mamba env create -f env.yaml --strict-channel-priority
mamba activate hf-train-cu118
python -m pip install --no-deps -r requirements-hf.txt

# deepspeed（选一）
DS_BUILD_OPS=0 python -m pip install "deepspeed==0.14.*"
# 或
mamba install -y cuda-toolkit=11.8 cmake ninja
python -m pip install "deepspeed==0.14.*"