Is the SAM Demo hosted in the https://segment-anything.com/demo use Pytorch models or ONNX models as the prompt encoder and mask decoder

Hi I just came across the facebook's segment anything model,

According to the web demo mentioned in GitHub, it uses the PyTorch model to generate embeddings and then uses onnx model to prompt the encoding part. but the accuracy is low when compared with the hosted demo in the meta's site https://segment-anything.com/demo.

appreciate an answer regarding this

Compare the hosted demo at https://segment-anything.com/demo with the PyTorch and ONNX model usage described in the issue. Confirm which models the demo uses and determine whether the reported accuracy difference is expected; done means providing a clear answer with any relevant model or configuration details.

Written by the indexing model from the issue text.