推論系統
HPD-Parsing Splits Document Generation into Parallel Branches, With the 1B Model Reaching Peak Throughput of 4,752 Tokens per Second
PaddlePaddle has released the 1B-parameter HPD-Parsing model, which uses a global layout branch to coordinate multiple block decoders and bypass the sequential bottleneck of generating an entire page token by token. Official tests show throughput 2.62 times that of the fastest existing comparison model, but the system currently depends on a custom vLLM package, and its peak figure cannot be directly equated with end-to-end PDF processing speed.
PaddlePaddle recently released the HPD-Parsing weights, model card, and deployment documentation, aiming to reshape how vision-language models decode documents for parsing. Traditional end-to-end document models first interpret an entire page image and then output text, tables, and formulas along a single autoregressive sequence. The longer the page, the longer later tokens must wait for earlier ones to finish. HPD-Parsing instead assigns the main branch responsibility only for the global layout and reading order. When it encounters a block, it emits a special dispatch signal, prompting the runtime to dynamically create multiple subrequests that generate the content of different regions concurrently. The results are then assembled according to the structure produced by the main branch.
The model also incorporates Progressive Multi-Token Prediction, which gradually increases the number of tokens predicted at a time to further reduce the number of decoding iterations. The paper reports peak throughput of 4,752 tokens/s on public document datasets—2.62 times that of the fastest comparison model and 3.06 times that of a standard autoregressive baseline. The Hugging Face model card lists an overall OmniDocBench v1.6 score of 94.91%. The model has approximately 1B parameters, is based on the InternVL3.5-1B family, and is released under the Apache-2.0 license.
Its technical value lies not only in its OCR score, but also in directly mapping the natural block structure of documents onto inference scheduling. For workloads involving large volumes of invoices, research papers, or RAG indexing, block-level parallelism could improve GPU utilization and reduce tail latency on long pages. However, deployment currently requires the custom `vllm-0.17.1+hpdparsing` package, indicating that this is not yet a standard vLLM capability. Dynamic branching also adds scheduling, memory, and result-reassembly overhead. Engineering teams should next evaluate per-page latency on real-world PDFs, batch throughput, peak VRAM usage, and errors involving complex reading order, rather than relying solely on peak token throughput.