Fix empty CUDA_ARCHITECTURES when SM120 is the only arch

[sudhakarsingh27]

When SM120 (Blackwell) is the only architecture in the build and gets moved to per-source COMPILE_OPTIONS, CMAKE_CUDA_ARCHITECTURES becomes empty, causing CMake to error with "CUDA_ARCHITECTURES is empty".

Set CMAKE_CUDA_ARCHITECTURES to OFF as a placeholder when it's empty. The actual arch flags are still added per-source via COMPILE_OPTIONS.

Description

Please include a brief summary of the changes, relevant motivation and context.

Fixes # (issue)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

This PR fixes a CMake build failure that occurs when CMAKE_CUDA_ARCHITECTURES is left empty after all user-specified architectures (e.g. SM120 alone) are removed from the list and promoted into per-source COMPILE_OPTIONS. CMake treats an empty CUDA_ARCHITECTURES as a fatal error, so the fix inserts a guard that sets the variable to OFF — a documented CMake sentinel meaning "do not add automatic -gencode flags" — while leaving the actual per-source arch flags intact.

Key points:

• The fix is correct and minimal: a 7-line if(NOT CMAKE_CUDA_ARCHITECTURES) block placed after all arch-removal loops and before add_library.
• It handles not just SM120-only builds but also SM100-only, SM101-only, and SM110-only builds, as all four arch families are removed from CMAKE_CUDA_ARCHITECTURES in preceding blocks.
• Setting CMAKE_CUDA_ARCHITECTURES OFF is standard CMake practice; per-source COMPILE_OPTIONS are orthogonal and continue to function correctly.
• No behavioral change for builds that include at least one non-promoted architecture (e.g., 70 80 89 90 120).

Confidence Score: 5/5

• Safe to merge — the change is a minimal, correct CMake build fix with no runtime or behavioral impact on existing multi-arch builds.
• The fix is 7 lines, uses documented CMake semantics (CMAKE_CUDA_ARCHITECTURES OFF), is placed at exactly the right point in the file (after all arch-removal loops, before add_library), and does not affect any builds where at least one non-promoted architecture remains. The only concern is that the PR checklist boxes and the GitHub issue reference were left unfilled, which is a process note rather than a code concern.
• No files require special attention.

Important Files Changed

Filename	Overview
transformer_engine/common/CMakeLists.txt	Adds a 7-line guard that sets CMAKE_CUDA_ARCHITECTURES to OFF when it becomes empty after all architectures are moved to per-source COMPILE_OPTIONS; fix is correct, minimal, and handles SM100/SM101/SM110/SM120 single-arch builds.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User sets CMAKE_CUDA_ARCHITECTURES\ne.g. 120] --> B{Is CMAKE_CUDA_ARCHITECTURES defined?}
    B -- No --> C[Auto-set based on CUDA toolkit version\ne.g. 70 80 89 90 100 120]
    B -- Yes --> D[Use user-supplied value]
    C --> E[Arch processing loop]
    D --> E
    E --> F{120 in list?}
    F -- Yes --> G[Remove 120 from CMAKE_CUDA_ARCHITECTURES\nAdd to NVTE_GENERIC_ARCHS and NVTE_SPECIFIC_ARCHS]
    F -- No --> H{100/101/110 in list?}
    G --> H
    H -- Yes --> I[Remove from CMAKE_CUDA_ARCHITECTURES\nAdd to NVTE_GENERIC/SPECIFIC_ARCHS]
    H -- No --> J{CMAKE_CUDA_ARCHITECTURES empty?}
    I --> J
    J -- Yes --> K["set(CMAKE_CUDA_ARCHITECTURES OFF)\n✅ PR Fix — avoids CMake error"]
    J -- No --> L[Keep remaining archs in CMAKE_CUDA_ARCHITECTURES]
    K --> M[add_library transformer_engine]
    L --> M
    M --> N[Per-source COMPILE_OPTIONS applied\nfor NVTE_GENERIC_ARCHS and NVTE_SPECIFIC_ARCHS]
    N --> O[Build succeeds with correct arch flags]

Loading

_{Reviews (2): Last reviewed commit: "Fix empty CUDA_ARCHITECTURES when SM120 ..." | Re-trigger Greptile}