repo-development
Maintain, test, regenerate, and troubleshoot a Kubeflow Pipelines source checkout without confusing checkout work with public KFP package usage.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- repo-development
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Maintain, test, regenerate, and troubleshoot a Kubeflow Pipelines source checkout without confusing checkout work with public KFP package usage.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/architecture-and-layout.md8.9 KB
- 参考文档references/development-testing.md11.0 KB
- 参考文档references/generated-code.md8.5 KB
- 参考文档references/troubleshooting.md12.8 KB
03 SKILL.md 正文
Kubeflow Pipelines Repo Development
Use this sub-skill only when the user is working on the Kubeflow Pipelines source repository: editing code, tests, generated files, CI, manifests, frontend assets, backend services, or checkout-local developer setup.
For public package use, route away instead of answering here:
- Writing KFP components, pipelines, artifacts, control flow, local execution, task modifiers ->
pipeline-authoring. - Compiling pipeline files,
Compiler().compile,kfp dsl compile,dsl-compile, PipelineSpec output ->compiler-and-cli. kfp.Client, uploads, runs, recurring runs, experiments, registry, service auth ->client-and-registry.from kfp import kubernetes, secrets, PVCs, tolerations, node selectors, pod labels, Kubernetes task configuration ->kubernetes-platform.- Other AI workflow libraries or different repositories -> use that repository's skill or ordinary repo inspection.
Operating Rules
- Confirm the task is checkout maintenance. If the user is trying to use an installed
kfppackage, route to the public-use sub-skill above. - Treat the bundled references as the operating baseline, then verify against the current checkout before running commands because Make targets, workflow matrices, and package versions may have changed.
- Obey any current checkout agent instructions and keep changes minimal. Preserve DCO/PR conventions and project style.
- Never hand-edit generated outputs. Change the source artifact and run the documented regeneration command from references/generated-code.md.
- Choose the narrowest safe command family from references/development-testing.md. Do not run cluster, Docker, GCP, release, cleanup, or credential-bound commands unless the user explicitly asked for that scope and prerequisites are available.
- When a repo edit changes public SDK behavior, update or test the corresponding public-use path too: authoring examples, compile behavior, client flows, or
kfp-kubernetestask config. - Record skipped checks honestly. Go, Node, Docker, Kind, kubectl, Ginkgo, and GPU/Kubernetes lanes are prerequisite-bound; a CPU import or compile check does not verify those lanes.
Reference Map
- references/architecture-and-layout.md: monorepo areas, package/version relationships, backend/frontend/runtime roles, and contribution conventions.
- references/development-testing.md: local setup, tool prerequisites, targeted test/format/build command families, frontend smoke testing, and dependency updates.
- references/generated-code.md: generated-file ownership, regeneration commands, CI drift checks, and frontend/backend API-client drift handling.
- references/troubleshooting.md: symptom-oriented fixes for generated drift, missing tools, cluster/e2e failures, frontend lockfile/API drift, GPU labels,
_KFP_RUNTIME=true, dependency locks, and formatting failures.
Safe Handoff Pattern
When answering a repo-maintenance task, include:
- changed area and owning paths;
- commands run and commands intentionally skipped with prerequisites;
- generated-code commands needed or not needed;
- public-package routes that may need examples/tests updated;
- CI lanes most likely to cover the change;
- any maintainer-only, credential-bound, destructive, or cluster/GPU-dependent gaps.
This sub-skill is self-contained. It names relative paths that should exist in a Kubeflow Pipelines checkout, but it does not require the original construction checkout or any absolute source path.
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).