S Aruna

Not consistent. I ran it 11 times. 10 times it failed to generate the same tokens when using the HF API vs NNSight's trace, but the 11th time it succeeded

I see. Thanks for looking and the comments. Noted about future bugs, will provide a script that runs off the shelf.

As a follow up, is a way to avoid this to not use multiple invokes, and use two separate `trace` calls?