Spaces:
Running
Every model is going nuts now
As of late HuggingChat is literally unusable, every other model starts printing gibberish as early as its second reply. What's going on? I can confirm it happens both with Llama-3.3-70B-Instruct and with Qwen2.5-72B-Instruct, maybe also with others but we are already talking about two different unrelated models/architectures so the problem goes probably beyond that
This is Llama:
And this is Qwen:
I left the system prompt empty (default) for both
Yep we're aware of the issue, trying to fix it, sorry for the inconvenience!
I have experienced a similar problem with Nemotron. It would work with the first turn, but on second turn it would spit double asterisks and other nonsensical tokens like crazy.
any update on how quickly it might be solved?
Hi! The issue should be fixed. We're still investigating the root cause so please report back if it happens again but it should work for now!