Submit, monitor, debug, and recover Iris jobs, controllers, and accelerator sessions.
npx skills add https://github.com/marin-community/marin --skill use-iris
Please help me install this Agent Skill. Skill: use-iris Source: https://github.com/marin-community/marin/tree/main/.agents/skills/use-iris Command: npx skills add https://github.com/marin-community/marin --skill use-iris
Operating the Iris job orchestration platform involves many distinct workflows—job submission, federated scheduling, controller rollouts, accelerator reservations, and stuck-pod recovery—and performing them incorrectly can kill running jobs or corrupt cluster state. This Skill routes each request to the correct operational procedure and enforces safety boundaries.
Ask the assistant to describe the status of job /myuser/my-training-run on the marin cluster using Iris.
High-intent search queries and answers about installing and using this skill.
Run `uv run iris --cluster=<cluster> job describe <job>` to see job state, and `task describe` or `task events` for individual tasks. For federated root jobs, also check the parent's job list, peer list, and the federated_jobs table.
Follow the monitor-job reference: keep one monitoring state file, check logs every 570 seconds after an initial 120-second wait, and confirm progress from live logs, W&B timestamps, and checkpoint movement. A controller RUNNING state alone is not completion.
Use scripts/iris/dev_gpu.py or dev_tpu.py with the cluster config to allocate, connect, and release a session. The default GPU allocation holds one H100 node; TPU sessions require an explicit --tpu-type such as v5p-8.
No. `iris cluster restart` kills all workers and jobs and requires explicit approval for the named cluster. Job recovery uses cancel and resubmit with the exact original command, never a cluster restart.
Start read-only with kubectl get pod and events using the kubeconfig from the cluster config. Cordon the node, try ordinary deletion first, and only force-delete after provider-confirmed node reboot for GPU-bound processes.
Write bounded diagnostics to $IRIS_OUTPUT_DIR, which Iris preserves as one outputs.tar.zst archive per attempt. Find the archive URI and its upload state with `attempt describe <task>:<attempt>`; use direct object-storage writes for large outputs.
open-thoughtsOrchestrate datagen jobs on Marin's Iris TPU cluster from launch to cleanup.
jaredcroxtonLog ML experiment metrics, configurations, artifacts, and lineage via W&B SDK.
fractaryQuery CloudWatch logs to investigate incidents and generate reports.
Mte90Retrieve Jira issue details and link Zephyr test cases.
