Model compression is progressing on Pruna's end.

Pruna v0.3.4 updates its open-source AI optimization framework with Llama.cpp integration, token merging, KVPress cache compression, and TorchAO.

With its v0.3.4 release, Pruna expands its open-source model optimization framework with three external contributions: Llama.cpp integration, token merging for image classification, and KVPress for KV cache compression. This version also improves compatibility with Transformers 5, TorchAO up to 0.17.0, and more recent Torch versions, while removing certain dependencies that had become unstable or obsolete.

The focus is less on spectacular new features and more on tool consolidation: more reliable installations, improved stability across backends, and fixes related to compilation, audio datasets, tests, and dependency management.

The documentation has also been improved, particularly to facilitate contributions, the addition of extras, and the AI-assisted contribution policy.