improvements

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

Author: ko3n1g

.github/workflows/_build.yml

@@ -161,7 +161,8 @@ jobs:
           export FLASH_ATTENTION_FORCE_BUILD="TRUE"
           export FLASH_ATTENTION_FORCE_CXX11_ABI=${{ inputs.cxx11-abi }}
 
-          timeout 20m python setup.py bdist_wheel --dist-dir=dist || EXIT_CODE=$?
+          # 5h timeout since GH allows max 6h and we want some buffer
+          timeout 5h python setup.py bdist_wheel --dist-dir=dist || EXIT_CODE=$?
 
           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")
@@ -174,14 +175,14 @@ jobs:
           # Do not fail the job if timeout killed the build
           exit $EXIT_CODE
 
-      - name: Log build logs
-        if: always()
+      - name: Log build logs after timeout
+        if: always() && steps.build_wheel.outputs.build_exit_code == 124
         run: |
           ls -al ./
           tar -cvf build.tar . --atime-preserve=replace
 
-      - name: Save build cache
-        if: always()
+      - name: Save build cache timeout
+        if: always() && steps.build_wheel.outputs.build_exit_code == 124
         uses: actions/cache/save@v4
         with:
           key: build-${{ inputs.release-version }}-${{ inputs.python-version }}-${{ inputs.cuda-version }}-${{ inputs.torch-version }}-${{ inputs.cxx11-abi }}-${{ github.run_number }}-${{ github.run_attempt }}