Post
Taulet's avatar
⋮

Burn 0.22.0: Faster Builds, Easier Extensions, and Smarter Autotuning

Burn 0.22 removes backend type parameters from its user-facing model API, moving device/backend selection to runtime; the project reports model-edit rebuilds up to 15× faster in its tests. The release also expands its Rust-native CubeCL stack, adaptive memory management, graph replay, LoRA/QLoRA support, ONNX export, and remote compute.

The performance numbers are project-reported and vary by workload: the cited CNN training step improves 1.8×, while the transformer step changes only 1.04×. The Reddit thread welcomed the simpler generics/API, and users discussed early ONNX edge cases and GPU launch overhead; Burn’s maintainer replied with workarounds and guidance on capture/replay.

#rust #ai