Closed kfertakis closed 2 months ago
We usually use docker to solve this problem, refer to our dockerfiles https://github.com/OpenLLMAI/OpenRLHF/tree/main/dockerfile
Hi, pip install packaging ninja then, you can download specified version from https://github.com/Dao-AILab/flash-attention/releases.
Hi,
building the project in a python venv using
setup.py
has the following issues.wheel
andpackaging
are missing from requirements.txt required byflash-attn
train_ppo_llama.sh
leads toflash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN2at4_ops5zeros4c...
error. The only way I could overcome this is by manually rebuilding flash-attn withFLASH_ATTENTION_FORCE_BUILD=TRUE pip install --force-reinstall flash-attn
which takes a long time. Is there a better way to setup the project?Thanks
Environment: Ubuntu 20.04 Python 3.10 CUDA 12.1