[Bug] `cutlass_profiler` fails to build with `CUTLASS_BUILD_SHARED_LIBS=OFF`

Bug: cutlass_profiler fails to build with CUTLASS_BUILD_SHARED_LIBS=OFF

Description

When building CUTLASS v4.3.5 (and main) with static-only configuration (-DCUTLASS_BUILD_SHARED_LIBS=OFF -DCUTLASS_BUILD_STATIC_LIBS=ON), the cutlass_profiler target fails to compile with:

fatal error: cutlass/library/library.h: No such file or directory

The header file exists at tools/library/include/cutlass/library/library.h, but the include path is never propagated to the profiler target.

Root Cause

In tools/library/CMakeLists.txt (lines 283–290), the cutlass_lib alias is only created when shared libs are enabled:

# For backward compatibility with the old name
if(CUTLASS_BUILD_SHARED_LIBS)
  add_library(cutlass_lib ALIAS cutlass_library)       # ← only when shared ON
endif()

if(CUTLASS_BUILD_STATIC_LIBS)
  add_library(cutlass_lib_static ALIAS cutlass_library_static)
endif()

However, tools/profiler/CMakeLists.txt (line 89) unconditionally links against cutlass_lib:

target_link_libraries(
  cutlass_profiler
  PRIVATE
  cutlass_lib          # ← always required, no fallback
  cutlass_tools_util_includes
  ...
)

When CUTLASS_BUILD_SHARED_LIBS=OFF, the cutlass_lib target does not exist. CMake silently treats the name as a plain linker flag (-lcutlass_lib) rather than a CMake target, so:

1. Transitive include directories from cutlass_library_includes (which provides tools/library/include/) are not propagated → compilation fails with missing headers.
2. Even if compilation somehow passed, linking would also fail because no libcutlass_lib.so or libcutlass_lib.a file exists.

Environment

• CUTLASS version: v4.3.5 (also verified on main)
• Platform: aarch64
• CMake: 3.31.6
• GCC: 13.3.0
• Base image: nvcr.io/nvidia/pytorch:25.08-py3

Steps to Reproduce

git clone --branch v4.3.5 --depth 1 https://github.com/NVIDIA/cutlass.git
cmake \
    -DCUTLASS_BUILD_SHARED_LIBS=OFF \
    -DCUTLASS_BUILD_STATIC_LIBS=ON \
    -DCUTLASS_NVCC_ARCHS="100a;103a" \
    -DCUTLASS_ENABLE_EXAMPLES=OFF \
    -DCUTLASS_ENABLE_TESTS=OFF \
    -DCUTLASS_LIBRARY_KERNELS="cutlass3x_sm100_tensorop_gemm_f16_*" \
    -S ./cutlass \
    -B ./cutlass/build
cmake --build ./cutlass/build --target cutlass_profiler

Why static-only?

On aarch64, even per-family shared libraries can overflow R_AARCH64_PREL32 relocations (±2 GB limit) in .eh_frame when thousands of SM100 GEMM kernel objects are linked into a single .so. Building static-only avoids this linker limitation entirely, since .a archives defer relocation to the final executable link.

CUTLASS already has -mcmodel=medium handling for x86_64 large library relocation issues, but has no equivalent workaround for aarch64.

Suggested Fix

Add a static fallback alias in tools/library/CMakeLists.txt:

if(CUTLASS_BUILD_STATIC_LIBS)
  add_library(cutlass_lib_static ALIAS cutlass_library_static)
  if(NOT TARGET cutlass_lib)
    add_library(cutlass_lib ALIAS cutlass_library_static)
  endif()
endif()

This ensures cutlass_lib exists regardless of the shared/static configuration, preserving backward compatibility.

Current Workaround

We apply a sed patch post-clone before running cmake:

sed -i '/add_library(cutlass_lib_static ALIAS cutlass_library_static)/a\
  if(NOT TARGET cutlass_lib)\n\
    add_library(cutlass_lib ALIAS cutlass_library_static)\n\
  endif()' cutlass/tools/library/CMakeLists.txt

Start in tools/library/CMakeLists.txt around the cutlass_lib and cutlass_lib_static aliases, then inspect the cutlass_profiler linkage in tools/profiler/CMakeLists.txt. Configure with shared libraries off and static libraries on, build the cutlass_profiler target, and confirm that compilation and linking succeed with the library include paths propagated.