We mostly figured it out on discord.
No theres no solution. Its something wrong with Tensor RT probably. Riva script is complex and there is no documentation.
Getting a models from NeMo and deploying into Tensor RT can get to level of complexity as high as messing with CUDA kernels.
I reccomend we open a Slack channel or talk to Edmar @edmar1 or Wendy @WendyGram to open a thread or discord so we can go over this.
We can go step by step until we hit the issue and explore solutions.