Hugging Face Launches AWS Trainium Model Compatibility Checker
TL;DR: Hugging Face has released a new Space and dataset that shows which popular models from the Hugging Face Hub can run on AWS Trainium chips. The torch-neuronx model support Space provides test results, run scripts, and performance metrics for the most popular models, currently focusing on text-generation and image-text-to-text models.
What's New
Hugging Face has launched a new tool to help developers determine whether their AI models can run on AWS Trainium chips. The torch-neuronx model support Space tests popular models from the Hugging Face Hub and provides detailed compatibility information for AWS Trainium hardware.
The Space builds on TorchNeuron, which was introduced in a previous post as the native PyTorch backend for AWS Trainium (the torch-neuronx package). While any Transformers model on the Hub may theoretically run on Trainium, this new tool shows which ones actually do in practice.
The tool offers three views: a Selected view featuring hand-picked top models, plus Trending and Downloads views that follow the Hub's own rankings. Each result shows whether a model runs unmodified on Trainium, the number of Trainium cores required, wall time including compilation, and peak device memory usage.
Currently, the Space tests only Transformers models, with a focus on text-generation and image-text-to-text models. Hugging Face plans to expand coverage to diffusers, sentence transformers, and other transformers tasks in the future.
For each compatible model, the Space provides ready-to-use run scripts following a standard recipe: loading the model with from_pretrained, applying sharding when necessary, then generating with a static cache and torch.compile(backend="neuron"). The scripts include proper setup for distributed computing when models are sharded across multiple cores.
All test results are stored in the public aws-neuron/torch-neuronx-model-support dataset, with one Parquet file per pair of neuronx-cc compiler and library versions. Each row contains the model's result, error messages for failures, full stack versions, and the exact script used. Developers can query this dataset directly using pandas for programmatic access.
Pricing and Availability
The torch-neuronx model support Space and dataset are publicly available as of October 23, 2026. The tool itself is free to use for checking model compatibility. Pricing for actually running models on AWS Trainium depends on AWS instance costs, which are not specified in the release.
How to Try It
To check if a model is compatible, visit the torch-neuronx model support Space on Hugging Face. Search for your model by name in the Trending or Downloads views, or browse the curated Selected view.
Once you find a compatible model, the Space provides everything needed to run it on a Trainium2 (trn2) instance. You can use the AWS Neuron Deep Learning Container and follow the provided installation and run commands. No torchrun is required—the scripts automatically launch distributed ranks when models are sharded.
For faster subsequent runs, you can skip compilation by copying pre-built graphs from the public NEFF cache bucket using the exact command provided by the Space, which includes the correct compiler version.
Users can also select different versions of the Neuron compiler (neuronx-cc) and Transformers to view compatibility across different stack configurations.
Who It's For
This tool is designed for ML engineers and researchers who want to deploy models on AWS Trainium hardware. It's particularly valuable for teams evaluating whether to use Trainium for their inference workloads, as it removes the guesswork about model compatibility.
The tool is most useful for those working with popular open-source models from the Hugging Face Hub who need to understand hardware requirements before committing to a deployment strategy.
Alternatives
Developers looking for model deployment options beyond AWS Trainium can explore other hardware backends. Hugging Face recently announced TorchTPU support for Google Cloud TPUs in an October 20, 2026 blog post. Traditional GPU deployments remain another option, though they may have different cost and performance characteristics compared to specialized AI chips like Trainium.
FAQ
What does a PASS result mean in the compatibility checker?
A PASS indicates that the model successfully loaded on a Trainium instance and generated 32 tokens that were verified for correctness. However, it doesn't provide information about throughput or performance with longer generations.
Can I query model compatibility programmatically?
Yes, all results are stored in the public aws-neuron/torch-neuronx-model-support dataset. You can query it directly using pandas with the command: pd.read_parquet('hf://datasets/aws-neuron/torch-neuronx-model-support/data')
What types of models are currently tested?
At the time of writing, only Transformers models are tested, specifically text-generation and image-text-to-text models. Models are tested on up to 4 Trainium2 devices (16 cores maximum). Hugging Face plans to expand to diffusers, sentence transformers, and other tasks.
How are models distributed across Trainium cores?
Most models use tensor parallelism and are sharded over the smallest number of Trainium cores that can hold the model's weights. Mixture-of-experts models without a tensor-parallel plan use expert parallelism only.