Flash-MinerU v1.0.1 Release Note

What's changed

Remove numpy<2 restrict from requirements

Full Changelog: v1.0.0...v1.0.1

Flash-MinerU v1.0.0 Release Note

Flash-MinerU v1.0.0 Release Note

English

• Default engine: MineruEngine now uses pipeline-parallel inference (overlapping pdf2img → VLM → Markdown).
• Performance: Up to 2.7× faster vs v0.0.4, 1.6× vs manual multi-process, 7.6× vs single GPU.
• Tuning: New inflight parameter to control concurrent batches in the DAG.
• Legacy API: Sequential batching moved to MineruEngineLegacy (deprecated).
• Docs: Added pipeline overview and benchmarks (docs/BENCHMARK.md).
• Fix: Resolved memory leak caused by unreleased PDFium objects (#11).

中文

• 默认引擎： MineruEngine 改为 流水线并行推理（pdf2img → VLM → Markdown 重叠执行）。
• 性能提升： 相比 v0.0.4 最高 2.7× 加速，相比手动多进程 1.6×，相比单卡 7.6×。
• 调参： 新增 inflight，控制 DAG 内并发 batch 数。
• 旧 API： 顺序执行迁移至 MineruEngineLegacy（已弃用）。
• 文档： 补充流水线说明与性能测试（docs/BENCHMARK.md）。
• 修复： 修复 PDFium 未释放导致的内存泄漏（#11）。

---

📊 Benchmark

Usage: [中文](./docs/BENCHMARK.zh.md) · [English](./docs/BENCHMARK.md)

Results (368 PDFs, single node with 8× A100)

Setup	Config	Time
Flash-MinerU v1.0.0	MineruEngine, 8 replicas, inflight=8	~8.5 min
MinerU (native)	8 manual processes (parallel mode, 1 GPU each)	~14 min
Flash-MinerU v0.0.4	MineruEngineLegacy, batch_size=16	~23 min
MinerU (native)	vLLM, single GPU	~65 min

Takeaways

• ⚡ ~2.7× faster vs v0.0.4
• ⚡ ~1.6× faster vs manual multi-process MinerU
• ⚡ ~7.6× faster vs single-GPU MinerU
• 🚀 Speedup mainly from pipeline-parallel execution (stage overlap + better utilization)

---

What's Changed

• fix: Fix memory leak from PDFium objects not being closed — @1773899415 — #11
• feat: Pipelined inference refactor with major speedup — @SunnyHaze — #12

---

New Contributors

• @1773899415 — first contribution in #11

---

Full Changelog:
v0.0.4...v1.0.0

Flash-MinerU v0.0.4 release note

What's Changed

• fix end_page_id bug in Pdf2ImageOp by @Lavender1 in #10

Full Changelog: v0.0.2...v0.0.4

Flash-MinerU v0.0.3 Release note

What's Changed

• fix end_page_id bug in Pdf2ImageOp by @Lavender1 in #10

Full Changelog: v0.0.2...v0.0.3

Flash-MinerU v0.0.2 Release Note

What's Changed

• [README] update readme by @SunnyHaze in #5
• [README] update some naive benchmark by @SunnyHaze in #6
• output layout.json and content_list.json by @Lavender1 in #7

Full Changelog: v0.0.1...v0.0.2

Flash-MinerU v0.0.1 Release Note

🎉 Flash-MinerU v0.0.1 — Initial Release

We’re excited to announce the first public release of Flash-MinerU 🎉
This version lays down the core foundation for Ray-based parallel acceleration of MinerU’s VLM inference pipeline, focusing on correctness, extensibility, and engineering clarity.

---

✨ Highlights

• 🚀 End-to-end MinerU parsing workflow implemented
  Core logic for MinerU PDF parsing is now fully integrated, including all essential parsing and processing functions.

• 🧩 Clean, minimal dependency setup
  Requirements have been streamlined to remove redundancy while keeping the minimal set needed for stable execution.

• 🧪 Compatibility with domestic computing environments
  Dependency versions have been adjusted to ensure smooth execution on special environments.

• 🧹 Codebase cleanup & structure refinement
  Improved readability and maintainability in preparation for future parallelization and scaling features.

---

🔧 What’s Changed

• Core Features

  • Added the main MinerU parsing logic and all related functions
    (PR #1 by @Lavender1)

• Dependencies & Environment

  • Added minimal required dependencies and removed redundant entries
    (PR #2 by @SunnyHaze)
  • Updated requirements.txt for compatibility with MUXI environments
    (PR #3 by @SunnyHaze)

• Refactoring & Cleanup

  • General code cleanup and structure optimization
    (PR #4 by @SunnyHaze)

---

🤝 New Contributors

Welcome to our first contributors — thanks for helping Flash-MinerU take off 🚀

• 🎉 @Lavender1 — first contribution via PR #1
• 🎉 @SunnyHaze — first contribution via PR #2

---

📜 Full Changelog

🔗 https://github.com/OpenDCAI/Flash-MinerU/commits/v0.0.1

---

🗺️ What’s Next?

This is just the beginning. Upcoming releases will focus on:

• Ray-based multi-replica VLM inference acceleration
• Performance benchmarks (single vs multi-GPU)
• Additional inference backends and service-oriented deployment

Stay tuned ⚡️📄