Define and launch Marin SkyRL reinforcement learning experiments on Iris. It guides configuration checks, model and data dependencies, small trial runs, and progress verification while keeping artifact versions and runtime revisions explicit.
中文简介正在准备中。当前显示英语简介。
Define, validate, submit, or restart a Marin SkyRL experiment through its artifact main. Use for RL launch files, SkyRLRolePlan or SkyRLTopology changes, MarinSkyRL configuration, RL smoke tests, and requests to launch or relaunch RL on Iris.
安装前可以查看交给 AI 代理的指令内容。
Use the experiment's artifact graph as the launch interface. For the configuration model, role arithmetic, and ownership boundaries, read RL launching.
main that constructs and returns the requested terminal ArtifactStep handles.@rl_build_options. Do not add another submission path to an experiment.SkyRLRolePlan in SkyRLSpec. Put cluster placement
in IrisSkyRLExecution.SkyRLLaunchConfig YAML as the only Marin-to-MarinSkyRL boundary. Extend that
config instead of adding request envelopes or internal per-field command-line flags.Run the main without --run first, using an immutable calendar version for a real experiment:
uv run python -m experiments.<module> --version YYYY.MM.DD <selection-options>
Treat a preflight failure as a recipe error. Fix the source config or role plan instead of weakening the validator. Check the printed runtime commit, physical allocation, role geometry, entrypoint, batch sizes, artifact identities, output paths, and terminal stage.
Use a fresh RL artifact version whenever SkyRLRuntime.commit changes. The temporary checkpoint
root follows the artifact name and version, so reusing a version can make resume_mode=latest load
private Torch or distributed state written by the old runtime. Reuse an RL version across a repin
only when checkpoint compatibility has been established explicitly or resume points at a fresh
checkpoint root.
Before a large launch, run a smoke through the same strategy, artifact handoffs, and role geometry. Confirm the runtime contains the required backend fixes and that host resources, credentials, context budget, checkpoint/export policy, and W&B identity are explicit.
Launch through the same main:
uv run python -m experiments.<module> --version YYYY.MM.DD <selection-options> --run
Choose the terminal stage explicitly when a main offers one. For example, use --stage rl when the
request is only to train; a default evaluation stage may add downstream GPU work.
For a live submission or inspection, use use-iris. Record the coordinator ID, child job ID,
artifact version, runtime commit, and topology in the task's durable surface. Verify all expected
tasks join, the configured model-loading path is active, optimizer steps advance, and terminal
checkpoint/export metadata exists before calling the smoke successful.
If a failure was deterministic from the artifact inputs, add a construction-time check with a behavior-focused regression test. Diagnose service outages, hardware faults, and data-dependent failures operationally. Cancel or resubmit only with current-thread authorization; a retry uses a fresh artifact version when the runtime commit changed.
还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Compare Zephyr changes using the same prepared dataset and matching execution settings. Analyze Finelog metrics for each processing stage to distinguish CPU efficiency, elapsed time, and peak memory changes, then report improvements, regressions, or tradeoffs.
中文简介正在准备中。当前显示英语简介。
Register Hugging Face datasets for Marin by inspecting their structure without downloading the full dataset. Identify text fields, configurations, and data splits, then add a dataset module and validate the mapping with a sample.
中文简介正在准备中。当前显示英语简介。
Turn a specified GCP access request into a reviewable Pulumi change and pull request. Select the narrowest suitable role and encrypt personal email identities, covering resource permissions and Cloud Run viewer access through IAP, an access gateway.
中文简介正在准备中。当前显示英语简介。
Add or update Pallas/Mosaic kernels, specialized computation routines for JAX. The skill covers reference implementations, output and gradient checks, public wrappers, fallback behavior, performance benchmarks, and hardware-specific tuning when requested.
中文简介正在准备中。当前显示英语简介。
Preserve old experiment scripts with a dated Git tag, a named snapshot of the code, and add comments linking related GitHub issues to the archived files and cleanup pull request.
中文简介正在准备中。当前显示英语简介。
Launch and continuously monitor a specified Zephyr pipeline on Iris when asked. It checks coordinator and worker health, tracks stage progress, identifies stalled or repeated failures, and requires explicit approval before stopping or restarting jobs.
中文简介正在准备中。当前显示英语简介。