#Mistral Models 502 errors
67 messages · Page 1 of 1 (latest)
They’re working on the playground- what parameters are you sending?
yh this is purely from API; all other models work fine which is weird.
{
model: 'mistralai/mistral-small',
messages: [ { role: 'user', content: [ { type: 'text', text: '12*13' } ] } ]
}
fails
{
model: 'openai/gpt-3.5-turbo-0125',
messages: [ { role: 'user', content: [ { type: 'text', text: '12*13' } ]
} ]
}```
while something like this works
hi, im getting a 502 error via sillytavern as well dunno if this is a good place to post this but i didnt wanna make a whole new thread
Endpoint response: { error: { message: '{"object":"error","message":"Service unavailable.","type":"internal_server_error","param":null,"code":"1000"}', code: 502
this is happening with all mistral models for me
Looking into it. (This error is coming from Mistral platform itself.) in the meantime try using Mixtral 8x7B Instruct, which is hosted outside of mistral platform
is there any ETA on this being fixed? i hope this isnt a permanent problem with mistral lol
wonder whats going on at their end
trying to reproduce it first - if i don't have any luck then we'll email them
cc @tall knot
I wasn't able to reproduce this, but I just deployed a possible mitigation of this. Can you guys try now?
If it didn't work, can you see if ST is sending a response_format parameter in your terminal for some reason? (it's not for me)
what countries are you two in?
try it to see if that helps
Which VPN region did you try?
the vpn was connected to japan
er okay hold on im using protonVPN and their free plan is annoying as shit
but i can confirm that the US vpn server doesnt work either
netherlands isn't working either
hmmm
should i try using another frontend? like venus.ai or something
maybe its a sillytavern problem or smth
working here
Is it possible that
when calling from ST, it's calling us directly from your machine
and your VPN doesn't shield outbound localhost request?....
That's my hunch...
oh what the hell
i just tried venus.ai out of curiosity, it works just fine there
thats so weird
yup
but it works on venus without a vpn
all good on that front
apparently theres an update, gonna update ST and see if that fixes anything
nope lol, damn
I'd check your VPN setting and see what's up :-?
hm ill try another one, see if that helps at all
hm yeah idk nothing really seems to be working
im using 8x7b rn but i really liked using mistral medium
@spring oak which parameters are you sending via api?
nothing crazy, just what is default in docs (all other models work) — Mistral: Mixtral 8x22B (base) works (as does Mistral 7B Instruct
It's mistral tiny/small/medium/large that fail.
Just tried and I get:
{
error: {
message: '{\n' +
' "message":"An unexpected error occurred",\n' +
' "request_id":"d26d4c2b14bcae3dd952431f33c4a00b"\n' +
'}',
code: 502
}
}
``` with those
code is some version of this, but again, works for all models but mistral tiny/small/medium/large @finite chasm
const openai = new OpenAI({
baseURL: "https://openrouter.ai/api/v1",
apiKey: API_KEY || process.env.OPENROUTER_API_KEY,
defaultHeaders: {
"HTTP-Referer": "XYZ",
"X-Title": "XYZ",
},
});
let request = {
model: model_name,
messages: [
{
role: "user",
content: [
{
type: "text",
text: textInput,
},
],
},
],
};
let completion = await openai.chat.completions.create(request);
What happens if you remove response_format?
ah sorry that shouldn't be there, only passed in for the few models that support it for evals; editing above
above @finite chasm seems to imply that it was coming from mistral ("This error is coming from Mistral platform itself."); I don't think it's an issue with how I'm querying the API
which would explain why Mistral: Mixtral 8x22B (base) works but not Mistral tiny/small/medium/large
@tall knot flagging that this is still an issue, getting
"\"{\\\"error\\\":{\\\"message\\\":\\\"{\\\\n \\\\\\\"message\\\\\\\":\\\\\\\"An unexpected error occurred\\\\\\\",\\\\n \\\\\\\"request_id\\\\\\\":\\\\\\\"6c0a0c05b22e0231ff34b16b608ca757\\\\\\\"\\\\n}\\\",\\\"code\\\":502}}\""
back for Mistral hosted models. All other models working as expected.
yeah that's an error coming from upstream :d...
We’ll email them