PaDoc lifts parsing throughput up to 118%

PaDoc increased valid-page throughput 67.4% to 118% and cut P95 latency 39.2% to 54.9% against a same-backbone sequential baseline on a 384-page subset and one Nvidia A800. The new paper branches layout regions over one shared page representation and reports a 94.24 overall OmniDocBench score, but its speed test covers only five concurrency levels on that subset. Builders comparing multimodal models for production workloads should benchmark shared-prefix parallel decoding before buying more parsing GPUs, then verify the gain on their own page layouts and serving backend.