Add PR test to build kernels and upload artifacts

.github/workflows/build-kernels.yml

@@ -0,0 +1,35 @@
+name: Build Kernels
+
+on:
+  pull_request:
+
+jobs:
+  build:
+
+    runs-on: [self-hosted, cpu]
+
+    steps:
+      - uses: actions/checkout@v3
+        with:
+          submodules: 'recursive'
+
+      - name: environment
+        run: |
+          python --version
+          nvcc --version
+          python -c "import torch; print('torch:', torch.__version__, torch)"
+          python -c "import torch; print('CUDA available:', torch.cuda.is_available())"
+
+      - name: build kernels
+        run: |
+          ts=$(date +%s)
+          DS_KERNELS_MAKE_JOBS=10 DS_KERNELS_BUILD_STRING=".dev${ts}" CUDA_ARCH_LIST="80;86;89;90" python setup.py bdist_wheel
+          fname=$(ls dist)
+          nname=$(echo $fname | sed 's/cp[0-9]\+-cp[0-9]\+/py3-none/' | sed 's/linux/manylinux1/')
+          mv "dist/$fname" "dist/$nname"
+          ls -al
+
+      - uses: actions/upload-artifact@v3
+        with:
+          name: deepspeed-kernels-whl
+          path: dist/*.whl

builder/builder.py

@@ -72,8 +72,12 @@ def build_extension(self, ext):
         abs_build_lib = os.path.join(os.path.abspath(self.build_lib),
                                      "dskernels")
 
-        subprocess.check_call(['cmake', '-B', abs_build_temp, 
+        subprocess.check_call(['cmake', '-B', abs_build_temp,
                                f'-DLIB_OUTPUT_DIR={abs_build_lib}',
                                f'-DCUDA_ARCH_LIST={cuda_arch_list}'],
                                cwd=ext.source)
-        subprocess.check_call(['make', '-j'], cwd=abs_build_temp)
+
+        # Allow user to specify degree of make parallelism
+        make_jobs = os.environ.get('DS_KERNELS_MAKE_JOBS', None)
+        make_cmd = f"make -j {make_jobs}" if make_jobs else "make -j"
+        subprocess.check_call(make_cmd.split(" "), cwd=abs_build_temp)

release/release.sh

@@ -2,6 +2,8 @@ set -ex
 
 rm -rf dist
 
+#export DS_KERNELS_MAKE_JOBS=10
+
 ts=$(date +%s)
 DS_KERNELS_BUILD_STRING=".dev${ts}" CUDA_ARCH_LIST="80;86" python setup.py bdist_wheel