For each tested model, the Space reports whether it ran unmodified, the number of Trainium cores used, wall-clock time including compilation, and peak device memory. It also exposes the tested Neuron compiler and Transformers versions, installation packages, and the run.py script used for the test. Users can select compiler and Transformers versions, and can copy matching compiled graphs from a public NEFF cache bucket.
The published results come from the public aws-neuron/torch-neuronx-model-support dataset, stored as Parquet files for pairs of compiler and library versions. Tests currently cover Transformers text-generation and image-text-to-text models, with a limit of four Trainium2 devices, or 16 cores. A PASS only records the described 32-token check; it does not establish throughput, behavior on longer generations, or compatibility beyond the tested stack and model versions.