hf-doc-build/translation-runner

Japanese Transformers documentation translation runner

This repository is the versioned runner described by the Japanese Transformers documentation synchronization plan. It is intentionally language-neutral in its implementation: language, model, prompt, glossary, paths, cache prefix, and validation thresholds live in configs/.

The runner has two safe local modes:

python -m hf_doc_translation.sync \
  --config transformers-ja \
  --checkout /Users/steven/hf/transformers \
  --base-ref ja-translation \
  --environment staging \
  --check-only \
  --no-publish

The same sync entry point performs a backfill when --force-backfill is added. A production run must use an immutable runner revision, a pinned model revision, and a clean checkout. The runner refuses to load the model when the configured TranslateGemma revision is missing or floating.

The GPU runtime loads the model through the exact Transformers checkout placed first on PYTHONPATH. Because documentation translation is text-only, it loads the checkpoint's Gemma 3 causal-language-model view and omits the unused vision tower; this gives native continuous batching the required text configuration while retaining the official TranslateGemma processor and chat template. It uses one persistent manager for cache misses and greedy decoding. The local test suite never downloads a model.

The initial rollout keeps CUDA graphs and asynchronous double-buffering disabled. They can be tuned after deterministic smoke and backfill parity pass with an attention implementation that supports their warmup requirements.

Repository contract

• English Markdown under docs/source/en is the only source of truth.
• The generated tree contains the same Markdown paths under docs/source/ja.
• Existing Japanese Markdown is regenerated, including shared paths.
• Japanese-only Markdown paths are removed from the generated tree.
• Unexpected non-Markdown assets are reported and are never deleted automatically.
• Protected source syntax is restored byte-for-byte before publication.
• Translation cache records and run manifests are ordinary JSON files under the mounted /translation-cache directory.
• The doc-builder HTML cache is read-only; the runner never passes --html_page_cache_write.

Runtime prerequisites

The current configuration pins TranslateGemma to 7d10f0b72f89a2d0f268cea30727d8b77c0d25c2 and the doc-builder checkout to e60a538eea9817ab312196d0d233604b01697265. Before a GPU Job is scheduled, an operator must accept the Gemma license for the Job service account, verify the current Jobs hardware name, and configure the scoped GitHub secret. These are external gates and are not silently inferred by this repository.

Jobs/submit.py resolves the immutable runner revision, downloads that exact snapshot inside the Job, verifies the config is present, and sets PYTHONPATH to the downloaded src/ directory. The runner dataset is not mounted as a Job volume.

Job progression

Run these commands from an operator environment with huggingface_hub 1.8 or newer and access to the hf-doc-build namespace:

export GITHUB_TOKEN="$(gh auth token)"
python Jobs/submit.py smoke
python Jobs/submit.py backfill
# after the 30-day staging soak:
python Jobs/submit.py schedule --soak-complete
unset GITHUB_TOKEN

The smoke Job must pass before the backfill is submitted. The launcher refuses to enable the daily schedule until the explicit --soak-complete acknowledgement is supplied; the 30-day staging soak and production review remain operational gates. A valid Jobs payment method and accepted TranslateGemma license are also required by the namespace/service account.

Before spending GPU time, the runtime import preflight can be run locally from a dedicated Python 3.10+ environment with pip, setuptools, wheel, and compatible PyTorch/CUDA dependencies. It clones the same staging branch, installs the exact checkout's hard dependencies, imports the pinned Transformers source, and does not download or load the 27B model:

PYTHONPATH=src python3 -m hf_doc_translation.sync preflight \
  --config configs/transformers-ja.yml \
  --repository https://github.com/stevhliu/transformers.git \
  --base-ref ja-translation \
  --environment staging

The preflight must report "passed": true before submitting a GPU smoke Job.

GITHUB_TOKEN is read only from the submitter's local environment and passed as an encrypted Job secret. It is never read from this repository or written to the Hub runner dataset. The smoke Job does not receive the GitHub secret.

The first staging PR must use the published stevhliu/transformers:ja-translation SHA. The runner publishes generated commits to the existing stevhliu/transformers fork: staging PRs target stevhliu/transformers:ja-translation, and production PRs target huggingface/transformers:main. No additional GitHub fork is required, but the Job's scoped GITHUB_TOKEN must be able to push to stevhliu/transformers and open the corresponding PRs.