Share and run elsewhere

A remote for runsCopy Anchor Link

Runs live in .louped/, which git ignores. To share them, or to keep them beside the code, give the project a remote. The quickest way is Runs → Connect, or a first louped push:

• Leave the remote empty to use a private Hugging Face Storage Bucket of your own, named after the project (hf://buckets/<you>/<project>). louped creates the bucket.
• Paste a Hugging Face token with write access if this machine is not signed in. It is kept where hf auth login keeps it, never in the project.

Either way, the remote is saved to louped.toml. You can also set it there by hand:

remote = "hf://buckets/<user>/<name>"   # a Hugging Face Storage Bucket
# remote = "runs"                        # or a folder, such as one in the repository
# remote = "s3://<bucket>/runs"          # or S3-compatible storage such as Cloudflare R2 (pip install s3fs)

LOUPED_REMOTE overrides it. Then:

louped push   # this louped's new runs, as one new bundle in the remote
louped pull   # every bundle this louped has not pulled

Runs has the same Push and Pull buttons. If an hf:// remote finds no token, they open Connect to ask for one. Each push writes a new folder and rewrites nothing, so two people or machines never conflict. A bundle holds Inspect logs, an MLflow store with its artifacts, and result.json saying where it ran. Model weights stay out.

Run on a clusterCopy Anchor Link

1. In the app, open Launch, set the options, choose Run on (ASU Sol, a Slurm cluster or another machine) and press Export.

2. Copy what downloads to the cluster and submit it:

   Copy Text

   sbatch louped-<id>.sh     # or: bash louped-<id>.sh on a plain machine

   When the project is a pushed git commit, the export is this one file: the job clones the project at that commit. Otherwise it is louped-<id>.tar.gz, and the app says why. Unpack it and run its job.sh instead.

3. With a remote set, the job pushes its results itself. Press Pull on Runs, or run louped pull. Without one, copy louped-result-<id>.tar.gz back and open Runs → Import result (or run louped import louped-result-<id>.tar.gz).

Pulled and imported runs show like local ones, marked with where they ran.

TipsCopy Anchor Link

• Hugging Face token: gated models (such as Llama) and hf:// remotes need one on the cluster. Run hf auth login there once, or export HF_TOKEN before sbatch. A job that pushes to an hf:// remote checks for the token before it runs, not after.
• Private repositories: the cluster needs access to clone it, such as an SSH key on GitHub.
• First run: run the job's install line once on a login node. The job then reuses the downloaded packages, which avoids timeouts and compute nodes without internet.
• Smaller installs: list the extras your experiment needs in its README (extras: tracking). The job installs only those.
• Sol: jobs use the public partition and QOS. Models and packages are cached under /scratch/$USER, since home space is small.

Publish a read-only dashboardCopy Anchor Link

louped publish site/

This writes the app and every answer its pages need as static files. Serve them at a domain's root:

• Hugging Face: a static Space, hf upload <user>/<space> site/ --repo-type space.
• Vercel: vercel deploy site/.
• GitHub Pages: a user or organization site (<user>.github.io).

Nothing runs there. Launching, the Playground, labelling and comparing two runs are off.