Deploy and run ML experiments on local, remote, Vast.ai, or Modal GPU environments.
npx skills add https://github.com/zjunlp/Mechanist --skill run-experiment-zjunlp
Please help me install this Agent Skill. Skill: run-experiment Source: https://github.com/zjunlp/Mechanist/tree/main/skills/run-experiment Command: npx skills add https://github.com/zjunlp/Mechanist --skill run-experiment-zjunlp
gpu: vast in CLAUDE.md, then asks to run the experiment — the Skill provisions a cost-optimized GPU, syncs code, runs training, downloads results, and destroys the instance to stop billing. ## Quick Start Tell the agent to run your training script as an experiment, for example: run my train.py experiment on the remote GPU server with seed 42.High-intent search queries and answers about installing and using this skill.
Add your SSH alias, conda environment, and code directory to the project's CLAUDE.md with `gpu: remote`, then invoke the run-experiment workflow. It checks GPU availability with nvidia-smi, syncs code via rsync, and launches the script in a screen session with CUDA_VISIBLE_DEVICES binding.
Set `gpu: vast` in CLAUDE.md after installing the vastai CLI and setting your API key. The workflow provisions a cost-optimized instance, syncs code to /workspace/project/, runs the experiment, downloads results, and auto-destroys the instance when auto_destroy is true.
Yes, set `gpu: modal` in CLAUDE.md and the deployment delegates to the serverless-modal workflow. Modal handles code sync and GPU allocation automatically, so no SSH, screen sessions, or manual cleanup are needed, and billing scales to zero when idle.
Yes, when `wandb: true` is set in CLAUDE.md, the workflow checks training scripts for existing wandb usage and injects wandb.init and wandb.log calls for training loss, eval metrics, GPU memory, and throughput. Project name and entity are read from CLAUDE.md.
The workflow launches each experiment in its own screen session or background process bound to a different GPU via CUDA_VISIBLE_DEVICES. For local parallel runs it uses nohup with shell wait, and each run gets its own runs/<run-id>/cost.json manifest.
Billing continues until the instance is explicitly destroyed. Set `auto_destroy: true` in CLAUDE.md so the workflow downloads results and logs, runs vastai destroy instance, and updates vast-instances.json automatically after completion.
HeXiao-55Automates deployment and execution of ML experiments on local or remote GPU servers.
raja21068Monitor running experiments and summarize JSON metrics with baseline deltas.
hve4638Rent and manage Vast.ai GPU instances for ML tasks.
NVIDIAProvides shared environment setup, credentials, remote execution, and SLURM guidance for ModelOpt workflows.