSDP. In Diffusers, image models including FLUX.1-dev, FLUX.2 and Qwen-Image-2.1 also support tensor parallelism, and custom kernels can be written in AWS's Neuron Kernel Interface and shared on the Hub.
Paired with TorchTPU for Google's chips the same week, this is a notable shift for anyone pricing training outside Nvidia. Both backends make the alternative accelerators look like any other PyTorch device to the libraries we review, PyTorch and Transformers among them, which lowers the cost of trying them.
It is a beta, and the rough edges show. Shapes still need to be kept static to avoid recompiles, so training scripts disable automatic dynamic shapes and pad batches; asynchronous weight loading in Transformers is not yet supported; and 89% coverage means some popular architectures still do not run. The post gives no speed or cost comparisons with GPUs, which is the number most teams will need. Test a representative workload before planning a move.