README / README.md
mrfakename's picture
Remove "Early Access" section as the Join button has been disabled
9e522ba verified
|
raw
history blame
4.57 kB
metadata
title: README
emoji: 🌍
colorFrom: indigo
colorTo: indigo
sdk: static
pinned: false

ZeroGPU Spaces

ZeroGPU is currently in beta. It's available to use for everyone and PRO users can host their own ZeroGPU Spaces.
Please share your feedback about ZeroGPU in the Community tab.
Browse ZeroGPU Spaces here: Dedicated list.

ZeroGPU is a new kind of hardware for Spaces.

It has two goals :

  • Provide free GPU access for Spaces
  • Allow Spaces to run on multiple GPUs

This is achieved by making Spaces efficiently hold and release GPUs as needed (as opposed to a classical GPU Space that holds exactly one GPU at any point in time)

ZeroGPU uses Nvidia A100 GPU devices under the hood (40GB of vRAM are available for each workloads)

Compatibility

ZeroGPU Spaces should mostly be compatible with any PyTorch-based GPU Space.
Compatibility with high level HF libraries like transformers or diffusers is slightly more guaranteed
That said, ZeroGPU Spaces are not as broadly compatible as classical GPU Spaces and you might still encounter unexpected bugs

Also, for now, ZeroGPU Spaces only works with the Gradio SDK

Supported versions:

  • Gradio: 4+
  • PyTorch: All versions from 2.0.0 to 2.2.0
  • Python: 3.10.13

Usage

In order to make your Space work with ZeroGPU you need to decorate the Python functions that actually require a GPU with @spaces.GPU
During the time when a decorated function is invoked, the Space will be attributed a GPU, and it will release it upon completion of the function.
Here is a practical example :

+import spaces
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(...)
pipe.to('cuda')

[email protected]
def generate(prompt):
    return pipe(prompt).images

gr.Interface(
    fn=generate,
    inputs=gr.Text(),
    outputs=gr.Gallery(),
).launch()
  1. We first import spaces (importing it first might prevent some issues but is not mandatory)
  2. Then we decorate the generate function by adding a @spaces.GPU line before its definition

Note that @spaces.GPU is effect-free and can be safely used on non-ZeroGPU environments

Duration

If you expect your GPU function to take more than 60s then you need to specify a duration param in the decorator like:

@spaces.GPU(duration=120)
def generate(prompt):
   return pipe(prompt).images

It will set the maximum duration of your function call to 120s.

You can also specify a duration if you know that your function will take far less than the 60s default.

The lower the duration, the higher priority your Space visitors will have in the queue