AGENT SKILL

cuml

cuML operating skill for GPU-accelerated classical ML, cuml.accel, Dask multi-GPU workflows, data utilities, and native source-build guidance.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
cuml
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
cuML operating skill for GPU-accelerated classical ML, cuml.accel, Dask multi-GPU workflows, data utilities, and native source-build guidance.
disable-model-invocation
true

02 资源文件 5

03 SKILL.md 正文

cuML

Use this repo skill for NVIDIA cuML tasks: direct GPU-native Python estimators, zero-code-change cuml.accel acceleration, cuml.dask multi-GPU workflows, supporting data/metrics/preprocessing utilities, or cuML source-build and C++/CUDA maintenance.

cuML is a CUDA/RAPIDS package. Direct cuml runtime proof requires a CUDA-capable installation and an NVIDIA GPU; CPU-only imports are not a substitute for estimator execution.

First checks

  1. Confirm the task is about cuML/RAPIDS GPU machine learning, cuml.accel, cuml.dask, libcuml, or source-build/C++ APIs.

  2. Read references/repo-provenance.md when deciding whether this skill is current for a checkout or when refreshing it.

  3. Read references/package-overview.md for package surfaces, install variants, required/optional dependencies, output/device/memory behavior, and supported-version notes.

  4. Run the bundled probe when an installed package is expected:

    python scripts/cuml_environment_probe.py --checks import cuda
    python scripts/cuml_environment_probe.py --checks import cuda health optional
    
  5. If any check fails, read references/troubleshooting.md before changing dependencies or broadening the scope.

Route by task

User task Read
Write, convert, debug, or validate direct single-GPU cuml estimator code with fit, predict, transform, output types, or model persistence. sub-skills/python-estimators/SKILL.md
Keep existing scikit-learn, UMAP, or HDBSCAN code unchanged and accelerate it with python -m cuml.accel, notebook magics, environment variables, or programmatic activation. sub-skills/sklearn-accel/SKILL.md
Build a cuml.dask workflow with Dask-CUDA, partitioned data, multi-GPU estimators, LocalCUDACluster, or distributed model collection. sub-skills/distributed-dask/SKILL.md
Generate synthetic data, split/encode/scale data, compute metrics, vectorize text, use explainers, or prepare time-series utility inputs. sub-skills/data-pipeline-utilities/SKILL.md
Build cuML from source, diagnose libcuml/CMake/CUDA toolchain issues, select native tests, or work with C++ examples/APIs. sub-skills/native-build-and-cpp/SKILL.md

Installation and backend guidance

  • Prefer the RAPIDS installation selector for current release/nightly commands and choose the CUDA/Python/package-manager combination that matches the host.
  • Pip package names are CUDA-suffixed in current RAPIDS wheels, for example cuml-cu12 or cuml-cu13; dependencies such as libcuml, cudf, rmm, pylibraft, and CuPy must match the same CUDA/RAPIDS generation.
  • cuml.dask requires additional Dask RAPIDS dependencies (rapids-dask-dependency, dask-cudf, raft-dask) matching the cuML version; dask-cuda/dask-ml are needed for many test or local-cluster workflows.
  • cuml.accel should be enabled before importing scikit-learn, UMAP, or HDBSCAN. Unsupported estimators, inputs, parameters, or versions fall back to CPU and should be diagnosed with logging/profiling rather than treated as cuML estimator failures.
  • Source builds are separate from ordinary package use. They need a full CUDA/toolchain/RAPIDS development environment and should start with the non-mutating probe in sub-skills/native-build-and-cpp/scripts/source_build_probe.py.

Validation policy

  • Always validate with tiny generated data before scaling to production-sized datasets.
  • Prefer quality metrics (accuracy_score, r2_score, adjusted Rand score, trustworthiness) over exact coefficient or label equality when comparing GPU and CPU algorithms.
  • For model serialization, load only trusted pickle/joblib artifacts. Do not unpickle data from untrusted sources.
  • For direct cuml, do not claim a CPU fallback. For cuml.accel, CPU fallback is expected for unsupported operations and should be made visible.
  • For Dask and source-build workflows, distinguish optional unverified capability from required CUDA runtime failures.

Generated helpers

  • scripts/cuml_environment_probe.py: shared import/CUDA/health/optional dependency check for installed cuML.
  • sub-skills/python-estimators/scripts/single_gpu_estimator_smoke.py: tiny KMeans, LinearRegression, RandomForest, and pickle checks.
  • sub-skills/sklearn-accel/scripts/cuml_accel_smoke.py: accelerator CLI/programmatic/profiling checks.
  • sub-skills/distributed-dask/scripts/dask_cuml_smoke.py: optional LocalCUDACluster + distributed KMeans smoke with missing-dependency diagnostics.
  • sub-skills/data-pipeline-utilities/scripts/data_utility_smoke.py: datasets, preprocessing, and metric checks.
  • sub-skills/native-build-and-cpp/scripts/source_build_probe.py: non-mutating source-build prerequisite probe.

Avoid using this skill when

  • The task is general scikit-learn usage with no GPU/RAPIDS/cuML surface.
  • The task is generic Dask administration rather than cuML estimators on Dask collections.
  • The task is CUDA kernel development unrelated to cuML/libcuml APIs.
  • The task requires downloading external datasets or running long benchmarks without explicit user approval.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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).

★ 351