search

runpod-workers / worker-a1111

Automatic1111 serverless worker.
MIT License
75 stars 109 forks source link

CUDA drivers? #8

Closed jadechip closed 1 year ago

jadechip commented 1 year ago

Is there a reason a base image with CUDA drivers is not used? Such as nvidia/cuda:11.7.1-cudnn8-runtime-ubuntu20.04? Would be happy to create a PR. Thanks

justinmerrell commented 1 year ago

Just to keep the docker size down, the drivers are installed, and the final built image uses the GPU.

If there is a particular issue you are running into please provide additional information, PRs that offer improvements are always welcomed :)

albicasty commented 1 year ago

sorry, it is not clear for me how cuda drivers get installed.

I am trying to build the image with docker locally, but I am getting errors because the NVidia drivers are not found.

albicasty commented 1 year ago

image image

These are some of the errors I get. I am on ubuntu 22.04 with nvidia drivers installed.

justinmerrell commented 1 year ago

Do you see that error during the docker build?

And pytorch includes everything needed to interact with the GPU, drivers are handled at a system level. https://discuss.pytorch.org/t/is-nvidia-driver-already-included-cuda-and-cuda-toolkit/184411

albicasty commented 1 year ago

ok thansk for the clarification.

Yes, I get the error during the build:

image

If I change RUN export COMMANDLINE_ARGS="--skip-torch-cuda-test --precision full --no-half" with ARG COMMANDLINE_ARGS="--skip-torch-cuda-test --precision full --no-half

I dunno get anymore the error during the build

justinmerrell commented 1 year ago

Are you making changes to the repo or building from a clean clone?

It looks like something needs those build time args vs the runtime args, but I'll consider this resolved since you have a resolution.