Changelog

Changelog

v1.1.1 — 2026-08-07

Fixed

• Name tidying no longer mangles the metadata sidecar. pdf2epub writes <document>_metadata.json, and the normal rules turned that underscore into a space and title-cased the word — Some Doc Metadata.json. It reads fine but quietly drops the suffix anything parsing those files would look for. The suffix is now held aside while the document part is tidied, then reattached verbatim: 26NV001 - Neat_metadata.json → 26NV001 Neat_metadata.json.

---

v1.1.0 — 2026-08-06

First version to have actually been run. v1.0.0 was written without a machine to test on; three of its assumptions turned out to be wrong, and conversion failed outright on every document.

Added

• GPU acceleration. The image ships a CUDA build of PyTorch (torch 2.13.0+cu130), so the model work can run on an NVIDIA card. USE_GPU (default true) probes once at startup — if the host has the NVIDIA container runtime it reports the card by name and passes --gpus all; if not, it prints the install commands and carries on using the CPU, because Docker fails the whole run on an unsupported --gpus rather than one document. Measured on an RTX 4070 Ti against the CPU baseline: an 8-page document went 90s → 10s, and a 77-page scan's layout pass went from ~9 seconds per page to ~34 pages per second.
• Dockerfile.gpu. GPU runs use a locally built variant of the image. PyTorch compiles its GPU kernels at runtime through Triton, and Triton shells out to a C compiler the upstream image doesn't contain — so a GPU run on the stock image dies with Failed to find C compiler before the first page, while CPU runs (which never take that path) are unaffected. The variant adds gcc and nothing else, and is built automatically the first time a GPU run needs it.

Fixed

• Every conversion failed with a permission error. pdf2epub creates a directory inside its own site-packages at runtime. That's fine for the image's default root user, but this script deliberately runs the container as your UID so the output tree doesn't come back root-owned — and site-packages is root-owned, so the mkdir failed with [Errno 13] Permission denied: .../site-packages/static and the document died before conversion started. A writable tmpfs is now mounted over that path (APP_STATIC_DIR), which costs nothing on the host, masks nothing in the image (the directory doesn't exist there), and keeps the non-root design intact.
• The header comment claimed the image was CPU-only and told you to install pdf2epub natively for GPU support. It isn't, and you don't.

Changed

• Two remaining assumptions confirmed by a real run rather than left as caveats: the :latest tag on ghcr.io is published and pulls cleanly (8.8 GB), and pdf2epub does write its results to OUTPUT/<document name>/.

---

v1.0.0 — 2026-08-06

First release.

Added

• PDF → Markdown/EPUB conversion via pdf2epub in Docker. One container per document, with staging mounted at /data/input and the output tree at /data/output so nothing is ever written next to the source. Containers run as the invoking user's UID, which keeps the output tree from coming back root-owned.
• Three-step pipeline — copy from the input folder to staging (preserving relative paths, collisions suffixed name(1).pdf), convert, then tidy names. Already-converted documents are skipped, including ones the tidy step renamed on an earlier run.
• Two conversion modes, chosen once per run: Markdown + EPUB, where pdf2epub prompts for EPUB metadata per document and so needs a terminal, and Markdown only (--skip-epub), which runs unattended.
• Name tidying — separators normalised and title case applied to each result folder and its top-level files. Files under images/ are left alone, since the generated markdown links to them by name.
• Guard rails per conversion — a hard memory ceiling (CONTAINER_MEM_MAX, default 8G) and a wall-clock timeout (CONTAINER_TIMEOUT, default 2h), so a malformed PDF that makes the model pipeline balloon is killed locally instead of taking the desktop session with it.
• Lazy error log — errors go to error_log.txt in the output folder with a per-run header (mode, image, paths, Docker version); the terminal only shows a pointer. The file isn't created on clean runs. Failures warn and continue rather than aborting the batch.
• Interrupt safety — an EXIT trap force-removes the running container and deletes the partially written result folder, so a cancelled run leaves nothing half-converted behind.
• Image bootstrap — uses a local image if present, otherwise pulls, otherwise builds from BUILD_CONTEXT if one is configured, and only then gives up.