v0.19.1

tracel-ai/burnv0.19.1Nov 6, 2025by laggui

AI Summary

A bug-fix focused release addressing memory leaks, CUDA compatibility, and Tensor API regressions to improve stability and performance.

Key Highlights

Fixed RAM memory leak in Autodiff with correct graph cleanup
Improved memory reuse with optimized sliced memory pool
Updated `cudarc` to auto-detect CUDA version and fix 12.8 features
Fixed Quantized Linear fusion configuration
Fixed PyTorch import regression with integer dictionary keys

New Features

Tensor API fixes including `*_like` preserving dtype and `RotaryEncoding` adjustments
CUDA 12.8 feature support
Bug fixes for SLURM environment variables and non-primitive field saving
Updated uvicorn version for SkyPilot compatibility

Full Release Notes

# Bug Fixes & Improvements

* **Autodiff:** fixed RAM memory leak with correct graph cleanup (#3957 #3982) @laggui
* **Better memory reuse:** improved sliced memory pool implementation (#3941) @nathanielsimard
* **Cuda:**  update `cudarc`, auto-detect CUDA version and fix some 12.8 features (CubeCL [#1008](https://github.com/tracel-ai/cubecl/pull/1008)) @wingertge
* **Quantized Linear:** fixed fusion configuration to fuse more precisions (#3941) @nathanielsimard
* **PyTorch import:** fixed pickle reader regression with integer dictionary keys (#3978) @laggui
* **Docs:** switched to `doc_cfg` to fix `docs.rs` builds (#3979) @laggui
* **Tensor API fixes**:
    - `*_like` preserves dtype (#3953) @crutcher
    - `RotaryEncoding` sum dimension for 3D input (#3954) @laggui
    - `squeeze` check for output rank > 0 (#3946) @laggui
    - `Linear` for input/output rank 1 (#3966) @lucasmdjl