I just saw this thread. The txt2kg project from Nvidia had a line in the docker yml file that actually prevented it from using the GPU (if you’re using the included Ollama container). This has been fixed now (see txt2kg playbook discrepancy).
I realize your post is now a month old almost, but if you didn’t have the magic line in your code that would explain the extremely slow performance.
I’m digging into the code to see if I can get it to use Ollama (or better, vllm) on a second machine, which would leave more room for the rest of the stack. It’s a rat’s nest so I don’t know whether or not I will succeed.
Let me know whether you’ve made progress on this.