gpubox
Skill8 starsChanged 5 months agoMITView on GitHub
Run or feasibility-check ML experiments on a personal idle GPU computer reached over SSH. Use when deciding whether a local/remote consumer GPU can run a training, evaluation, fine-tuning, or prototype job before launching it.
- Reads credentials
- Installs packages
In this repository58 other files it gives its agents.
- skills/writing-polish/humanizer/AGENTS.mdAGENTS.md
- frontend-designSkill
- ask-mattSkill
- codebase-designSkill
- code-reviewSkill
- diagnosing-bugsSkill
- domain-modelingSkill
- github-codex-keySkill
- github-ssh-key-pushSkill
- implementSkill
- improve-codebase-architectureSkill
- prototypeSkill
- setup-matt-pocock-skillsSkill
- tddSkill
- to-issuesSkill
- to-prdSkill
- triageSkill
- huggingface-accelerateSkill
- quantizing-models-bitsandbytesSkill
- evaluating-llms-harnessSkill
- ml-training-recipesSkill
- peft-fine-tuningSkill
- experiment-tracking-swanlabSkill
- unslothSkill
- matlab-add-awgnSkill
- matlab-agentic-toolkit-setupSkill
- matlab-build-appSkill
- matlab-create-live-scriptSkill
- matlab-debuggingSkill
- matlab-design-digital-filterSkill
- matlab-display-imageSkill
- matlab-install-productsSkill
- matlab-list-productsSkill
- matlab-map-database-objectsSkill
- matlab-model-serdes-systemsSkill
- matlab-modernize-codeSkill
- matlab-read-databaseSkill
- matlab-review-codeSkill
- matlab-testingSkill
- matlab-use-duckdbSkill
- matlab-write-databaseSkill
- grillingSkill
- grill-meSkill
- grill-with-docsSkill
- handoffSkill
- teachSkill
- writing-great-skillsSkill
- ara-compilerSkill
- ara-research-managerSkill
- research-paper-notesSkill
- ara-rigor-reviewerSkill
- brainstorming-research-ideasSkill
- creative-thinking-for-researchSkill
- academic-plottingSkill
- ml-paper-writingSkill
- paper-spineSkill
- humanizerSkill
- prompt-optimizerSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 18
Create an SSH alias on the agent machine:
- Distinctive: line 40
Before launching a job, collect hardware and environment facts:
- Distinctive: line 72
Use the documented Python environment explicitly.
- Distinctive: line 100
Do not leave long jobs running silently.
- Distinctive: line 112
Keep public examples generic and ask users to fill in their own machine details.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
gpubox - Personal Idle GPU Box
This skill is a template for using a spare personal computer as a small remote GPU box for AI experiments. It is intentionally hardware-agnostic: the machine could be a Windows laptop with WSL, a Linux desktop, or any always-on computer reachable over SSH.
The main rule is simple: check feasibility before running anything. Consumer GPUs are useful, but VRAM is usually the binding constraint.
Suggested Setup
Create an SSH alias on the agent machine:
Host gpubox
HostName <tailscale-or-lan-hostname>
Port 22
User <remote-user>
IdentityFile ~/.ssh/<your-gpubox-key>
IdentitiesOnly yesRecommended remote environment:
- NVIDIA GPU with working CUDA drivers.
- Python environment named
ml, or another documented environment name. - PyTorch installed with the matching CUDA build.
tmuxorscreenfor long-running jobs.- Fast local disk for datasets and checkpoints.
- Optional private network such as Tailscale, ZeroTier, or a LAN-only SSH setup.
Step 1 - Feasibility Gate
Before launching a job, collect hardware and environment facts:
ssh -o ConnectTimeout=12 gpubox 'nvidia-smi --query-gpu=name,memory.total,memory.free,memory.used,utilization.gpu --format=csv,noheader'
ssh gpubox 'python --version; which python; python - <<PY
import torch
print("torch", torch.__version__)
print("cuda", torch.cuda.is_available())
if torch.cuda.is_available():
print(torch.cuda.get_device_name(0))
PY'Then estimate:
- Model parameter count and dtype.
- Batch size, sequence length, image size, or sample length.
- Optimizer state memory.
- Activation memory.
- Dataset and checkpoint disk requirements.
- Expected runtime.
Give a clear verdict before running:
fits: likely safe to run.fits with changes: reduce batch size, use LoRA/QLoRA, enable gradient checkpointing, shorten context, or use smaller inputs.does not fit: recommend CPU-only preprocessing, cloud GPU, or a smaller experiment.
Step 2 - Run Pattern
Use the documented Python environment explicitly. Non-interactive SSH sessions often do not source shell startup files.
Examples:
ssh gpubox 'conda run -n ml python train.py'
ssh gpubox 'conda run -n ml pip install <package>'For longer runs:
ssh gpubox 'tmux new -d -s job "conda run -n ml python train.py 2>&1 | tee ~/job.log"'
ssh gpubox 'tail -n 40 ~/job.log; nvidia-smi --query-gpu=memory.used,utilization.gpu --format=csv,noheader'Step 3 - Practical Defaults
For small GPUs, prefer:
- LoRA or QLoRA instead of full fine-tuning.
- Small batch sizes with gradient accumulation.
- Mixed precision where stable.
- Gradient checkpointing when activation memory dominates.
- Small validation subsets before full evaluation.
- Checkpointing less often if disk is limited.
Do not leave long jobs running silently. Record the command, environment, dataset, checkpoint, and result location so the experiment can be reproduced.
Privacy Notes
Do not publish:
- Real SSH hostnames, private IP addresses, usernames, or ports.
- Private key filenames tied to a real machine.
- Tailscale node names.
- Absolute local paths that identify a person or workstation.
Keep public examples generic and ask users to fill in their own machine details.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

