Each version of torch is only available for specific versions of CUDA and ROCm.
The Invoke installer and dockerfile try to install torch 2.4.1 with ROCm 5.6
support, which does not exist. As a result, the installation falls back to the
default CUDA version so AMD GPUs aren't detected. This commits fixes that by
bumping the ROCm version to 6.1, as suggested by the PyTorch documentation. [1]
The specified CUDA version of 12.4 is still correct according to [1] so it does
need to be changed.
Closes#7006Closes#7146
[1]: https://pytorch.org/get-started/previous-versions/#v241
With these changes, the Docker image can be built and executed
successfully on hosts with AMD devices with ROCm acceleration.
Previously, a ROCm-enabled version of torch would be installed, but
later removed during installation of InvokeAI itself. This was caused by
InvokeAI needing a newer torch version than was previously installed.
The fix consists of multiple components:
* Update the hardcoded versions of torch and torchvision to the versions
currently used in pyproject.toml, so that a new version need not be
installed during installation of InvokeAI.
* Specify --extra-index-url on installation of InvokeAI so that even if
a verison mismatch occurs, the correct torch version should still be
installed. This also necessitates changing --index-url to
--extra-index-url for the Torch repo. Otherwise non-torch dependencies
would not be found.
* In run.sh, build the image for the selected service.
* ci: add docker build timout; log free space on runner before and after build
* docker: bump frontend builder to node=20.x; skip linting on build
* chore: gitignore .pnpm-store
* update code owners for docker and CI
---------
Co-authored-by: Millun Atluri <Millu@users.noreply.github.com>
CONTAINER_UID is used for the user ID within the container, however I noticed the UID was hard coded to 1000 in the Dockerfile chown -R command.
This leaves the default as 1000, but allows it to be overrriden by setting CONTAINER_UID.
- rewrite Dockerfile
- add a stage to build the UI
- add docker-compose.yml
- add docker-entrypoint.sh such that any command may be used at runtime
- docker-compose adds .env support - add a sample .env file
- dont build frontend since complications with QEMU
- set pip cache dir
- add pip cache to all pip related build steps
- dont lock pip cache
- update dockerignore to exclude uneeded files
env.sh:
- move check for torch to CONVTAINER_FLAVOR detection
Dockerfile
- only mount `/var/cache/apt` for apt related steps
- remove `docker-clean` from `/etc/apt/apt.conf.d` for BuildKit cache
- remove apt-get clean for BuildKit cache
- only copy frontend to frontend-builder
- mount `/usr/local/share/.cache/yarn` in frountend-builder
- separate steps for yarn install and yarn build
- build pytorch in pyproject-builder
build.sh
- prepare for installation with extras
* remove non maintained Dockerfile
* adapt Docker related files to latest changes
- also build the frontend when building the image
- skip user response if INVOKE_MODEL_RECONFIGURE is set
- split INVOKE_MODEL_RECONFIGURE to support more than one argument
* rename `docker-build` dir to `docker`
* update build-container.yml
- rename image to invokeai
- add cpu flavor
- add metadata to build summary
- enable caching
- remove build-cloud-img.yml
* fix yarn cache path, link copyjob
