pyro-ppl / numpyro / 27173159644

92%

Build Type

push

github

Committed by web-flow

Commit Message

Fix enable_validation (#2201)

* fix validation context

* set default to False

* set default to True

* empty

* Revert "set default to True"

This reverts commit c66e2fbb0.

* Reapply "set default to True"

This reverts commit 955d50f63.

* fix test_get_dependencies

* fix categorical_probs

* better params and disable validation for contributed examples

* ruff format

* Fix tests for validation enabled by default

With validation defaulting to True, fix the remaining failures using valid
parameters where possible, and skip validation only where a test or library
path intentionally uses invalid / off-manifold values:

- test_distribution_constraints: build out-of-bounds params under
  validation_enabled(False) (the kwarg does not reach compound dists' internals).
- test_log_prob_gradient / test_sample_gradient: scope validation off around the
  finite-difference reference, which perturbs params off their constraint manifold.
- AutoLaplaceApproximation.get_posterior: build its MVN with validate_args=False
  (singular-Hessian fallback intentionally zeroes scale_tril).
- EulerMaruyama / GaussianRandomWalk log_prob: internal Normals built from the
  input value use validate_args=False; the public log_prob still validates value.
- HSGP centered approximation: Normal(0, spd) with validate_args=False since spd
  underflows to 0 for high-frequency basis functions.
- compat.param: project a concrete constraint-violating init onto its constraint
  (matches Pyro semantics), fixing the unnormalized-simplex pyroapi case.
- enum_elbo: fix unnormalized guide_probs_a; drop validation_enabled(False) wraps.
- HSGP tests: drop validation_enabled(False) wraps.
- test_subposterior_structure: use non-degenerate subposteriors.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* Stabilize test_functional_map convergence

pmap-NUTS landed just over rtol=0.06 on the mean (0.937 vs 1.0) because pmap
and v... (continued)

Coverage Stats

5 of 5 new or added lines in 4 files covered. (100.0%)

1 existing line in 1 file now uncovered.

14404 of 15639 relevant lines covered (92.1%)

1.68 hits per line