Hugging Face adds a Trainium model lookup
The lookup pairs model test results with runnable scripts.
Hugging Face’s announcement describes a lookup for developers checking whether models run unchanged on Trainium. Search by name and select compiler and Transformers versions. A selected model provides its tested software stack, extra packages and execution script. On Trainium2 with that stack, install the listed packages, then run the script; it starts distributed processes when the model is sharded. Reported run time includes compilation.
The maintained dataset defines PASS narrowly: loading and generation completed, and the output check passed or could not run. Tests generate 32 tokens; success says nothing about throughput or longer outputs. A missing result for a version pair means untested, not failed. EXCLUDED also means untested. Failure labels distinguish loading, generation and output checking. Single-core output is compared with the same checkpoint on a CPU, allowing near ties; with more cores, or an incomplete CPU run, a model judges English fluency.
The backend introduction explains the mechanism: TorchNeuron is a PyTorch plugin exposing Trainium as a native device. Hugging Face recommends using its libraries directly instead of Optimum Neuron. Its Transformers inference example recommends a static generation cache to prevent frequent recompilation.
