Hugging Face publishes a lookup table for open models on AWS Trainium

Does our open model run on Amazon's Trainium chips? That question used to need a weekend of trial compiles. Hugging Face has now published a lookup table that answers it for a growing list of models. The post went up on the Hugging Face blog on 23 October. The author is David Corvoysier.

The data lives in two places. There is a Space called aws-neuron/torch-neuronx-model-support. A team can search it by model id. There are three views: Selected, Trending and Downloads. There is also a Parquet file at the same path. The team can read it with one line of Python:

df = pd.read_parquet("hf://datasets/aws-neuron/torch-neuronx-model-support/data")

For a working model, the table lists four facts. The smallest number of Trainium cores that hold the weights. The wall time of the test. The peak device memory. And the exact script that produced those numbers. A team can run that script on its own trn2 instance and get the same result.

How the tests work

Each model is loaded on a trn2 instance. The test shards it over the fewest Trainium cores that fit the weights. It then asks the model to generate 32 tokens. For a single-core model, the output is checked against the same model on CPU. For a sharded model, a judge model reads the output. It checks that the text is fluent English. The ceiling is four Trainium2 devices. That gives 16 cores. Anything that needs more is out of range for this test.

The passing graphs are also cached. The post describes a public cache of Neuron Executable Format files. These are the files the Neuron compiler produces for each model and setting. A team that pulls from the cache skips the compile step. That is where a first Trainium run on a new model usually stalls.

What a team changes

Before this, the answer to "does our model run on Trainium" needed a chat with the Neuron engineers. Or an experiment in the team's own account. The data cannot catch every case. It records the Neuron SDK version at the time of the test. It also records the tokenizer and the runtime. The team's own stack can differ from any of them. But the lookup names the smallest instance a passing model needs. That is what a buyer needs to turn an open weight into a price per thousand tokens.

The post also names a worked example in the Python snippet. A query for Qwen3 0.6B returns five fields. The Neuron compiler version. The library name and version used in the test. The number of Neuron cores. And whether the run passed. The same shape of row holds for every tested model in the table.

Source

Hugging Face blog.