v0.3.5

The juiciest bits 🧃

FP8 quantization lands in the garden 🌱

This release brings two new FP8 quantization algorithms: a static one, and a time-aware one built for diffusion pipelines. Time-aware quantization already works with more pipelines, and there is a tutorial to walk you through both. All of it comes from @kfachikov, who joined us as a first-time contributor this release and did not slow down once!

• feat: add static fp8 quantization algorithm by @kfachikov in #733
• feat: add time-aware fp8 quantization for diffuser by @kfachikov in #737
• feat: support for more pipelines for time-aware quantization by @kfachikov in #746
• feat: tutorial on static and time-aware quantizrion algorithms by @kfachikov in #741

A new metric for the evaluation toolbox 📊

• feat: add MSE evaluation metric by @Hungreeee in #716

Locking the garden gate 🔒

We tightened security across the repo. Checkpoints now load with weights_only, GitHub Actions are pinned to commit SHAs, and dependency updates wait 7 days before we pick them up:

• fix: set weights_only on torch.load and document pickle use by @davidberenstein1957 in #757
• fix(security): load the aesthetic head state dict with weights_only by @davidberenstein1957 in #776
• ci: pin GitHub Actions to commit SHAs by @davidberenstein1957 in #755
• chore(deps): add a 7-day cooldown to dependency updates by @davidberenstein1957 in #758
• chore: use safety@pruna.ai for code of conduct reports by @davidberenstein1957 in #726

Some Bug Fixing 🐞 and Maintenance

Metric evaluation, device handling and diffusers pipelines all got a round of fixes:

• fix: unallocate metric model prior restoring original model by @kfachikov in #690
• fix: file exists on consecutive metric evaluations by @kfachikov in #699
• fix: make CallSequenceTracker unwrap idempotent for bound methods by @FAdrien in #701
• fix: parameters order assumption in diffuser by @kfachikov in #706
• fix: use dtype instead of deprecated torch_dtype by @xyf5432 in #735
• fix: always move diffusers pipelines in move_to_device by @begumcig in #778

We also capped a few dependencies to keep installs stable, and kept the test suite up to date with new library versions:

• build: cap pyarrow below 26 in the vbench extra by @begumcig in #785
• fix: cap huggingface-hub below 2.0 by @begumcig in #790
• test: change test fixure for new diffusers version by @begumcig in #777
• test: stop patching OTLP exporter internals in telemetry log-level tests by @begumcig in #779
• build: bump pruna version to v0.3.5 by @begumcig in #789

Documentation updates 📚

A new tutorial shows how to measure the energy and emissions side of your optimized models:

• docs: add sustainability benchmarking tutorial by @zinichakraborty in #711

New Faces in the Garden 👩‍🌾

Five first-time contributors this release, our biggest batch yet 💐

• @kfachikov made their first contribution in #690 (and they did not stop there: both FP8 quantization algorithms, a tutorial and a handful of fixes followed!)
• @FAdrien made their first contribution in #701
• @zinichakraborty made their first contribution in #711
• @Hungreeee made their first contribution in #716
• @xyf5432 made their first contribution in #735

Full Changelog: v0.3.4...v0.3.5