build: bypass Torch CUDA discovery without nvcc

[njzjz-bot]

Summary

• Add a DEEPMD_BYPASS_TORCH_CUDA_CHECK CMake option, enabled by default.
• When DeePMD-kit is not building CUDA support and nvcc is unavailable, predefine an imported torch::cudart target before find_package(Torch).
• Document the new CMake option.

Motivation

CUDA-enabled PyTorch wheels can force CMake CUDA toolkit discovery from TorchConfig.cmake, even for CPU-only DeePMD-kit PyTorch builds. This mirrors the workaround used in deepmd-gnn so CPU-only builds do not require a local CUDA compiler.

Tests

• uvx pre-commit run --files source/CMakeLists.txt doc/install/install-from-source.md
• cmake -S source -B /tmp/deepmd-kit-cmake-bypass-test -DENABLE_PYTORCH=ON -DBUILD_CPP_IF=OFF -DBUILD_PY_IF=ON -DCMAKE_PREFIX_PATH=/tmp/deepmd-kit-fake-torch -DUSE_CUDA_TOOLKIT=OFF -DCMAKE_BUILD_TYPE=Release
• Verified the same fake Torch configuration fails with -DDEEPMD_BYPASS_TORCH_CUDA_CHECK=OFF, confirming the shim is what unblocks find_package(Torch).

Authored by OpenClaw (model: custom-chat-jinzhezeng-group/gpt-5.5)

Summary by CodeRabbit

• Documentation

  • Added an installation guide entry documenting a new configuration option that controls PyTorch CUDA toolkit discovery behavior.

• New Features

  • Introduced a configuration option to better support CPU-only builds by bypassing PyTorch’s CUDA toolkit discovery when CUDA is unavailable.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: 211707cd-3264-45c1-9b02-2f9e89af6194

📥 Commits

Reviewing files that changed from the base of the PR and between 97ed2e7 and ed36315.

📒 Files selected for processing (2)

• doc/install/install-from-source.md
• source/CMakeLists.txt

✅ Files skipped from review due to trivial changes (1)

• doc/install/install-from-source.md

🚧 Files skipped from review as they are similar to previous changes (1)

• source/CMakeLists.txt

---

📝 Walkthrough

Walkthrough

A new CMake boolean option DEEPMD_BYPASS_TORCH_CUDA_CHECK (default ON) is added to source/CMakeLists.txt. When enabled and CUDA support is not being built, it searches for nvcc; if not found and torch::cudart is absent, it creates a stub INTERFACE IMPORTED library target to prevent PyTorch's CUDA toolkit discovery from failing. The option is also documented in the installation guide.

Changes

PyTorch CUDA bypass option

Layer / File(s)	Summary
CMake option declaration and torch::cudart stub source/CMakeLists.txt	Declares DEEPMD_BYPASS_TORCH_CUDA_CHECK option (default ON) at the beginning of the file, then conditionally defines torch::cudart as an INTERFACE IMPORTED target when USE_CUDA_TOOLKIT is false, PyTorch is enabled, and nvcc cannot be found.
Installation guide documentation doc/install/install-from-source.md	Adds a table entry for DEEPMD_BYPASS_TORCH_CUDA_CHECK describing its type (BOOL), default value (ON), and the behavior it controls: bypassing PyTorch CUDA toolkit discovery when a local CUDA compiler is unavailable.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Suggested reviewers

• wanghan-iapcm

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately and concisely summarizes the main change: introducing a mechanism to bypass PyTorch's CUDA discovery when nvcc is unavailable.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.14%. Comparing base (1050e46) to head (ed36315).
⚠️ Report is 10 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #5558      +/-   ##
==========================================
- Coverage   82.16%   82.14%   -0.03%     
==========================================
  Files         896      900       +4     
  Lines      102643   104140    +1497     
  Branches     4340     4474     +134     
==========================================
+ Hits        84341    85550    +1209     
- Misses      16965    17180     +215     
- Partials     1337     1410      +73     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[njzjz]

LGTM

— Opus 4.8