EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Lobsters·5d ago·5 min read
Burn 0.22.0: Faster Builds, Easier Extensions, and Smarter Autotuning
TL;DR
Burn 0.22.0 introduces significant enhancements for building applications, including faster rebuilds and improved model integration, achieving up to 15× faster compilation times.
✦ Why It Matters
Upgrade to Burn 0.22.0 today to streamline your application builds and improve performance.
Key Takeaways
- Remove backend type parameters from your application code.
- Utilize the CubeCL Environment for efficient compilation caching.
- Adapt existing models using LoRA and QLoRA.
- Leverage improved autotuning for better CPU and GPU performance.
Full Summary
Burn 0.22.0 focuses on simplifying application development and enhancing performance by removing backend generics from the user API. Previously, backend types were tied to models and tensor functions, complicating the compilation process.
With this release, execution is determined by the device rather than backend types, leading to rebuilds that are up to 15 times faster. Additionally, improvements in adaptive memory management and smarter autotuning optimize execution across CPUs and GPUs.
The new CubeCL Environment allows applications to reuse compilation and autotuning caches, further enhancing efficiency. Support for adapting existing models with LoRA and QLoRA, as well as ONNX export, expands the framework's capabilities for remote computing.