Expose torch.cuda.current_solver_handle for cuSOLVER handle sharing

[IvanYashchuk]

Summary

This PR adds torch.cuda.current_solver_handle() and the backing C binding torch._C._cuda_getCurrentSolverHandle so Python users can access PyTorch’s current cuSOLVER handle in the same style as torch.cuda.current_blas_handle(). During validation, importing torch exposed a missing link dependency for this symbol, so the PR also links torch_python with torch_cuda_linalg in CUDA builds to ensure the handle symbol resolves correctly at runtime.

Motivation

nvmath-python cuSOLVER bindings operate on intptr_t handles and expose stream management APIs such as get_stream and set_stream. Without a way to retrieve PyTorch’s existing cuSOLVER handle, users and library authors must create and manage separate handles, which adds lifecycle and stream coordination complexity. Exposing PyTorch’s current cuSOLVER handle enables direct interoperability with nvmath-python while keeping handle ownership and configuration in PyTorch.

nvmath-python interop example

import torch
import nvmath.bindings.cusolverDn as cusolverDn

assert torch.cuda.is_available()

handle = torch.cuda.current_solver_handle()

# Save current stream associated with the PyTorch-managed cuSOLVER handle.
orig_stream = cusolverDn.get_stream(handle)

# Move the handle to a non-default CUDA stream.
alt_stream_obj = torch.cuda.Stream()
alt_stream = alt_stream_obj.cuda_stream
cusolverDn.set_stream(handle, alt_stream)
assert cusolverDn.get_stream(handle) == alt_stream

# Run a CUDA linalg op to confirm the handle remains usable.
_ = torch.linalg.cholesky(torch.eye(4, device="cuda"))

# Restore original stream.
cusolverDn.set_stream(handle, orig_stream)
assert cusolverDn.get_stream(handle) == orig_stream

cc @ptrblck @msaroufim @eqy @jerryzh168 @tinglvv @nWEIdia @jianyuh @nikitaved @mruberry @walterddr @xwang233 @lezcano @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo @azahed98

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/176705

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Run pull request jobs on OSDC runners in shadow mode

✅ No Failures

As of commit 922ab32 with merge base da74fec ():
💚 Looks good so far! There are no failures yet. 💚

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

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

torch/csrc/cuda/Module.cpp:1636

• The PR description mentions adding a build-system link dependency (linking torch_python with torch_cuda_linalg) to resolve a missing symbol at import time, but no corresponding build/CMake changes are included in this diff. If that link dependency is still required for some configurations (e.g., split/lazy CUDA linalg builds), please add the build change; otherwise, update the PR description to reflect the actual fix path so reviewers/packagers aren’t misled.

#endif
  END_HANDLE_TH_ERRORS
}

PyObject* THCPModule_cuda_tunableop_enable(PyObject* _unused, PyObject* arg) {
  HANDLE_TH_ERRORS

[nikitaved]

Thank you, @IvanYashchuk ! Looks like the failures are unrelated -- but could you please rebase?

@malfet, could you, please, have a look?

[IvanYashchuk]

@malfet, could you please review the changes or find someone else who could help merge this work?

[IvanYashchuk]

@pytorchbot merge

[pytorchmergebot]

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here