Upgrading FL4Health Libraries#558
Conversation
Bumps [codecov/codecov-action](https://github.com/codecov/codecov-action) from 5 to 6. - [Release notes](https://github.com/codecov/codecov-action/releases) - [Changelog](https://github.com/codecov/codecov-action/blob/main/CHANGELOG.md) - [Commits](codecov/codecov-action@v5...v6) --- updated-dependencies: - dependency-name: codecov/codecov-action dependency-version: '6' dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com>
Bumps [pypa/gh-action-pypi-publish](https://github.com/pypa/gh-action-pypi-publish) from 1.13.0 to 1.14.0. - [Release notes](https://github.com/pypa/gh-action-pypi-publish/releases) - [Commits](pypa/gh-action-pypi-publish@ed0c539...cef2210) --- updated-dependencies: - dependency-name: pypa/gh-action-pypi-publish dependency-version: 1.14.0 dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com>
Bumps [astral-sh/setup-uv](https://github.com/astral-sh/setup-uv) from 7.6.0 to 8.1.0. - [Release notes](https://github.com/astral-sh/setup-uv/releases) - [Commits](astral-sh/setup-uv@v7.6.0...v8.1.0) --- updated-dependencies: - dependency-name: astral-sh/setup-uv dependency-version: 8.1.0 dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com>
updates: - [github.com/astral-sh/uv-pre-commit: 0.10.12 → 0.11.13](astral-sh/uv-pre-commit@0.10.12...0.11.13) - [github.com/astral-sh/ruff-pre-commit: v0.15.7 → v0.15.12](astral-sh/ruff-pre-commit@v0.15.7...v0.15.12) - [github.com/pre-commit/mirrors-mypy: v1.19.1 → v2.0.0](pre-commit/mirrors-mypy@v1.19.1...v2.0.0)
| super().__init__(optimizer, -1, False) | ||
| super().__init__(optimizer, -1) | ||
|
|
||
| # mypy incorrectly infers get_lr returns a float |
There was a problem hiding this comment.
This has been fixed 🙂
| self.num_windows = ceil(max_steps / self.steps_per_lr) | ||
| self._step_count: int | ||
| super().__init__(optimizer, -1, False) | ||
| super().__init__(optimizer, -1) |
There was a problem hiding this comment.
Torch no longer supports the verbose argument for schedulers (this is similar to the issue that was plaguing the nnunet schedulers).
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #558 +/- ##
==========================================
+ Coverage 79.65% 79.71% +0.05%
==========================================
Files 166 166
Lines 10554 10573 +19
==========================================
+ Hits 8407 8428 +21
+ Misses 2147 2145 -2 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
[pre-commit.ci] pre-commit autoupdate
…odecov/codecov-action-6
…s/codecov/codecov-action-6 Bump codecov/codecov-action from 5 to 6
…ypa/gh-action-pypi-publish-1.14.0
…s/pypa/gh-action-pypi-publish-1.14.0 Bump pypa/gh-action-pypi-publish from 1.13.0 to 1.14.0
…stral-sh/setup-uv-8.1.0
…s/astral-sh/setup-uv-8.1.0 Bump astral-sh/setup-uv from 7.6.0 to 8.1.0
| GHSA-rf74-v2fm-23pw | ||
| CVE-2026-33230 | ||
| CVE-2026-33231 | ||
| CVE-2026-34073 |
lotif
left a comment
lotif
left a comment
There was a problem hiding this comment.
Approving with a couple of minor comments :)
| # of the class signature) into the param groups called "decoupled_weight_decay" which cases an error in the | ||
| # kwargs below. See: https://github.com/pytorch/pytorch/blob/v2.11.0/torch/optim/adamw.py#L57 | ||
| if optim_class == torch.optim.AdamW: | ||
| optimizer_kwargs.pop("decoupled_weight_decay") |
There was a problem hiding this comment.
Weird that we need a workaround for their workaround. What is the case where this happens?
There was a problem hiding this comment.
Basically, they got rid of this in the constructor, but if you follow the link, you'll see they "inject" it into the dictionary. To make this mixin work, Andrei was reconstructing the optimizer arguments to create a new one from the old one. The injection means that there is a key corresponding to the argument that used to be called decoupled_weight_decay. So it gets into the kwargs which are passed to the constructor which no longer wants decoupled_weight_decay.
The way Andrei was doing this is fairly brittle, but I don't want to mess with it much in this PR.
There was a problem hiding this comment.
Yeah, sounds a little hacky, but I agree with you about not messing with it here.
| "flwr>=1.18.0,<1.19.0", | ||
| "opacus>=1.3.0,<2.0.0", | ||
| "torch==2.6.0", | ||
| "torch>=2.6.0", |
2 participants
PR Type
Other
Short Description
Updating the libraries in the FL4Health library. In particular, this updates many of the libraries with vulnerabilities, but also finally upgrading Torch and Tensorflow with the HPC upgrades.
Note: This required a few small workarounds. One of them is a monkey patch for nnunet, of which both Marcelo and I aren't huge fans, but it seems as though they are not planning to move to newer torch implementations soon. So it felt like the least invasive approach at the moment.
Tests Added
None added but made sure the unit and smoke tests all pass locally and on the cluster.