remote-gpu

Remote GPU Orchestrator

Operate a headless GPU server from your Mac. Submit jobs (training, inference, agents), monitor progress, and retrieve results — all over SSH or HTTP API.

Architecture

┌─────────────────┐         SSH / HTTP API        ┌──────────────────────┐
│  Mac (Control)   │ ──────────────────────────▶  │  NUC / GPU Server    │
│                  │                               │                      │
│  - Claude Code   │  Commands:                    │  - RTX 4090 (12GB)   │
│  - This skill    │    submit_job                  │  - gpu-server.py     │
│  - gpu-remote.sh │    check_status               │  - Job queue         │
│                  │    stream_logs                 │  - Claude Code       │
│                  │    download_results            │  - autoany / symphony│
│                  │    run_claude_session          │  - LTX-2 / training  │
└─────────────────┘                               └──────────────────────┘

Quick Setup

1. Configure SSH Access

# On Mac — set up passwordless SSH to NUC
ssh-keygen -t ed25519 -f ~/.ssh/nuc_gpu
ssh-copy-id -i ~/.ssh/nuc_gpu.pub user@NUC_IP

# Add to ~/.ssh/config
cat >> ~/.ssh/config << 'EOF'
Host nuc-gpu
  HostName NUC_IP_ADDRESS
  User YOUR_USER
  IdentityFile ~/.ssh/nuc_gpu
  Port 22
  ServerAliveInterval 60
EOF

# Test
ssh nuc-gpu "nvidia-smi"

2. Install Server on NUC

# SSH into NUC
ssh nuc-gpu

# Copy and start the server
pip install fastapi uvicorn psutil
python gpu-server.py --port 8420 --workdir ~/gpu-jobs

Or run scripts/setup-nuc.sh nuc-gpu from Mac to automate.

3. Use from Mac

# Via SSH (simplest)
source scripts/gpu-remote.sh
gpu-submit "python train.py --epochs 10" --workdir ~/project
gpu-status
gpu-logs job-abc123
gpu-download job-abc123

# Via HTTP API (if gpu-server.py running)
curl http://nuc-gpu:8420/submit -d '{"command":"python train.py"}'
curl http://nuc-gpu:8420/jobs

Job Types

Training Runs

# Submit a training job
gpu-submit "cd ~/project && python train.py --config config.yaml" \
  --name "lora-training-v2" \
  --workdir ~/project

# Monitor GPU usage during training
gpu-watch  # streams nvidia-smi every 5s