ci: narrow CUDA wheel architectures

Author: abetlen

.github/workflows/build-wheels-cuda.yaml

@@ -133,7 +133,10 @@ jobs:
             $env:CMAKE_ARGS = ''
           }
           $env:VERBOSE = '1'
-          $env:CMAKE_ARGS = "-DGGML_CUDA_FORCE_MMQ=ON -DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=all -DCMAKE_CUDA_FLAGS=--allow-unsupported-compiler $env:CMAKE_ARGS"
+          # Match upstream llama.cpp CI by targeting a conservative set of
+          # CUDA architectures instead of `all`, which now includes newer
+          # virtual architectures some toolkit versions cannot assemble.
+          $env:CMAKE_ARGS = "-DGGML_CUDA_FORCE_MMQ=ON -DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=61;70;75;80;86;89 -DCMAKE_CUDA_FLAGS=--allow-unsupported-compiler $env:CMAKE_ARGS"
           # if ($env:AVXVER -eq 'AVX') {
           $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX2=off -DGGML_FMA=off -DGGML_F16C=off'
           # }