opennmt-py
Routes OpenNMT-py workflows for data preparation, training, inference, server deployment, and checkpoint/model conversion.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- opennmt-py
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Routes OpenNMT-py workflows for data preparation, training, inference, server deployment, and checkpoint/model conversion.
- disable-model-invocation
- true
02 资源文件 9
- 参考文档references/cli-reference.md2.8 KB
- 参考文档references/compatibility.md1.2 KB
- 参考文档references/repo-provenance.md652 B
- 参考文档references/repo-routing-metadata.json535 B
- 参考文档references/troubleshooting.md2.3 KB
- 参考文档references/workflows.md1.8 KB
- 脚本scripts/check_cli_help.py678 B
- 脚本scripts/check_cuda.py644 B
- 脚本scripts/check_install.py700 B
03 SKILL.md 正文
OpenNMT-py
Use this skill for OpenNMT-py tasks that involve corpus preparation, vocabulary building, training, translation, REST serving, checkpoint maintenance, or model conversion.
Install
For a local checkout, prefer an editable install:
python -m pip install -e .
For a published install:
python -m pip install "OpenNMT-py==3.5.1"
Core runtime facts:
- Python 3.9+
- PyTorch 2.1 to <2.3
- The package expects
torch,ctranslate2,pyonmttok, andpyyamlto be present. - Optional workflows add
sentencepiece,safetensors,pandas,gradio, andbitsandbytes.
Minimal smoke check:
python -c "import onmt, torch; print('onmt_import=ok'); print(torch.__version__, torch.cuda.is_available())"
If you need CUDA paths, read references/compatibility.md and run scripts/check_cuda.py.
Route map
sub-skills/data-preparation/— build vocabularies, validate corpus YAML, apply transforms, handle source features, and debug data layout issues.sub-skills/training/— train seq2seq or language models, inspect model configs, handle multi-GPU, embeddings, alignment, LoRA, and checkpointing.sub-skills/inference/— translate, score, align, run the REST server, and use CTranslate2 or LLM-style inference configs.sub-skills/conversion/— convert external checkpoints, upgrade old checkpoints, extract or merge weights, and release models.
Start here
- Read
references/repo-provenance.mdto check whether this skill still matches the current checkout. - Read
references/cli-reference.mdfor the main console commands. - Read the owning sub-skill for the workflow you want to run.
Common entry points
- Build vocabulary:
onmt_build_vocab - Train:
onmt_train - Translate:
onmt_translate - Serve:
onmt_server - Average models:
onmt_average_models - Release models:
onmt_release_model
When to read the bundled references
references/workflows.mdfor a quick route overview and the standard quickstart sequence.references/compatibility.mdfor backend, Python, and optional-dependency notes.references/troubleshooting.mdfor install, import, CUDA, config, and runtime failures.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).