The TensorRT conversion process will automatically fuse ops if it can yield better performance. We highly recommend configuring TensorRT to do so too, but there seems to be a gap where it’s hard to tell which operations / layers are fused especially when experimenting with different deep learning optimization workflows.

Cellulose plans to leverage the TensorRT API to annotate these from within the model editor in our dashboard.

Have questions / need help?

Please reach out to, and we’ll get back to you as soon as possible.