Hugging Face's October 23 Trainium lookup reports model-support results for selected checkpoints.
The big change
The October 19 TorchNeuron announcement introduced a native PyTorch backend. The newer lookup gives engineers a tested checkpoint and stack for an initial trial. A PASS establishes a 32-token run; longer generation, task quality and throughput still need separate evaluation.
How a PASS is checked
On Trainium2, the project uses the fewest cores that hold model weights and generates 32 tokens. It compares a single-core run with CPU output from the same checkpoint; a small judge checks sharded output for fluent English. The ceiling is four devices, or 16 cores.
What to take from a row
Hugging Face lists core count, compilation-inclusive time, memory, stack and script. BIG CHANGE reviewed documentation only.
Coverage is limited to Transformers text-generation and image-text-to-text models. Live counts remain unknown.
