fix(ci): align cuda toolkit roots and tags

Author: abetlen

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

@@ -100,9 +100,11 @@ jobs:
         run: |
           $cudaVersion = $env:CUDAVER
           if ($IsLinux) {
-            mamba install -y --override-channels -c "nvidia/label/cuda-$cudaVersion" -c conda-forge "cuda-toolkit=$cudaVersion" "cuda-nvcc_linux-64=$cudaVersion"
+            # Keep nvcc, cudart, and headers on the same NVIDIA label so the
+            # detected toolkit version matches the published wheel tag.
+            mamba install -y --override-channels -c "nvidia/label/cuda-$cudaVersion" "cuda-toolkit=$cudaVersion" "cuda-nvcc_linux-64=$cudaVersion" cuda-cudart cuda-cudart-dev
           } else {
-            mamba install -y --override-channels -c "nvidia/label/cuda-$cudaVersion" -c conda-forge "cuda-toolkit=$cudaVersion"
+            mamba install -y --override-channels -c "nvidia/label/cuda-$cudaVersion" "cuda-toolkit=$cudaVersion"
           }
           if ($LASTEXITCODE -ne 0) {
             exit $LASTEXITCODE
@@ -115,28 +117,25 @@ jobs:
           $env:CUDA_HOME = $env:CONDA_PREFIX
           $env:CUDA_TOOLKIT_ROOT_DIR = $env:CONDA_PREFIX
           $cudaHostCompilerArg = ''
+          $env:CMAKE_ARGS = ''
           if ($IsLinux) {
             if (Test-Path '/usr/bin/g++-12') {
               $env:CC = '/usr/bin/gcc-12'
               $env:CXX = '/usr/bin/g++-12'
               $env:CUDAHOSTCXX = '/usr/bin/g++-12'
               $cudaHostCompilerArg = " -DCMAKE_CUDA_HOST_COMPILER=$env:CUDAHOSTCXX"
             }
-            $cudaTargetRoot = Join-Path $env:CONDA_PREFIX 'targets/x86_64-linux'
-            if (Test-Path (Join-Path $cudaTargetRoot 'include/cuda_runtime.h')) {
-              $env:CUDAToolkit_ROOT = $cudaTargetRoot
-              $env:CUDA_TOOLKIT_ROOT_DIR = $cudaTargetRoot
-              $env:CMAKE_ARGS = "-DCUDAToolkit_ROOT=$cudaTargetRoot -DCUDA_TOOLKIT_ROOT_DIR=$cudaTargetRoot$cudaHostCompilerArg"
-              $env:CPATH = "$cudaTargetRoot/include:$env:CPATH"
-              $env:CPLUS_INCLUDE_PATH = "$cudaTargetRoot/include:$env:CPLUS_INCLUDE_PATH"
-              $env:LIBRARY_PATH = "$cudaTargetRoot/lib:$env:LIBRARY_PATH"
-              $env:LD_LIBRARY_PATH = "$cudaTargetRoot/lib:$env:CONDA_PREFIX/lib:$env:LD_LIBRARY_PATH"
+            if (Test-Path (Join-Path $env:CONDA_PREFIX 'include/cuda_runtime.h')) {
+              $env:CUDAToolkit_ROOT = $env:CONDA_PREFIX
+              $env:CUDA_TOOLKIT_ROOT_DIR = $env:CONDA_PREFIX
+              $env:CMAKE_ARGS = "-DCUDAToolkit_ROOT=$env:CONDA_PREFIX -DCUDA_TOOLKIT_ROOT_DIR=$env:CONDA_PREFIX$cudaHostCompilerArg"
+              $env:CPATH = "$env:CONDA_PREFIX/include:$env:CPATH"
+              $env:CPLUS_INCLUDE_PATH = "$env:CONDA_PREFIX/include:$env:CPLUS_INCLUDE_PATH"
+              $env:LIBRARY_PATH = "$env:CONDA_PREFIX/lib:$env:LIBRARY_PATH"
+              $env:LD_LIBRARY_PATH = "$env:CONDA_PREFIX/lib:$env:LD_LIBRARY_PATH"
             } else {
-              $env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH
               $env:CMAKE_ARGS = $cudaHostCompilerArg.Trim()
             }
-          } else {
-            $env:CMAKE_ARGS = ''
           }
           $nvccPath = Join-Path $env:CONDA_PREFIX 'bin/nvcc'
           if (-not (Test-Path $nvccPath)) {
@@ -145,14 +144,16 @@ jobs:
           if (-not (Test-Path $nvccPath)) {
             throw 'Failed to find nvcc in the conda environment'
           }
+          $env:CUDACXX = $nvccPath
+          $env:PATH = "$(Split-Path $nvccPath):$env:PATH"
           $nvccVersion = ((& $nvccPath --version) | Select-String 'release ([0-9]+\.[0-9]+)').Matches[0].Groups[1].Value
           if (-not $nvccVersion) {
             throw 'Failed to detect the installed CUDA toolkit version'
           }
           $cudaTagVersion = $nvccVersion.Replace('.','')
           $env:VERBOSE = '1'
-          # Pin a portable architecture list instead of CMake's `all`, which now
-          # includes future targets the hosted-runner toolchains cannot assemble.
+          # Keep a portable SM set, including sm_70, instead of CMake's `all`,
+          # which now pulls in future targets the hosted-runner toolchains cannot assemble.
           $env:CMAKE_ARGS = "-DGGML_CUDA_FORCE_MMQ=ON -DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=70;75;80;86;89;90 -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'