#Started receiving "stop_sequences: each stop sequence must contain non-whitespace"
14 messages · Page 1 of 1 (latest)
Potentially related or not, we also have been receiving "Reading the request body timed out due to data arriving too slowly.".
Also I seem to have indeed tons of timeouts on both MythoMax standard and Mistral 7B nitro with both Together and Lepton providers, both from the East US region and Germany, both in locally in dev and in prod. So it sounds like something might be wrong...
And it started working again just now, on both the local dev and remote environments. The openrouter.ai chat didn't seem affected.
So, it happened between 1:20 AM UTC and 3 AM UTC, it seems to be back up now? The original post error happened between 1:36 and 1:50 AM UTC. Since now it looks like it's working, there's not much more I can do but if you do figure out what happened, I'd be curious to know.
cc @light radish - there was a weird issue with newline being a stop word :-?
I still have the issue, by the way. Not much but I do.
I would need to log the provider that was used to help diagnose it further, which I'm not doing currently.
But if you're able to run tests quickly on all providers, they should all support \n as a stop word.
Oh you can check the provider detail now in activity page FYI
I think I figured it out. I was using anthropic/claude-3-opus:beta, when I remove :beta it works. I still think this error is wrong, but I guess because it's a beta version it doesn't matter now.
@supple wren sorry for the ping, I thought reviving this thread could be more efficient. I started receiving this error again (stop_sequences: each stop sequence must contain non-whitespace) for anthropic/claude-3.5-sonnet and opus. (Sept. 2, 3 and 4). I can't reproduce locally with either of the providers Anthropic or Google (BTW the page https://openrouter.ai/models/anthropic/claude-3.5-sonnet says "Google Vertex" but the Provider field should be Google). Let me know if that rings a bell. I'd love to move away from the line break as the stop word, but it's very efficient when you actually need... one line of text 🙂
Thanks for the flag, seems like the stop sequence error is from upstream??
We can remove/trim out the stop words that's simply empty for this model :-?