Built-in Benchmarks now have dedicated pages. Every page includes a workload explanation, Mermaid topology, runnable configuration, output fields, and the performance dimensions the case can expose.

Real-model workloads

• MinerU Scale — real workload: production-hardened local/HDFS PDF processing, configurable large actor pools, atomic output, and resume.
• MinerU PDF: dynamic page fan-out, cross-PDF GPU batching, and ordered document reconstruction.
• Panda-70M Video Caption: source-owned clip fan-out, four temporal Qwen teachers, and ordered source reduction.
• YOLO → SAM: a multi-stage image pipeline with two persistent vision-model pools.
• Dual vLLM: generation and refinement through two persistent LLM engines.
• SGLang → vLLM: one Pipeline whose inference backends run in separate Conda environments.

Application integrations

• Flash-MinerU: a maintained PDF → pages → GPU VLM/OCR → ordered document application, with the current pipeline-parallel benchmark context.

Dependency-free topology references

• Nested Document: two levels of Document → Page → TableJob expansion and reduction.
• Video Caption: the basic Video → Frame → Video 1 → M → 1 relation.
• Video Multimodal: audio and vision sibling branches joined by parent Video.

See the Benchmarks Overview for selection guidance and Interpret Performance Results for metric semantics.

Contributors: Ma, Xiaochen