The Marin Project
Official@marin-community · United States of America
Offers specialized diagnostic and lifecycle management for JAX-based distributed training frameworks and repository maintenance.
Agent Skills by The Marin Project
Showing 51 vetted skills indexed across 1 GitHub repositories.
commit
Automates linting, committing, PR creation, and monitoring for GitHub branches.
wandb-reporting
Publishes W&B runs, reports, and artifacts for experiment results too dense for GitHub comments.
ab-test-zephyr
Runs control/treatment Zephyr benchmarks and compares Finelog stage metrics for CPU, elapsed time, and memory.
lint-review
Runs the infra/lint catalog review on a pull request and posts findings as inline GitHub comments.
update-docs
Updates durable documentation when code changes behavior or establishes reusable operational patterns.
refresh-fork
Rebase Marin external fork pins onto newer upstream bases and validate them with end-to-end tests.
noslop
Reviews branch diffs to delete low-value tests and filler prose.
consult-echo
Search and capture shared Marin knowledge through the Echo CLI.
manage-hero-run
Launch, monitor, resume, and seal production-critical Marin training runs above 1e22 FLOPs.
query-finelog
Query Finelog logs and telemetry for Iris tasks, workers, and vLLM metrics.
run-research
Coordinates multi-session research programs with logbooks, hypothesis queues, and durable snapshots.
triage-canary
Diagnose failed canary ferry runs and file GitHub issues with Slack summaries.
write-tests
Writes and reviews Marin tests for behavior changes, regressions, and test-quality requests.
task-logbook
Maintains append-only Markdown logbooks and coordinating GitHub issues for reproducible experiments.
review-grant
Decrypt, confirm, and apply IAM grant pull requests on the marin Pulumi stack.
run-ferries
Launch, monitor, and seal Marin canary and daily ferry training runs.
fix-docs
Updates and archives stale Markdown documentation in lib/iris, lib/zephyr, and lib/fray.
background-research
Searches internal Marin artifacts and external literature to produce cited prior-work research briefs.
review-pr
Runs a multi-agent correctness review of GitHub pull requests with inline comments.
write-design-doc
Create Marin design docs, research digests, and spec contracts through a seven-phase review workflow.
scan-logs
Scan oversized logs with Gemini-backed grep and summarization modes.
add-grant
Implements GCP IAM and Cloud Run IAP grants as reviewable Pulumi changes in marin-iac.
use-iris
Submit, monitor, debug, and recover Iris jobs, controllers, and accelerator sessions.
babysit-zephyr
Launch and monitor Zephyr pipeline jobs on Iris clusters with staged health checks.
Frequently Asked Questions About The Marin Project
FAQPage SchemaWhat specific tasks can be performed using The Marin Project?▼
The project enables diagnostic analysis of Zephyr pipelines, profiling of Iris controller processes via RPC, and systematic management of JAX-based training experiments. It also facilitates repository maintenance through automated GitHub issue triage, pull request reviews, and documentation parity checks.
Which technical personas benefit from these capabilities?▼
These capabilities are designed for machine learning engineers, distributed systems researchers, and performance engineers working with JAX, Ray, and high-performance computing clusters. It is specifically suited for teams managing complex experiment lifecycles and hardware-level kernel optimization.
What are the prerequisites for implementing these diagnostic features?▼
Implementation requires an existing environment configured with JAX and Ray, alongside access to the marin-community repository. Users must have appropriate permissions for GitHub repository management and access to hardware profiling data generated by Xprof or Perfetto.