Integrate AMD GPU in CI/CD environment #26007
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
| @@ -0,0 +1,29 @@ | |||
| FROM rocm/pytorch:rocm5.6_ubuntu20.04_py3.8_pytorch_2.0.1 | |||
There was a problem hiding this comment.
There should be a workflow file that build this image. So far I don't see it, and I am wondering where is transformers-pytorch-latest-amdgpu-push-ci created/pushed? Do you do this manually somewhere?
There was a problem hiding this comment.
This image is built and published by AMD directly on the Docker Hub.
I can add a job in the build-docker-images.yml workflow to build transformers-pytorch-latest-amdgpu-push-ci.
It has been pushed to our Docker Hub organisation by myself 🙂
fix dockerfile Co-authored-by: Felix Marty <felix@hf.co>
AMDGPU CI lives in an other workflow
|
@ydshieh @LysandreJik I think we are in a good shape for review and merging. What we did:
What we cannot ensure as of today:
|
|
Hi @mfuntowicz Looking at the runs in https://github.com/huggingface/transformers/actions/workflows/self-push-amd.yml, you will see no test job (Model test) is being triggered (as no test is being collected). Also the slack report won't work as the tag is sitll using |
…g `amd-gpu` and `miXXX` labels.
|
@LysandreJik in case you want to take a final look :-) |
|
Merge now so @mfuntowicz can show progress to AMD team today. |
* Add a Dockerfile for PyTorch + ROCm based on official AMD released artifact * Add a new artifact single-amdgpu testing on main * Attempt to test the workflow without merging. * Changed BERT to check if things are triggered * Meet the dependencies graph on workflow * Revert BERT changes * Add check_runners_amdgpu to correctly mount and check availability * Rename setup to setup_gpu for CUDA and add setup_amdgpu for AMD * Fix all the needs.setup -> needs.setup_[gpu|amdgpu] dependencies * Fix setup dependency graph to use check_runner_amdgpu * Let's do the runner status check only on AMDGPU target * Update the Dockerfile.amd to put ourselves in / rather than /var/lib * Restore the whole setup for CUDA too. * Let's redisable them * Change BERT to trigger tests * Restore BERT * Add torchaudio with rocm 5.6 to AMD Dockerfile (huggingface#26050) fix dockerfile Co-authored-by: Felix Marty <felix@hf.co> * Place AMD GPU tests in a separate workflow (correct branch) (huggingface#26105) AMDGPU CI lives in an other workflow * Fix invalid job name is dependencies. * Remove tests multi-amdgpu for now. * Use single-amdgpu * Use --net=host for now. * Remote host networking. * Removed duplicated check_runners_amdgpu step * Let's tag machine-types with mi210 for now. * Machine type should be only mi210 * Remove unnecessary push.branches item * Apply review suggestions moving from `x-amdgpu` to `x-gpu` introducing `amd-gpu` and `miXXX` labels. * Remove amdgpu from step names. * finalize * delete --------- Co-authored-by: fxmarty <9808326+fxmarty@users.noreply.github.com> Co-authored-by: Felix Marty <felix@hf.co> Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
No description provided.