fix(build): Handle namespace packages for PyPI CUDA detection

[sbhavani]

Description

The NCCL package from PyPI (e.g., nvidia-nccl-cu12/cu13) is a PEP 420 namespace package that doesn't have __file__ attribute. This caused get_cuda_include_dirs() to fail with TypeError when building with only NCCL from PyPI.

Use nvidia.__path__[0] which works for both namespace packages and regular packages.

Fixes: #2331

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:

• Fixed get_cuda_include_dirs() to handle namespace packages (which lack __file__)
• Added clear error message when nvidia package directory cannot be located

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 Overview

Greptile Summary

This PR fixes a bug where get_cuda_include_dirs() would crash when the nvidia package is a PEP 420 namespace package (which doesn't have a __file__ attribute). This occurs when users install NCCL and other CUDA packages from PyPI.

Key Changes:

• Replaced Path(nvidia.__file__).parent with Path(list(nvidia.__path__)[0])
• Added proper error handling when the nvidia package directory cannot be located
• Solution works for both regular packages (with __init__.py) and namespace packages (without __init__.py)

How it works:

• All Python packages (both regular and namespace) have a __path__ attribute containing their directory locations
• Regular packages also have __file__, but namespace packages do not
• By using __path__[0], the code handles both package types uniformly

The fix is minimal, targeted, and maintains backward compatibility with existing installations.

Confidence Score: 5/5

• Safe to merge - fixes a real bug without introducing new issues
• The change correctly handles both namespace packages (PEP 420) and regular packages by using path[0] instead of file.parent. The truthiness check on path prevents IndexError from empty lists. The logic is sound and maintains backward compatibility.
• No files require special attention

Important Files Changed

File Analysis

Filename	Score	Overview
build_tools/utils.py	4/5	Fixed namespace package handling by using path[0] instead of file.parent - handles both PEP 420 namespace packages and regular packages correctly

Sequence Diagram

sequenceDiagram
    participant Builder as Build System
    participant Utils as get_cuda_include_dirs()
    participant Python as Python Import System
    participant FS as File System

    Builder->>Utils: Call to get CUDA headers
    Utils->>Utils: Check cuda_toolkit_include_path()
    alt CUDA toolkit found
        Utils-->>Builder: Return toolkit include paths
    else No toolkit, use PyPI packages
        Utils->>Python: import nvidia
        alt Import fails
            Python-->>Utils: ModuleNotFoundError
            Utils-->>Builder: RuntimeError("CUDA not found")
        else Import succeeds
            Python-->>Utils: nvidia module
            Utils->>Utils: Check hasattr(nvidia, "__path__")
            alt Has __path__ and not empty
                Utils->>Python: Access nvidia.__path__[0]
                Python-->>Utils: First path in __path__
                Utils->>Utils: cuda_root = Path(__path__[0])
                Utils->>FS: cuda_root.iterdir()
                FS-->>Utils: List of subdirectories
                Utils->>Utils: Filter for dirs with /include
                Utils-->>Builder: Return list of include paths
            else No __path__ or empty
                Utils-->>Builder: RuntimeError("Could not locate nvidia package directory")
            end
        end
    end

Loading

[greptile-apps]

Greptile Overview

Greptile Summary

This PR fixes a TypeError when building with PyPI CUDA packages (like nvidia-nccl-cu12/cu13) by handling PEP 420 namespace packages that lack the __file__ attribute. The fix correctly switches from using nvidia.__file__ to nvidia.__path__[0] in get_cuda_include_dirs().

Key Changes

• Modified get_cuda_include_dirs() in build_tools/utils.py to use nvidia.__path__[0] instead of nvidia.__file__
• Added proper attribute checking with hasattr(nvidia, "__path__") before accessing
• Added clear error message when nvidia package directory cannot be located

Issues Found

• Incomplete fix: transformer_engine/common/__init__.py at lines 249-252 has the same nvidia.__file__ pattern that will fail with namespace packages. Line 249 checks if nvidia.__file__ is None, but namespace packages raise AttributeError (attribute doesn't exist), not return None. Consider extending this fix to that location.
• Minor optimization: list(nvidia.__path__)[0] can be simplified to nvidia.__path__[0] since __path__ is already indexable.

Confidence Score: 3/5

• This PR is safe to merge with moderate confidence - it fixes the stated issue but leaves similar issues unaddressed
• The fix is technically correct and solves the immediate problem with namespace packages in get_cuda_include_dirs(). However, the same pattern exists in transformer_engine/common/__init__.py which will still fail with namespace packages. Score of 3 reflects that this is a partial fix - the change works but doesn't address all instances of the problem in the codebase.
• Consider reviewing transformer_engine/common/__init__.py lines 249-252 for the same namespace package issue

Important Files Changed

File Analysis

Filename	Score	Overview
build_tools/utils.py	3/5	Fixed get_cuda_include_dirs() to handle PEP 420 namespace packages using __path__, but missed fixing similar issue in transformer_engine/common/__init__.py and has minor code optimization opportunity

Sequence Diagram

sequenceDiagram
    participant Build as Build System
    participant Utils as build_tools/utils.py
    participant Nvidia as nvidia package
    participant FS as File System

    Build->>Utils: get_cuda_include_dirs()
    
    alt CUDA Toolkit Installed
        Utils->>Utils: cuda_toolkit_include_path()
        Utils-->>Build: [toolkit_include_path]
    else PyPI CUDA Packages
        Utils->>Nvidia: import nvidia
        alt ModuleNotFoundError
            Nvidia-->>Utils: Error
            Utils-->>Build: RuntimeError("CUDA not found.")
        else Success
            Nvidia-->>Utils: nvidia module
            
            alt Has __path__ attribute (NEW)
                Utils->>Nvidia: check hasattr(__path__) && __path__
                alt __path__ exists and not empty
                    Utils->>Nvidia: nvidia.__path__[0]
                    Nvidia-->>Utils: First package path
                    Utils->>FS: iterate cuda_root subdirs
                    FS-->>Utils: subdirectories
                    Utils->>Utils: filter dirs with include/
                    Utils-->>Build: [subdir/include, ...]
                else No __path__ or empty
                    Utils-->>Build: RuntimeError("Could not locate nvidia package directory.")
                end
            end
        end
    end

Loading

[greptile-apps]

Redundant list() conversion - nvidia.__path__ is already indexable, can use nvidia.__path__[0] directly

Suggested change

	cuda_root = Path(list(nvidia.__path__)[0])
	cuda_root = Path(nvidia.__path__[0])

[greptile-apps]

This fix correctly handles namespace packages here, but transformer_engine/common/__init__.py line 249-252 has the same issue with nvidia.__file__:

• Line 249 checks if nvidia.__file__ is None, but namespace packages will raise AttributeError when accessing __file__ (the attribute doesn't exist, it doesn't return None)
• Line 252 uses Path(nvidia.__file__).parent which will also fail with namespace packages

Consider applying the same fix pattern (__path__) to _nvidia_cudart_include_dir() in that file. Should this PR also fix the similar nvidia.__file__ usage in transformer_engine/common/__init__.py line 249-252, or will that be addressed in a separate PR?