Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

fixed rocm permission error and updated rocm containerfile #376

Closed
wants to merge 1 commit into from
Closed

fixed rocm permission error and updated rocm containerfile #376

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 12 additions & 7 deletions container-images/rocm/Containerfile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,23 +16,28 @@ RUN curl --retry 8 --retry-all-errors -o \
cat /etc/pki/rpm-gpg/RPM-GPG-KEY-CentOS-Official
RUN rpm --import /etc/pki/rpm-gpg/RPM-GPG-KEY-CentOS-Official

# Set amd gpu architecture for RDNA3
# https://llvm.org/docs/AMDGPUUsage.html#processors
ENV AMDGPU_TARGETS=gfx1100
Copy link
Collaborator

@ericcurtin ericcurtin Oct 26, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

gfx1010 \
gfx1030 \
gfx1100 \
gfx1101 \
gfx1102

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

Copy link
Collaborator

@ericcurtin ericcurtin Oct 26, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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?


RUN dnf install -y rocm-dev hipblas-devel rocblas-devel && \
Copy link
Collaborator

@ericcurtin ericcurtin Oct 26, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.

dnf clean all && \
git clone https://github.com/ggerganov/llama.cpp && \
cd llama.cpp && \
git reset --hard ${LLAMA_CPP_SHA} && \
cmake -B build -DCMAKE_INSTALL_PREFIX:PATH=/usr -DGGML_CCACHE=0 \
-DGGML_HIPBLAS=1 && \
cmake --build build --config Release -j $(nproc) && \
-DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=${ROCM_DOCKER_ARCH} && \
cmake --build build --config Release -j$(nproc) && \
cmake --install build && \
cd / && \
git clone https://github.com/ggerganov/whisper.cpp.git && \
cd whisper.cpp && \
git reset --hard ${WHISPER_CPP_SHA} && \
make -j $(nproc) GGML_HIPBLAS=1 && \
mv main /usr/bin/whisper-main && \
mv server /usr/bin/whisper-server && \
cmake -B build -DCMAKE_INSTALL_PREFIX:PATH=/usr -DGGML_CCACHE=0 \
-DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=${ROCM_DOCKER_ARCH} && \
cmake --build build --config Release -j$(nproc) && \
mv build/bin/main /usr/bin/whisper-main && \
Copy link
Collaborator

@ericcurtin ericcurtin Oct 26, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

Copy link
Collaborator

@ericcurtin ericcurtin Oct 26, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

mv build/bin/server /usr/bin/whisper-server && \
cd / && \
rm -rf /var/cache/*dnf* /opt/rocm-*/lib/llvm \
/opt/rocm-*/lib/rocblas/library/*gfx9* llama.cpp whisper.cpp

/opt/rocm-*/lib/rocblas/library/*gfx9* llama.cpp whisper.cpp
2 changes: 2 additions & 0 deletions ramalama/cli.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -679,6 +679,8 @@ def run_container(args):
if gpu_type == "HIP_VISIBLE_DEVICES":
conman_args += ["-e", f"{gpu_type}={gpu_num}"]
if args.image == default_image():
# https://github.com/containers/podman/issues/10166
conman_args += ["--group-add", "keep-groups"]
conman_args += ["quay.io/ramalama/rocm:latest"]
else:
conman_args += [args.image]
Expand Down
Loading