1-bit Adam uses advanced communication schemes that are not yet supported by PyTorch distributed and NCCL. We rely on Message Passing Interface (MPI) for these advanced communication primitives.
We package the necessary dependencies in the DeepSpeed docker images. However, if you are using a different build system, please install MPI and mpi4py on your system. We have tested CUDA-Aware MPI communication using the [MVAPICH2-GDR](http://mvapich.cse.ohio-state.edu/userguide/gdr/) library. However, any CUDA-Aware communication library including [OpenMPI](https://www.open-mpi.org/) should work fine with these examples.
We package the necessary dependencies in the DeepSpeed docker images. However, if you are using a different build system, please install MPI and mpi4py on your system. To install the prerequisites run:
```shell
pip install deepspeed[1bit_adam]
```
We have tested CUDA-Aware MPI communication using the [MVAPICH2-GDR](http://mvapich.cse.ohio-state.edu/userguide/gdr/) library. However, any CUDA-Aware communication library including [OpenMPI](https://www.open-mpi.org/) should work fine with these examples.
An example launch command for 1-bit Adam using the `deepspeed` launcher is as follows:
