Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/president-not-sure/rocm
ROCm container
https://github.com/president-not-sure/rocm
Last synced: 9 days ago
JSON representation
ROCm container
- Host: GitHub
- URL: https://github.com/president-not-sure/rocm
- Owner: president-not-sure
- License: mit
- Created: 2024-10-17T19:53:18.000Z (28 days ago)
- Default Branch: master
- Last Pushed: 2024-10-17T22:49:42.000Z (28 days ago)
- Last Synced: 2024-10-19T11:17:47.857Z (26 days ago)
- Size: 5.86 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# ROCm
ROCm container for linux
> [!NOTE]
> More than 30GB when built. So you have to build it locally as it would not fit a regular runner.
> Hence why there is no built package in the repo.
> This container image could be considered bloated, but it should not miss anything in it.
> If you have access to larger github runners the `publish.yaml` is functional with a few simple modifications like the `runs-on:`.
> It works right now, but it runs out of space.
> You would also need to generate a new key pair with `skopeo` and add a new github secret with the private key to the repo or remove signing altogether.
## Building
Minimal build command
```shell
podman build --tag=localhost/rocm:latest rocm.Containerfile .
```
## Running
Example run command
```shell
podman --runtime=/usr/bin/crun run \
--env=PYTORCH_HIP_ALLOC_CONF=backend:native,max_split_size_mb:128,expandable_segments:False,garbage_collection_threshold:0.5 \
--rm \
--interactive \
--tty \
--group-add=keep-groups \
--device=/dev/dri:rwm \
--device=/dev/kfd:rwm \
localhost/rocm:latest
```
`--env=PYTORCH_HIP_ALLOC_CONF=...`
Pytorch settings that will reduce the amount of crashes with big AI models.
[More info](https://pytorch.org/docs/stable/notes/cuda.html)
`--env=HSA_OVERRIDE_GFX_VERSION=10.3.0`
This specific to your gpu and only necessary for older gpus (6700XT and under I believe).
`-runtime=/usr/bin/crun` `--group-add=keep-groups` `--device=/dev/dri:rwm` `--device=/dev/kfd:rwm`
Required device permissions.
You will need to add yourself in the `video` and `render` groups on your host machine and create them if they don't exist.
If SELinux is running on your host machine, you will need to enable the `container_use_devices` boolean for rootless containers.
`--ipc=host` `--cap-add=SYS_PTRACE`
AMD documentation specifies to use these, but I have never needed them for any AI applications.
> [!NOTE]
> This command will only bring up a ROCMm container bash command line interface.
> You will need to attach a `--volume=` and run a custom `entrypoint.sh` to make it useful.