Skip to content

Document parsing pipeline 0.1.0-rc.4

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 03 Nov 11:09
· 19 commits to main since this release

Quantized dots GPU warmup fix · 0.1.0-rc.4

The real parsing worker now skips the native server's fixed-size dummy vision
warmup. On the tested GTX 1660 Ti 6 GB, rc.3 exhausted VRAM while reserving a
3,645.66 MiB dummy workspace; disabling that step lets the real synthetic page
complete. This keeps Q4_K_M language / Q8_0 projector weights, actual page pixel
limits and full GPU offload unchanged. It does not guarantee arbitrary pages fit.

The synthetic GPU check covers preparation, text/table extraction, canonical
exports, review HTML and server teardown, with no cloud calls. It is a runtime
regression probe, separate from OCR-quality evaluation. Earlier rc.3 failure
evidence is retained; CPU-hosted checks are not credited as GPU inference.

The default remains quantized dots for the 6 GB laptop. Original BF16 dots scored
best on selected handwritten mathematics (32/37 vs Q4 29/37), with 8.53 GiB peak
PyTorch allocation on Colab T4; this is not total VRAM or a universal ranking.
Optional visual descriptions use Gemma through Ollama Cloud. Weights and secrets
are not bundled. The public four-profile demo remains version 4.1.

Docker setup ·
Validation.

Public GPU parsing image: bmaged23/multimodal-document-parser@sha256:3cc1432bb586a40dffeb3757e1be0587094ecd9d4deb33fd5c0fbe5ab13ad7a4.

Built from e810b9f36f032ebdc0ce37c97f7d9730229ff43a after source/browser CI and image smoke tests. Includes SBOM and build provenance. Gemma executes through Ollama Cloud; its weights are not bundled.

Docker setup

Published-image GPU verification

The exact public rc.4 image passed a synthetic text/table page through preparation and parsing workers on GTX 1660 Ti, including canonical exports, review HTML and server teardown, with no cloud calls. Two earlier vision-weight allocation failures under Windows host-memory pressure are retained; the unchanged image succeeded after releasing WSL file caches. This is a runtime check, separate from corpus OCR quality.

Download the validation record.