dfdx
A deep learning library in Rust featuring shape-checked tensors and neural networks with compile-time safety.
What is dfdx?
dfdx is a deep learning library for Rust that prioritizes ergonomics and safety through compile-time shape checking. It provides a tensor library with GPU acceleration and a comprehensive set of neural network building blocks, enabling robust and performant machine learning development.
Rust developers building deep learning models who prioritize compile-time correctness, safety, and performance, particularly those working on projects where runtime shape errors are unacceptable.
Developers choose dfdx for its unique combination of ergonomic API, compile-time shape and type checking that prevents runtime errors, and performance through GPU acceleration with minimal unsafe code and dynamic borrow checks.
Overview
Deep learning in Rust, with shape checked tensors and neural networks
Use Cases
Best For
- Building neural networks in Rust with compile-time shape safety to eliminate runtime tensor dimension errors.
- Developing GPU-accelerated deep learning models in Rust using CUDA for high-performance computation.
- Creating modular neural network architectures using ergonomic building blocks like Linear, Conv2D, and Transformer layers.
- Implementing training loops with standard optimizers (SGD, Adam, AdamW, RMSprop) in a type-safe Rust environment.
- Research or production projects requiring precise control over gradient tape management without Rc/RefCell overhead.
- Validating deep learning operations against PyTorch for correctness while leveraging Rust's safety guarantees.
Not Ideal For
- Projects requiring stable, production-ready APIs without frequent breaking changes.
- Teams heavily reliant on Python's ML ecosystem for pre-trained models or rapid prototyping.
- Environments without NVIDIA GPUs or where CUDA toolkit installation is not feasible.
- Applications with tensor shapes that are unknown or dynamically determined at runtime.
Pros & Cons
Pros
All tensor operations are validated at compile time, preventing runtime dimension errors and ensuring type safety, as emphasized in the design goals.
Supports high-performance computation with an optional CUDA backend for tensors up to 6 dimensions, detailed in the feature flags documentation.
Uses a flexible Module trait and tuples for sequential models, enabling clean, type-safe network definitions without Rc/RefCell overhead.
Avoids dynamic borrow checks by moving gradient tapes through computation graphs, giving precise control over recorded tensors, as explained in the implementation details.
Cons
The library is explicitly marked as pre-alpha with planned breaking releases, making it risky for production use or long-term projects.
GPU acceleration requires NVIDIA's CUDA toolkit, limiting cross-platform compatibility and adding setup complexity for non-NVIDIA systems.
As a Rust library, it lacks the extensive tooling and community support of Python frameworks, with fewer pre-trained models and third-party plugins.
Frequently Asked Questions
Related Projects
candle-wasm-examplesMinimalist ML framework for Rust
burnBurn is a next generation tensor library and Deep Learning Framework that doesn't compromise on flexibility, efficiency and portability.
tokenizers💥 Fast State-of-the-Art Tokenizers optimized for Research and Production
mistral.rsFast, flexible LLM inference
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.