mamba env create -f hf-train-cu118.yaml --strict-channel-priority
mamba activate hf-train-cu118

pip install --upgrade pip
pip install --no-deps -r requirements-hf.txt --constraint constraints-cu118.txt

# 需要 deepspeed 时再装：
DS_BUILD_OPS=0 pip install "deepspeed==0.14.*"  # 先不编译 CUDA 内核