Skip to content

Flash-MinerU v1.0.0 Release Note

Choose a tag to compare

@SunnyHaze SunnyHaze released this 03 Apr 18:39
· 7 commits to main since this release
d6d6d56

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