#deep-learning #middleware #backpressure #batched-prediction


Middleware for serving deep learning models with batched prediction

14 releases

0.2.2 Dec 17, 2020
0.2.1 Sep 15, 2020
0.2.0 Apr 7, 2020
0.1.10 Apr 3, 2020
0.1.7 Mar 31, 2020

#9 in Machine learning

Download history 3/week @ 2020-09-25 15/week @ 2020-10-02 12/week @ 2020-10-16 16/week @ 2020-10-23 17/week @ 2020-10-30 1/week @ 2020-11-06 16/week @ 2020-11-13 15/week @ 2020-11-20 31/week @ 2020-11-27 40/week @ 2020-12-04 32/week @ 2020-12-11 16/week @ 2020-12-18 5/week @ 2020-12-25 18/week @ 2021-01-01 11/week @ 2021-01-08

63 downloads per month


188 lines


Rust middleware for serving deep learning models with batched prediction.

Build License Crates Docs

Deep learning models are usually implemented to make efficient use of a GPU by batching inputs together in "mini-batches". However, applications serving these models often receive requests one-by-one. So using a conventional single or multi-threaded server approach will under-utilize the GPU and lead to latency that increases linearly with the volume of requests.

batched-fn is middleware for deep learning services that queues individual requests and provides them as a batch to your model. It can be added to any application with minimal refactoring simply by inserting the batched_fn! macro into the function that runs requests through the model. The trade-off is a small delay incurred while waiting for a batch to be filled, though this can be tuned with the max_delay and max_batch_size config parameters.


  • 🚀 Easy to use: drop the batched_fn! macro into existing code.
  • 🔥 Lightweight and fast: queue system implemented on top of the blazingly fast flume crate.
  • 🙌 Easy to tune: simply adjust max_delay and max_batch_size.
  • 🛑 Back pressure mechanism included: just set the channel_cap config parameter.


Suppose you have a model API that look like this:

// `Batch` could be anything that implements the `batched_fn::Batch` trait.
type Batch<T> = Vec<T>;

struct Input {
    // ...

struct Output {
    // ...

struct Model {
    // ...

impl Model {
    fn predict(&self, batch: Batch<Input>) -> Batch<Output> {
        // ...

    fn load() -> Self {
        // ...

Without batched-fn a webserver route would need to call Model::predict on each individual input, resulting in a bottleneck from under-utilizing the GPU:

use once_cell::sync::Lazy;

static MODEL: Lazy<Model> = Lazy::new(Model::load);

fn predict_for_http_request(input: Input) -> Output {
    let mut batched_input = Batch::with_capacity(1);

But by dropping the batched_fn macro into your code you automatically get batched inference behind the scenes without changing the one-to-one relationship between inputs and outputs:

async fn predict_for_http_request(input: Input) -> Output {
    let batch_predict = batched_fn! {
        handler = |batch: Batch<Input>, model: &Model| -> Batch<Output> {
        config = {
            max_batch_size: 16,
            max_delay: 50,
        context = {
            model: Model::load(),

❗️ Note that the predict_for_http_request function now has to be async.

Here we set the max_batch_size to 16 and max_delay to 50 milliseconds. This means the batched function will wait at most 50 milliseconds after receiving a single input to fill a batch of 16. If 15 more inputs are not received within 50 milliseconds then the partial batch will be ran as-is.

Tuning max batch size and max delay

The optimal batch size and delay will depend on the specifics of your use case, such as how big of a batch you can fit in memory (typically on the order of 8, 16, 32, or 64 for a deep learning model) and how long of a delay you can afford. In general you want to set both of these as high as you can.

It's worth noting that the response time of your application might actually go down under high load. This is because the batch handler will be called as soon as either a batch of max_batch_size is filled or max_delay milliseconds has passed, whichever happens first. So under high load batches will be filled quickly, but under low load the response time will be at least max_delay milliseconds (adding the time it takes to actually process a batch and respond).


~23K SLoC