Skip to content

Before you get started

1. get a hugging face account

Be sure to create a huggingface account, and visit this webpage:

2. get a hugging face API token

Go to your profile and look under settings and then go to Access Tokens.

3. read + accept license (or you will be blocked)

Here is the model card, the description + developer: https://huggingface.co/CompVis/stable-diffusion-v1-4 link

You need to: read, accept, and acknowledge the LICENSE. if done correctly this will allow the following to be run:

from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4",
    revision="fp16",
    torch_dtype=torch.float16
).to("cuda")

The above will pull the stable diffusion from the centralized huggingface modewl repository. This will also save the following information to a caching directory. This is passed during run_docker as an environmental variable.

4. Check your pytorch version

import torch

torch.__version__

The following means its pytorch intended for CUDA 10.2 which is old 

>>> 1.12.1+cu102

5. Common driver errors:

  • you have a gpu but there's no NVIDIA drivers
  • the container has CUDA==10.2, but pytorch==cu11.3, meaning the versions don't match

6. Can pytorch see the gpu?

import torch

print(torch.cuda.is_available())
>>> True

Note: if the answer is False there is often a warning that is logged to give a hint at what is missing

7. is the GPU even available?

$ nvidia-smi