Skip to content

architecture.md

Latest commit

 

History

History
50 lines (41 loc) · 2.55 KB

architecture.md

File metadata and controls

50 lines (41 loc) · 2.55 KB
Raw

Architecture

MinerU-ROCm is an evaluation and reproducibility layer around upstream MinerU. It does not fork model weights or replace the upstream model APIs.

Runtime layers

  1. mineru-rocm parses the command and validates user inputs.
  2. The driver selects pipeline or vlm-vllm and runs the resumable page runner.
  3. A backend converts one input page to Markdown. Per-page failures are recorded rather than aborting the complete dataset.
  4. The validation and manifest layers enforce page conservation, non-empty predictions, stable filenames, and SHA-256 identity.
  5. The pinned OmniDocBench scorer produces EditDist, TEDS, CDM, and the final Overall score.
  6. OmniDocBench-ROCm packages self-contained model cards and evidence bundles.

Backend boundaries

Backend Inference process Accelerator path
Linux pipeline In-process MinerU 3.4.4 PyTorch ROCm; pipeline ONNX components may use CPU
Linux VLM OpenAI-compatible vLLM server vLLM ROCm on one gfx1100 GPU
Windows pipeline In-process MinerU 3.4.4 Windows ROCm PyTorch plus ONNX Runtime DirectML
Windows VLM OpenAI-compatible llama.cpp server llama.cpp HIP on one gfx1151 GPU

On Windows, all compatible ONNX sessions are DirectML-first. The known DirectML-incompatible slanet-plus.onnx control-flow model is routed through an explicit CPU override. Runtime DirectML failures are counted and surfaced in _run_stats.json; they are never silent.

Evidence boundary

Published scores are derived only from committed metric_result.json files. Each formal bundle also contains run statistics, provenance, dataset identity, and a prediction manifest. Bulk Markdown predictions remain outside Git, but every formal prediction is represented by filename and SHA-256 in its manifest.

The primary VLM card is model_card.json. The pipeline is a supplementary result with model_card.pipeline.json and platform-specific cards. The legacy model-card badges remain only for compatibility and are deprecated by the central ROCmDoc standard. The current trust model keeps producer assurance separate from central platform_review; see ROCMDOC_STANDARD.md and COMPATIBILITY.md. This repository does not assign central review outcomes. The Hardware matrix displays the two dimensions without creating another normative vocabulary; neither denotes vendor or upstream support.