The juiciest bits 🧃
New algorithms joining the garden 🌱
We added a fresh batch of algorithm support this release, and this one deserves an extra round of applause: all of these new algorithm additions came from external contributors! Huge thank you to @krishjp, @rensortino, and @kschwethelm for helping expand what Pruna can optimize and for bringing more community-built improvements into the repo
- feat: integrate Llama.cpp and enhance engine stability for cross-platform usage by @krishjp in #616
- feat: token merging for image classification by @rensortino in #537
- feat: integrate KVPress for KV cache compression (#366) by @kschwethelm in #623
Compatibility updates and pruning some old branches: ✂️
Transformers 5 compatibility is here, TorchAO support now goes up to 0.17.0, and stable-fast wheel handling has been updated for newer Torch versions. At the same time, we cleaned up outdated or unreachable dependencies to keep installs smoother and more reliable:
- feat: compatibility with transformers 5 by @sdiazlor in #685
- build: bump max torchao from 0.16.0 to 0.17.0 by @gsprochette in #692
- fix: drop ipex support because their index is unreachable by @gsprochette in #667
- fix: drop outdated whisper-s2t by @gsprochette in #689
Some Bug Fixing 🐞 and Maintenance
A lot of work went into making Pruna more stable across backends, datasets, compilation flows, and test environments:
- fix: torch compile default to inductor on cpu by @gsprochette in #673
- fix: torch compile bugs hqq by @gsprochette in #668
- fix: patch torchmetrics (transformers 5) by @sdiazlor in #683
- fix: update moe tiny model for testing (transformers 5) by @sdiazlor in #684
- fix: stop auto decoding for audio datasets by @begumcig in #686
- style: move type ignore to its correct place by @gsprochette in #681
- test: fix breaking nightly tests for ONEIG dataset by @begumcig in #687
- fix: add kernel version by @sdiazlor in #675
- fix: device casting for llama cpp by @begumcig in #693
- fix: stop audio decoding by @begumcig in #694
- fix: switch torch_dynamic and quanto saves to save_before_apply by @gsprochette in #696
We’ve also continued improving the developer experience around CI, dependency handling, pre-commit checks, and test stability:
- ci: add no extra flag by @begumcig in #653
- ci: automatically assign new reviewers by @sdiazlor in #620
- build: add dependencies cooldown by @gsprochette in #660
- fix: ruff check pre-commit fails due to "Failed to parse pyproject.toml" by @psyberck in #661
- build: restrict dependency cooldown to pypi index by @gsprochette in #672
- build: switch back to unsafe-best-match for easy torch install by @gsprochette in #674
- ci: add stale issue handling by @begumcig in #682
- refactor: add requires stable fast mark by @begumcig in #688
- ci: change uv sync to uv pip install by @gsprochette in #691
- docs: remove reference to pruna[full] by @gsprochette in #695
Documentation updates 📚
We cleaned up and expanded contribution docs so it’s easier to understand how to add extras, manage dependencies, and contribute safely:
- docs: add contribution guidlines for extras by @begumcig in #654
- docs: removed first prune from PR template by @minettekaum in #652
- docs: update with AI Contribution Policy by @sdiazlor in #621
- docs: improve documentation for adding dependencies by @gsprochette in #659
New Faces in the Garden 👩🌾
- @kschwethelm made their first contribution in #623
- @psyberck made their first contribution in #661
- @krishjp made their first contribution in #616
Full Changelog: v0.3.3...v0.3.4
