Fix clang CI build: use clang as nvcc host compiler

[xwang233]

Summary

Fix CI clang-build-23 job failure where nvcc used gcc as host compiler but received clang-specific flags (-fclang-abi-compat=17) from PyTorch's CMake config.

Two issues were fixed:

• setup-env.sh: Switch from CC="ccache clang" to CC=clang + CMAKE_*_COMPILER_LAUNCHER=ccache. The old pattern caused CMake to resolve the CUDA host compiler (-ccbin) to ccache instead of clang. Set CUDAHOSTCXX to tell CMake to use clang++ as nvcc's host compiler.
• CMakeLists.txt: Read CUDAHOSTCXX into CMAKE_CUDA_HOST_COMPILER before enable_language(CUDA) to ensure it takes effect before PyTorch's TorchConfig.cmake can override it.

Test plan

• CI clang-build-23 job passes

🤖 Generated with Claude Code

[github-actions]

Review updated until commit 3dac2c5

Description

• Fix CI clang build failure by ensuring nvcc uses clang++ as host compiler instead of defaulting to gcc

• Replace CC="ccache clang" pattern with CMAKE_*_COMPILER_LAUNCHER to properly handle ccache with CUDA

• Set CUDAHOSTCXX environment variable in setup-env.sh to direct nvcc to clang++ binary

• Add CMAKE_CUDA_HOST_COMPILER configuration in CMakeLists.txt before enable_language(CUDA)

Changes walkthrough

Relevant files

Bug fix

setup-env.sh Configure CUDA host compiler and ccache launcher                  tools/setup-env.sh Change CC/CXX to use raw compiler paths instead of "ccache clang" pattern Use CMAKE_C_COMPILER_LAUNCHER, CMAKE_CXX_COMPILER_LAUNCHER, CMAKE_CUDA_COMPILER_LAUNCHER for ccache Set CUDAHOSTCXX=$CLANGXX_PATH to ensure nvcc uses clang++ as host compiler +14/-6    CMakeLists.txt Set CUDA host compiler in CMake configuration                        CMakeLists.txt Add CMAKE_CUDA_HOST_COMPILER configuration before enable_language(CUDA) Set CMAKE_CUDA_HOST_COMPILER from CUDAHOSTCXX environment variable if not already set Ensure nvcc uses clang++ as host compiler when CXX is clang +8/-0

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review

CUDAHOSTCXX path correctness Verify that CLANGXX_PATH is correctly set and points to a valid clang++ compiler before exporting CUDAHOSTCXX. The script checks for clang-19 existence earlier, but ensure the path variables are properly initialized. export CUDAHOSTCXX=$CLANGXX_PATH CMake configuration timing The CUDA host compiler setting is placed before enable_language(CUDA), which is correct. However, verify that this placement works correctly with PyTorch's TorchConfig.cmake that may be included later in the build process. The conditional check ensures we don't override an explicitly set compiler, which is good. if(NOT CMAKE_CUDA_HOST_COMPILER AND DEFINED ENV{CUDAHOSTCXX}) set(CMAKE_CUDA_HOST_COMPILER $ENV{CUDAHOSTCXX}) endif()

[greptile-apps]

Greptile Summary

Fixed the CI clang-build-23 job where nvcc was incorrectly dispatching to gcc instead of clang as its host compiler.

Changes:

• tools/setup-env.sh: Exports CUDAHOSTCXX to explicitly tell nvcc to use clang++ as the host compiler, and refactors ccache setup to use CMAKE_*_COMPILER_LAUNCHER variables instead of wrapping CC/CXX with ccache directly
• CMakeLists.txt: Adds CMake logic to set CMAKE_CUDA_HOST_COMPILER from the CUDAHOSTCXX environment variable before enable_language(CUDA) is called, working around PyTorch's TorchConfig.cmake potentially overriding the environment variable

This fix addresses the root cause where gcc doesn't understand clang-specific compiler flags like -fclang-abi-compat=17 that PyTorch adds to CMAKE_CXX_FLAGS.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The changes are well-targeted fixes that address a specific CI failure by ensuring nvcc uses the correct host compiler. The implementation follows best practices: setting environment variables in the shell script and using CMake's proper mechanisms before enable_language(CUDA). The refactoring to use CMAKE_*_COMPILER_LAUNCHER for ccache is also a cleaner approach that avoids the original issue where ccache was being incorrectly identified as the compiler.
• No files require special attention

Important Files Changed

Filename	Overview
CMakeLists.txt	Added logic to set CMAKE_CUDA_HOST_COMPILER from CUDAHOSTCXX env var before enable_language(CUDA) to ensure nvcc uses the correct host compiler
tools/setup-env.sh	Reorganized environment setup to export CUDAHOSTCXX and changed ccache configuration to use CMAKE_*_COMPILER_LAUNCHER instead of wrapping CC/CXX

_{Last reviewed commit: 3dac2c5}

[greptile-apps]

_{1 file reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{2 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{1 file reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{2 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[xwang233]

!build

[wujingyue]

Why not export CMAKE_CUDA_HOST_COMPILER=$CLANGXX_PATH? cc @xwang233