Flash-MinerU ⚡️📄

简体中文 | English

Accelerating the VLM inference pipeline of MinerU with Ray, turning PDF parsing into a scalable data infrastructure component

Flash-MinerU is a lightweight, low-intrusion acceleration layer for MinerU. Beyond speeding up VLM inference, it upgrades PDF parsing into a high-throughput, distributed data pipeline: a useful building block for modern AI systems.

PDFs are one of the most important high-quality knowledge sources for AI workflows, including papers, reports, and manuals. Converting them into structured, model-ready data such as Markdown and JSON is a foundational step for:

• 📊 Data governance and curation
• 🧪 Synthetic data generation pipelines
• 🧠 LLM / MLLM training and evaluation

Flash-MinerU focuses on making this stage scalable, efficient, and production-ready:

• Minimal dependencies, lightweight installation
  • One-line install via pip install flash-mineru
  • Works in constrained or domestic environments such as METAX
• System-level acceleration, not reimplementation
  • Fully reuses MinerU’s logic and data structures
  • Preserves output consistency
• Designed for scale
  • Multi-GPU / multi-process / multi-node ready
  • Built on Ray as a unified execution layer

---

✨ Features

• 🚀 Ray-powered distributed execution
  Turns PDF parsing into a scalable data pipeline, from single-node multi-GPU setups to clusters

• 🧠 High-throughput VLM inference
  Focuses on the bottleneck stage and currently defaults to vLLM

• 🔄 Pipeline-parallel execution (core improvement)
  Uses an asynchronous pipeline with cross-stage overlap for sustained high utilization

• 🧩 Low-intrusion, composable design
  Retains MinerU’s middle_json and downstream logic for easy integration

---

🎯 How pipeline parallelism helps

Flash-MinerU turns MinerU’s sequential pipeline into an asynchronous pipelined system:

• 🟢 Much higher GPU utilization
  Keeps GPUs busy more than 90% of the time, while vanilla MinerU is often around 40-50% because stages block each other

• 🔄 Cross-stage overlap (key speedup)
  Different batches run in different stages at the same time, such as render / VLM / Markdown, instead of waiting for full completion

• ⚡ Result: much higher throughput
  Less idle time plus more overlap leads to significantly faster end-to-end processing

Left — bubble schedule (before) Batched sequential execution; GPU idle gaps.

Right — pipelined (Flash-MinerU) Asynchronous pipeline; high utilization.

---

📦 Installation

Basic installation (lightweight mode)

Suitable if you have already installed the inference backend manually (e.g., vLLM), or are using an image with a prebuilt environment:

pip install flash-mineru

Install with vLLM backend enabled (optional)

If you want Flash-MinerU to install vLLM as the inference backend for you:

pip install flash-mineru[vllm]

---

🚀 Quickstart

Minimal Python API example

from flash_mineru import MineruEngine

# Path to PDFs
pdfs = [
    "resnet.pdf",
    "yolo.pdf",
    "text2sql.pdf",
]

engine = MineruEngine(
    model="<path_to_local>/MinerU2.5-2509-1.2B",
    # Model can be downloaded from https://huggingface.co/opendatalab/MinerU2.5-2509-1.2B
    batch_size=16,             # PDFs per logical batch; often choose a multiple of GPU count
    replicas=8,                # Parallel vLLM / model instances; often match GPU count
    num_gpus_per_replica=0.9,  # GPU memory fraction for vLLM KV cache per instance; 1.0 uses full VRAM headroom
    save_dir="outputs_mineru", # Output directory for parsed results
    inflight=4,                # Pipeline depth (v1.0.0 path); can raise on high-memory hosts with diminishing returns
)

# Legacy v0.0.4 sequential batching (deprecated): from flash_mineru import MineruEngineLegacy

results = engine.run(pdfs)
print(results)  # list[list[str]], dir name of the output files

Output structure

• Each PDF’s parsing results will be generated under:

  <save_dir>/<pdf_name>/
  

• The Markdown file is located by default at:

  <save_dir>/<pdf_name>/vlm/<pdf_name>.md
  

---

📊 Benchmark

Scripts: English · 简体中文

Results (368 PDFs, single-node ~8× A100 class)

Method	Inference configuration	Total time
Flash-MinerU v1.0.0	MineruEngine, 8 replicas, inflight=8, pipeline parallelism	~8.5 min
MinerU (vanilla)	Hand-spawned pool of 8 mineru processes (Benchmark-mineru.py parallel mode, one GPU per process, vlm-auto-engine)	~14 min
Flash-MinerU v0.0.4	MineruEngineLegacy, 8 replicas × 1 GPU, batch_size=16, batch-sequential	~23 min
MinerU (vanilla)	vLLM, single GPU	~65 min

Commands: docs/BENCHMARK.md.

Summary

• v1.0.0 is about ~1.7× faster wall time than the eight-process baseline (~8.5 min vs ~14 min)
• v0.0.4 (MineruEngineLegacy) is slower than that baseline (~23 min), which highlights what pipeline parallelism adds versus “many full stacks in parallel”
• ~65 min single-GPU is the same-corpus reference baseline

Experimental setup (expand)

• Dataset: 23 paper PDFs (≈9–37 pages each) × 16 copies → 368 files; default folder test/sample_pdfs
• Versions: MinerU v2.7.5; Flash-MinerU v0.0.4 = MineruEngineLegacy (sequential stages per batch); v1.0.0 = MineruEngine (pipeline parallelism, default API)
• Hardware: single host, 8 × NVIDIA A100

Note: Throughput-focused. Output shape matches MinerU. Upstream does not ship a polished official multi-GPU “one click” path; the eight-process row is our benchmark script sharding eight separate mineru runs.

---

🗺️ Roadmap

• Benchmark scripts & docs — docs/BENCHMARK.md
• Support for more inference backends (e.g., sglang)
• Service-oriented deployment (HTTP API / task queue)
• Sample datasets and more comprehensive documentation

---

🤝 Acknowledgements

• MinerU This project is built upon MinerU’s overall algorithm design and engineering practices, and parallelizes its VLM inference pipeline. The mineru_core/ directory contains code logic copied from and adapted to the MinerU project. We extend our sincere respect and gratitude to the original authors and all contributors of MinerU. 🔗 Official repository / homepage: https://github.com/opendatalab/MinerU

• Ray Provides powerful abstractions for distributed and parallel computing, making multi-GPU and multi-process orchestration simpler and more reliable. 🔗 Official website: https://www.ray.io/ 🔗 Official GitHub: https://github.com/ray-project/ray

• vLLM Provides a high-throughput, production-ready inference engine (currently the default backend). 🔗 Official website: https://vllm.ai/ 🔗 Official GitHub: https://github.com/vllm-project/vllm

---

📜 License

AGPL-3.0

Notes: The mineru_core/ directory in this project contains derivative code based on MinerU (AGPL-3.0). In accordance with the AGPL-3.0 license requirements, this repository as a whole is released under AGPL-3.0 as a derivative work. For details, please refer to the root LICENSE file and mineru_core/README.md.