release-go-live-binary-build-matrix
Update tools/scripts/generate_binary_build_matrix.py when a PyTorch release goes live. Advances CURRENT_STABLE_VERSION to the new stable, promotes the…
npx skills add https://github.com/pytorch/test-infra --skill release-go-live-binary-build-matrixRelease Go-Live: Binary Build Matrix
Updates tools/scripts/generate_binary_build_matrix.py and the associated test fixtures when a PyTorch release goes live. This mirrors prior PRs such as:
Release 2.10 go live. Update release matrix (#7668)Release 2.11 promotion script changes (#7868)Release 2.12 go live. Update release matrix (#8070)
When to use this skill
Use when the user asks to:
- Advance stable version to X.Y.0 in the binary build matrix
- Promote release CUDA/ROCm arches now that release X.Y is going live
- Update
generate_binary_build_matrix.pyfor a new PyTorch release - Do "release go live" changes
Target files
| File | Why |
|---|---|
tools/scripts/generate_binary_build_matrix.py | Source of truth for version + channel arches |
tools/tests/assets/build_matrix_*.json | Snapshot fixtures (compact JSON, must match script output) |
tools/tests/update_test_assets.sh | Regeneration script (canonical) |
tools/tests/test_generate_binary_build_matrix.py | Test runner |
Instructions
Step 1: Determine the new stable version
Ask the user (if not clear) what version is going live (e.g. 2.12.0). The current state usually looks like:
CURRENT_NIGHTLY_VERSION = "2.13.0"
CURRENT_CANDIDATE_VERSION = "2.12.0" # about to become stable
CURRENT_STABLE_VERSION = "2.11.0" # being replaced
So a "2.12 go live" advances CURRENT_STABLE_VERSION to match CURRENT_CANDIDATE_VERSION. Do NOT touch CURRENT_NIGHTLY_VERSION or CURRENT_CANDIDATE_VERSION in this PR — those are advanced by separate PRs (e.g. Update nightly version to X.Y.Z, [Release X.Y] advance candidate version).
Step 2: Advance CURRENT_STABLE_VERSION
In tools/scripts/generate_binary_build_matrix.py, change the single line:
CURRENT_STABLE_VERSION = "<old>"
to the new stable version.
Step 3: Promote release-channel arches to match test
Compare the three channels in each arches dict:
CUDA_ARCHES_DICT = {
"nightly": [...],
"test": [...],
"release": [...], # promote this to match "test"
}
ROCM_ARCHES_DICT = {
"nightly": [...],
"test": [...],
"release": [...], # promote this to match "test"
}
If release differs from test, update release to match test. This is the actual "what CUDA/ROCm versions ship with this release" decision — the candidate-channel arches that survived the release cycle become the official release arches.
Also check STABLE_CUDA_VERSIONS — usually already aligned across channels, but verify all three entries point at the same default for the new release.
Step 4: Regenerate test fixtures (MUST use the shell script)
The fixtures are compact (single-line) JSON. The script tools/tests/update_test_assets.sh is the canonical regenerator. Do NOT use python3 -m tools.tests.test_generate_binary_build_matrix --update-reference-files — that pretty-prints with indent=2 and produces a massive churn diff that does not match the file format on main.
Run from the repo root (note the script uses relative paths so cd matters):
cd tools/tests
python3 ../scripts/generate_binary_build_matrix.py --build-python-only disable --with-xpu disable > assets/build_matrix_linux_wheel_cuda.json
python3 ../scripts/generate_binary_build_matrix.py --build-python-only disable --with-rocm disable --with-xpu disable > assets/build_matrix_linux_wheel_cuda_norocm.json
python3 ../scripts/generate_binary_build_matrix.py --build-python-only disable --with-cpu disable --with-xpu disable > assets/build_matrix_linux_wheel_nocpu.json
python3 ../scripts/generate_binary_build_matrix.py --build-python-only disable --with-cpu disable --with-rocm disable --with-xpu enable > assets/build_matrix_linux_wheel_xpu.json
python3 ../scripts/generate_binary_build_matrix.py --build-python-only disable --operating-system="macos" --with-cuda disable --with-rocm disable > assets/build_matrix_macos_wheel.json
python3 ../scripts/generate_binary_build_matrix.py --build-python-only disable --operating-system="windows" > assets/build_matrix_windows_wheel_cuda.json
python3 ../scripts/generate_binary_build_matrix.py --build-python-only disable --with-rocm disable --with-cuda disable --operating-system="windows" > assets/build_matrix_windows_wheel_xpu.json
Skip the conda lines from update_test_assets.sh — --package-type conda is broken on main (noted in PR #8065) and the conda fixtures are no longer maintained.
Step 5: Run tests
python3 -m tools.tests.test_generate_binary_build_matrix
Expected: Ran 7 tests in <time>s OK.
Step 6: Verify the diff is minimal
git diff --stat should show roughly:
tools/scripts/generate_binary_build_matrix.py: ~4 lines (stable version + release arches list)- 7 fixture JSON files: 1 line each (single-line files, change is in-place)
Sanity check: the only diff in fixtures should be "stable_version": "<old>" → "stable_version": "<new>". If a fixture has multi-line / pretty-printed diff, you used the wrong regenerator — revert with git checkout upstream/main -- tools/tests/assets/ and re-run Step 4.
If the CUDA release arches changed, expect new entries in fixtures (e.g. new cu132 rows when adding 13.2 to release).
Step 7: Commit and PR
Suggested commit/PR title pattern (matches history):
Release X.Y go live. Update release matrix
PR body should call out:
- Advance
CURRENT_STABLE_VERSIONfrom<old>to<new> - (If applicable) Advance
CUDA_ARCHES_DICT["release"]/ROCM_ARCHES_DICT["release"] - Regenerate test fixtures via
tools/tests/update_test_assets.sh
Common pitfalls
- Don't pretty-print fixtures. The
--update-reference-filestest flag writesindent=2. The fixtures onmainare single-line. Always use the shell script approach. - Don't touch nightly/candidate versions. Those are advanced by separate PRs in the release cycle.
upstreamispytorch/test-infra. Branch fromupstream/main, push toorigin(your fork), open the PR againstpytorch/test-infra:main.- Stash unrelated WIP.
generate_binary_build_matrix.pychanges should go in a clean branch offupstream/main— don't pile them onto an unrelated feature branch.