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

Video Generation (LTX-2)

gpu-submit "cd ~/LTX-2 && source .venv/bin/activate && \
  python -m ltx_pipelines.run \
    --config configs/ltx-2.3-22b-distilled-2stage.yaml \
    --quantization fp8-cast \
    --prompt 'A drone shot over mountains at dawn' \
    --height 704 --width 1216 --num_frames 97 \
    --output /tmp/output.mp4" \
  --name "ltx-mountains" \
  --download /tmp/output.mp4

Claude Code Sessions

# Start a Claude Code session on the NUC
gpu-claude "Fix the failing tests in ~/project" --workdir ~/project

# Start with a specific branch
gpu-claude "Implement the feature described in PLAN.md" \
  --workdir ~/project --branch feature/new-api

Autoany EGRI Loops

# Run an EGRI optimization loop on GPU
gpu-submit "cd ~/autoany && cargo run -- \
  --config egri.toml \
  --max-iterations 50 \
  --target-metric accuracy" \
  --name "egri-optimization"

Symphony Orchestrations

# Launch a symphony workflow on GPU
gpu-submit "cd ~/symphony && cargo run -- \
  orchestrate workflow.toml" \
  --name "symphony-pipeline"

Commands Reference

All commands work via the gpu-remote.sh shell functions:

Options

--name NAME        Human-readable job name
--workdir DIR      Working directory on NUC
--branch BRANCH    Git branch to checkout before running
--download FILE    Auto-download this file when job completes
--gpu GPU_ID       Target GPU index (default: 0)
--timeout SECS     Job timeout (default: 3600)

HTTP API (gpu-server.py)

If running the Python API server on the NUC:

See references/api-reference.md for full API documentation.

Configuration

Create ~/.config/gpu-remote/config.toml on Mac:

[server]
host = "nuc-gpu"          # SSH host alias or IP
port = 8420               # API server port (if using HTTP)
user = "your-user"        # SSH user
mode = "ssh"              # "ssh" or "api"

[defaults]
workdir = "~/gpu-jobs"
timeout = 3600
gpu_id = 0

[sync]
exclude = [".git", "node_modules", "__pycache__", ".venv"]

Troubleshooting

• SSH timeout: Add ServerAliveInterval 60 to SSH config
• CUDA OOM: Check gpu-status for other jobs using VRAM, cancel or wait
• Job stuck: Use gpu-logs JOB_ID to check output, gpu-cancel JOB_ID to kill
• Server down: SSH in and restart: ssh nuc-gpu "python gpu-server.py &"
• File transfer slow: Use gpu-sync (rsync) instead of individual downloads