Hugging Face has released a public lookup tool showing whether popular Transformers models run on AWS Trainium2 through TorchNeuron, its native PyTorch backend for Amazon’s AI accelerator.
Each tested model includes the compiler and library versions, number of Trainium cores, wall-clock time including compilation, peak device memory and the script used for the run. Developers can copy cached compiled graphs built with the same compiler version, avoiding the first-run compilation step when reproducing a result.
The current test loads each model on the smallest number of cores that can hold its weights, up to 16 cores across four Trainium2 devices, then generates 32 tokens. Single-core output is compared with a CPU run of the same checkpoint. For sharded models, a small judge model checks whether the output is fluent English. Results and errors are also published as a queryable dataset.
A passing result has a deliberately limited meaning. It does not measure throughput, longer generations or application quality, and current coverage is restricted to text-generation and image-to-text Transformers models. Hugging Face plans to add diffusion models, sentence encoders and more tasks. The tool therefore answers an initial deployment question—whether a model starts and produces plausible short output—rather than whether Trainium is the best production target.