[Iluvatar GPU] Optimize attention performance and fix moe load ckpt error (#3651)

This commit is contained in:
yzwu
2025-09-22 21:13:59 +08:00
committed by GitHub
parent 5532e8a323
commit 504461b6b5
17 changed files with 1344 additions and 363 deletions

View File

@@ -13,10 +13,10 @@ python -m pip install -r requirements_iluvatar.txt
echo "uninstall org"
python -m pip uninstall paddlepaddle -y
python -m pip uninstall paddle-iluvatar-gpu -y
python -m pip install --pre paddlepaddle==3.0.0.dev20250708 -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
# TODO: Change to open access URL
python -m pip install --pre paddle-iluvatar-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/ixuca/
# python -m pip install /data1/fastdeploy/packages/paddle_iluvatar_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
# python -m pip install --pre paddlepaddle==3.0.0.dev20250708 -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
# python -m pip install --pre paddle-iluvatar-gpu==3.0.0.dev20250806 -i https://www.paddlepaddle.org.cn/packages/nightly/ixuca/
python -m pip install paddlepaddle==3.1.1 -i https://www.paddlepaddle.org.cn/packages/stable/cpu/
python -m pip install paddle-iluvatar-gpu==3.1.1 -i https://www.paddlepaddle.org.cn/packages/stable/ixuca/
# Patch, remove if image updated
cp /data1/fastdeploy/packages/cusolver.h /usr/local/lib/python3.10/site-packages/paddle/include/paddle/phi/backends/dynload/cusolver.h
echo "build whl"