Report torch/torchvision versions and binary sizes for every validated build

[atalman]

Every validated build now reports what it installed and how big it is.

Why

check_wheel_size exists to enforce an 850 MB ceiling, and its scope is deliberately narrow: linux and linux-aarch64 pip wheels only, skipping ROCm and libtorch. That means windows, macos and ROCm builds report no size at all, and even where the check does run the number only appears as an annotation.

What this adds

A write_build_report step after the smoke tests, on every OS, that never fails the job:

--- Build report
| field | value |
| --- | --- |
| build | linux / py3.12 / cuda 12.8 |
| torch | 2.10.0.dev20260824+cu128 |
| torch wheel | 812.4 MB |
| torch installed | 2143.7 MB |
| CUDA | 12.8 |
| cuDNN | 9.24.0 |
| NCCL | 2.30.7 |
| torchvision | 0.25.0.dev20260824+cu128 |
| torchvision wheel | 8.1 MB |
| torchvision installed | 24.3 MB |

Two sizes, because they answer different questions:

• wheel -- compressed download size, what users actually pull from the index. Read off pip's Downloading/Using cached line, so it is - on the uv/wheel-variants path and when the wheel was already satisfied.
• installed -- unpacked bytes on disk, walked from the installed package, so it is available on every install path and every OS.

Versions are read from the installed packages rather than from the matrix, so a mismatch between what was requested and what pip resolved shows up in the report.

The report goes to stdout always, and is appended to the job summary when the runner exposes a writable one -- it is not reachable from inside the validation container, so that append is guarded rather than assumed.

Refactor

The size parsing is factored out of check_wheel_size into parse_wheel_size_mb <log> <dist> and reused for both torch and torchvision. check_wheel_size keeps its current scope, threshold and behaviour; only the parsing moved.

Not covered

• libtorch builds return before this point (they go through validate_libtorch.py), so they still report nothing. Worth a follow-up if libtorch size matters.
• The report is per job. If a single aggregated table across all ~170 builds is wanted, that needs an artifact per job plus a final aggregation job -- deliberately left out of this change.

Test plan

• bash -n clean.
• parse_wheel_size_mb unit-checked against a captured pip log: torch -> 812.4, torchvision -> 8.1, absent distribution -> empty (torch-[0-9] does not match torchvision-/torchaudio-).
• cuDNN integer decoding checked: 91002 -> 9.10.2, 92400 -> 9.24.0, 91701 -> 9.17.1, matching the values published in pytorch RELEASE.md.
• write_build_report with no torch installed renders import failed: ... and exits 0 rather than aborting the job.
• With GITHUB_STEP_SUMMARY pointing at an unwritable path, the function emits zero bytes on stderr and still exits 0.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
torchci	Ignored		Aug 25, 2026 1:33am

[jeanschmidt]

would be better to have this as a standalone script, so lint and other rules would apply and it would be easier to maintain the quality.

[atalman]

Agreed, done in 3edf8f3 — it is now .github/scripts/build_report.py.

That puts it under FLAKE8 and PYFMT (**/*.py) and MYPY (.github/scripts/**/*.py). mypy.ini gets one exemption, following the existing per-module sections in that file:

[mypy-.github.scripts.build_report]
# torch/torchvision are not installed in the lint environment; the script
# imports them defensively at runtime and degrades when they are absent.
disable_error_code = import-not-found, import-untyped

One bonus that only became clear while moving it: the heredoc version had to cd out of the repo before running python, because python - puts the cwd on sys.path and the validation runs from the pytorch/pytorch checkout — so import torch could pick up the source tree instead of the installed package. A script file makes sys.path[0] the script's own directory, so that hazard is gone and the cd with it.

validate_binaries.sh is 73 lines lighter; the shell function now just invokes the script and handles the job-summary append.