mirror of
https://github.com/BillyOutlast/flash-attention-prebuild-wheels-rocm.git
synced 2026-07-01 01:27:54 -04:00
Junya Morioka
82 lines
2.8 KiB
Bash
Executable File
82 lines
2.8 KiB
Bash
Executable File
#!/bin/bash
|
|
|
|
set -e
|
|
|
|
# Parameters with defaults
|
|
FLASH_ATTN_VERSION=$1
|
|
PYTHON_VERSION=$2
|
|
TORCH_VERSION=$3
|
|
CUDA_VERSION=$4
|
|
|
|
echo "Building Flash Attention with parameters:"
|
|
echo " Flash-Attention: $FLASH_ATTN_VERSION"
|
|
echo " Python: $PYTHON_VERSION"
|
|
echo " PyTorch: $TORCH_VERSION"
|
|
echo " CUDA: $CUDA_VERSION"
|
|
|
|
# Set CUDA and PyTorch versions
|
|
MATRIX_CUDA_VERSION=$(echo $CUDA_VERSION | awk -F \. {'print $1 $2'})
|
|
MATRIX_TORCH_VERSION=$(echo $TORCH_VERSION | awk -F \. {'print $1 "." $2'})
|
|
|
|
echo "Derived versions:"
|
|
echo " CUDA Matrix: $MATRIX_CUDA_VERSION"
|
|
echo " Torch Matrix: $MATRIX_TORCH_VERSION"
|
|
|
|
# Install PyTorch
|
|
TORCH_CUDA_VERSION=$(python get_torch_cuda_version.py $MATRIX_CUDA_VERSION $MATRIX_TORCH_VERSION)
|
|
|
|
echo "Installing PyTorch $TORCH_VERSION+cu$TORCH_CUDA_VERSION..."
|
|
if [[ $TORCH_VERSION == *"dev"* ]]; then
|
|
pip install --force-reinstall --no-cache-dir --pre torch==$TORCH_VERSION --index-url https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}
|
|
else
|
|
pip install --force-reinstall --no-cache-dir torch==$TORCH_VERSION --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
|
|
fi
|
|
|
|
# Verify installation
|
|
echo "Verifying installations..."
|
|
nvcc --version
|
|
python -V
|
|
python -c "import torch; print('PyTorch:', torch.__version__)"
|
|
python -c "import torch; print('CUDA:', torch.version.cuda)"
|
|
python -c "from torch.utils import cpp_extension; print(cpp_extension.CUDA_HOME)"
|
|
|
|
# Checkout flash-attn
|
|
echo "Checking out flash-attention v$FLASH_ATTN_VERSION..."
|
|
git clone https://github.com/Dao-AILab/flash-attention.git -b "v$FLASH_ATTN_VERSION"
|
|
|
|
# Determine MAX_JOBS and NVCC_THREADS based on system resources
|
|
NUM_THREADS=$(nproc)
|
|
RAM_GB=$(free -g | awk '/^Mem:/{print $2}')
|
|
echo "System resources:"
|
|
echo " CPU threads: $NUM_THREADS"
|
|
echo " RAM: ${RAM_GB}GB"
|
|
|
|
# Calculate max product based on constraints:
|
|
# - MAX_JOBS x NVCC_THREADS <= NUM_THREADS
|
|
# - 4GB x MAX_JOBS x NVCC_THREADS <= RAM_GB
|
|
MAX_PRODUCT_CPU=$NUM_THREADS
|
|
MAX_PRODUCT_RAM=$((RAM_GB / 4))
|
|
MAX_PRODUCT=$((MAX_PRODUCT_CPU < MAX_PRODUCT_RAM ? MAX_PRODUCT_CPU : MAX_PRODUCT_RAM))
|
|
|
|
# Set MAX_JOBS = NVCC_THREADS = floor(sqrt(MAX_PRODUCT))
|
|
# This balances parallelism across both dimensions
|
|
MAX_JOBS=$(awk -v max="$MAX_PRODUCT" 'BEGIN {print int(sqrt(max))}')
|
|
NVCC_THREADS=$MAX_JOBS
|
|
|
|
# Ensure minimum values of 1
|
|
MAX_JOBS=$((MAX_JOBS < 1 ? 1 : MAX_JOBS))
|
|
NVCC_THREADS=$((NVCC_THREADS < 1 ? 1 : NVCC_THREADS))
|
|
|
|
echo "Build parallelism settings:"
|
|
echo " MAX_JOBS: $MAX_JOBS"
|
|
echo " NVCC_THREADS: $NVCC_THREADS"
|
|
|
|
# Build wheels
|
|
echo "Building wheels..."
|
|
cd flash-attention
|
|
LOCAL_VERSION_LABEL="cu${MATRIX_CUDA_VERSION}torch${MATRIX_TORCH_VERSION}"
|
|
NVCC_THREADS=$NVCC_THREADS MAX_JOBS=$MAX_JOBS FLASH_ATTENTION_FORCE_BUILD=TRUE FLASH_ATTN_LOCAL_VERSION=${LOCAL_VERSION_LABEL} \
|
|
python setup.py bdist_wheel --dist-dir=dist
|
|
wheel_name=$(basename $(ls dist/*.whl | head -n 1))
|
|
echo "Built wheel: $wheel_name"
|