[CUDA] Heuristics for Hopper QMM (#3173)

2026-02-27 09:22:10 +09:00
parent 5c4abd2f06
commit 0c8107ce8a
13 changed files with 189 additions and 40 deletions
@@ -21,7 +21,7 @@ runs:
      run: |
        if ${{ startsWith(inputs.toolkit, 'cuda') && runner.arch == 'arm64' }} ; then
          # There is no GPU in arm64 runner, use a common arch.
-          CMAKE_ARGS="$CMAKE_ARGS -DMLX_CUDA_ARCHITECTURES=90a"
+          CMAKE_ARGS="$CMAKE_ARGS -DMLX_CUDA_ARCHITECTURES=80"
          # Can not build tests and stubs when the built executables can not run.
          CMAKE_ARGS="$CMAKE_ARGS -DMLX_BUILD_TESTS=OFF -DMLX_BUILD_PYTHON_STUBS=OFF"
        fi