Spaces:

zero-gpu-explorers

/

README

like 88

Running

It's not the crux of this issue, but if the durarion designation exceeds 120 seconds, it's basically buggy.

It's not the crux of this issue, but if the durarion designation exceeds 120 seconds, it's basically buggy.

NO, I just do a test to find the quota.
Actually, the duration maybe 5~10s

Actually, the duration maybe 5~10s

Oh. If so, it's definitely a bug. 😅
I'm also using Zero GPU space, and sometimes the login quota doesn't work properly.
However, in my case, for some reason it's a different pattern from yours, and the login status doesn't work properly in the web browser. If I sign in explicitly using the OAuth sign-in button on the space, it works.
I've never used the Gradio Client...

I have exactly the same need as Nerva1228, and thanks John666 but it seems that, sound like a meme but, it's not a bug it's a feature. Not being able to use my PRO user quota programatically is not good, I want to test my own spaces with different combinations from within my code and I'm as limited as a normal user, is this intended to be that way or there is a solution, or maybe a feature for the future.

I got it to work!!!!

• I duplicated the black-forest-labs/FLUX.1-dev (!!!Important!!!! Go to their Model Card Page and get access granted to their gated model first. Then you can duplicate the model)
• Set the environment variables:
  HF_TOKEN =Use your HF_token (I used read and write credentials for this)
  ZEROGPU_V2=true
  ZERO_GPU_PATCH_TORCH_DEVICE=1

Then in your own (duplicated) space:

• Navigate to the Files
• Click on app.py
• Change

@spaces.GPU(duration=75)  # The duration max value can be 120 but this wasn't enough and still didn't work for me

to

@spaces.GPU()

Make sure in your python code that uses the gradio_client python library your HF_TOKEN is set in the environment or you set the parameter hf_token when creating the client
example:

from gradio_client import Client 
 client = CLIENT("your_duplicated_space/FLUX.1-dev", hf_token=os.getenv("HF_TOKEN"))

I got it to work!!!!

• I duplicated the black-forest-labs/FLUX.1-dev (!!!Important!!!! Go to their Model Card Page and get access granted to their gated model first. Then you can duplicate the model)
• Set the environment variables:
  HF_TOKEN =Use your HF_token (I used read and write credentials for this)
  ZEROGPU_V2=true
  ZERO_GPU_PATCH_TORCH_DEVICE=1

Then in your own (duplicated) space:

• Navigate to the Files
• Click on app.py
• Change

@spaces.GPU(duration=75)  # The duration max value can be 120 but this wasn't enough and still didn't work for me

to

@spaces.GPU()

Make sure in your python code that uses the gradio_client python library your HF_TOKEN is set in the environment or you set the parameter hf_token when creating the client
example:

from gradio_client import Client 
 client = CLIENT("your_duplicated_space/FLUX.1-dev", hf_token=os.getenv("HF_TOKEN"))

And what about the usage of the quota, you are able to use your PRO User quota??

I got it to work!!!!

• I duplicated the black-forest-labs/FLUX.1-dev (!!!Important!!!! Go to their Model Card Page and get access granted to their gated model first. Then you can duplicate the model)
• Set the environment variables:
  HF_TOKEN =Use your HF_token (I used read and write credentials for this)
  ZEROGPU_V2=true
  ZERO_GPU_PATCH_TORCH_DEVICE=1

Then in your own (duplicated) space:

• Navigate to the Files
• Click on app.py
• Change

@spaces.GPU(duration=75)  # The duration max value can be 120 but this wasn't enough and still didn't work for me

to

@spaces.GPU()

Make sure in your python code that uses the gradio_client python library your HF_TOKEN is set in the environment or you set the parameter hf_token when creating the client
example:

from gradio_client import Client 
 client = CLIENT("your_duplicated_space/FLUX.1-dev", hf_token=os.getenv("HF_TOKEN"))

And what about the usage of the quota, you are able to use your PRO User quota??

Yes. Try it out and see if it works for you.

I tried it and it doen't work, one thing is ti reduce the duration like this: @spaces.GPU(duration=75), it's ok.

But that doesn't change the fact that we are consuming quota as a normal user and not as a PRO User when we use it programatically.

The problem stated by Nerva1228 and me, it's still there, we want to be able to use our 25 minutes quota via our python code, not only in the hf spaces interface.

Perhaps related:
https://huggingface.co/spaces/BestWishYsh/ConsisID-preview-Space/discussions/1#674969a022b99c122af5d407

I am stacked with the same problem.
Is any way to upvote this issue?

@hysts Is this a problem specific to Zero GPU space?
Is it better to raise the issue with HF or Gradio?

@John6666 There’s no need to create an issue for this. My understanding is that our infra team is aware that ZeroGPU quota does not work with API use case and that there's a request for it, but it simply hasn’t been implemented yet. cc @cbensimon

Thank you! I understand. I will avoid the issue so as not to cause any unnecessary trouble.😅

@John6666 There’s no need to create an issue for this. My understanding is that our infra team is aware that ZeroGPU quota does not work with API use case and that there's a request for it, but it simply hasn’t been implemented yet. cc @cbensimon

If it is indeed a bug recognized internally within HF, will Hugging face compensate the PRO subscribers for not being able to use quota for which they have paid for? @hysts

@asr143r I'm not in a position to answer refund-related questions, so I've forwarded your inquiry internally.

That’s great, the important thing is HF notice and tell us if it’s an error or is the way is has to be.

Is there a public ticket link we can look at so we can be aware of any changes? I have asked for a refund a couple of hours after I got PRO because of this, and I am waiting for them to fix it. In the meantime, open to other options.

Ditto!

@hysts It's not just the api. I get this as a PRO user on my own Zero GPU space (https://huggingface.co/spaces/MegaTronX/Llama-3_1-8B-Abliterated) all the time. My subscription fee is basically a donation ;) to Hugging Face at this point since Zero GPU spaces are also the reason I originally subscribed.

If the infrastructure team is too busy at the moment to tackle this I'd be happy to help. I've been doing DevOps since before they had a name for it and I've built cloud infrastructure for enterprise clients from Intuit (only first year 3rd party vendor in their company history that didn't crash during war room week, aka tax week) to Walmart (oversaw the infrastructure for the inception of WalmartOne, their internal social media site that employees view their paychecks on). There are also a number of AI cloud infrastructure and SaaS companies (e.g. Runpod, Fireworks AI, etc.), that are trying, and failing miserably in many cases, to interface with Hugging Face Repos, including for their own Serverless (and Dedicated in some cases) cloud GPU offerings. I'm sure I could help with that as well, assuming Hugging Face isn't purposefully eschewing those companies for any reason, and is willing to negotiate a beneficial and compensatory solution with them. I'm more than confident that they'd jump at the chance to throw money at Hugging Face's Hugging Face to get their already being sold, never mind advertised, services that don't work, working. I don't need to consult with them to know that they're extremely desperate to get that stuff working.

Hi @MegaTronX It seems that the SDK version of your Space is gradio==4.31.3 https://huggingface.co/spaces/MegaTronX/Llama-3_1-8B-Abliterated/blob/2af4a53bd398c40b66faaed62b251e87bf953ab7/README.md?code=true#L7, but can you try upgrading it to the latest gradio==5.9.1? IIRC, there was a bug in old gradio where the ZeroGPU quota was not properly handled when using gr.ChatInterface, but it's fixed in the latest gradio.

@hysts Yup, you nailed it, upgrading to the latest gradio version got rid of the GPU quota errors I was getting on the site. Thanks for solving that mystery for me.

For anyone else that is getting this error, in this case when using their space on the Hugging Face site, there is a potential caveat to be aware of when upgrading gradio that tripped me up initially. Sometimes when you change a number, you don't actually change the number, i.e. check the logs. After updating the gradio version in the requirements file, I was still getting the error. When I looked under the container logs section I saw "IMPORTANT: You are using gradio version 4.31.3, however version 4.44.1 is available, please upgrade." Digging through the builds log it showed that 5.9.1 was being installed initially, but later in the logs it was being uninstalled, followed by 4.31.3 being reinstalled. I didn't see any entries indicative of why, but I assume it's probably just too big of a leap between the two versions, especially where 5.0 was a major release with fundamental changes to the code base. Rather than trying my luck with incremental upgrades to see if I could get there (which if I had 5 likes on the space I might have done), I just created a new space and after working through the deprecated code etc., I then made my chatbot talk to me for an excessive amount of time with no quota errors in site.