ROCm Image Size

[futursolo]

See: ROCm/ROCm#4224
See: ROCm/ROCm-docker#120

The ROCm image is very big comparing to the intel variant.

For example, the rocm variant is almost double the size of the ipex variant.

ghcr.io/futursolo/pai-apps/comfyui ipex 4af61988e3dc   23 hours ago   9.37GB
ghcr.io/futursolo/pai-apps/comfyui rocm 8a0c80645cdb   21 hours ago   16.6GB

[sylcas]

@futursolo I posted a Dockerfile here, that you can build yourself. It reduces the image to a very moderate amount of uncompressed 6.8 GB by removing applications that are not needed and also stripping all components that are not for my specific gfx architecture. You may need to fiddle with the image, but it shows a good starting point for people that like to get their hands "dirty".

[futursolo]

This project intends to package generic container images that supports as many graphics cards as possible.
So I guess the technique of only leaving certain cards enabled might not be applicable to images built here.
But I will checkout the other parts of the dockerfile and see if any additional improvements can be made.

In addition, I am more interested in a long term solution on the AMD side, such as: ROCm/ROCm#3985 (comment)

Anyways, thank you for the tips @sylcas!

[sylcas]

You are welcome @futursolo!

If you want all cards to be supported you just have to remove one line from the Dockerfile I provided:

find /pytorch/lib/ -type f -regextype posix-extended -regex '.*gfx[0-9]+[^/]*\.(dat|co|hsaco)$'            \
        -not -regex '.*gfx120[0-9]+[^/]*\.(dat|co|hsaco)$' -delete

After that all gfx profiles are active. The image will be larger, but still a lot smaller than the original rocm/pytorch image since there is a lot of other stuff that I have removed from the image.

[sylcas]

Today I had a little more time for experiments, so I created a new container for the current rocm version 6.4.3. When keeping all gfx card profiles (by removing the find/delete I mentioned yesterday) the uncompressed size of the container is ~15.3GB.

The numbers you mentioned @futursolo are probably compressed sizes, right?

When removing the gfx card profiles, that I do not need the image shrinks down to a moderate 8GB (uncompressed). To be clear, this is not a fully featured rocm, but that is also not needed if all you want is to run some pytorch application which I think comfyui is, right? I am running faster-whisper with this container and so far everything seams to be running smoothly.

This might be interesting for issue #120 ROCm/ROCm-docker#120

[futursolo]

The numbers you mentioned @futursolo are probably compressed sizes, right?

They are the number shown by docker images so I guess they are uncompressed size.
GitHub Container Registry doesn’t show image sizes, but I think the last time I had a full pull it was about 3~5GB-ish.
So that should be the compressed image size.

The current uncompressed image size for pytorch only is:

ghcr.io/futursolo/pai-apps/base pytorch-v2.8.0-rocm-v6.4.3-20250823   75a5012f8e65   6 days ago     16GB