burn

---

Training and inference usually live in separate worlds. Models are typically trained in Python then exported to an open format like ONNX or optimized for production engines like vLLM, ONNX Runtime, or TensorRT. This export step is often brittle and lossy, ruling out complex architectures and advanced deployment use cases.

Burn unifies the two. By executing multi-platform tensor operations via a single, unified API, the exact code used for training is the exact code that runs in production. This makes workloads like on-device personalization and federated learning straightforward, while enabling teams to go from prototype to deployment in a single codebase.

Burn preserves the intuitive ergonomics of PyTorch, with dynamic shapes and graphs, but JIT-compiles streams of tensor operations, performing automatic kernel fusion. You get the flexibility of dynamic graphs without the performance drop.

Rust for Research?

Rust used to be a tough sell for research: long compilation times disrupted the fast edit-compile-run loop that draws researchers to Python. Burn changes this paradigm. Designed around incremental compilation, modifying model code recompiles in under 5 seconds, even in release mode. This delivers a Python-like feedback loop with the speed and safety of Rust.

Ecosystem

Burn is the core of a growing, fully open-source Rust AI ecosystem. You are not adopting a single library, you are joining a stack that spans GPU compute, model interop and domain toolkits, with plenty of room to help shape what comes next.

Category	Project	Description
Compute	CubeCL	GPU compute language and compiler behind Burn's accelerated backends. Write kernels once in Rust, run on CUDA, ROCm, Metal, Vulkan and WebGPU. Usable standalone.
Model interop	burn-onnx	Import ONNX models into Burn as native Rust code
	burn-store	Save, load and import model weights, including PyTorch and Safetensors
Domains	burn-vision	Computer vision operators and building blocks
	burn-rl	Reinforcement learning building blocks
	burn-dataset	Dataset loading, transforms and ready-made sources
Models	models	Curated pre-trained models and examples built with Burn
Tooling	burn-bench	Benchmark and compare backends, tracking performance over time

Burn's CubeCL backends (CUDA, ROCm, Metal, Vulkan, WebGPU, CPU) compose with autodiff, fusion and remote-execution decorators, while the pure-Rust Flex CPU/no_std backend composes with autodiff only. See Supported Backends below for the full matrix.

Every project here is open-source and actively developed. Want to help build the Rust AI ecosystem? The good first issues are a great place to start, and the Contributing guide will get you set up.

These crates are not maintained by Tracel, but they are part of the same Rust AI story. Anything that helps you load data, build environments, or ship models belongs here. Built something that fits? Open a PR to add it!

Category	Crate	Description
Data & loading	polars	Fast DataFrames for tabular data
	arrow-rs	Apache Arrow columnar memory format
	image	Image decoding, encoding and processing
	hf-hub	Download models and datasets from the Hugging Face Hub
Tokenization & NLP	tokenizers	Fast, production-ready tokenizers
	rust-bert	Ready-to-use NLP pipelines and transformer models
Numerical & linear algebra	ndarray	N-dimensional arrays
	nalgebra	Linear algebra
Classical ML	linfa	Classical ML toolkit, in the spirit of scikit-learn
	smartcore	Classical ML algorithms, no BLAS/LAPACK required
Inference & runtimes	candle	Minimalist ML framework with a focus on LLM inference
	mistral.rs	Fast, multimodal LLM inference engine
	ort	ONNX Runtime bindings for hardware-accelerated inference
	tract	Pure-Rust inference for ONNX and NNEF models
	wonnx	100% Rust, WebGPU-accelerated ONNX runtime for native and the web
LLM apps & RAG	rig	Build modular LLM applications and agents
	langchain-rust	LangChain-style chain orchestration
Embeddings & vector search	fastembed	Generate text embeddings and rerank locally
	qdrant	Vector search engine, written in Rust
	lancedb	Embedded, developer-friendly vector database
Computer vision	kornia-rs	Low-level 3D computer vision library
Simulation & environments	rapier	Physics engine for robotics and RL environments
Visualization	rerun	Multimodal data and CV/robotics visualization
	plotters	Plotting and charting

Backend

Burn strives to be as fast as possible on as many hardwares as possible, with robust implementations. We believe this flexibility is crucial for modern needs where you may train your models in the cloud, then deploy on customer hardwares, which vary from user to user.

Supported Backends

Most backends support all operating systems, so we don't mention them in the tables below.

GPU Backends:

	CUDA	ROCm	Metal	Vulkan	WebGPU
Nvidia	☑️	-	-	☑️	☑️
AMD	-	☑️	-	☑️	☑️
Apple	-	-	☑️	-	☑️
Intel	-	-	-	☑️	☑️
Qualcom	-	-	-	☑️	☑️
Wasm	-	-	-	-	☑️

CPU Backends:

	Cpu (CubeCL)	Flex
X86	☑️	☑️
Arm	☑️	☑️
Wasm	-	☑️
no-std	-	☑️

The two native CPU backends are independent. Cpu (cpu feature) is the CubeCL runtime for the CPU: it JIT-compiles the same kernels as the GPU backends through LLVM and supports fusion. Flex (flex feature) is a pure-Rust eager backend with no native dependencies that also runs on Wasm and no_std.

Migration: LibTorch was deprecated in 0.22.0 and has been removed from main. For GPU acceleration, use a CubeCL backend (CUDA, ROCm, Metal, Vulkan, WebGPU). For CPU execution, use the CubeCL CPU backend or burn-flex.

Burn's backend architecture lets you swap backends while keeping the same model code. You can enable multiple backends in the same application and choose the device for your tensors and modules at runtime through Device. This gives you the freedom to use different backends side by side and select the hardware best suited to each workload.

Autodifferentiation and automatic kernel fusion integrate with the same tensor and module APIs, so models benefit from these capabilities on supported backends without changing their implementation.

In application code, autodiff is runtime context carried by tensors. Devices provide the default context for newly created tensors, and each tensor can later enable or remove autodiff independently. Enabling autodiff permits graph recording; it does not by itself make a tensor retain gradients.

Internally, Burn implements this by decorating a concrete backend, so autodiff cannot execute by itself. Enable it on a device before creating tensors or initializing a model. With the autodiff and wgpu features enabled:

use burn::tensor::{Device, Distribution, Tensor};

fn main() {
    let device = Device::wgpu(Default::default()).autodiff();

    let x: Tensor<2> = Tensor::random([32, 32], Distribution::Default, &device);
    let y: Tensor<2> = Tensor::random([32, 32], Distribution::Default, &device).require_grad();

    let tmp = x.clone() + y.clone();
    let tmp = tmp.matmul(x);
    let tmp = tmp.exp();

    let grads = tmp.backward();
    let y_grad = y.grad(&grads).unwrap();
    println!("{y_grad}");
}

backward() checks graph participation at runtime. Enable autodiff before the forward pass and call require_grad() on source leaves whose gradients you need. is_autodiff(), is_tracked(), and is_require_grad() inspect autodiff association, graph participation, and gradient retention respectively. See the autodiff guide.

See the Autodiff Backend README for more details.

This backend decorator enhances a backend with kernel fusion, provided that the inner backend supports it. Note that you can compose this backend with other backend decorators such as Autodiff. All first-party accelerated backends (like WGPU and CUDA) use Fusion by default (burn/fusion feature flag), so you typically don't need to apply it manually.

#[cfg(not(feature = "fusion"))]
pub type Cube = burn_cubecl::CubeBackend;

#[cfg(feature = "fusion")]
pub type Cube = burn_fusion::Fusion<burn_cubecl::CubeBackend>;

Device::autodiff().gradient_checkpointing() enables the balanced gradient-checkpointing strategy, which trades recomputation for reduced activation storage during training.

See the Fusion Backend README for more details.

Remote execution has a client and a server. The server's devices select the compute backend; clients use a remote Device with the same tensor API. Iroh, the default transport, reaches a server across any network, authenticated and encrypted; see the server example and the distributed computing guide. On a trusted network, WebSocket is the simplest setup: enable remote-server, remote-websocket, and cuda on the server, and remote-websocket plus autodiff on the client:

use burn::remote::RemoteHost;
use burn::server::{RemoteServer, ServeError, WebSocketTransport};
use burn::tensor::{Device, Distribution, Tensor};

fn main_server() -> Result<(), ServeError> {
    RemoteServer::new([Device::cuda(0)]).serve(WebSocketTransport::new(3000))
}

fn main_client() -> Result<(), burn::remote::ConnectError> {
    let host = RemoteHost::websocket("ws://localhost:3000");
    let device = Device::remote_options(&host).init()?.autodiff();
    let tensor_gpu = Tensor::<2>::random([3, 3], Distribution::Default, &device);
    Ok(())
}

Training & Inference

The whole deep learning workflow is made easy with Burn, as you can monitor your training progress with an ergonomic dashboard, and run inference everywhere from embedded devices to large GPU clusters.

Burn was built from the ground up with training and inference in mind. It's also worth noting how Burn, in comparison to frameworks like PyTorch, simplifies the transition from training to deployment, eliminating the need for code changes.

Click on the following sections to expand 👇

As you can see in the previous video (click on the picture!), a new terminal UI dashboard based on the Ratatui crate allows users to follow their training with ease without having to connect to any external application.

You can visualize your training and validation metrics updating in real-time and analyze the lifelong progression or recent history of any registered metrics using only the arrow keys. Break from the training loop without crashing, allowing potential checkpoints to be fully written or important pieces of code to complete without interruption 🛡

Burn supports importing ONNX (Open Neural Network Exchange) models through the burn-onnx crate, allowing you to easily port models from TensorFlow or PyTorch to Burn. The ONNX model is converted into Rust code that uses Burn's native APIs, enabling the imported model to run on any Burn backend (CPU, GPU, WebAssembly) and benefit from all of Burn's optimizations like automatic kernel fusion.

Our ONNX support is further described in this section of the Burn Book 🔥.

Note: This crate is in active development and currently supports a limited set of ONNX operators.

You can load weights from PyTorch or Safetensors formats directly into your Burn-defined models. This makes it easy to reuse existing models while benefiting from Burn's performance and deployment features.

Learn more in the Saving & Loading Models section of the Burn Book.

Several of our backends can run in WebAssembly environments: Flex for CPU execution, and WGPU for GPU acceleration via WebGPU. This means that you can run inference directly within a browser. We provide several examples of this:

• MNIST where you can draw digits and a small convnet tries to find which one it is! 2️⃣ 7️⃣ 😰
• Image Classification where you can upload images and classify them! 🌄

Burn's core components support no_std. This means it can run in bare metal environment such as embedded devices without an operating system.

As of now, only the Flex backend can be used in a no_std environment.

Benchmarks

To evaluate performance across different backends and track improvements over time, we provide a dedicated benchmarking suite.

Run and compare benchmarks using burn-bench.

Getting Started

Just heard of Burn? You are at the right place! Just continue reading this section and we hope you can get on board really quickly.

To begin working effectively with Burn, it is crucial to understand its key components and philosophy. This is why we highly recommend new users to read the first sections of The Burn Book 🔥. It provides detailed examples and explanations covering every facet of the framework, including building blocks like tensors, modules, and optimizers, all the way to advanced usage, like coding your own GPU kernels.

The project is constantly evolving, and we try as much as possible to keep the book up to date with new additions. However, we might miss some details sometimes, so if you see something weird, let us know! We also gladly accept Pull Requests 😄

Let's start with a code snippet that shows how intuitive the framework is to use! In the following, we declare a neural network module with some parameters along with its forward pass.

use burn::nn;
use burn::module::Module;
use burn::tensor::Tensor;

#[derive(Module, Debug)]
pub struct PositionWiseFeedForward {
    linear_inner: nn::Linear,
    linear_outer: nn::Linear,
    dropout: nn::Dropout,
    gelu: nn::Gelu,
}

impl PositionWiseFeedForward {
    pub fn forward<const D: usize>(&self, input: Tensor<D>) -> Tensor<D> {
        let x = self.linear_inner.forward(input);
        let x = self.gelu.forward(x);
        let x = self.dropout.forward(x);

        self.linear_outer.forward(x)
    }
}

We have a somewhat large amount of examples in the repository that shows how to use the framework in different scenarios.

Following the book:

• Basic Workflow : Creates a custom CNN Module to train on the MNIST dataset and use for inference.
• Custom Training Loop : Implements a basic training loop instead of using the Learner.
• Custom WGPU Kernel : Learn how to create your own custom operation with the WGPU backend.

Additional examples:

• Custom CSV Dataset : Implements a dataset to parse CSV data for a regression task.
• Regression : Trains a simple MLP on the California Housing dataset to predict the median house value for a district.
• Custom Image Dataset : Trains a simple CNN on custom image dataset following a simple folder structure.
• Custom Renderer : Implements a custom renderer to display the Learner progress.
• Image Classification Web : Image classification web browser demo using Burn, WGPU and WebAssembly.
• MNIST Inference on Web : An interactive MNIST inference demo in the browser.
• MNIST Training : Demonstrates how to train a custom Module (MLP) with the Learner configured to log metrics and keep training checkpoints.
• PyTorch Import Inference : Imports a PyTorch model pre-trained on MNIST to perform inference on a sample image with Burn.
• Text Classification : Trains a text classification transformer model on the AG News or DbPedia dataset. The trained model can then be used to classify a text sample.
• Text Generation : Trains a text generation transformer model on the DbPedia dataset.
• Wasserstein GAN MNIST : Trains a WGAN model to generate new handwritten digits based on MNIST.

For more practical insights, you can clone the repository and run any of them directly on your computer!

We keep an updated and curated list of models and examples built with Burn, see the tracel-ai/models repository for more details.

Don't see the model you want? Don't hesitate to open an issue, and we may prioritize it. Built a model using Burn and want to share it? You can also open a Pull Request and add your model under the community section!

Deep Learning is a special form of software where you need very high level abstractions as well as extremely fast execution time. Rust is the perfect candidate for that use case since it provides zero-cost abstractions to easily create neural network modules, and fine-grained control over memory to optimize every detail. To this day, the mainstream solution has been to offer APIs in Python but rely on bindings to low-level languages such as C/C++. This reduces portability, increases complexity and creates friction between researchers and engineers. Rust's approach to abstractions is versatile enough to tackle this two-language dichotomy, and Cargo makes it easy to build, test and deploy from any environment, which is usually a pain in Python.

Rust's AI ecosystem is young, but it is real and growing quickly. Foundational pieces are already here: Burn and CubeCL for training and compute, candle for inference, Hugging Face's tokenizers and safetensors, and polars and ndarray for data. Betting on Rust today means betting on a stack that is growing, and one where contributors still shape the direction. The pieces that don't exist yet are opportunities rather than dead-ends (see Contributing).

Rust is also what makes one-stack-everywhere possible: a single self-contained binary with no Python runtime to ship, running from servers down to no_std embedded targets.

Deprecation NoteSince 0.14.0, the internal structure for tensor data has changed. The previous Data struct was deprecated and officially removed since 0.17.0 in favor of the new TensorData struct, which allows for more flexibility by storing the underlying data as bytes and keeping the data type as a field. If you are using Data in your code, make sure to switch to TensorData.

Burn 0.22 uses burnpack for native records and cannot directly read legacy Recorder formats such as .mpk, .bin, or JSON. Load the checkpoint in a compatible older Burn project, export the model weights through burn-store, and import them into your 0.22 model. See Migrating checkpoints for an example and the distinction between transferring weights and resuming training state.

For records saved before 0.14.0, an earlier migration step may also be needed: use Burn 0.14, 0.15, or 0.16 with the record-backward-compat feature to load and re-save the record with the newer tensor-data representation. For legacy binary records, first use the version that wrote them to export a self-describing format such as NamedMpkFileRecorder. This historical migration does not convert records to the format required by 0.22; follow the checkpoint migration guide afterward.

Community

If you are excited about the project, don't hesitate to join our Discord! We try to be as welcoming as possible to everybody from any background. You can ask your questions and share what you built with the community!

Contributing

Before contributing, please read the Contributing Guidelines and our Code of Conduct. The Contributor Book covers architecture, environment setup, and guides for common tasks.

Status

Burn is currently in active development, and there will be breaking changes. While any resulting issues are likely to be easy to fix, there are no guarantees at this stage.

License

Burn is distributed under the terms of either the MIT License or the Apache License, Version 2.0. Unless otherwise noted, this applies to all contents of this repository, including source code, examples, documentation, and code snippets in the Burn Book and Contributor Book. See the licenses for details.

Third-party materials and externally downloaded datasets or models remain subject to their respective licenses.

Opening a pull request is assumed to signal agreement with these licensing terms.