Hugging Face Launches Model Compatibility Checker for AWS Trainium

Hugging Face has published a dedicated Space that answers a practical question for teams evaluating AWS Trainium: which of the thousands of models on the Hub actually run on the hardware without modification? The torch-neuronx model support Space tests popular Transformers checkpoints against the TorchNeuron backend and surfaces the results in a searchable table.

What the Space shows

Each entry reports whether the model passes, the number of Trainium cores used for tensor or expert parallelism, wall-clock time including compilation, and peak device memory. Users can filter by trending or download rank, select specific Neuron compiler and Transformers versions, and see why a model was skipped — usually deprioritization, occasionally true incompatibility.

Beyond a simple pass/fail, the Space provides the exact run.py script used in testing, the Neuron stack version, and the extra packages required. On a trn2 instance (for example, inside the AWS Neuron Deep Learning Container) engineers can install the listed packages and execute the script directly — no torchrun wrapper needed because the script spawns its own distributed ranks when sharding is required.

A notable time-saver: the first run compiles the model graphs, but the Space also publishes the command to pull pre-compiled NEFF artifacts from a public Hugging Face bucket, keyed to the exact compiler version. This can cut bring-up time dramatically for CI/CD pipelines or rapid prototyping.

Current coverage focuses on text-generation and image-text-to-text tasks, limited to models that fit within four Trainium2 devices (16 cores). The team plans to expand to Diffusers, Sentence Transformers, and additional tasks. Results are backed by the open aws-neuron/torch-neuronx-model-support dataset, one Parquet file per compiler/library version pair.