Thanks for sharing that but If I'm not mistaken, I couldn't share my exact issue here. Shared memory will also utilize the same communications as HTTP/RPC. However, instead of transferring the entire image, for example, to the Triton server, it will bind the image's address to some shared memory and then send the address using HTTP to the Triton server. By doing this, we can save the cost of transferring data. Please correct me if I'm wrong about this. I want to know if clearml can support such a thing at the moment?