Environment:

1. Framework: PyTorch
2. Framework version: 1.9.1
3. Horovod version: v0.23.0
4. CUDA version: 11.2
5. NCCL version: 2.11.4
6. Python version: 3.9

Hello,
I have installed Horovod with a customer nccl library. But everytime I modify the nccl library, I need to rebuild horovod with following steps to make sure it uses modified nccl:

cd ~/horovod
python setup.py sdict
HOROVOD_NCCL_HOME=/path/to/nccl HOROVOD_GPU_OPERATIONS=NCCL pip install --no-cache-dir dist/horovod-0.23.0.tar.gz

Is this a must? Any suggestion to simplify the rebuild process?