aim
Use Aim for experiment tracking SDK instrumentation, local/remote run storage, CLI/UI/server workflows, storage maintenance, and ML framework logging integrations.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- aim
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use Aim for experiment tracking SDK instrumentation, local/remote run storage, CLI/UI/server workflows, storage maintenance, and ML framework logging integrations.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/package-overview.md3.9 KB
- 参考文档references/repo-provenance.md2.1 KB
- 参考文档references/repo-routing-metadata.json300 B
- 参考文档references/troubleshooting.md5.0 KB
- 脚本scripts/check_aim_environment.py6.0 KB
03 SKILL.md 正文
Aim repo skill
Use this skill when a task involves Aim experiment tracking: Python SDK instrumentation, local Aim repositories, run/metric/media/artifact logging, query expressions, Aim CLI/UI/server operation, remote tracking, storage/run maintenance, watcher notifications, or ML framework callback integrations.
First checks
- Install the base public package when Aim is not already available:
python -m pip install aim - Confirm the package and entry points:
python -c "import aim; print(getattr(aim, '__version__', 'unknown'))" aim version - For a reusable diagnostic, run:
python scripts/check_aim_environment.py --check-optional - Prefer explicit repository paths in both CLI and SDK workflows. Avoid relying on whatever current directory the agent or job scheduler happens to use.
Route by task
- Python instrumentation, SDK APIs, metrics/media/params/artifacts, local repo lifecycle, query language, or missing tracked data: read
sub-skills/tracking-sdk/SKILL.md. - CLI commands, local UI, remote tracking server, notebook UI, run/storage maintenance, conversion command discovery, or watcher/notifier operation: read
sub-skills/cli-and-services/SKILL.md. - PyTorch/Lightning/Hugging Face/Keras/XGBoost/CatBoost/LightGBM/Optuna/other framework callbacks, optional dependency errors, direct
Run.trackfallbacks, or TensorBoard migration/sync: readsub-skills/framework-integrations/SKILL.md.
Root references
references/package-overview.mdexplains Aim's repo/run/sequence/context model and common end-to-end flow.references/troubleshooting.mdcovers package-level install/import, version, repository path, cleanup, optional dependency, service, and storage-risk issues.references/repo-provenance.mdrecords the source commit, package versions, evidence paths, and refresh baseline.references/repo-routing-metadata.jsonis structured router metadata for managed repo-skill import tooling.
Root script
scripts/check_aim_environment.pychecks Aim import/version/signatures, safe CLI help/version commands, and optional dependency availability without installing packages, starting services, or mutating repositories.
Operating guardrails
- Do not run
aim up,aim server, oraim-watcher startunless the user asked for a long-running service and gave host/port/lifetime expectations. - Do not run destructive or storage-mutating commands (
aim runs rm,aim storage restore,aim storage prune,aim storage reindex, or similar) without listing targets, checking backup/restore context, and getting explicit confirmation. - Do not install broad ML framework stacks for examples by default. Install or validate only the specific optional dependency required by the user's chosen integration.
- Close Aim
RunandReporesources explicitly in scripts, especially before deleting temporary repositories. - Keep generated guidance self-contained. If a workflow needs executable help, use the bundled scripts in this skill tree rather than original repository examples or tests.
Minimal SDK pattern
from aim import Repo, Run
repo = Repo.from_path("./aim-repo", init=True)
run = Run(repo=repo, experiment="demo", system_tracking_interval=None, capture_terminal_logs=False)
try:
run["hparams"] = {"lr": 1e-3, "batch_size": 32}
run.track(0.5, name="loss", step=0, epoch=0, context={"subset": "train"})
finally:
run.close()
repo.close()
Minimal CLI pattern
aim init --repo ./aim-repo
aim up --repo ./aim-repo --host 127.0.0.1 --port 43800
For remote training, route to cli-and-services: usually start aim server on the storage host and point SDK clients at an aim://... URL.
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).