Derive package version from git tags via hatch-vcs - #6
Merged
Conversation
The version in pyproject.toml was never bumped past 1.0.2, so the v2.0.0 and v2.0.1 release publishes both rebuilt 1.0.2 and were rejected by PyPI. Switch the build backend to hatchling + hatch-vcs so the version is derived from the release tag (v2.0.1 -> 2.0.1) and a release only requires pushing a tag; there is no version string left in the repo. - uv remains the frontend (sync/lock/build); uv_build rejects dynamic = ["version"], which is why the backend changes - publish workflow checks out full history so git describe sees tags - sdist is restricted to the package (tests/fixtures is ~90MB) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The [tables] extra existed solely to reproduce the pymupdf4llm table ablation in benchmarks/parsebench/RESULTS.md; the shipped package never imports those dependencies (the native table engine is the default and self-contained), and publishing an AGPL-licensed ablation dependency as a package extra invites accidental installs. The benchmark provider already degrades gracefully when the packages are absent; reproducing the ablation is a plain `pip install pymupdf4llm markdown2`. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Follow-through on dropping the [tables] extra: delete the ablation code itself, not just its published dependency. The pymupdf4llm delegation path in table_providers.py, the pymupdf4llm engine mode in tbl_engine_spike.py, and the Fast-vs-Quality timing script existed only to compare against the external parser the native table engine already ties; an unknown WARP_TABLE_PROVIDER value now simply disables the provider. The pymupdf4llm *baseline pipeline* (a rival parser in the leaderboard field run) is unaffected. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Python 3.14 changed the Linux multiprocessing default from fork to forkserver; ParseBench's evaluation workers then no longer inherit the in-process warp pipeline/adapter registrations, every layout example fails with "no provider adapter matched", and Visual Grounding silently collapses to ~10 while the run exits 0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
JSv4
added a commit
that referenced
this pull request
Jul 7, 2026
Full official faithful run from the v2.0.1 merge commit (native table engine, OCR extra installed, VG scored in-run via the fork fix): Tables 57.45 / Charts 7.04 / Content 70.81 / Sem.Format 45.81 / Visual 20.71 -> Overall 40.36 (+1.39 vs the 2026-07-02 baseline, from PR #5's content-faithfulness root fixes). A back-to-back rerun after PR #6's ablation removal reproduced every dimension to within 0.04, confirming the removal is score-neutral. Also corrects the stale claim that the 42 image-only layout failures pad the VG headline (padding is product-type-scoped and never reaches avg_layout_element_rule_pass_rate). User-authorized direct push to main (docs-only change). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
8 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description of the change
Fixes PyPI publishing:
pyproject.tomlwas never bumped past1.0.2, so the v2.0.0 and v2.0.1 release workflows both rebuilt1.0.2and were rejected by PyPI (both runs failed; PyPI is stuck at 1.0.2).1. Tag-derived versioning (hatchling + hatch-vcs). The package version is derived from the release tag (
v2.0.1→2.0.1). Cutting a release now only requires pushing a tag — there is no version string left anywhere in the repo.uv_build: it rejectsdynamic = ["version"]outright (missing field 'version') — dynamic/VCS metadata is deliberately unsupported. uv remains the frontend everywhere (uv sync/lock/buildunchanged; lockfile now records the root project version as(dynamic)).fetch-depth: 0) sogit describecan see tags.tests/fixturesis ~90MB and was at risk of being shipped by hatchling's defaults.2. Drop the benchmark-only
[tables]extra from published metadata. It existed solely to reproduce the pymupdf4llm table ablation; the shipped package never imports those deps (the native table engine is default and self-contained), and an AGPL ablation dependency has no business in PyPI metadata.3. Remove the pymupdf4llm ablation code entirely (per review): the delegation path in
table_providers.py, the spike-script engine mode, and the Fast-vs-Quality timing script. The native table engine already ties the external parser it was compared against. The pymupdf4llm baseline pipeline (rival parser in the leaderboard field run) is unaffected.Verification
uv buildat an untagged commit →2.0.2.dev0+g6884782a1(correct setuptools-scm derivation); wheel-from-sdist path (no.git) resolves via PKG-INFO.pyfiles; sdist has no tests/fixturesuvx twine check dist/*PASSED for both artifactsuv sync --frozenworks (Docker parity; images copy.gitand thepython:3.14-bookwormbase has git)pytest tests/test_package_metadata.py tests/test_service.py tests/test_table_engine.py— 76 passedget_table_provider()default remains the native engine;WARP_TABLE_PROVIDER=pymupdf4llmnow maps to no provider (warp's raw tables)Type of change
🤖 Generated with Claude Code