Converge the SDK, Gradio, remote CLI, Doclib and Router on Tasks with hash-based uploads, per-file result URLs, task listing and optional model output. Preserve Files and the synchronous file_parse wrapper while removing standalone Uploads/Jobs and local-path HTTP inputs.
Migrate contract, lifecycle, transport and real-document regression coverage; update documentation and HTTP examples. Validation: 1452 passed, 3 conditional skips, plus targeted final PDF checks.
BREAKING CHANGE: clients and workers must use Tasks; standalone Uploads/Jobs endpoints and local-source/inline HTTP options have been removed.
Reuse platform ONNX decisions when no discrete GPU is available. Bind and verify the actual Vulkan devices before applying the legacy Intel submission workaround in engine and server paths.
Preserve explicit device settings and offline behavior, and include regression coverage and hardware validation records.
Accept form manifests and per-file uploads with resumable task state and optional source hashes. Publish result URLs only after final ZIP artifacts are ready, and align Router handling with the same protocol.
Remove synchronous parsing and task result routes, drop fixed local upload limits, and update the HTTP client, documentation, and regression coverage.
Add auto/cpu/gpu configuration, Windows adapter discovery, isolated rec calibration, atomic preference persistence, and vendor session presets. Route layout, rec, and full MFR batches through DirectML with CPU fallback.
Use mutually exclusive Windows x64 DirectML dependencies and guard ORT/Magika distribution identity. Include calibration fixtures, three-machine results, the 0.8 threshold with ten UHD 770 repeats, and installation documentation.
Validation: 762 MinerU tests and 71 DocVortex tests passed; dependency checks and Windows Python 3.10-3.14 resolution passed. The 19-page OCR result matches the CPU baseline. DocVortex PyPI acceptance remains pending a compatible release; pinned source acceptance passed.
Freeze real detection inputs and replay isolated device, fusion, shape, microbatch, and NNPACK strategies with output parity checks and memory accounting. Add demo PDF Basic/OCR stage timing and keep explicit NNPACK-on controls independent of the production default.
Refs #5632
Disable NNPACK during the first lazy Torch initialization while preserving later explicit overrides and GPU/ONNX execution paths. Verify the process setting across worker threads without loading Torch on uninitialized lightweight paths.
Apple M4 CPU FP32 Basic/OCR on demo1 and demo2 retained byte-identical outputs across two order-rotated comparisons. Median parsing time fell from 216.67s to 105.63s, primarily in MFR. Related regression checks: 148 passed; default SDK demo2 replay matched the accepted outputs.
Refs #5632
Filter narrow formula-mask fragments before cropping and keep valid boxes aligned with recognition results. Add coverage for zero-size crops, formula-mask endpoints, and empty recognition batches.
Recover unclaimed oversized initial letters only at the start of a text paragraph, restore PDFium U+0002 line-end hyphens before text assembly, and consume DocVortex ligature-space geometry without changing OCR routing.
Refs opendatalab/MinerU#5635 and opendatalab/MinerU#5636
Validation: 431 focused regression tests passed. The three-page issue sample passes standard auto/txt and Flash txt with identical Python/Rust ModelJson, MiddleJson and Markdown outputs; source and before/after line boxes were reviewed.
Invalidate stale local capabilities before config changes return, wake the health worker, and reject superseded config/probe results until the target tier is ready. Preserve the actual child startup state for graceful shutdown.
Refs #5272
Apply filename prefix matching to the final query token and paginate the
complete filtered candidate set with stable ordering and accurate totals.
Reuse input page-range syntax in read locators, including rN and all, and
preserve selected-page boundaries during block and character continuation.
Update CLI/SDK/HTTP contracts, bilingual help, documentation, and regressions.
Validation: 667 Doclib/CLI tests passed; Ruff and formatting checks passed.
Refs #5366, #5337, #5365.
Use DocVortex generated-space evidence in Python span filling and rotated-line paths, while keeping the Rust fast path aligned. Document list-tail continuation semantics and add spacing, code protection, rotation, and Python/Rust parity regressions.
Align output filtering with reading-order candidate validity in Torch and ONNX. Remove the uniform conf override and cover classification threshold boundaries and multi-label candidates.
myhloli