[pingsutw]

Tracking issue

N/A — master CI has been red since the last green run on 2026-10-02 (run 37051324786) due to upstream releases, not repo changes.

Why are the changes needed?

Every Build run on master currently fails for several independent upstream reasons:

1. pyarrow 26.0.0 (released 2026-10-09) requires numpy>=2. dev-requirements.in leaves pyarrow unpinned, while tensorflow<=2.15.1 forces numpy 1.26.4. The test-serialization, build-with-pandas and both integration jobs die at collection time with:

   ImportError: pyarrow requires NumPy 2.0 or newer, found 1.26.4
   

2. torch>=2.6 defaults torch.load(weights_only=True), which refuses to unpickle nn.Module objects. All four build-with-extras jobs fail test_native.py::test_wf and test_transformations.py::test_to_python_value_and_literal[...PyTorchModule...] with:

   _pickle.UnpicklingError: Weights only load failed. ... Unsupported global: torch.nn.modules.linear.Linear
   

3. pandas 3.x wheels need numpy>=2.0.2 at import time, but the build-plugins job installs numpy<2.0.0 for the onnx, sqlalchemy and pandera plugins (and whylogs pins numpy<2.0 itself). build-plugins (3.12, ...) for whylogs, pandera, onnx-scikitlearn fail with:

   ImportError: Please upgrade numpy to >= 2.0.2 to use this pandas version.
   

   sqlalchemy fails the same way, surfaced as '_LazyModule' object has no attribute 'DataFrame' because pandas is lazily imported.
4. leptonai 0.47 (released 2026-10-20) removed leptonai.api.v1, which the dgxc-lepton connector imports:

   ModuleNotFoundError: No module named 'leptonai.api.v1'
   

What changes were proposed in this pull request?

• dev-requirements.in: pin pyarrow<26 with a comment explaining the tensorflow/numpy conflict. Resolution now picks pyarrow 25.0.1 + numpy 1.26.4 + tensorflow 2.15.1.
• flytekit/extras/pytorch/native.py: add a TORCH_LOAD_KWARGS class attribute on PyTorchTypeTransformer (empty by default) that is forwarded to torch.load. Only PyTorchModuleTransformer sets {"weights_only": False}, since it pickles a whole nn.Module. The kwarg is only passed when torch supports it (torch<1.13 on py3.10 does not). PyTorchTensorTransformer and PyTorchCheckpointTransformer keep torch's safe default: tensors, state dicts and plain hyperparameter values all load fine with weights_only=True. This restores the pre-torch-2.6 behavior for the module transformer only.
• plugins/flytekit-{whylogs,pandera,sqlalchemy,onnx-scikitlearn}/dev-requirements.in: pin pandas<3, following the existing scipy<1.18 workaround in onnx-scikitlearn for the same CI numpy<2 forcing.
• plugins/flytekit-dgxc-lepton/setup.py: pin leptonai<0.47.

Follow-up worth considering (not done here, since it is a CI config change): the numpy<2.0.0 forcing in pythonbuild.yml is commented as an onnx workaround but also applies to sqlalchemy and pandera, which likely no longer need it.

How was this patch tested?

Reproduced locally in an arm64 Python 3.12 venv:

# torch 2.14.1 (what CI installs)
pytest tests/flytekit/unit/extras/pytorch -q   # master: 2 failed, 11 passed; branch: 13 passed

# pandas 2.3.3 + numpy 1.26.4 + pyarrow 25.0.1 (what the pins resolve to)
cd plugins/flytekit-sqlalchemy && pytest tests -q   # 11 passed

# leptonai 0.27.3
cd plugins/flytekit-dgxc-lepton && pytest tests -q  # 23 passed

uv pip install --dry-run --python-version 3.11 -r dev-requirements.in resolves to pyarrow 25.0.1 / numpy 1.26.4 / tensorflow 2.15.1. Dry-run resolution of each pinned plugin with numpy<2.0.0 picks pandas 2.3.3.

No new tests: the existing tests that were failing are the regression tests for these changes.

Setup process

Screenshots

Check all the applicable boxes

• I updated the documentation accordingly.
• All new and existing tests passed.
• All commits are signed-off.

Related PRs

Docs link

[codecov]

Codecov Report

❌ Patch coverage is 0% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 47.77%. Comparing base (3fbaf5e) to head (d6614ed).

Files with missing lines	Patch %	Lines
flytekit/extras/pytorch/native.py	0.00%	4 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (3fbaf5e) and HEAD (d6614ed). Click for more details.

HEAD has 9 uploads less than BASE

Flag	BASE (3fbaf5e)	HEAD (d6614ed)
	10	1

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #3477       +/-   ##
===========================================
- Coverage   80.83%   47.77%   -33.06%     
===========================================
  Files         217      217               
  Lines       22960    22963        +3     
  Branches     3038     3037        -1     
===========================================
- Hits        18559    10970     -7589     
- Misses       3547    11385     +7838     
+ Partials      854      608      -246     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.