ci: probe cu124 real-arch stripped wheel size

Author: abetlen

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

@@ -21,8 +21,8 @@ jobs:
         run: |
           $matrix = @{
               'os' = @('ubuntu-22.04') #, 'windows-2022')
-              'pyver' = @("3.9", "3.10", "3.11", "3.12")
-              'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") #, "12.5.1", "12.6.1")
+              'pyver' = @("3.12")
+              'cuda' = @("12.4.1") #, "12.5.1", "12.6.1")
               'releasetag' = @("basic")
           }
 
@@ -153,9 +153,9 @@ jobs:
           }
           $cudaTagVersion = $nvccVersion.Replace('.','')
           $env:VERBOSE = '1'
-          # 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"
+          # Build real cubins for the supported GPUs and keep a single PTX target
+          # for forward compatibility instead of embedding PTX for every SM.
+          $env:CMAKE_ARGS = "-DGGML_CUDA_FORCE_MMQ=ON -DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=70-real;75-real;80-real;86-real;89-real;90-real;90-virtual -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'
           # }
@@ -166,6 +166,24 @@ jobs:
           #  $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX=off -DGGML_AVX2=off -DGGML_FMA=off -DGGML_F16C=off'
           # }
           python -m build --wheel
+          Write-Output 'Wheel size before repack:'
+          Get-ChildItem dist/*.whl | ForEach-Object {
+            Write-Output ("{0}`t{1:N2} MiB" -f $_.Name, ($_.Length / 1MB))
+          }
+          $wheelDir = Join-Path $env:RUNNER_TEMP 'cuda-wheel-unpacked'
+          if (Test-Path $wheelDir) {
+            Remove-Item -Recurse -Force $wheelDir
+          }
+          python -m wheel unpack (Get-ChildItem dist/*.whl | Select-Object -First 1).FullName -d $wheelDir
+          Get-ChildItem $wheelDir -Recurse -File | Where-Object { $_.Name -match '\.so(\..*)?$' } | ForEach-Object {
+            & strip --strip-unneeded $_.FullName
+          }
+          Remove-Item dist/*.whl
+          python -m wheel pack (Get-ChildItem $wheelDir | Select-Object -First 1).FullName -d dist
+          Write-Output 'Wheel size after repack:'
+          Get-ChildItem dist/*.whl | ForEach-Object {
+            Write-Output ("{0}`t{1:N2} MiB" -f $_.Name, ($_.Length / 1MB))
+          }
           # Publish tags that reflect the actual installed toolkit version.
           Write-Output "CUDA_VERSION=$cudaTagVersion" >> $env:GITHUB_ENV