Dependency clash for numpy when installing from scratch using pip

Hello,

Thanks for this library. I have access to a SLURM-based system that doesn't support Docker images. Hence, I have been trying to install your library directly using PIP using the Docker images as reference. However, it seems that the numpy version that you state (see https://github.com/alibaba/ROLL/blob/main/docker/Dockerfile.torch260.vllm#L20) is incompatible with the numpy version required by other libraries that are installed in the following lines, such as `transformer-engine` (i.e., transformer-engine seems to require numpy > 2, while you install numpy < 2).

How shall I solve this? What is the recommended way to install this library?

Start by reading docker/Dockerfile.torch260.vllm around line 20 and the subsequent dependency installation lines, then compare the pinned numpy constraint with transformer-engine's requirement. Done means there is a documented or reproducible pip installation path for SLURM users without Docker with compatible dependency constraints.

Written by the indexing model from the issue text.