[Guix packaging] kha-white/manga-ocr #134

New issue

Open

opened 2026-08-14 13:17:47 +00:00 by htayj · 1 comment

htayj commented

2026-08-14 13:17:47 +00:00

(Migrated from github.com)

Candidate

Upstream canonical URL: https://github.com/kha-white/manga-ocr
Source pinned commit/release when known: default-branch HEAD c333b5d36e88d539d6b040b4c4cf90ad5ecd4f69 (checked 2026-08-14)
Target concrete installed deliverable: the manga-ocr local Japanese OCR command
Primary category: ai-tool
Tags: ai-ml, image-processing, ocr
Primary language normalized: Python
Build system: pyproject-build-system
SPDX expression: Apache-2.0
License status: confirmed-free
License evidence: Inspected the upstream license text at this pinned commit; GitHub identifies it as Apache-2.0 (Apache License 2.0). No GPL only-or-later term is inferred beyond the upstream evidence.
Difficulty: hard — Transformer model/runtime dependencies and offline model handling need substantial closure work.
Workflow state: research
Existing Guix coverage: Checked 2026-08-14: GNU Guix, Nonguix, Guix Science, Guix HPC, Guix Past, Guix 'R Us, and RDE; no equivalent package with the same upstream origin was found.

Scope and blockers

Package local OCR inference only; model assets must be separately source-pinned and fetched outside the build.

Acceptance checks

guix lint -L . manga-ocr
guix build -L . manga-ocr
OCR a bundled fixture with a preinstalled local model and networking disabled.

Imported from GitHub issue/PR. Originally posted by htayj on 2026-08-14T13:17:47Z.

## Candidate - Upstream canonical URL: https://github.com/kha-white/manga-ocr - Source pinned commit/release when known: default-branch HEAD `c333b5d36e88d539d6b040b4c4cf90ad5ecd4f69` (checked 2026-08-14) - Target concrete installed deliverable: the manga-ocr local Japanese OCR command - Primary category: ai-tool - Tags: ai-ml, image-processing, ocr - Primary language normalized: Python - Build system: pyproject-build-system - SPDX expression: Apache-2.0 - License status: confirmed-free - License evidence: Inspected the upstream license text at this pinned commit; GitHub identifies it as `Apache-2.0` (Apache License 2.0). No GPL only-or-later term is inferred beyond the upstream evidence. - Difficulty: hard — Transformer model/runtime dependencies and offline model handling need substantial closure work. - Workflow state: research - Existing Guix coverage: Checked 2026-08-14: GNU Guix, Nonguix, Guix Science, Guix HPC, Guix Past, Guix 'R Us, and RDE; no equivalent package with the same upstream origin was found. ## Scope and blockers Package local OCR inference only; model assets must be separately source-pinned and fetched outside the build. ## Acceptance checks - `guix lint -L . manga-ocr` - `guix build -L . manga-ocr` - OCR a bundled fixture with a preinstalled local model and networking disabled. --- Imported from [GitHub issue/PR](https://github.com/htayj/guix-channel/issues/134). Originally posted by [htayj](https://github.com/htayj) on 2026-08-14T13:17:47Z.

tay commented

2026-09-21 14:26:14 +00:00

Owner

Goocastle recorded disposition: blocked.

A distinct manga-ocr package is not yet safely deliverable because the required Python inference closure and native noninteractive runtime contract are unresolved. Canonical upstream is https://github.com/kha-white/manga-ocr at revision c333b5d36e88d539d6b040b4c4cf90ad5ecd4f69; its LICENSE is Apache-2.0, pyproject.toml declares setuptools >= 61, setuptools_scm, Python >= 3.9, fire, fugashi, jaconv, loguru, numpy, Pillow >= 10.0.0, pyperclip, torch >= 1.0, transformers >= 4.45.0, and unidic_lite, and exposes only the manga_ocr console script. The pinned model is kha-white/manga-ocr-base at Hugging Face revision aa6573bd10b0d446cbf622e29c3e084914df9741; its card/metadata declares Apache-2.0 and its pytorch_model.bin has Git OID 5de3f9ccbbaa00931f69b6a72d0f5028e2563374 with LFS SHA-256 c63e0bb5b3ff798c5991de18a8e0956c7ee6d1563aca6729029815eda6f5c2eb. It is separate source and must be source-pinned outside the application build. The local channel contains only kha-white-manga-ocr-source in guix/tay/packages/starred-i-m.scm, not the intended manga-ocr package, so this is not a delivered duplicate. Target module guix/tay/packages/manga-ocr.scm and pyproject-build-system are appropriate. However, targeted inspection found no python-transformers, python-torch, python-fire, python-fugashi, python-unidic-lite, python-jaconv, python-loguru, or python-pyperclip definitions in the local channel, and the reviewed official Guix module check did not establish available compatible definitions for the complete closure. A Guix daemon is unavailable, so daemon-backed closure/build validation could not be performed; that limitation is not itself claimed as an upstream build failure. Upstream run.py only supports an indefinitely looping clipboard/directory CLI, and MangaOcr initialization performs a mandatory warm-up on bundled assets/example.jpg; no native one-shot stdout mode exists. A delivery can add a deterministic package-provided smoke executable that loads a preinstalled local model in offline mode, recognizes the bundled fixture, writes a text output, and emits a fixed single-line success marker, but the executable, exact argv, expected marker, and model/cache placement must be defined. Exact blockers are: (1) obtain or package a mutually compatible complete closure, including transformers >= 4.45.0, torch, fugashi/unidic_lite, and every remaining declared runtime input; (2) obtain the seven model assets at revision aa6573bd10b0d446cbf622e29c3e084914df9741 with the completed Apache-2.0/redistribution provenance record; and (3) validate a daemon-built offline package proof using the deterministic smoke wrapper. Once those prerequisites are present, this candidate should be reassessed as implementation-ready.

Goocastle recorded disposition: blocked. A distinct manga-ocr package is not yet safely deliverable because the required Python inference closure and native noninteractive runtime contract are unresolved. Canonical upstream is https://github.com/kha-white/manga-ocr at revision c333b5d36e88d539d6b040b4c4cf90ad5ecd4f69; its LICENSE is Apache-2.0, pyproject.toml declares setuptools >= 61, setuptools_scm, Python >= 3.9, fire, fugashi, jaconv, loguru, numpy, Pillow >= 10.0.0, pyperclip, torch >= 1.0, transformers >= 4.45.0, and unidic_lite, and exposes only the manga_ocr console script. The pinned model is kha-white/manga-ocr-base at Hugging Face revision aa6573bd10b0d446cbf622e29c3e084914df9741; its card/metadata declares Apache-2.0 and its pytorch_model.bin has Git OID 5de3f9ccbbaa00931f69b6a72d0f5028e2563374 with LFS SHA-256 c63e0bb5b3ff798c5991de18a8e0956c7ee6d1563aca6729029815eda6f5c2eb. It is separate source and must be source-pinned outside the application build. The local channel contains only kha-white-manga-ocr-source in guix/tay/packages/starred-i-m.scm, not the intended manga-ocr package, so this is not a delivered duplicate. Target module guix/tay/packages/manga-ocr.scm and pyproject-build-system are appropriate. However, targeted inspection found no python-transformers, python-torch, python-fire, python-fugashi, python-unidic-lite, python-jaconv, python-loguru, or python-pyperclip definitions in the local channel, and the reviewed official Guix module check did not establish available compatible definitions for the complete closure. A Guix daemon is unavailable, so daemon-backed closure/build validation could not be performed; that limitation is not itself claimed as an upstream build failure. Upstream run.py only supports an indefinitely looping clipboard/directory CLI, and MangaOcr initialization performs a mandatory warm-up on bundled assets/example.jpg; no native one-shot stdout mode exists. A delivery can add a deterministic package-provided smoke executable that loads a preinstalled local model in offline mode, recognizes the bundled fixture, writes a text output, and emits a fixed single-line success marker, but the executable, exact argv, expected marker, and model/cache placement must be defined. Exact blockers are: (1) obtain or package a mutually compatible complete closure, including transformers >= 4.45.0, torch, fugashi/unidic_lite, and every remaining declared runtime input; (2) obtain the seven model assets at revision aa6573bd10b0d446cbf622e29c3e084914df9741 with the completed Apache-2.0/redistribution provenance record; and (3) validate a daemon-built offline package proof using the deterministic smoke wrapper. Once those prerequisites are present, this candidate should be reassessed as implementation-ready.

tay added the

state:blocked

label

2026-09-21 14:26:15 +00:00