Fix lint formatting and mypy checks by ordabayevy · Pull Request #3453 · pyro-ppl/pyro

ordabayevy · 2026-05-25T22:25:49Z

Main Changes

CI/tooling compatibility
- Removed global CC=gcc-9 / CXX=g++-9 from GitHub Actions so PyTorch C++ extensions can use a newer runner compiler.
- Added --warn-unused-ignores to make lint.
- Updated mypy target from Python 3.9 to 3.10.
- Disabled pytest-benchmark in test-funsor because it warns under xdist, and warnings are errors.
Dependency cleanup
- Removed visdom from broad extras/test dependencies.
- Made Visdom imports lazy in AIR/VAE examples, so CI can run examples without installing Visdom unless visualization flags are used.
- Pinned funsor[torch] to the remote fixed commit in setup.py and docs/requirements.txt.
PyTorch compatibility fixes
- Added Uniform.infer_shapes() to avoid newer Torch’s property-backed arg_constraints issue.
- Replaced list-based tensor indexing with tuple indexing in stable log-prob, stats, and DCT Adam code.
- Removed the obsolete lazy_property.__call__ patch that was causing inspect/doctest issues with newer Torch.
Typing/lint cleanup
- Removed stale type: ignore comments now caught by --warn-unused-ignores.
Other cleanup
- Removed the optional gpytorch.lazy.NonLazyTensor path in SpanningTree.log_partition_function, using torch.linalg.cholesky directly.

Why

The branch mainly makes the project work with newer PyTorch/tooling: PyTorch changed arg_constraints, warns on list indexing and tensor scalar conversion, and now requires a newer C++ compiler path for extensions. It also reduces CI fragility by removing Visdom from the default test dependency set and tightening lint so stale ignores don’t accumulate.

fritzo · 2026-05-25T22:37:39Z

Thank you, Yerdos!

Preinstall visdom without build isolation to avoid pkg_resources failures on modern setuptools, require explicit context in ConditionalAutoRegressiveNN, and replace cast-based mypy suppressions with type ignores.

martinjankowiak · 2026-06-02T15:45:04Z

@ordabayevy thanks! what else needs to be done here to merge?

ordabayevy · 2026-06-02T16:37:25Z

There are unit tests that are failing due to the new PyTorch version. I will try to get them fixed today.

ordabayevy

@fritzo @martinjankowiak ready for a review!

ordabayevy · 2026-06-04T15:38:56Z

      run: |
        python -m pip install --upgrade pip wheel setuptools
        pip install ruff black mypy nbstripout nbformat
+        pip install -e .


Added this so that the type checker can get dependency libraries stubs.

ordabayevy · 2026-06-04T15:39:34Z

 pyro-api>=0.1.1
 tqdm>=4.36
-funsor[torch]
+funsor[torch] @ git+https://github.com/pyro-ppl/funsor.git@8a8ca81fcc61733ad82e163028c7ff78420df9f0


Will be pinned to a released funsor version before pyro release.

ordabayevy · 2026-06-04T15:40:28Z


    # Viz sample from prior.
    if args.viz:
+        import visdom


Visdom is not actively supported - made the import lazy.

should we just replace this with some matplotlib equivalent?

In another PR? This one already has a lot of changes

ordabayevy · 2026-06-04T15:41:45Z

    test_data = torch.tensor(
        pd_dataframe[["SeasonAt-Bats", "SeasonHits"]].values,
        dtype=torch.float,
-        device=device,


The constructor should use the default device so no need to pass it.

ordabayevy · 2026-06-04T15:42:27Z

-            permutation = torch.randperm(channels, device="cpu").to(
-                torch.Tensor().device
-            )
+            permutation = torch.randperm(channels)


Seems redundant.

ordabayevy · 2026-06-04T15:42:59Z

-            log_det = gpytorch.lazy.NonLazyTensor(truncated).logdet()
-        except ImportError:
-            log_det = torch.linalg.cholesky(truncated).diag().log().sum() * 2
+        log_det = torch.linalg.cholesky(truncated).diag().log().sum() * 2


NonLazyTensor doesn't exist in new gpytorch releases.

ordabayevy · 2026-06-04T15:44:45Z

    def integrate(fn, domain):
        sl = [slice(None)] + (len(domain.shape) - 1) * [None]
-        half_roots_sl = half_roots[sl]
+        half_roots_sl = half_roots[tuple(sl)]


In new pytorch releases list of tensors index gets wrapped in torch.tensor().

ordabayevy · 2026-06-04T15:46:06Z


+    @staticmethod
+    def infer_shapes(low, high):
+        return torch.Size(broadcast_shape(low, high)), torch.Size()


Superclass infer_shapes relies on cls.arg_constraints but in new pytorch Uniform.arg_constraints became an instance property.

ordabayevy · 2026-06-04T15:47:54Z

                    assert chain_id < self.num_chains
                self.exec_traces.append(tr)
-                self.log_weights.append(logit)
+                self.log_weights.append(torch.as_tensor(logit).detach())


The old torch.tensor(self.log_weights) would remove grads and emit warning so detaching here should be okay?

ordabayevy · 2026-06-04T15:49:01Z

+    alpha_atol = 0.1 if alpha_near_one else 0.04
+    beta_atol = 0.12 if alpha_near_one else 0.06
+    assert_close(alpha, pyro.param("alpha").item(), atol=alpha_atol)
+    assert_close(beta, pyro.param("beta").item(), atol=beta_atol)


Had to increase tolerance for tests to pass.

because of your precision change?

I don't think so because torch.as_tensor should return double but I'm re-running the stricter test with explicit double() to verify.

It fails with explicit double() too

martinjankowiak · 2026-06-04T17:10:27Z


    # Viz sample from prior.
    if args.viz:
+        import visdom


should we just replace this with some matplotlib equivalent?

martinjankowiak · 2026-06-04T17:13:52Z

    roots, weights = roots_legendre(num_points)
-    roots = torch.Tensor(roots).double()
-    weights = torch.Tensor(weights).double()
+    roots = torch.as_tensor(roots)


keep doouble?

roots_legendre returns float64 and torch.as_tensor preserves the double dtype.

martinjankowiak · 2026-06-04T17:15:24Z

+    alpha_atol = 0.1 if alpha_near_one else 0.04
+    beta_atol = 0.12 if alpha_near_one else 0.06
+    assert_close(alpha, pyro.param("alpha").item(), atol=alpha_atol)
+    assert_close(beta, pyro.param("beta").item(), atol=beta_atol)


because of your precision change?

ordabayevy · 2026-06-04T18:00:12Z

I don't remember having the DCO check before. Was it enabled automatically? Do I have to sign and push a commit?

martinjankowiak · 2026-06-04T18:06:45Z

not sure about DCO. but yeah try signing your commit...

Signed-off-by: Yerdos Ordabayev <yerdos2030@gmail.com>

martinjankowiak · 2026-06-05T18:20:13Z

thanks yerdos!

Fix lint formatting and mypy checks

52ae7d4

ordabayevy added 4 commits May 26, 2026 03:00

upgrade versions

a4cdef1

Fix visdom install in CI and tighten NN typing

afdc732

Preinstall visdom without build isolation to avoid pkg_resources failures on modern setuptools, require explicit context in ConditionalAutoRegressiveNN, and replace cast-based mypy suppressions with type ignores.

Fix Funsor test and doctest compatibility

5f3e3f7

Pin Funsor dependency to fix branch

b40db87

martinjankowiak mentioned this pull request Jun 2, 2026

Stabilize Gaussian mixture backward passes #3452

Open

ordabayevy added 12 commits June 3, 2026 01:39

tuple sequence

b6814a0

Stop forcing GCC 9 in CI

850c82c

Fix PyTorch distribution compatibility

fb8668c

Remove Visdom from test dependencies

524b979

Enforce unused type ignore checks

b06eb00

Install Pyro before CI lint

ea50d79

Preserve gradients during posterior runs

5c698ff

Detach posterior log weights before storage

f4709fb

Stack predictive log probabilities

c99a897

Relax near-one stable fit tolerances

c0ee224

Detach autoregressive Jacobian scalar

5507926

Modernize tensor constructors

ff3dd89

ordabayevy commented Jun 4, 2026

View reviewed changes

ordabayevy requested review from fritzo and martinjankowiak June 4, 2026 17:06

martinjankowiak approved these changes Jun 4, 2026

View reviewed changes

add back double()

5aea156

Signed-off-by: Yerdos Ordabayev <yerdos2030@gmail.com>

fritzo enabled auto-merge (squash) June 5, 2026 18:10

fritzo merged commit 6cc3ecd into dev Jun 5, 2026
9 checks passed

Uh oh!

Conversation

ordabayevy commented May 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Main Changes

Why

Uh oh!

fritzo commented May 25, 2026

Uh oh!

martinjankowiak commented Jun 2, 2026

Uh oh!

ordabayevy commented Jun 2, 2026

Uh oh!

ordabayevy left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ordabayevy commented Jun 4, 2026

Uh oh!

martinjankowiak commented Jun 4, 2026

Uh oh!

Uh oh!

martinjankowiak commented Jun 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

ordabayevy commented May 25, 2026 •

edited

Loading