Thor platform compilation issue

[gfvvz]

ptxas /tmp/tmpxft_0000909b_00000000-6_tgv_gqa.ptx, line 2331; error : Instruction 'redux.f32' not supported on .target 'sm_110'

cmake -B build -DCMAKE_TOOLCHAIN_FILE=cmake/aarch64.cmake
-DCUTLASS_NVCC_ARCHS="100a;110"
-DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda/targets/sbsa-linux

[ajavanma]

The 100a entry in "100a;110" is what pulls this example into the build. Example 93's CMake condition enables it when the architecture list contains 100a, 100f, 103a, or 103f; the target then inherits the requested architectures, including 110.

Its warp_reduce emits redux.sync.max/min.NaN.f32. The PTX target requirements restrict that floating-point instruction to SM100 architecture/family-specific targets. Changing only 110 to 110a while retaining 100a will therefore still try to compile an unsupported instruction for Thor.

For a Thor-only build with CUDA 13+, use only -DCUTLASS_NVCC_ARCHS=110a and configure a new build directory, e.g. change -B build to -B build-thor, keeping your existing toolchain arguments. This skips example 93. A fresh directory matters because CUTLASS_NVCC_ARCHS_ENABLED is cached separately.

If you only need the library/profiler, -DCUTLASS_ENABLE_EXAMPLES=OFF also excludes this example. If your goal is to run this particular GQA kernel on Thor, it needs a port, including a supported reduction implementation; the architecture flag alone cannot make its SM100-specific PTX run there.

This is based on the pinned source above; I haven't cross-compiled a Thor build.