Remove support for older architectures for PyPI releases (#5181)

Summary:
X-link: facebookresearch/FBGEMM#2181

PyPI has a file upload limitation of 600MB for the fbgemm project, and the admins are unwilling to increase the limit.  As such, we remove sm 12.0a support from the build, to put the file size back to under 600MB.  This onnly affects PyPI packages, as Nova packages will continue to be published with support for SM 8.0, 9.0a, 10.0a, 12.0a


Differential Revision: D88325954

Pulled By: q10

Author: q10

.github/scripts/fbgemm_gpu_build.bash

@@ -312,7 +312,15 @@ __configure_fbgemm_gpu_build_cuda () {
     elif  [[ $cuda_version_nvcc == *"V13.0"* ]] ||
           [[ $cuda_version_nvcc == *"V12.9"* ]] ||
           [[ $cuda_version_nvcc == *"V12.8"* ]]; then
-      local arch_list="8.0;9.0a;10.0a;12.0a"
+      # NOTE: If we reach this point, then we are building the package for
+      # publishing to PyPI
+      if [ "${PYPI_PUBLISH_CHANNEL:-}" = "release" ]; then
+        # FBGEMM non-nightly releases can be built with a different version of
+        # CUDA, which migh result in a larger binary size than what PyPI allows.
+        local arch_list="8.0;9.0a;10.0a"
+      else
+        local arch_list="8.0;9.0a;10.0a;12.0a"
+      fi
 
     elif  [[ $cuda_version_nvcc == *"V12.6"* ]] ||
           [[ $cuda_version_nvcc == *"V12.4"* ]] ||

.github/workflows/fbgemm_gpu_release_cuda.yml

@@ -65,6 +65,10 @@ jobs:
       matrix: ${{ needs.generate-build-matrix.outputs.matrix }}
       repo-ref: ${{ github.ref }}
       pytorch-channel-version: ${{ (github.event_name == 'workflow_dispatch' && github.event.inputs.pytorch-channel-version) || 'release' }}
+      extra-env: >-
+        {
+          "PYPI_PUBLISH_CHANNEL": "release"
+        }
 
   generate-test-matrix:
     needs: build