Import flash_mla from sgl-kernel (#12135)

This commit is contained in:
Baizhou Zhang
2025-10-29 23:54:21 -07:00
committed by GitHub
parent fb52d35f63
commit 621dfb8886
8 changed files with 9 additions and 63 deletions

View File

@@ -3,7 +3,6 @@
set -euxo pipefail
IS_BLACKWELL=${IS_BLACKWELL:-0}
RUN_DEEPSEEK_V32=${RUN_DEEPSEEK_V32:-0}
CU_VERSION="cu129"
if [ "$CU_VERSION" = "cu130" ]; then
@@ -113,22 +112,6 @@ if [ "$IS_BLACKWELL" != "1" ]; then
$PIP_CMD install xformers --index-url https://download.pytorch.org/whl/${CU_VERSION} --no-deps $PIP_INSTALL_SUFFIX
fi
# Install dependencies for deepseek-v3.2
if [ "$RUN_DEEPSEEK_V32" = "1" ]; then
# Install flashmla
FLASHMLA_COMMIT="1408756a88e52a25196b759eaf8db89d2b51b5a1"
FLASH_MLA_DISABLE_SM100="0"
if [ "$IS_BLACKWELL" != "1" ]; then
FLASH_MLA_DISABLE_SM100="1"
fi
git clone https://github.com/deepseek-ai/FlashMLA.git flash-mla
cd flash-mla
git checkout ${FLASHMLA_COMMIT}
git submodule update --init --recursive
FLASH_MLA_DISABLE_SM100=${FLASH_MLA_DISABLE_SM100} $PIP_CMD install -v . $PIP_INSTALL_SUFFIX --no-build-isolation
cd ..
fi
# Show current packages
$PIP_CMD list
python3 -c "import torch; print(torch.version.cuda)"