Add portable Docker build workflow using public TheRock wheels#3149
Open
ethanwee1 wants to merge 28 commits intoROCm:developfrom
Open
Add portable Docker build workflow using public TheRock wheels#3149ethanwee1 wants to merge 28 commits intoROCm:developfrom
ethanwee1 wants to merge 28 commits intoROCm:developfrom
Conversation
Bring in the Dockerfile, install scripts, and workflow from rocm-npi-dev adapted for the public ROCm nightlies index at rocm.nightlies.amd.com. Pushes images to rocmshared/pytorch-private on DockerHub.
Adds a temporary push trigger scoped to add-portable-docker-build so the workflow can be tested on a fork before merge. Also resolves all inputs through a defaults step so push-triggered runs use sensible values (gfx94X-dcgpu, ROCm 7.13.0a20260413, Python 3.12).
Public nightlies use git-hash local versions for triton (e.g. 3.7.0+git20a46016.rocm7.13.0a20260413) that pip cannot match via ==. Let torch's dependency resolver pull the correct triton instead.
Runs test_binary_ufuncs (CPU-only) inside the built image to validate the PyTorch installation before pushing to Docker Hub.
|
Jenkins build for 0d1061bf0ad4f40123d7f76ece2d4f72969221f0 commit finished as NOT_BUILT |
The previous version piped through tail which swallowed the exit code, and pytest was not installed in the image.
…rsion
Tag format: pytorch-{branch}-{commit}-rocm{version}-{os}-py{python}-{gfx}
If the requested PYTHON_VERSION is not the system default (3.12 on Ubuntu 24.04), install it from the deadsnakes PPA before creating the venv.
|
Jenkins build for 0d1061bf0ad4f40123d7f76ece2d4f72969221f0 commit finished as NOT_BUILT |
|
Jenkins build for 3360f8d4c3252d9e9d9b7ee667f040b7e41f6666 commit finished as NOT_BUILT |
- Nightly schedule (daily 06:00 UTC) builds 4 images: pytorch/pytorch main (nightly), ROCm/pytorch release/2.11, 2.10, 2.9 - ROCm version is no longer hardcoded; discovered from the latest available torch wheel on the staging index for each matrix entry - Manual dispatch also auto-discovers ROCm version when left empty - Added --torch-version-prefix support to install_pytorch_wheels.py for filtering wheels by major.minor version - Dockerfile supports Python 3.10-3.14 via deadsnakes PPA
- schedule (cron) → 4-build nightly matrix - push → single build with defaults - workflow_dispatch → single build with user inputs
- Added pytorch_repo input (default: pytorch/pytorch) - Changed pytorch_branch default from develop to main - Nightly matrix unchanged (already uses correct repos per entry)
- Repo and branch first, followed by python, gfx, rocm, torch prefix, index URL - Removed rarely-used inputs (exact versions, base_image) that auto-discovery handles
- Removed torch_version_prefix input - Prefix is now extracted from the branch: release/2.11 → "2.11", nightly → empty - Applies to both nightly matrix and manual dispatch
|
Jenkins build for 14ce620d192c6a461ca44d38410f1b8cf1d78dc5 commit finished as NOT_BUILT |
The key function was only comparing the base torch version (e.g. 2.9.1) and ignoring the ROCm version after +. This caused it to pick old 2025 wheels over latest 2026 ones when the base versions were identical.
|
Jenkins build for 4b63c1c933dfad75aae068aa83ba98144bb28373 commit finished as NOT_BUILT |
…ABI mismatches torch import is required; torchaudio, torchvision, and triton are best-effort — logged but won't fail the build if there's a version mismatch on older release branches.
In auto-discovery mode, install torch and rocm[devel] with exact versions first, then install torchaudio/torchvision/triton unpinned so pip picks versions compatible with the installed torch.
|
Jenkins build for 4b63c1c933dfad75aae068aa83ba98144bb28373 commit finished as FAILURE |
ethanwee1
force-pushed
the
add-portable-docker-build
branch
from
d70128b to
4b63c1c
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Copied from https://github.com/AMD-ROCm-Internal/rocm-npi-dev/actions/workflows/build_portable_linux_pytorch_dockers.yml
Latest run and docker generated: docker.io/rocm/pytorch-private:pytorch-nightly-f8d08404-rocm7.13.0a20260413-ubuntu24.04-py3.12-gfx950-dcgpu
https://github.com/ethanwee1/pytorch/actions/runs/24362482659