HPD-Parsing

### TL;DR

HPD-Parsing is a lightweight 1B hierarchical parallel document parsing model designed for high-throughput processing of complex document structures. By utilizing a main layout branch for global coordination and localized parallel branches for content generation, it achieves state-of-the-art inference efficiency while maintaining high accuracy.

Key Insights & Metrics

Key Features

• Hierarchical Parallel Decoding (HPD) for high-throughput, non-sequential generation
• Progressive Multi-Token Prediction (P-MTP) to reduce total decoding steps
• Staged adaptation pipeline with automated difficulty-aware data curation
• Dynamic tile-based cropping supporting resolutions up to 24 tiles of 448x448
• Customized vLLM support with dynamic request forking capabilities

Discussion