Distributed Communication#

MLX supports distributed communication operations that allow the computational cost of training or inference to be shared across many physical machines. At the moment we support two different communication backends:

MPI a full-featured and mature distributed communications library
A ring backend of our own that uses native TCP sockets and should be faster for thunderbolt connections.

The list of all currently supported operations and their documentation can be seen in the API docs.

Note

Some operations may not be supported or not as fast as they should be. We are adding more and tuning the ones we have as we are figuring out the best way to do distributed computing on Macs using MLX.

Getting Started#

A distributed program in MLX is as simple as:

import mlx.core as mx

world = mx.distributed.init()
x = mx.distributed.all_sum(mx.ones(10))
print(world.rank(), x)

The program above sums the array mx.ones(10) across all distributed processes. However, when this script is run with python only one process is launched and no distributed communication takes place. Namely, all operations in mx.distributed are noops when the distributed group has a size of one. This property allows us to avoid code that checks if we are in a distributed setting similar to the one below:

import mlx.core as mx

x = ...
world = mx.distributed.init()
# No need for the check we can simply do x = mx.distributed.all_sum(x)
if world.size() > 1:
    x = mx.distributed.all_sum(x)

Running Distributed Programs#

MLX provides mlx.launch a helper script to launch distributed programs. Continuing with our initial example we can run it on localhost with 4 processes using

$ mlx.launch -n 4 my_script.py
array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
array([4, 4, 4, ..., 4, 4, 4], dtype=float32)

We can also run it on some remote hosts by providing their IPs (provided that the script exists on all hosts and they are reachable by ssh)

$ mlx.launch --hosts ip1,ip2,ip3,ip4 my_script.py
array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
array([4, 4, 4, ..., 4, 4, 4], dtype=float32)

Consult the dedicated usage guide for more information on using mlx.launch.

Selecting Backend#

You can select the backend you want to use when calling init() by passing one of {'any', 'ring', 'mpi'}. When passing any, MLX will try to initialize the ring backend and if it fails the mpi backend. If they both fail then a singleton group is created.

Note

After a distributed backend is successfully initialized init() will return the same backend if called without arguments or with backend set to any.

The following examples aim to clarify the backend initialization logic in MLX:

# Case 1: Initialize MPI regardless if it was possible to initialize the ring backend
world = mx.distributed.init(backend="mpi")
world2 = mx.distributed.init()  # subsequent calls return the MPI backend!

# Case 2: Initialize any backend
world = mx.distributed.init(backend="any")  # equivalent to no arguments
world2 = mx.distributed.init()  # same as above

# Case 3: Initialize both backends at the same time
world_mpi = mx.distributed.init(backend="mpi")
world_ring = mx.distributed.init(backend="ring")
world_any = mx.distributed.init()  # same as MPI because it was initialized first!

Training Example#

In this section we will adapt an MLX training loop to support data parallel distributed training. Namely, we will average the gradients across a set of hosts before applying them to the model.

Our training loop looks like the following code snippet if we omit the model, dataset and optimizer initialization.

model = ...
optimizer = ...
dataset = ...

def step(model, x, y):
    loss, grads = loss_grad_fn(model, x, y)
    optimizer.update(model, grads)
    return loss

for x, y in dataset:
    loss = step(model, x, y)
    mx.eval(loss, model.parameters())

All we have to do to average the gradients across machines is perform an all_sum() and divide by the size of the Group. Namely we have to mlx.utils.tree_map() the gradients with following function.

def all_avg(x):
    return mx.distributed.all_sum(x) / mx.distributed.init().size()

Putting everything together our training loop step looks as follows with everything else remaining the same.

from mlx.utils import tree_map

def all_reduce_grads(grads):
    N = mx.distributed.init().size()
    if N == 1:
        return grads
    return tree_map(
        lambda x: mx.distributed.all_sum(x) / N,
        grads
    )

def step(model, x, y):
    loss, grads = loss_grad_fn(model, x, y)
    grads = all_reduce_grads(grads)  # <--- This line was added
    optimizer.update(model, grads)
    return loss

Utilizing `nn.average_gradients`#

Although the code example above works correctly; it performs one communication per gradient. It is significantly more efficient to aggregate several gradients together and perform fewer communication steps.

This is the purpose of mlx.nn.average_gradients(). The final code looks almost identical to the example above:

model = ...
optimizer = ...
dataset = ...

def step(model, x, y):
    loss, grads = loss_grad_fn(model, x, y)
    grads = mlx.nn.average_gradients(grads) # <---- This line was added
    optimizer.update(model, grads)
    return loss

for x, y in dataset:
    loss = step(model, x, y)
    mx.eval(loss, model.parameters())

Getting Started with MPI#

MLX already comes with the ability to “talk” to MPI if it is installed on the machine. Launching distributed MLX programs that use MPI can be done with mpirun as expected. However, in the following examples we will be using mlx.launch --backend mpi which takes care of some nuisances such as setting absolute paths for the mpirun executable and the libmpi.dyld shared library.

The simplest possible usage is the following which, assuming the minimal example in the beginning of this page, should result in:

$ mlx.launch --backend mpi -n 2 test.py
1 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
0 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)

The above launches two processes on the same (local) machine and we can see both standard output streams. The processes send the array of 1s to each other and compute the sum which is printed. Launching with mlx.launch -n 4 ... would print 4 etc.

Installing MPI#

MPI can be installed with Homebrew, using the Anaconda package manager or compiled from source. Most of our testing is done using openmpi installed with the Anaconda package manager as follows:

$ conda install conda-forge::openmpi

Installing with Homebrew may require specifying the location of libmpi.dyld so that MLX can find it and load it at runtime. This can simply be achieved by passing the DYLD_LIBRARY_PATH environment variable to mpirun and it is done automatically by mlx.launch.

$ mpirun -np 2 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python test.py
$ # or simply
$ mlx.launch -n 2 test.py

Setting up Remote Hosts#

MPI can automatically connect to remote hosts and set up the communication over the network if the remote hosts can be accessed via ssh. A good checklist to debug connectivity issues is the following:

ssh hostname works from all machines to all machines without asking for password or host confirmation
mpirun is accessible on all machines.
Ensure that the hostname used by MPI is the one that you have configured in the .ssh/config files on all machines.

Tuning MPI All Reduce#

Note

For faster all reduce consider using the ring backend either with Thunderbolt connections or over Ethernet.

Configure MPI to use N tcp connections between each host to improve bandwidth by passing --mca btl_tcp_links N.

Force MPI to use the most performant network interface by setting --mca btl_tcp_if_include <iface> where <iface> should be the interface you want to use.

Getting Started with Ring#

The ring backend does not depend on any third party library so it is always available. It uses TCP sockets so the nodes need to be reachable via a network. As the name suggests the nodes are connected in a ring which means that rank 1 can only communicate with rank 0 and rank 2, rank 2 only with rank 1 and rank 3 and so on and so forth. As a result send() and recv() with arbitrary sender and receiver is not supported in the ring backend.

Defining a Ring#

The easiest way to define and use a ring is via a JSON hostfile and the mlx.launch helper script. For each node one defines a hostname to ssh into to run commands on this node and one or more IPs that this node will listen to for connections.

For example the hostfile below defines a 4 node ring. hostname1 will be rank 0, hostname2 rank 1 etc.

[
    {"ssh": "hostname1", "ips": ["123.123.123.1"]},
    {"ssh": "hostname2", "ips": ["123.123.123.2"]},
    {"ssh": "hostname3", "ips": ["123.123.123.3"]},
    {"ssh": "hostname4", "ips": ["123.123.123.4"]}
]

Running mlx.launch --hostfile ring-4.json my_script.py will ssh into each node, run the script which will listen for connections in each of the provided IPs. Specifically, hostname1 will connect to 123.123.123.2 and accept a connection from 123.123.123.4 and so on and so forth.

Thunderbolt Ring#

Although the ring backend can have benefits over MPI even for Ethernet, its main purpose is to use Thunderbolt rings for higher bandwidth communication. Setting up such thunderbolt rings can be done manually, but is a relatively tedious process. To simplify this, we provide the utility mlx.distributed_config.

To use mlx.distributed_config your computers need to be accessible by ssh via Ethernet or Wi-Fi. Subsequently, connect them via thunderbolt cables and then call the utility as follows:

mlx.distributed_config --verbose --hosts host1,host2,host3,host4

By default the script will attempt to discover the thunderbolt ring and provide you with the commands to configure each node as well as the hostfile.json to use with mlx.launch. If password-less sudo is available on the nodes then --auto-setup can be used to configure them automatically.

To validate your connection without configuring anything mlx.distributed_config can also plot the ring using DOT format.

mlx.distributed_config --verbose --hosts host1,host2,host3,host4 --dot >ring.dot
dot -Tpng ring.dot >ring.png
open ring.png

If you want to go through the process manually, the steps are as follows:

Disable the thunderbolt bridge interface
For the cable connecting rank i to rank i + 1 find the interfaces corresponding to that cable in nodes i and i + 1.
Set up a unique subnetwork connecting the two nodes for the corresponding interfaces. For instance if the cable corresponds to en2 on node i and en2 also on node i + 1 then we may assign IPs 192.168.0.1 and 192.168.0.2 respectively to the two nodes. For more details you can see the commands prepared by the utility script.

Distributed Communication

Contents

Distributed Communication#

Getting Started#

Running Distributed Programs#

Selecting Backend#

Training Example#

Utilizing nn.average_gradients#

Getting Started with MPI#

Installing MPI#

Setting up Remote Hosts#

Tuning MPI All Reduce#

Getting Started with Ring#

Defining a Ring#

Thunderbolt Ring#

Utilizing `nn.average_gradients`#