Limit to MAX_JOBS=1 with CUDA 12.2

.github/workflows/publish.yml

@@ -168,7 +168,8 @@ jobs:
           export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
           export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
           # Limit MAX_JOBS otherwise the github runner goes OOM
-          MAX_JOBS=2 FLASH_ATTENTION_FORCE_BUILD="TRUE" FLASH_ATTENTION_FORCE_CXX11_ABI=${{ matrix.cxx11_abi}} python setup.py bdist_wheel --dist-dir=dist
+          # CUDA 11.8 can compile with 2 jobs, but CUDA 12.2 goes OOM
+          MAX_JOBS=$([ "$MATRIX_CUDA_VERSION" == "122" ] && echo 1 || echo 2) FLASH_ATTENTION_FORCE_BUILD="TRUE" FLASH_ATTENTION_FORCE_CXX11_ABI=${{ matrix.cxx11_abi}} python setup.py bdist_wheel --dist-dir=dist
           tmpname=cu${MATRIX_CUDA_VERSION}torch${MATRIX_TORCH_VERSION}cxx11abi${{ matrix.cxx11_abi }}
           wheel_name=$(ls dist/*whl | xargs -n 1 basename | sed "s/-/+$tmpname-/2")
           ls dist/*whl |xargs -I {} mv {} dist/${wheel_name}

flash_attn/__init__.py

-__version__ = "2.5.9"
+__version__ = "2.5.9.post1"
 
 from flash_attn.flash_attn_interface import (
     flash_attn_func,

training/Dockerfile

@@ -85,7 +85,7 @@ RUN pip install transformers==4.25.1 datasets==2.8.0 pytorch-lightning==1.8.6 tr
 RUN pip install git+https://github.com/mlcommons/logging.git@2.1.0
 
 # Install FlashAttention
-RUN pip install flash-attn==2.5.9
+RUN pip install flash-attn==2.5.9.post1
 
 # Install CUDA extensions for fused dense
-RUN pip install git+https://github.com/HazyResearch/flash-attention@v2.5.9#subdirectory=csrc/fused_dense_lib
+RUN pip install git+https://github.com/HazyResearch/flash-attention@v2.5.9.post1#subdirectory=csrc/fused_dense_lib