Remove flash-attn GIT_TAG override to main — causes FLASHATTENTION_FP8_TWO_LEVEL_INTERVAL undefined error
v0.19.0 pins a compatible flash-attn commit (2921022). The sed that forced GIT_TAG to main pulled in newer code that references FLASHATTENTION_FP8_TWO_LEVEL_INTERVAL which isn't defined in v0.19.0's build config. Use the pinned commit instead.
This commit is contained in:
@@ -166,7 +166,9 @@ RUN cd vllm && \
|
||||
echo "========================================\n\n" && \
|
||||
git submodule sync && \
|
||||
git submodule update --init --recursive -j 8 && \
|
||||
sed -i 's/GIT_TAG [a-f0-9]\{40\}/GIT_TAG main/' cmake/external_projects/vllm_flash_attn.cmake && \
|
||||
# NOTE: Removed the sed that forced flash-attn GIT_TAG to main.
|
||||
# v0.19.0 pins a compatible commit; building from main causes
|
||||
# FLASHATTENTION_FP8_TWO_LEVEL_INTERVAL undefined errors.
|
||||
sed -i 's/register_opaque_type(ModuleName, typ="value", hoist=True)/register_opaque_type(ModuleName, typ="value")/' vllm/utils/torch_utils.py && \
|
||||
export MAX_JOBS=8 && \
|
||||
export CMAKE_BUILD_PARALLEL_LEVEL=$MAX_JOBS && \
|
||||
|
||||
Reference in New Issue
Block a user