fix: align Torch wheels with LibTorch minor ABI

Author: Asixa

.github/workflows/pypi.yml

@@ -496,7 +496,7 @@ jobs:
           if-no-files-found: error
 
   test-torch-full-wheel-compatibility:
-    name: Full wheel / ${{ matrix.os }} / py${{ matrix.python-version }} / torch2.11
+    name: Full wheel / ${{ matrix.os }} / py${{ matrix.python-version }} / torch2.10
     needs: [build-torch-linux, build-windows-wheels]
     runs-on: ${{ matrix.os }}
     strategy:
@@ -522,13 +522,13 @@ jobs:
         with:
           name: ${{ matrix.artifact }}
           path: ${{ runner.temp }}/wheel-artifact
-      - name: Validate full wheel with Torch 2.11
+      - name: Validate full wheel with Torch 2.10
         shell: pwsh
         working-directory: ${{ runner.temp }}
         run: |
           python -m pip install --upgrade pip
           $nativeExit=$LASTEXITCODE; if ($nativeExit -ne 0) { exit $nativeExit }
-          python -m pip install --no-cache-dir torch==2.11.0 `
+          python -m pip install --no-cache-dir torch==2.10.0 `
             --index-url https://download.pytorch.org/whl/cu128
           $nativeExit=$LASTEXITCODE; if ($nativeExit -ne 0) { exit $nativeExit }
           $artifactDir = Join-Path $env:RUNNER_TEMP "wheel-artifact"
@@ -539,6 +539,8 @@ jobs:
           $wheel = $wheels[0].FullName
           python -m pip install --no-deps --force-reinstall $wheel
           $nativeExit=$LASTEXITCODE; if ($nativeExit -ne 0) { exit $nativeExit }
+          python -m pip check
+          $nativeExit=$LASTEXITCODE; if ($nativeExit -ne 0) { exit $nativeExit }
           python -I "$env:GITHUB_WORKSPACE/tests/packaging/wheel_lifecycle_smoke.py" installed torch
           $nativeExit=$LASTEXITCODE; if ($nativeExit -ne 0) { exit $nativeExit }
           python -m pip uninstall -y rayd-torch

CHANGELOG.md

@@ -22,6 +22,8 @@ All notable changes to RayD are documented in this file.
   implemented.
 - Expanded single- and multi-GPU CUDA/OptiX CI, Stable ABI checks, wheel
   install/uninstall lifecycle validation, and release-artifact verification.
+- Bound the complete `rayd-torch` wheel to its PyTorch 2.10 build minor; the
+  separately audited `_stable_ops` slice retains cross-version validation.
 
 ### Fixed
 

README.md

@@ -42,8 +42,10 @@ Release artifacts cover CPython 3.10 through 3.14 on Windows x86-64 and
 `manylinux_2_28_x86_64`. The native backend wheels are CPython-specific while
 the Torch backend still contains the transitional `_C` extension. The
 `_stable_ops` library inside `rayd-torch` is untagged and uses the LibTorch 2.10
-Stable ABI boundary. The `rayd` meta-distribution is a universal pure-Python
-wheel and is the only distribution that also publishes an sdist.
+Stable ABI boundary. The complete `rayd-torch` wheel requires PyTorch 2.10
+(`torch>=2.10,<2.11`) while that separately audited Stable ABI library is
+validated through PyTorch 2.13. The `rayd` meta-distribution is a universal
+pure-Python wheel and is the only distribution that also publishes an sdist.
 
 > [!IMPORTANT]
 > Since 0.6.0, RayD uses explicit backend namespaces. The parent `rayd` namespace does

docs/dev/ci_build_matrix.md

@@ -50,13 +50,20 @@ the two backend projects use the unique `pypi-rayd-drjit` and
 | `build-drjit-linux` | Five parallel CPython 3.10-3.14 cibuildwheel jobs | Build and repair five `rayd-drjit` `manylinux_2_28_x86_64` wheels inside CUDA-enabled manylinux images |
 | `build-torch-linux` | Five parallel CPython 3.10-3.14 cibuildwheel jobs | Build and repair five full `rayd-torch` wheels; audit `_legacy_ops`, `_stable_ops`, external framework dependencies, and CUDA images |
 | `build-windows-wheels` | 2 backends x Python 3.10-3.14 on `windows-2022` | Build and audit ten `win_amd64` wheels |
-| `test-torch-full-wheel-compatibility` | Ubuntu and Windows x Python 3.10-3.14 | Install every matching `rayd-torch` wheel with PyTorch 2.11/cu128 and run the complete installed/uninstalled native lifecycle |
+| `test-torch-full-wheel-compatibility` | Ubuntu and Windows x Python 3.10-3.14 | Install every matching `rayd-torch` wheel with its PyTorch 2.10/cu128 build baseline and run the complete installed/uninstalled native lifecycle |
+| `test-stable-torch-abi` | Ubuntu and Windows across Torch 2.10-2.13 | Extract and load the same `_stable_ops` library under every supported Stable ABI runtime |
 | `build-meta` | Python 3.12 on Ubuntu | Build and check the pure Python `rayd` wheel and sdist |
 | `validate-wheel-set` | Ubuntu, after all build and compatibility jobs | Validate the complete release artifact set via `tests.packaging.test_release_artifact_matrix`; gates every publish job |
 | `publish-*` | published GitHub Releases only, on `ubuntu-latest` | Publish backend wheels first, then the meta distribution, using PyPI trusted publishing |
 
 Both native backends keep `manylinux_2_28` rather than changing to the witwin `manylinux_2_35` tag. This is a stricter backward-compatibility target and matches the Dr.Jit and PyTorch 2.10/cu128 Linux wheels used by the builds.
 
+The complete `rayd-torch` wheel uses non-Stable-ABI `at::` and `c10::` APIs in
+`_legacy_ops`, so its runtime dependency is restricted to the PyTorch 2.10
+minor used to build it. PyTorch requires separate wheels per version for such
+extensions; only the audited `_stable_ops` slice is cross-version, and its
+matrix continues to load the same binary under PyTorch 2.10 through 2.13.
+
 Every Linux matrix entry uses cibuildwheel's isolated test environment to
 load the installed public backend and native extension, then uninstalls the
 distribution and confirms that the backend import is absent. Every Windows
@@ -82,7 +89,7 @@ Torch native source builds require CUDA Toolkit 11.3 or newer. The legacy
 library resolves `cuCtxGetCurrent` through the CUDA runtime only when an OptiX
 context is requested, so installed-wheel imports have no direct
 `libcuda.so`/`nvcuda.dll` dependency. The hosted full-wheel lifecycle injects
-no driver stub on either OS, then loads every wheel with PyTorch 2.11 before
+no driver stub on either OS, then loads every wheel with PyTorch 2.10 before
 uninstalling it. Actual CUDA and OptiX execution remains owned by the
 self-hosted GPU acceptance workflows.
 

tests/governance/test_gpu_ci_contract.py

@@ -131,7 +131,7 @@ def test_release_torch_build_and_lifecycle_use_one_pinned_version(self) -> None:
         self.assertIn("python torch/scripts/verify_driver_independence.py $wheel", windows)
         self.assertEqual(self.release.count("torch/scripts/verify_driver_independence.py"), 2)
 
-    def test_torch_211_full_wheel_lifecycle_covers_every_published_wheel(self) -> None:
+    def test_torch_baseline_full_wheel_lifecycle_covers_every_published_wheel(self) -> None:
         lifecycle = self.release.split("  test-torch-full-wheel-compatibility:", 1)[1].split(
             "  test-stable-torch-abi:", 1
         )[0]
@@ -149,16 +149,18 @@ def test_torch_211_full_wheel_lifecycle_covers_every_published_wheel(self) -> No
             "name: ${{ matrix.artifact }}",
             "path: ${{ runner.temp }}/wheel-artifact",
             "working-directory: ${{ runner.temp }}",
-            "torch==2.11.0",
+            "torch==2.10.0",
             "https://download.pytorch.org/whl/cu128",
             "pip install --no-deps --force-reinstall $wheel",
+            "python -m pip check",
             'wheel_lifecycle_smoke.py" installed torch',
             "pip uninstall -y rayd-torch",
             'wheel_lifecycle_smoke.py" absent torch',
         ):
             with self.subTest(marker=marker):
                 self.assertIn(marker, lifecycle)
 
+        self.assertNotIn("torch==2.11.0", lifecycle)
         validation = self.release.split("  validate-wheel-set:", 1)[1].split("  publish-drjit:", 1)[0]
         self.assertIn("- test-torch-full-wheel-compatibility", validation)
 
@@ -172,7 +174,7 @@ def test_windows_native_release_commands_fail_immediately(self) -> None:
         )[0]
 
         self.assertEqual(windows.count(fail_fast), 13)
-        self.assertEqual(compatibility.count(fail_fast), 6)
+        self.assertEqual(compatibility.count(fail_fast), 7)
 
         windows_command_tails = (
             'python -m pip install --upgrade pip build twine "scikit-build-core>=0.10" "cmake>=3.22" ninja',
@@ -196,6 +198,7 @@ def test_windows_native_release_commands_fail_immediately(self) -> None:
             "python -m pip install --upgrade pip",
             "--index-url https://download.pytorch.org/whl/cu128",
             "python -m pip install --no-deps --force-reinstall $wheel",
+            "python -m pip check",
             'python -I "$env:GITHUB_WORKSPACE/tests/packaging/wheel_lifecycle_smoke.py" installed torch',
             "python -m pip uninstall -y rayd-torch",
             'python -I "$env:GITHUB_WORKSPACE/tests/packaging/wheel_lifecycle_smoke.py" absent torch',

tests/packaging/test_torch_project_metadata.py

@@ -28,8 +28,11 @@ def test_default_dependencies_require_torch_not_dr_jit(self):
     def test_transitional_wheels_cover_supported_python_and_torch_baseline(self):
         data = tomllib.loads((TORCH_ROOT / "pyproject.toml").read_text())
         self.assertEqual(data["project"]["requires-python"], ">=3.10,<3.15")
-        self.assertIn("torch>=2.10,<2.12", data["project"]["dependencies"])
+        self.assertEqual(data["project"]["dependencies"], ["torch>=2.10,<2.11"])
         self.assertIn("torch==2.10.0", data["build-system"]["requires"])
+        readme = (TORCH_ROOT / "README.md").read_text(encoding="utf-8")
+        self.assertIn("require PyTorch 2.10", readme)
+        self.assertIn("validated across PyTorch 2.10 through 2.13", readme)
 
     def test_public_python_source_has_no_obsolete_product_name(self):
         source_roots = (TORCH_ROOT / "python" / "rayd" / "torch", WORKSPACE_ROOT / "python" / "rayd" / "_impl")

torch/README.md

@@ -10,6 +10,11 @@ import rayd.torch as rt
 Install the `rayd-torch` distribution. It owns only `rayd/torch/**` and can
 coexist with the independently installed `rayd-drjit` backend.
 
+Published complete wheels are built for and require PyTorch 2.10
+(`torch>=2.10,<2.11`) because `_legacy_ops` uses the non-Stable-ABI
+LibTorch C++ surface. The separately audited `_stable_ops` library remains
+validated across PyTorch 2.10 through 2.13.
+
 Published wheels contain both native backends and do not require an OptiX SDK
 at runtime. Building this dual-backend distribution from source does require
 the OptiX SDK headers, even when the target machine will select CUDA at

torch/pyproject.toml

@@ -10,7 +10,7 @@ readme = "README.md"
 authors = [{ name = "Xingyu Chen", email = "xic063@ucsd.edu" }]
 license = "BSD-3-Clause"
 requires-python = ">=3.10,<3.15"
-dependencies = ["torch>=2.10,<2.12"]
+dependencies = ["torch>=2.10,<2.11"]
 
 [tool.scikit-build]
 build-dir = "build/{wheel_tag}"