CUDA wheel: check every dependency, and give torch_cpu_library targets torch's compile settings back by shoumikhin · Pull Request #24087 · pytorch/executorch

shoumikhin · 2026-10-22T18:51:44Z

Problem

#24060 made the Linux CUDA wheel load next to a CPU-only PyTorch by linking torch_cpu_library instead of torch. Three things were too weak.

The check looked only for libtorch_cuda.so and libc10_cuda.so. libnvrtc.so.13 and other CUDA libraries slipped through.
The relinked targets lost torch's compile settings. Linking torch also brings its include directories as system headers, TORCH_CXX_FLAGS, and on recent torch a C++ standard. torch_cpu_library misses most of them. So warnings inside torch's headers are reported as ours, and a configured TORCH_CXX_FLAGS is dropped.
The comments at the five link sites did not say why torch_cpu_library is used. At aten_bridge the reason was wrong.

Fix

Allow list in the CUDA wheel smoke test. Every dependency a shipped library records must be in the wheel, a torch library that a CPU-only torch also has, a system library the manylinux policy allows, or from a package a plain install requires.
Comments. Each site now says torch records libtorch_cuda with a CUDA build of PyTorch, and torch_cpu_library does not. aten_bridge is now STATIC, as its comment assumes.
Compile settings. find_package_torch copies torch's interface include directories, compile options and compile features to torch_cpu_library once, so every target in the build that links torch_cpu_library gets them. The CUDA settings a CUDA build of PyTorch adds to torch are not copied. If there is no torch_cpu_library, as with a static libtorch, nothing is copied.

Testing

A new CMake test checks that targets linking torch_cpu_library, also through a PUBLIC link, get the C++ API include directory, TORCH_CXX_FLAGS and, when torch sets one (the 2.16 nightly does, 2.14 does not), the C++ standard, also when torch was found first. A second test checks that find_package_torch configures when only a torch target exists. Both pass on Linux x86-64 with torch 2.14.0 (CPU, CUDA 13.0) and a 2.16 CUDA 13.2 nightly, and on macOS arm64 with 2.14.0 CPU. Each fails without its part of the change.
On Linux x86-64, the new wheel check passes on a CUDA wheel built with Link torch_cpu_library, not torch, so the CUDA wheel imports with a CPU-only torch #24060, next to a CPU-only torch. It fails on a CUDA 13.2 nightly from before Link torch_cpu_library, not torch, so the CUDA wheel imports with a CPU-only torch #24060, naming libnvrtc.so.13 and three more.
With torch 2.14.0+cpu and GCC 11, the six relinked targets build. op_allclose.cpp with -Wall -Wextra -Werror: 25 errors in torch headers without the change, none with it.
With BUILD_SHARED_LIBS=ON, aten_bridge is now static and _training_lib builds.

pytorch-bot · 2026-10-22T18:51:48Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24087

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (1 Unrelated Failure)

As of commit 9cb582b with merge base b578b2e ():

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

Test CUDA Builds / test-muse-glimmer-cuda-e2e-kquant-17gb-solo-text-batching / linux-job (gh) (trunk failure)
RuntimeError: Command bash /__w/_temp/exec_script failed with exit code 1

This comment was automatically generated by Dr. CI and updates every 15 minutes.

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

Copilot

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

Copilot

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

The CUDA wheel smoke test checked that no shipped library records libtorch_cuda.so or libc10_cuda.so. That misses other libraries that only a CUDA build of torch installs. The wheel before the linking fix also recorded libnvrtc.so.13 on the LLM runner module. With the two names removed, the old check passed while that module still failed to import next to a CPU-only torch. Turn it into an allow list. Every recorded dependency must be shipped in the wheel, be one of the torch libraries a CPU-only torch also has, be part of the C and C++ system runtime, or be installed by a package that a plain install of the wheel requires (today that is the CUDA runtime). A requirement behind an extra, or for another platform, is skipped, so an optional package that happens to be installed does not widen the list. The check compares names only. It does not load anything.

The same three-line comment sat at five link sites. It explained why the line does not link torch, but not what torch_cpu_library does differently, and "in a CUDA build" read as our own CUDA option rather than how PyTorch was built. Each site now says that torch records libtorch_cuda with a CUDA build of PyTorch, so a CPU-only torch could not load the library, and that torch_cpu_library does not. At aten_bridge the old reason did not apply. aten_bridge is a static library and is never loaded on its own. Its link matters because it is passed on to _training_lib. But it was only static by default: with BUILD_SHARED_LIBS=ON it became a shared library. Mark it STATIC so the comment holds in every layout. Nothing installs it and only _training_lib links it.

Linking the torch target does more than link. TorchConfig also gives it TORCH_INCLUDE_DIRS, which CMake treats as system headers because the target is imported, TORCH_CXX_FLAGS, and on recent torch a minimum C++ standard (C++20). torch_cpu_library, which our targets now link instead so that a CPU-only torch can load them, only passes on torch/include. So those targets changed. torch/csrc/api/include became a normal include directory, so warnings inside torch's headers are now reported as if they were our code. And a TORCH_CXX_FLAGS set at configure time no longer reached them. For example, compiling kernels/portable/cpu/op_allclose.cpp with -Wall -Wextra -Werror went from no errors to 25 errors, all inside torch's headers. Default builds were not affected. find_package_torch now copies torch's interface include directories, compile options and compile features to torch_cpu_library. It does this once, also when the project found torch before calling it. Every target that links torch_cpu_library, directly or through a PUBLIC link such as custom_ops_aot_lib, now gets these settings, and a new link site cannot forget them. With a CUDA build of PyTorch, torch also carries the CUDA include directory and definitions of torch_cuda_library. Those are not copied, since these targets do not use CUDA. Changing torch_cpu_library itself also affects any other target in the build that links it. When there is no torch_cpu_library, for example with a static libtorch or a parent project that defines its own torch target, nothing is copied. A new test configures a small project and checks that a target linking torch_cpu_library, directly or through a PUBLIC link, gets torch's C++ API include directory, TORCH_CXX_FLAGS and, when torch sets one, its C++ standard. It checks both when find_package_torch finds torch and when the project found it first. A second test checks that find_package_torch still configures when only a torch target exists. Both pass with CPU and CUDA builds of torch. The first fails without this change, and the second fails if the copy does not first check that torch_cpu_library exists.

Copilot

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

shoumikhin requested a review from larryliu0820 as a code owner October 22, 2026 18:51

shoumikhin added the release notes: build Changes related to build, including dependency upgrades, build flags, optimizations, etc. label Oct 22, 2026

Copilot AI balanced review requested due to automatic review settings October 22, 2026 18:51

shoumikhin added the release notes: build Changes related to build, including dependency upgrades, build flags, optimizations, etc. label Oct 22, 2026

meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Oct 22, 2026

Copilot AI reviewed Oct 22, 2026

Copilot AI balanced review requested due to automatic review settings October 22, 2026 20:48

shoumikhin force-pushed the cuda-wheel-dep-followups branch from dad4443 to c520aa1 Compare October 22, 2026 20:48

Copilot AI reviewed Oct 22, 2026

Copilot AI balanced review requested due to automatic review settings October 23, 2026 10:15

shoumikhin force-pushed the cuda-wheel-dep-followups branch from c520aa1 to 76d4404 Compare October 23, 2026 10:15