Serving encapsulates client and server work to perform batched requests.

Servings can be executed on the fly, without starting a server, but most often they are used to run servers that batch requests until a given size or timeout is reached.

More specifically, servings are a mechanism to apply a computation on a Nx.Batch, with hooks for preprocessing input from and postprocessing output for the client. Thus we can think of an instance of Nx.Serving.t() (a serving) as something that encapsulates batches of Nx computations.

inline-serverless-workflow

Inline/serverless workflow

First, let's define a simple numerical definition function:

defmodule MyDefn do
  import Nx.Defn

  defnp print_and_multiply(x) do
    print_value({:debug, x})
    x * 2
  end
end

The function prints the given tensor and doubles its contents. We can use new/1 to create a serving that will return a JIT or AOT compiled function to execute on batches of tensors:

iex> serving = Nx.Serving.new(fn opts -> Nx.Defn.jit(&print_and_multiply/1, opts) end)
iex> batch = Nx.Batch.stack([Nx.tensor([1, 2, 3])])
iex> Nx.Serving.run(serving, batch)
{:debug, #Nx.Tensor<
  s64[1][3]
  [
    [1, 2, 3]
  ]
>}
#Nx.Tensor<
  s64[1][3]
  [
    [2, 4, 6]
  ]
>

We started the serving by passing a function that receives compiler options and returns a JIT or AOT compiled function. We called Nx.Defn.jit/2 passing the options received as argument, which will customize the JIT/AOT compilation.

You should see two values printed. The former is the result of Nx.Defn.Kernel.print_value/1, which shows the tensor that was actually part of the computation and how it was batched. The latter is the result of the computation.

When defining a Nx.Serving, we can also customize how the data is batched by using the client_preprocessing as well as the result by using client_postprocessing hooks. Let's give it another try:

iex> serving = (
...>   Nx.Serving.new(fn opts -> Nx.Defn.jit(&print_and_multiply/1, opts) end)
...>   |> Nx.Serving.client_preprocessing(fn input -> {Nx.Batch.stack(input), :client_info} end)
...>   |> Nx.Serving.client_postprocessing(&{&1, &2, &3})
...> )
iex> Nx.Serving.run(serving, [Nx.tensor([1, 2]), Nx.tensor([3, 4])])
{:debug, #Nx.Tensor<
  s64[2][2]
  [
    [1, 2],
    [3, 4]
  ]
>}
{#Nx.Tensor<
   s64[2][2]
   [
     [2, 4],
     [6, 8]
   ]
 >,
 :server_info,
 :client_info}

You can see the results are a bit different now. First of all, notice that we were able to run the serving passing a list of tensors. Our custom client_preprocessing function stacks those tensors into a batch of two entries and returns a tuple with a Nx.Batch struct and additional client information which we represent as the atom :client_info. The default client preprocessing simply enforces a batch was given and returns no client information.

Then the result is a triplet tuple, returned by the client postprocessing function, containing the result, the server information (which we will later learn how to customize), and the client information. From this, we can infer the default implementation of client_postprocessing simply returns the result, discarding the server and client information.

So far, Nx.Serving has not given us much. It has simply encapsulated the execution of a function. Its full power comes when we start running our own Nx.Serving process. That's when we will also learn why we have a client_ prefix in some of the function names.

stateful-process-workflow

Stateful/process workflow

Nx.Serving allows us to define an Elixir process to handle requests. This process provides several features, such as batching up to a given size or time, partitioning, and distribution over a group of nodes.

To do so, we need to start a Nx.Serving process with a serving inside a supervision tree:

children = [
  {Nx.Serving,
   serving: Nx.Serving.new(Nx.Defn.jit(&print_and_multiply/1)),
   name: MyServing,
   batch_size: 10,
   batch_timeout: 100}
]

Supervisor.start_child(children, strategy: :one_for_one)

Now you can send batched runs to said process:

iex> batch = Nx.Batch.stack([Nx.tensor([1, 2, 3]), Nx.tensor([4, 5, 6])])
iex> Nx.Serving.batched_run(MyServing, batch)
{:debug, #Nx.Tensor<
  s64[2][3]
  [
    [1, 2, 3],
    [4, 5, 6]
  ]
>}
#Nx.Tensor<
  s64[2][3]
  [
    [2, 4, 6],
    [8, 10, 12]
  ]
>

In the example, we pushed a batch of 2 and eventually got a reply. The process will wait for requests from other processes, for up to 100 milliseconds or until it gets 10 entries. Then it merges all batches together and once the result is computed, it slices and distributes those responses to each caller.

If there is any client_preprocessing function, it will be executed before the batch is sent to the server. If there is any client_postprocessing function, it will be executed after getting the response from the server.

partitioning

Partitioning

You can start several partitions under th same serving by passing partitions: true when starting the serving. The number of partitions will be determined according your compiler and for which host it is compiling.

For example, when creating the serving, you may pass the following defn_options:

Nx.Serving.new(computation, compiler: EXLA, client: :cuda)

Now when booting up the serving:

children = [
  {Nx.Serving,
   serving: serving,
   name: MyServing,
   batch_size: 10,
   batch_timeout: 100,
   partitions: true}
]

If you have two GPUs, batched_run/3 will now gather batches and send them to the GPUs as they become available to process requests.

distribution

Distribution

All Nx.Servings are distributed by default. If the current machine does not have an instance of Nx.Serving running, batched_run/3 will automatically look for one in the cluster. The nodes do not need to run the same code and applications. It is only required that they run the same Nx version.

The load balancing between servings is done randomly, however, the number of partitions are considered if the partitioned: true option is also given. For example, if you have a node with 2 GPUs and another with 4, the latter will receive the double of requests compared to the former.

batched_run/3 receives an optional distributed_preprocessing callback as third argument for preprocessing the input for distributed requests. When using libraries like EXLA or Torchx, the tensor is often allocated in memory inside a third-party library so it is necessary to either transfer or copy the tensor to the binary backend before sending it to another node. This can be done by passing either Nx.backend_transfer/1 or Nx.backend_copy/1 as third argument:

Nx.Serving.batched_run(MyDistributedServing, input, &Nx.backend_copy/1)

Use backend_transfer/1 if you know the input will no longer be used.

Similarly, the serving has a distributed_postprocessing callback which can do equivalent before sending the reply to the caller.

The servings are dispatched using Erlang Distribution. You can use Node.connect/1 to manually connect nodes. In a production setup, this is often done with the help of libraries like libcluster.

advanced-notes

Advanced notes

module-based-serving

Module-based serving

In the examples so far, we have been using the default version of Nx.Serving, which executes the given function for each batch.

However, we can also use new/2 to start a module-based version of Nx.Serving which gives us more control over both inline and process workflows. A simple module implementation of a Nx.Serving could look like this:

defmodule MyServing do
  @behaviour Nx.Serving

  defnp print_and_multiply(x) do
    print_value({:debug, x})
    x * 2
  end

  @impl true
  def init(_inline_or_process, :unused_arg, [defn_options]) do
    {:ok, Nx.Defn.jit(&print_and_multiply/1, defn_options)}
  end

  @impl true
  def handle_batch(batch, 0, function) do
    {:execute, fn -> {function.(batch), :server_info} end, function}
  end
end

It has two functions. The first, init/3, receives the type of serving (:inline or :process) and the serving argument. In this step, we capture print_and_multiply/1as a jitted function.

The second function is called handle_batch/3. This function receives a Nx.Batch and returns a function to execute. The function itself must return a two element-tuple: the batched results and some server information. The server information can be any value and we set it to the atom :server_info.

Now let's give it a try by defining a serving with our module and then running it on a batch:

iex> serving = Nx.Serving.new(MyServing, :unused_arg)
iex> batch = Nx.Batch.stack([Nx.tensor([1, 2, 3])])
iex> Nx.Serving.run(serving, batch)
{:debug, #Nx.Tensor<
  s64[1][3]
  [
    [1, 2, 3]
  ]
>}
#Nx.Tensor<
  s64[1][3]
  [
    [2, 4, 6]
  ]
>

From here on, you use start_link/1 to start this serving in your supervision and even customize client_preprocessing/1 and client_postprocessing/1 callbacks to this serving, as seen in the previous sections.

Note in our implementation above assumes it won't run partitioned. In partitioned mode, init/3 may receive multiple defn_options as the third argument and handle_batch/3 may receive another partition besides 0.