Run whole suite with CVM and Numba linkers #7993

ricardoV94 · 2025-12-14T11:35:22Z

Related to #7926
Related to PyMC V6 roadmap

Relies on next PyTensor release

PyMC will now be tested on both C and Numba backends, antecipating the switch to Numba as the default backend from PyTensor

codecov · 2025-12-21T13:24:45Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.45%. Comparing base (056e80c) to head (cd3983e).
⚠️ Report is 10 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #7993   +/-   ##
=======================================
  Coverage   91.44%   91.45%           
=======================================
  Files         116      116           
  Lines       19002    19025   +23     
=======================================
+ Hits        17377    17399   +22     
- Misses       1625     1626    +1

Files with missing lines	Coverage Δ
pymc/dims/distributions/scalar.py	`95.08% <100.00%> (+0.21%)`	⬆️
pymc/distributions/continuous.py	`98.26% <100.00%> (+0.01%)`	⬆️
pymc/testing.py	`90.64% <100.00%> (+0.09%)`	⬆️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

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

ricardoV94 · 2026-01-03T02:49:41Z

Seems like only one meaningful test failure left: https://github.com/pymc-devs/pymc/actions/runs/20662736046/job/59344923930?pr=7993#step:5:1066

Less noisy when doing interactive debugging

Remove useless os paramerizations and update codecov name

Most warnings are filtered by PyTensor but pytest reenables them

ricardoV94 · 2026-01-05T14:31:55Z

Tests are passing, CI didn't blow up, although it's a tad slower. This is the slowest one: https://github.com/pymc-devs/pymc/actions/runs/20717356014/job/59471703297?pr=7993#step:5:994

We may want to tweak the first two which take 2m (8% of the whole CI time)

92.50s call     tests/logprob/test_transform_value.py::test_weakref_leak
35.06s call     tests/logprob/test_transform_value.py::test_default_value_transform_logprob[pt_dist17-dist_params17-<lambda>-size17]

And these, as the CI had a large delta from the one in CVM: https://github.com/pymc-devs/pymc/actions/runs/20717356014/job/59471703228?pr=7993#step:5:1251

57.98s call     tests/distributions/test_multivariate.py::TestLKJCorr::test_distribution
32.25s call     tests/distributions/test_continuous.py::TestMatchesScipy::test_interpolated
26.81s call     tests/distributions/test_multivariate.py::TestMatchesScipy::test_mvnormal_indef
22.84s call     tests/distributions/test_multivariate.py::TestLKJCholeskyCov::test_distribution
22.21s call     tests/distributions/test_multivariate.py::TestLKJCholeskCov::test_dist
21.85s call     tests/distributions/test_continuous.py::TestInterpolated::test_distribution

Finally, this is a tad too much for a progress_bar

34.92s call     tests/test_progress_bar.py::test_progressbar_nested_compoundtest

Copying comment to #7686

jessegrabowski · 2026-01-05T14:34:07Z

The performance of the linalg distributions (lkj, mvn) should improve a bit after we cut a release with the numba caching

ricardoV94 · 2026-01-05T14:36:39Z

The performance of the linalg distributions (lkj, mvn) should improve a bit after we cut a release with the numba caching

Yup. I thought we already released it, so that's good news. Also pymc-devs/pytensor#1821 as those tend up to show in our pdf/cdfs

jessegrabowski · 2026-01-05T14:51:24Z

tests/distributions/test_multivariate.py

    expected_rv_op_params = {
        "mu": np.array([1.0, 2.0]),
-        "cov": quaddist_matrix(chol=pymc_dist_params["chol"]).eval(),
+        "cov": quaddist_matrix(chol=pymc_dist_params["chol"]).eval(mode="FAST_COMPILE"),


Why does this mode need to be specified? Does it work if we're in numba mode?

it does, but bothered me during interactive debugging, because we end up triggering the slow-compilation codebase for a simple constant-fold

jessegrabowski · 2026-01-05T14:51:51Z

pymc/testing.py


+    def check_pymc_draws_match_reference_not_numba(self):
+        # This calls `check_pymc_draws_match_reference` but only if the default linker is NOT numba.
+        # It's used when the draws aren't expected to match in that backend.


Why aren't draws expected to match in numba? Some rng state issue?

Not an issue, we just don't copy the exact PRNG algorithm that scipy/numpy does for these specific distributions. If JAX were the default backend we wouldn't expect for them to ever match.

This is an exact test, not a distributional one

jessegrabowski · 2026-01-05T14:55:29Z

tests/model/test_core.py

+            assert "This does not respect one of the following constraints: sigma > 0" in out
+        else:  # "random"
+            if isinstance(get_default_mode().linker, NumbaLinker):
+                # Numba doesn't raise for negative sigma in NormalRV


Maybe, didn't want to block on that decision though

pymc-devs/pytensor#1823

ricardoV94 added this to the v6 milestone Dec 14, 2025

ricardoV94 added pytensor numba labels Dec 14, 2025

ricardoV94 force-pushed the pytensor_numba branch from 11964ce to ab53d82 Compare December 14, 2025 11:53

ricardoV94 changed the title ~~Run whole suite with numba linker~~ Run whole suite with C and Numba linkers Dec 14, 2025

ricardoV94 force-pushed the pytensor_numba branch 2 times, most recently from 2e3466e to 82ee1d9 Compare December 21, 2025 13:06

ricardoV94 mentioned this pull request Dec 21, 2025

Allow numba cache path to create parent directory pymc-devs/pytensor#1798

Merged

ricardoV94 changed the title ~~Run whole suite with C and Numba linkers~~ Run whole suite with CVM and Numba linkers Dec 21, 2025

ricardoV94 force-pushed the pytensor_numba branch 3 times, most recently from a3f98e2 to a471e75 Compare December 23, 2025 23:18

This was referenced Dec 26, 2025

Numba Dirichlet: Fix case with discrete alpha parameter pymc-devs/pytensor#1801

Merged

Numba RVs: Include output_bc_pattern in cache key pymc-devs/pytensor#1804

Merged

ricardoV94 force-pushed the pytensor_numba branch from a471e75 to 93a552f Compare December 27, 2025 08:56

This was referenced Jan 2, 2026

Clone inner graph ops pymc-devs/pytensor#1809

Merged

Consider numba fastmath config in cache key pymc-devs/pytensor#1810

Merged

ricardoV94 force-pushed the pytensor_numba branch from 93a552f to a0b679a Compare January 2, 2026 14:57

ricardoV94 mentioned this pull request Jan 2, 2026

Fix non-inplace IfElse on numba backend pymc-devs/pytensor#1765

Merged

11 tasks

ricardoV94 force-pushed the pytensor_numba branch 2 times, most recently from b36a9d9 to 9c28ebd Compare January 2, 2026 17:09

ricardoV94 added 3 commits January 5, 2026 13:35

Make HalfCauchy a symbolic RV

bb05c80

Use FAST_COMPILE in test param

84bb8fa

Less noisy when doing interactive debugging

Add pytensor linker matrix ("cvm", "numba") to tests

9b2925b

Remove useless os paramerizations and update codecov name

ricardoV94 force-pushed the pytensor_numba branch 2 times, most recently from 4d98fe3 to c88bc7a Compare January 5, 2026 12:47

ricardoV94 added 2 commits January 5, 2026 14:46

Ignore numba warnings in tests

64cbc77

Most warnings are filtered by PyTensor but pytest reenables them

Numba RV draws don't always match exactly with numpy/scipy

b7d264f

ricardoV94 added 5 commits January 5, 2026 14:46

Test doesn't xfail in numba

638b732

Precision test requires fasthmath=False in numba backend

31ea2c9

Numba NormalRV does not raise for negative sigma

af319c5

Xfail test in numba backend

d74b1e6

Bump minor PyTensor dependency

cd3983e

ricardoV94 force-pushed the pytensor_numba branch from c88bc7a to cd3983e Compare January 5, 2026 13:46

ricardoV94 requested a review from jessegrabowski January 5, 2026 13:46

ricardoV94 marked this pull request as ready for review January 5, 2026 13:46

ricardoV94 mentioned this pull request Jan 5, 2026

Speed up test times 🚀 #7686

Open

jessegrabowski approved these changes Jan 5, 2026

View reviewed changes

ricardoV94 merged commit bbcb81d into pymc-devs:main Jan 5, 2026
42 checks passed

ricardoV94 deleted the pytensor_numba branch January 5, 2026 15:28

ricardoV94 added the no releasenotes Skipped in automatic release notes generation label Jan 5, 2026

Run whole suite with CVM and Numba linkers #7993

Run whole suite with CVM and Numba linkers #7993

Uh oh!

Conversation

ricardoV94 commented Dec 14, 2025

Uh oh!

codecov bot commented Dec 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

ricardoV94 commented Jan 3, 2026

Uh oh!

ricardoV94 commented Jan 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jessegrabowski commented Jan 5, 2026

Uh oh!

ricardoV94 commented Jan 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jessegrabowski Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

ricardoV94 Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

jessegrabowski Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

ricardoV94 Jan 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jessegrabowski Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

ricardoV94 Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

ricardoV94 Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

codecov bot commented Dec 21, 2025 •

edited

Loading

ricardoV94 commented Jan 5, 2026 •

edited

Loading

ricardoV94 commented Jan 5, 2026 •

edited

Loading

ricardoV94 Jan 5, 2026 •

edited

Loading