Spaces:

zero-gpu-explorers

/

README

like 88

Running

I'm considering upgrading to the PRO account for ZeroGPU. Are there any other restrictions besides the 10-space limit I should know about? Specifically, how many minutes per generation and the total generating hours allowed? I don't plan to deploy my own space; I just want to use the existing ones as I've reached my free quota limit.

The system has a rate limit in place to ensure fair usage. This limit allows a maximum of 300 seconds of usage per person or IP address at a time.
When it comes to refilling, 1 ZeroGPU second is equivalent to 30 seconds in real life. This means that if you've used up your allocated time, you'll need to wait for 30 seconds in real life to refill 1 ZeroGPU second.

This limit allows a maximum of 200 seconds of usage per person or IP address at a time.

In my own spaces, it is okay to request 300 seconds at once. I think the maximum quota is around 600 seconds.

Got it, assuming it takes about 5 minutes (300 seconds) per generation, that works for me. But I'm still confused about the quota. So, I have unlimited generating time as long as I'm willing to wait for my quota to refill, right? I get 1 ZeroGPU second every 30 real-life seconds, so if I wait 12 hours, I get 1,440 ZeroGPU seconds? Sorry for asking such basic questions.

No, if you wait for 12 hours you got 600 sec because it also has maximum limit.

Got it, thanks for the quick reply

For me it was worthless to move to PRO User, as PRO User you get 1500 seconds quota BUT ONLY to use via the HF spaces. If you want to access that quota via programatically you'll be considered a normal user, only 300 seconds. So, for me it has absolutely no use to be PRO user since I do all via programatically :(

For me it was worthless to move to PRO User, as PRO User you get 1500 seconds quota BUT ONLY to use via the HF spaces. If you want to access that quota via programatically you'll be considered a normal user, only 300 seconds. So, for me it has absolutely no use to be PRO user since I do all via programatically :(

That is weird policy. Do you have any link to documentation stating this?

I have no link stating this, the lack of that information is my problem. If there is information about it an I hasn't found it I'll gladly receive it. Then, how can I know? Well, because I tested it, I have a PRO account and it give me 1500 seconds if I use mine or other spaces via the web browser. But if I use it via API calls with a python program, it simply gives me just 300 seconds. That's it because it has no way to know it's me, it only accepts the HF_TOKEN than can say you (any user) has access to that API but there is no way to tell that yo are a PRO User, that is the main problem, I d0n't know if Hugging Face just missed that part or it is correct and it's the bussiness model to not allow many programatic calls.

I have no link stating this, the lack of that information is my problem. If there is information about it an I hasn't found it I'll gladly receive it. Then, how can I know? Well, because I tested it, I have a PRO account and it give me 1500 seconds if I use mine or other spaces via the web browser. But if I use it via API calls with a python program, it simply gives me just 300 seconds. That's it because it has no way to know it's me, it only accepts the HF_TOKEN than can say you (any user) has access to that API but there is no way to tell that yo are a PRO User, that is the main problem, I d0n't know if Hugging Face just missed that part or it is correct and it's the bussiness model to not allow many programatic calls.

I totally understand you! I just purchased $9 worth of PRO subscription assuming that 1500 seconds of usage would be applicable to API as well (as that's what I was gonna use anyway). I was just browsing to find the actual documentation and came across your post. It's definitely not a right design and discouraging people from creating.
@huggingface

Maybe ongoing issue?
https://huggingface.co/spaces/zero-gpu-explorers/README/discussions/129#675dae2dda45b7d6f73d7e47

I have a quick question about ZeroGPU limitations.
Next year, (or perhaps that one aswell before the end lol), I plan to switch on a PRO account, the reason?
Well, I'm stuck with free Inference-API (which is really good), but I want have the free hassle to test models (while converting/fine-tuning them) in a full AUTOMATIC1111 webUI.

So.. i'm asking myself.
As I see, 25 minutes for ZeroGPU space, (I don't really understand what its mean), it's in the account billing tab (?) when clicked!

Some question..

• 1. Is the ZeroGPU is limited to 25 minutes of use by DAY ???
• 2. Is the ZeroGPU is limited to 25 minutes of use in a single inference/request/generation (don't know how you will call them if not passing by Free inference-API or Paid Inference-endpoint-API.)
     And then quota is decreased while using it, but refilled like a cooldown with waiting a little of time (as I though reading from there).

So thankful, if you can answer me.

If you want to access that quota via programatically you'll be considered a normal user, only 300 seconds.

@Moibe , this behavior is not a feature, it is a bug.

@FashionStash yes it is limited to 25 minutes per day in total, it is not per inference but for any inference or space used. It is refellied inmediatly, I don't know if second by second, but I built a program myself to measure it and refill my counter every 5 minutes. The equivalent would be 1 second refilled every five minutes.

I'm new to all this too, but I have been testing it intesively so feel free to ask any question, doesn't matter how simple you think it is. It's very helpful to count on someone that has gone through it.

@Moibe : hello, thank you for the information. Very useful information!
25mn is sufficient to you?
Because it seems too much smaller than some it's needed for hard task.
By the way I accepted your demand to my fs-blu* model. But take in your mind, that it doesn't work as expected. In fact if I gated it, it's because it needs a lot of new stuff to do with it for making it working. Because once it would be functional, I would like release in public totally ungated.

Sorry for the inconvenience. I'm working on another models which they clearly works better but under SD 1.5 which is not causing me too much headaches to make customs cookies with SD 1.5.

I hope it is clear.
(if you have tips on how why my gated model won't get work, then, use Discussions tab in model to open your answer!
Thanks! 😁)

@FashionStash thanks for the accept. And yes, for me 25 mins is enough right now, but in the long run it will fall short. The problem that I have, is that when calling it programatically for example from a python app, it only give me 300 seconds. It seems that it has been a problem for several developers here.

@Moibe ; Thank you for sharing your thoughts. Hmm.. Is there any way to pass a HuggingFace Token when doing requests to ZeroGPU programmatically?

@FashionStash I hasn't figured out any. What I do simply is to count the quota that I still have and continually also add the refill. if I hit zero quota, then programatically I change the hardware from zero gpu to a paid hardware, to keep going on automatically.