[CUTLASS C++] Possible SM121 sparse GEMM unit-test CMake coverage gap

[yashb98]

I reproduced what looks like a CMake coverage gap on main at 59e3a333. With CUTLASS_ENABLE_TESTS=ON and CUTLASS_NVCC_ARCHS=121a, the neighboring dense SM120 gate generates its aggregate target, but the sparse and blockscaled-sparse gates do not.

Reproduction

cmake -S . -B build-sm121-tests \
  -DCUTLASS_NVCC_ARCHS=121a \
  -DCUTLASS_ENABLE_TESTS=ON \
  -DCUTLASS_ENABLE_EXAMPLES=OFF \
  -DCUTLASS_ENABLE_TOOLS=OFF \
  -DCUTLASS_ENABLE_LIBRARY=OFF \
  -DCUTLASS_ENABLE_PROFILER=OFF

cmake --build build-sm121-tests --target help \
  | grep -E 'cutlass_test_unit_gemm_device_sm120_(tensorop|sptensorop|bssp)'

The only matching aggregate target is:

... cutlass_test_unit_gemm_device_sm120_tensorop

Why this looks unintentional

The two sparse gates currently match only 120a. All 15 .cu files in those directories guard their test definitions with CUTLASS_ARCH_MMA_SM120_SUPPORTED || CUTLASS_ARCH_MMA_SM121_SUPPORTED, and Example 80 already enables its sparse GEMMs for 120a|121a.

As additional evidence, merged PR #3427 reports building cutlass_test_unit_gemm_device_sm120_bssp_general with CUTLASS_NVCC_ARCHS=120a;121a and passing 7/7 tests on GB10.

Proposed scope

I can prepare a focused PR that adds 121a to these two gates, then build and run every newly generated target on my GB10 (compute capability 12.1) with CUDA 13.0.88.

Is this coverage difference intentional? If not, is anyone already working on it?