fixed rocm permission error and updated rocm containerfile

[bmahabirbu]

I updated the rocm container to include the ability to build with a selected GPU architecture. The original rocm container didnt work for me until I built it with the RDNA3 gfx code=1100 (supports 7800 and 7900).

I followed the steps in the build doc for hipblas
https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md

Also I used-DGGML_HIPBLAS=ON instead of -DGGML_HIPBLAS=1 when using cmake. The container would build but llama.cpp would not utilize the gpu even with -ngl enabled.

Also, I'm currently running Ubuntu 24.04 and I ran into a peculiar issue where podman wasn't giving the correct permission for passing -device path inside the container. It turned out to be an issue with running in rootless. I followed this thread [https://github.com/containers/podman/issues/10166]
and added

--group-add keep-groups

and it worked! Not sure if this issue is unique to my environment. For this to work podman oci runtime should be crun. Not sure if runc still has this issue, I'll dig into this some more.

I'll write up a doc based on my troubleshooting. It might help speed debugging for others!

[ericcurtin]

DCO build failed @bmahabirbu for all commits in the https://github.com/containers namespace you need to sign your commits with git commit -s. You can amend and push previous commits with:

git commit --amend -s
git push -f

[ericcurtin]

Might be simpler to do a:

cmake --install build

and do two mv's

haven't tried.

The reason the mv's are there is these two binaries are badly named. The llama.cpp equivalents used to also be named main and server, so you couldn't have both installed in the same directory because they have the same names.

[bmahabirbu]

That makes sense I tried it with cmake --install build and then the mv's but it didnt work. could be the second cmake --install overrides it still

[ericcurtin]

Maybe something like this:

+    cmake -B build -DCMAKE_INSTALL_PREFIX:PATH=/usr -DGGML_CCACHE=0 \
+      -DGGML_HIPBLAS=1 && \
+    cmake --build build --config Release -j $(nproc) && \
+    mv build/bin/main /usr/bin/whisper-main && \
+    mv build/bin/server /usr/bin/whisper-server && \
+    cmake --install build && \

the install doesn't cover the examples

[ericcurtin]

How does this work without the dnf install's for the ROCm libs/headers hmmm

[bmahabirbu]

Apologies and good catch! I realized this just after making the PR! I decided to make a change manually to keep as close to main as possible but accidentally deleted those lines.

[ericcurtin]

This only makes sense if you run it in the same RUN statement as where these things are added, to reduce the size of the image.

[bmahabirbu]

Gotcha! I'll reconfigure my containerfile to have it all under the same RUN statement.

[ericcurtin]

I wonder should we add like llama.cpp upstream containers if it doesn't cause issues:

gfx1010 \
gfx1030 \
gfx1100 \
gfx1101 \
gfx1102

[bmahabirbu]

Good idea, I checked the llama.cpp repo docs and that's how the official rocm dockerfile does it

[ericcurtin]

If you are curious why we leave out gfx9* ones, it's because the container images were getting humongous and we were hitting limits.

It would also take an age for users to download huge images, so we had to trim.

I wouldn't be against creating a rocm-gfx9 image in future to solve that problem though, if people complained their older AMD GPUs don't work.

[bmahabirbu]

That makes sense, I hooked up an old radeon 5450 gfx1036 to my rig and edited the llama.cpp arguments to use it but it ran out of memory it only had 512 mb. I have a vega 64 lying around I can give that a shot if it fits into my pc!

In general @ericcurtin sorry for the PR mess I'm wondering once I confirm compatibility should I make a new PR?

[rhatdan]

@bmahabirbu still working on this one?

[bmahabirbu]

@bmahabirbu still working on this one?

I'm all set with these changes!

Right now I'm looking into how InstructLab and Ollama create their respective container files and I'm hoping to integrate more useful changes

[rhatdan]

@bmahabirbu any update?

[bmahabirbu]

@bmahabirbu any update?

The Containerfiles for Ramalama have gotten really neat and the version I made here is messy comparatively.

I recently tested the updated rocm container on a new Fedora 40 Workstation install on my PC and I still needed to point the target to gfx1100 (which is a change I made in this PR) to correspond with my GPU architecture for it to work.

For now, we can close this PR and I can make a new one adding those architectures to the new rocm containerfile.