Hugging Face adds lookup tool for AWS Trainium model compatibility

Hugging Face introduced a tool on October 23 that shows which popular models have passed tests on Amazon’s Trainium chips. The tool provides compatibility results and scripts for reproducing the tests.

Users can search model names and filter results by software version. Entries show core usage, execution time including compilation, and peak device memory. Initial testing covers text generation and models that generate text from image and text inputs.

The release follows Hugging Face’s integration of TorchNeuron, which lets its libraries use Trainium through a native PyTorch backend. The lookup tool adds model-specific evidence, but passing has a narrow meaning: tests generate just 32 tokens. Single-core results are compared against CPU output, while distributed runs receive an English-fluency check from another model.

Testing currently covers models that fit within four Trainium2 devices, or 16 cores. A passing result does not establish throughput or reliability for longer responses. Hugging Face plans to broaden coverage to additional tasks and libraries, including Diffusers and sentence transformers, and makes the underlying results publicly available.