build: lock torch 2.12 and cap <2.13

.github/workflows/test.yml

@@ -0,0 +1,20 @@
+name: test
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+  workflow_dispatch:
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install uv
+        uses: astral-sh/setup-uv@v6
+        with:
+          python-version: "3.12"
+          enable-cache: true
+      - name: Run tests
+        run: uv run --frozen pytest -q

pyproject.toml

@@ -13,7 +13,7 @@ authors = [
 ]
 classifiers = [ "Programming Language :: Python :: 3 :: Only", "Programming Language :: Python :: 3.12" ]
 dependencies = [
-  "torch>=2.8",
+  "torch>=2.8,<2.13",
 ]
 
 [project.urls]
@@ -27,6 +27,15 @@ package-dir = {"" = "src"}
 where = ["src"]
 include = ["wherobots_export*"]
 
+[dependency-groups]
+dev = [
+  "numpy",
+  "pytest>=8",
+]
+
 [[tool.uv.index]]
 name = "pytorch-cpu"
 url = "https://download.pytorch.org/whl/cpu"
+
+[tool.uv.sources]
+torch = { index = "pytorch-cpu" }

tests/test_export.py

@@ -0,0 +1,64 @@
+from pathlib import Path
+
+import pytest
+import torch
+from torch.export.pt2_archive._package import load_pt2
+
+from wherobots_export.torch.export import create_example_input_from_shape, save
+
+
+class TinyModel(torch.nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+        self.conv = torch.nn.Conv2d(3, 4, kernel_size=3, padding=1)
+
+    def forward(self, pixels: torch.Tensor) -> torch.Tensor:
+        return self.conv(pixels).mean(dim=(2, 3))
+
+
+class Normalize(torch.nn.Module):
+    def forward(self, pixels: torch.Tensor) -> torch.Tensor:
+        return (pixels - 0.5) / 0.5
+
+
+def test_save_creates_nonempty_pt2(tmp_path: Path) -> None:
+    output_file = tmp_path / "model.pt2"
+    save(model=TinyModel(), output_file=output_file, input_shape=[1, 3, 32, 32], device="cpu")
+    assert output_file.exists()
+    assert output_file.stat().st_size > 0
+
+
+def test_save_dynamic_batch_roundtrip(tmp_path: Path) -> None:
+    model = TinyModel().eval()
+    output_file = tmp_path / "model.pt2"
+    save(model=model, output_file=output_file, input_shape=[-1, 3, 32, 32], device="cpu")
+
+    contents = load_pt2(str(output_file))
+    assert set(contents.exported_programs) == {"model"}
+
+    exported = contents.exported_programs["model"].module()
+    pixels = torch.randn(4, 3, 32, 32)  # batch differs from the export-time example
+    torch.testing.assert_close(exported(pixels), model(pixels))
+
+
+def test_save_with_transforms_packages_both(tmp_path: Path) -> None:
+    output_file = tmp_path / "model.pt2"
+    save(
+        model=TinyModel(),
+        output_file=output_file,
+        input_shape=[-1, 3, 32, 32],
+        device="cpu",
+        transforms=Normalize(),
+    )
+    contents = load_pt2(str(output_file))
+    assert set(contents.exported_programs) == {"model", "transforms"}
+
+
+def test_create_example_input_rejects_all_dynamic() -> None:
+    with pytest.raises(ValueError):
+        create_example_input_from_shape([-1, -1, -1, -1])
+
+
+def test_create_example_input_fills_dynamic_dims() -> None:
+    example = create_example_input_from_shape([-1, 3, -1, -1], shape_default_value=64)
+    assert example.shape == (2, 3, 64, 64)