Some questions about usage

[fighterhit]

1. Are there no PyTorch or Tensorflow frameworks in the Triton Inference Server image provided by NGC? e.g. nvcr.io/nvidia/tritonserver:22.12-py3
2. I have a PyTorch model file saved through torch.save(model.state_dict(), 'model.pt'). How could I provide services through Triton Inference Server? Is there any recommended usage? I tried to use the torch2trt library, but it was not installed successfully.
3. In the past, my model provided http services through flask. The input was a self-defined MessagePack serialized byte array. The Content-Type was application/octet-stream. Before the data was input into the model, deserialization and image conversion operations were performed. The model output tensor will also be converted into a python list and returned to the client. How could I use Triton Inference Server to meet the above requirements and provide http services?