メニュー
AstroAI only: drive a CANFAR Ray cluster with astroai — cluster start/status/stop, dashboard, run/jobs submit/list/logs. Use when the user wants Ray batch compute on AstroAI images, not generic CANFAR batch (see canfar-batch).
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/astroai/canfar-lab --skill astroai-rayコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
Reframe an asserted conclusion as a neutral question before evaluating it. Use when a user presents a plan, interpretation, diagnosis, or preference and may be seeking confirmation; do not use for ordinary direct execution.
Apply a compact evidence-first integrity check to scientific claims, research ideas, analyses, and technical decisions. Use when correctness, causal interpretation, reproducibility, or an important project decision is at stake; agreement is never the objective.
AstroAI session workflow: pixi/uv under $WORK, astroai save/resume, Ray cluster start/run, agents. Use for new users or "how do I work here".
| name | astroai-ray |
| description | AstroAI only: drive a CANFAR Ray cluster with astroai — cluster start/status/stop, dashboard, run/jobs submit/list/logs. Use when the user wants Ray batch compute on AstroAI images, not generic CANFAR batch (see canfar-batch). |
astroai (AstroAI images only)Scope: Requires AstroAI session images with
astroaiCLI. For platform headless/batch without Ray, usecanfar-batch.
One CLI. Installed on AstroAI session images.
astroai --help
Usual path: one autoscaling manager, then a job with --cpus.
astroai cluster start
export ASTROAI_RAY_JOBS_ADDRESS=… # skip inside the manager
astroai run train.py --cpus 2
cluster start writes ~/.config/canfar/lab/ray-manager.env, creates the
manager if needed, and lets Ray add ray-as-* workers when the job needs
CPUs. Same as AstroAI hub Start batch compute.
Do not call ray job submit. The job command is astroai run.
astroai cluster start
astroai cluster start --max-workers 8 --cores 2 --ram 8
astroai cluster start --min-workers 1 --gpus 1 --timeout 1800
Prints export ASTROAI_RAY_JOBS_ADDRESS=…. The caller must export it in
their shell (a CLI cannot export into the parent). --json returns
manager_url, jobs_address, dashboard_url, cluster_phase,
joined_workers, and autoscaling.
If a manager was already running, restart_manager is true: stop it and
re-run cluster start so the new manager sources the env file.
start is safe to call again. It does not create a second manager.
export ASTROAI_RAY_JOBS_ADDRESS=… # from start; skip inside the manager
astroai run train.py --cpus 2
astroai jobs submit --cmd 'python -m mosaic.stack --in /arc/projects/g/in' --wait
astroai jobs list
astroai jobs logs <run-id>
--input / --output URIs are stored on the Ray job. They are not copied.
Put data on /arc. /scratch dies with the session.
astroai cluster status
astroai cluster stop # destroys workers AND the manager
astroai cluster dashboard # Ray Dashboard URL (jobs, nodes, logs)
astroai cluster dashboard iframe # notebook / marimo
joined: N / M is the health number. auth: ok means CANFAR credentials
are present.
astroai status is session CPU/disk/quota, not the cluster. Use
cluster status for the cluster.
--max-workers, never by launching
workers yourself.--json when you will parse. Plain text when showing the user.start is safe to call again. It does not create a second manager.start, jobs are astroai run (or astroai jobs submit --cmd).cluster stop when the user is done.start already waits. If join is slow, give the user
cluster dashboard instead of polling forever.job_* tools need Ray. Cluster tools do not. Same CLI functions
either way.ray-manager.env by hand. cluster start
and the hub button do that.