Skip to content

Latest commit

 

History

History
77 lines (61 loc) · 2.65 KB

README.md

File metadata and controls

77 lines (61 loc) · 2.65 KB
Raw

pygloo

Pygloo provides Python bindings for gloo. It is implemented using pybind11.

It is currenlty used in Ray for collective communication between CPUs.

Requirements

Python >= 3.6

Installation

Install From Wheels

We provide prepackaged Python wheels (manylinux2014_x86_64,manylinux_2_24_x86_64). To install from wheels:

pip install pygloo

Building from source

One can build pygloo from source if none of released wheels fit with the development environment.

Pygloo uses Bazel to automatically manange dependencies and compilation. To compile from source, install Bazel>=2.0.0 following the Bazel installation guide. After installing Bazel, build and install pygloo following this command:

python setup.py install

Testing

Pygloo uses Ray to create multiple, distributed processes for collective communication tests. See tests directory.

Example

An example for allreduce.

import os
import ray
import pygloo
import numpy as np

@ray.remote(num_cpus=1)
def test_allreduce(rank, world_size, fileStore_path):
    '''
    rank  # Rank of this process within list of participating processes
    world_size  # Number of participating processes
    fileStore_path # The path to create filestore
    '''
    context = pygloo.rendezvous.Context(rank, world_size)
    # Prepare device and store for rendezvous
    attr = pygloo.transport.tcp.attr("localhost")
    dev = pygloo.transport.tcp.CreateDevice(attr)
    fileStore = pygloo.rendezvous.FileStore(fileStore_path)
    store = pygloo.rendezvous.PrefixStore(str(world_size), fileStore)

    context.connectFullMesh(store, dev)

    sendbuf = np.array([[1,2,3],[1,2,3]], dtype=np.float32)
    recvbuf = np.zeros_like(sendbuf, dtype=np.float32)
    sendptr = sendbuf.ctypes.data
    recvptr = recvbuf.ctypes.data

    pygloo.allreduce(context, sendptr, recvptr,
                    sendbuf.size, pygloo.glooDataType_t.glooFloat32,
                    pygloo.ReduceOp.SUM, pygloo.allreduceAlgorithm.RING)

if __name__ == "__main__":
    ray.init(num_cpus=6)
    world_size = 2
    fileStore_path = f"{ray.worker._global_node.get_session_dir_path()}" + "/collective/gloo/rendezvous"
    os.makedirs(fileStore_path)
    ray.get([test_allreduce.remote(rank, world_size, fileStore_path) for rank in range(world_size)])

License

Gloo is licensed under the Apache License, Version 2.0.