ninja

Signed-off-by: oliver könig <okoenig@nvidia.com>

Author: ko3n1g

.github/workflows/_build.yml

@@ -123,7 +123,7 @@ jobs:
           path: build-cache.tar.gz
           key: build-${{ inputs.release-version }}-${{ inputs.python-version }}-${{ inputs.cuda-version }}-${{ inputs.torch-version }}-${{ inputs.cxx11-abi }}-${{ github.run_number }}-${{ github.run_attempt }}
           restore-keys: |
-            build-${{ inputs.release-version }}-${{ inputs.python-version }}-${{ inputs.cuda-version }}-${{ inputs.torch-version }}-${{ inputs.cxx11-abi }}-${{ github.run_number }}-
+            build-${{ inputs.release-version }}-${{ inputs.python-version }}-${{ inputs.cuda-version }}-${{ inputs.torch-version }}-${{ inputs.cxx11-abi }}-
 
       - name: Log build logs
         run: |
@@ -150,7 +150,13 @@ jobs:
           export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
           # Limit MAX_JOBS otherwise the github runner goes OOM
           # nvcc 11.8 can compile with 2 jobs, but nvcc 12.3 goes OOM
+          if [ -d "build/temp.linux-x86_64-cpython-312" ]; then
+            cd build/temp.linux-x86_64-cpython-312
+            ninja
+            
+          fi
           MAX_JOBS=$([ "$MATRIX_CUDA_VERSION" == "129" ] && echo 1 || echo 2) NVCC_THREADS=2 FLASH_ATTENTION_FORCE_BUILD="TRUE" FLASH_ATTENTION_FORCE_CXX11_ABI=${{ inputs.cxx11_abi}} python setup.py bdist_wheel --dist-dir=dist
+
           tmpname=cu${WHEEL_CUDA_VERSION}torch${MATRIX_TORCH_VERSION}cxx11abi${{ inputs.cxx11_abi }}
           wheel_name=$(ls dist/*whl | xargs -n 1 basename | sed "s/-/+$tmpname-/2")
           ls dist/*whl |xargs -I {} mv {} dist/${wheel_name}