June 2026 monthly summary for vllm-project/llm-compressor: Delivered end-to-end REAP pruning support for Mixture-of-Experts (MoE) models within the LLM Compressor framework, enabling calibration-based expert reduction, model saving after compression, and improved calibration-flow handling. Added a What's New entry documenting the REAP Pruning Modifier and the associated VRAM reduction. Improved pipeline reliability so pruning uses the correct calibration flow (disabling moe_calibrate_all_experts where unnecessary and switching pruning to sequential_epoch_end), with alignment to LinearExperts2D and LLM Compressor v5 changes. Documented usage patterns and provided end-to-end pruning examples for model validation. Impact-focused accomplishments include practical VRAM reductions enabling larger models to fit on available hardware, and preserved/inferred accuracy under moderate sparsity on representative MoE models. These changes position the team to deliver cost-efficient inference at scale and reduce GPU costs for MoE workloads.

July 2025: Delivered key API enhancements in DS4SD/docling-core to support robust document manipulation, including table edits and insertion of document items, paired with refactoring and tests to boost reusability and reliability. This work underpins MCP document workflows, improves automation, and reduces downstream maintenance costs.