hummingbird
Use Hummingbird to convert trained traditional ML models into tensor backends, validate parity, choose backends, and troubleshoot optional dependencies.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- hummingbird
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Use Hummingbird to convert trained traditional ML models into tensor backends, validate parity, choose backends, and troubleshoot optional dependencies.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/api-overview.md4.0 KB
- 参考文档references/repo-provenance.md2.7 KB
- 参考文档references/repo-routing-metadata.json343 B
- 参考文档references/troubleshooting.md4.4 KB
- 脚本scripts/check_hummingbird_env.py5.1 KB
03 SKILL.md 正文
Hummingbird repo skill
Use this skill when a task involves the hummingbird-ml Python package: converting trained classical ML models into PyTorch, TorchScript, ONNX, or TVM-style tensor computations; validating prediction parity; saving/loading converted containers; or diagnosing optional model-source/back-end dependencies.
Quick identification
- Import package:
hummingbird,hummingbird.ml. - Distribution:
hummingbird-ml. - Core APIs:
hummingbird.ml.convert(...),hummingbird.ml.convert_batch(...), andhummingbird.ml.load(...). - Typical source models: scikit-learn estimators/pipelines, optional LightGBM/XGBoost/Prophet sklearn-style models, optional SparkML models, and ONNX-ML
ModelProtoinputs. - Typical targets:
"torch"/"pytorch","torch.jit"/"torchscript","onnx", and optional"tvm".
Install and smoke-check
Install the package and extras that match the user's workflow:
python -m pip install hummingbird-ml
python -m pip install 'hummingbird-ml[onnx]' # ONNX output or ONNX-ML input workflows
python -m pip install 'hummingbird-ml[extra]' # LightGBM, XGBoost, Prophet source workflows
python -m pip install 'hummingbird-ml[sparkml]' # SparkML source workflows
Run the bundled environment probe before deeper troubleshooting:
python scripts/check_hummingbird_env.py --json
Read repo provenance when deciding whether this skill matches the current package version or should be refreshed.
Route by task
| User task shape | Read next |
|---|---|
Build a basic convert(...) call, convert a fitted sklearn estimator to PyTorch, or validate predictions quickly |
core conversion |
Use convert_batch, handle uneven fixed-batch inference, or recover from NotFittedError, MissingBackend, or MissingConverter in the core path |
core conversion |
Decide whether a scikit-learn estimator, transformer, Pipeline, ColumnTransformer, or tree strategy is supported |
sklearn pipelines and operators |
Work with ONNX output, ONNX-ML source models, onnxruntime, onnxmltools, skl2onnx, saved containers, digests, or override_flag |
ONNX and model I/O |
| Convert LightGBM, XGBoost, SparkML, or Prophet models, or diagnose optional-library install/import failures | optional source models |
| Choose TorchScript, TVM, CUDA/GPU, threading, batching, or benchmark/performance options | advanced backends and performance |
Core operating rules
- Hummingbird converts trained models. If the source estimator must be fitted first, route the user to fit/validate it before calling
convert. - Prefer
backend="torch"for the lowest-friction CPU parity check unless the user explicitly needs ONNX, TorchScript, TVM, or GPU execution. - Provide representative
test_inputwhen tracing/exporting (torch.jit, ONNX, TVM), when shapes are not obvious, or when source packages such as XGBoost need feature-count inference. - Validate parity on the method the downstream code will call:
predict,predict_proba,transform,decision_function, orscore_samples. - Treat CUDA, TVM, SparkML, LightGBM, XGBoost, and Prophet as optional surfaces. Check imports and backend availability before promising runtime behavior.
- Do not use benchmark-scale scripts as smoke tests. Use bundled smoke/probe helpers and tiny data unless the user explicitly requests a benchmark run.
Shared references and helpers
- API overview summarizes package entry points, backend aliases, constants, containers, and dependency surfaces.
- Cross-cutting troubleshooting covers install/import, backend, optional dependency, parity, and artifact-loading failures.
- Environment checker reports package version, available backend aliases, optional imports, and torch CUDA status without installing anything.
Verification note
This skill was generated with CPU PyTorch and ONNX workflows verified. CUDA, TVM, SparkML, LightGBM, XGBoost, and Prophet are documented as optional surfaces unless the active environment separately proves those dependencies/backends.
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).