remote-run-ssh

# Remote Run over SSH Operate CVlization examples on the remote GPU reachable as `ssh l1`. This playbook keeps the remote copy minimal—just the target example folder (e.g., `examples/perception/multimodal_multitask/recipe_analysis_torch`) and the `cvlization/` library—then relies on the example’s own `build.sh` / `train.sh` Docker helpers. The user’s long-lived checkout on the remote stays untouched. ## When to Use - Heavy trainings or evaluations that require CUDA (CIFAR10 speed runs, multimodal pipelines, etc.). - Performance or regression measurements on the remote GPU after local code changes. - Producing reproducible logs / artifacts for discussions or CI baselines without pushing a branch first. ## Prerequisites - Local repo state ready to sync (uncommitted changes acceptable). - SSH config already maps the GPU machine to `l1`. - Remote host provides NVIDIA GPU (currently A10) and Docker with GPU runtime enabled. - At least ~15 GB free under `/tmp` for the slim workspace, Docker context, and caches. - Hugging Face tokens or other creds available locally if the example pulls hub assets. ## Quick Reference 1. Choose the example to run. 2. `rsync` only the example folder, `cvlization/`, and any required helper dirs to `/tmp/cvlization_remote` on `l1`. 3. On `l1`, run `./build.sh` inside the example folder to build the Docker image. 4. Run `./train.sh` (or the example’s equivalent) to launch the job with GPU access. 5. Collect logs / metrics and record the run in `var/s...