Spaces:

John6666

/

votepurchase-multiple-model

like 60

Running on Zero

Hi.
I am facing the same problem.🤢
And this is even though I haven't tinkered with any part of it that involves infreferences.
I believe it has something to do with the fact that the specifications for the Zero GPU space have changed (or are in the process of changing) significantly.
Yesterday I had another error, and the day before yesterday yet another, and so on.

In other words, it is more likely that the re-build was triggered and fatal than the contents of the commit.
It's okay to restart, but when you rebuild, you're usually swamped.
https://huggingface.co/posts/bartowski/524900219749834#66cfaa865594967fbdedc542
https://discuss.huggingface.co/t/flowise-space-stuck-on-building/103813/5

What is even more puzzling is that some spaces can be re-built and others cannot.
Perhaps there is some conflict between the "spaces" and "stablepy" libraries used in DiffuseCraft and votepurchase.

https://huggingface.co/r3gm
https://github.com/R3gm/stablepy
By the way, r3gm, the author of stablepy, is in HF, so if I know what is going on I will report to him.
Yes, if anyone knows what is going on...😓
No kidding, I'm starting to think that even the HF staff doesn't know the technical reason for the glitch, except for the server's virtual machine-related administrators, the spaces library developers and a few more people who are actively tinkering with Spaces (like Mr. multimodalart)...

I also don't see any catastrophic changes in the inference related code

I tried a few other spaces and they all work flawlessly for some reason

I've also seen a few with these set environment variables

ZERO_GPU_PATCH_TORCH_DEVICE(I do not know what that even is)

ZEROGPU_V2(which from what I heard should be set to true)

Maybe those are the problem?
I'd say try setting them and see if things work. (I believe those variables were pushed in a new spaces package update)
Since you mentioned that maybe the space got rebuilt

ZEROGPU_V2 is probably the main culprit in this case. Because as far as the behavior of the space is concerned, it became true by default sometime between the day before yesterday and this morning or so.
I've tried setting it to false and so on since a while ago, but there is no response in any way. It stays on.
I guess I'll have to wait and see about improving this one.

Never heard of ZERO_GPU_PATCH_TORCH_DEVICE. It might be worth looking into.

Please do report to him if you find anything!

I also tried checking the HF discord, no one is talking about this there for some reason.

If only we had a real way to bug report stuff that are related to spaces, to HF staff that actually work on spaces.

Another part of the problem is that there is no way to contact staff that work on zeroGPU other than creating a discussion on the zeroGPU org and hoping one of them sees it lol

I believe ZEROGPU_V2 cannot be turned off since it's going to become a stable version of zeroGPU (not that anyone knows the difference between V2 and V1 in the first place)

ZERO_GPU_PATCH_TORCH_DEVICE is also a strange one that I discovered on https://huggingface.co/spaces/multimodalart/FLUX.1-merged

But it's not mentioned anywhere in the code

I was able to fix it.
I figured out how to fix it, but I can't determine if this is a bug or a temporary spec during the change.🤯
Specifically, I had been adding @spaces.GPU decorator to the function I was calling inside the function, but when I added it to the outermost function, it fixed it.
This method is wasteful and should not normally be recommended.

I just synced with the latest commit and now the space functions as intended!

The Zero GPU space repo is on HF, but with this number of people, I can't even send a mentions...
https://huggingface.co/zero-gpu-explorers

Thank you for the help!
I'll be closing this now and I'll report back to you if I find anything weird.

I think I might be able to tell you the list of admins of the org

ZERO_GPU_PATCH_TORCH_DEVICE is also a strange one that I discovered on https://huggingface.co/spaces/multimodalart/FLUX.1-merged

But it's not mentioned anywhere in the code

I just googled "ZERO_GPU_PATCH_TORCH_DEVICE", what a surprise, zero hits.🙀

I just googled "ZERO_GPU_PATCH_TORCH_DEVICE", what a surprise, zero hits.🙀

Yeah this is Not documented anywhere, I hate the fact that zeroGPU does not document anything at all

Also here is the list of admins in the org:
https://huggingface.co/akhaliq
https://huggingface.co/cbensimon
https://huggingface.co/victor
https://huggingface.co/julien-c
https://huggingface.co/ysharma
https://huggingface.co/sayakpaul
https://huggingface.co/sbrandeis
https://huggingface.co/hysts
https://huggingface.co/merve

I'm not sure which one you should mention but i think its either victor or Julien

I've opened Discussion for now.
https://huggingface.co/spaces/zero-gpu-explorers/README/discussions/104

Of those I've only seen sayakpaul and victor...
I've heard of some of the others, though.

I just mentioned to victor.

they should probably take a few hours to reply but atleast we might get some solution to such issues.

There is also this problem where the "No GPU available for you after 60 seconds" and the "GPU task aborted" deducts the full set duration (which was not even complete and resulted in an error) from the quota for some reason.
I should probably open a discussion for this

Sorry, I'm really sleepy. See you tomorrow.🌚