Stuck on 4bit?

#17
by xzuyn - opened

anon8231489123/gpt4-x-alpaca-13b-native-4bit-128g is RUNNING, but now nothing is updating. Has this broken things?

also this isnt 4bit, but maybe this is a problem model too? Intel/gpt-j-6B-int8-dynamic

I'm surprised how little attention it's been given from HF that the leaderboard has been basically stuck since a little into day two. You'd think someone from the team would poke at it or give it a non-passing glance. Might need a report to get someone's attention?

They released an article a few days ago about how '8-bit' is the future like it was some sort of new hotcake, so maybe they legitimately don't know that moderately low pequliarity 4-6bit GPTQ and ggml have been a thing for months?

They once released it and then just walked away?

Open LLM Leaderboard org

Hi, we have not forgotten the leaderboard. The initial release overloaded the cluster and I had to kill a lot of jobs. I need to manually restart some of them. Be patient, the models will eventually be evaluated.

Can you clean up the RUNNING status so that we know what's up next?

Hi, we have not forgotten the leaderboard. The initial release overloaded the cluster and I had to kill a lot of jobs. I need to manually restart some of them. Be patient, the models will eventually be evaluated.

Ok its good to know you guys are aware of it.

Hi all, thanks for your attention to this project! We are monitoring all the discussions in this Space and fixing bugs and making improvements.
We're committed to making this helpful for our users โ€” so we'll make sure to reflect your feedbacks :)

Just curious. Why run 20+ jobs simultaneously instead of just a few (or maybe even 1 at a time) to avoid the stuck? I check the leaderboard every several hours and found a lot of eval jobs stay in "running" status for very long time.

Just curious. Why run 20+ jobs simultaneously instead of just a few (or maybe even 1 at a time) to avoid the stuck? I check the leaderboard every several hours and found a lot of eval jobs stay in "running" status for very long time.

It will be good if they could also add the expected completion time for each model that is running, so we don't have to keep coming back to check...

Hi all, we have added a CHANGELOG component so that you can follow which bugs are being fixed and features are being added.

clefourrier changed discussion status to closed

Sign up or log in to comment