basilica-cli-helper: Скилл для управления GPU и ML

--- name: basilica-cli-helper description: This skill should be used when users need to rent GPUs, run ML training jobs, or manage compute resources on Basilica's decentralized GPU marketplace. Use it for PyTorch/TensorFlow training, distributed training setup, GPU rental management, cost monitoring, or any Basilica CLI workflows. Includes workaround for non-TTY environments like Claude Code. --- # Basilica CLI Helper **Rent GPUs and run ML training jobs on Basilica's decentralized compute platform.** This skill helps access high-performance GPUs through Basilica's CLI. Use this when needing to: - Rent GPUs for machine learning training - Run distributed training jobs - Manage compute resources and costs - Execute code on remote GPU instances ## Claude Code TTY Limitation **IMPORTANT**: The `basilica up` command requires interactive TTY selection, which fails in Claude Code's non-terminal environment. ### Workaround: Use the bundled script ```bash # List available GPUs (non-interactive) python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py list # Filter by GPU type and cloud python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py list --gpu-type h100 --compute secure-cloud # Rent by selection number (from list output) python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py rent --select 1 # Rent by offering ID directly python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py rent hyperstack-165 ``` This script uses the Basilica REST API directly, bypassing the interactive selection. ## Quick Reference ### Essential Commands | Command | Description | |---------|-------------| | `basilica login` | Authenticate with Basilica | | `basilica ls` | List available GPUs with pricing | | `basilica ps` | List your active rentals | | `basilica status <uid>` | Check rental status | | `basilica exec "cmd" --target <uid>` | Execute command on rental | | `basilica ssh <uid>` | SSH into instance | | `basilica cp <src> <dst>` | Copy files to/from instance | | `basilica down <uid>` | Terminate rental | | `basilica balance` | Check account balance | ### Cloud Types Basilica offers two compute sources: | Cloud | Flag | Description | |-------|------|-------------| | Secure Cloud | `--compute secure-cloud` | Datacenter GPUs (Hyperstack, DataCrunch, Lambda) | | Community Cloud | `--compute community-cloud` | Decentralized miner GPUs (Bittensor network) | ## Command Reference ### Authentication ```bash # Standard login (opens browser) basilica login # Device code flow (for WSL, SSH, containers) basilica login --device-code # Logout basilica logout # Check balance basilica balance ``` ### Listing GPUs ```bash # List all available GPUs basilica ls # Filter by GPU type basilica ls h100 basilica ls a100 # Filter by cloud type basilica ls --compute secure-cloud basilica ls --compute community-cloud # Additional filters basilica ls --gpu-min 2 --gpu-max 8 basilica ls --price-max 5.00 basilica ls --memory-min 80 basilica ls --country US # JSON output basilica ls --json ``` ### Starting Rentals **Note**: `basilica up` requires interactive TTY. In Claude Code, use the workaround script above. In a terminal: ```bash # Interactive GPU selection basilica up # Filter to GPU type (still requires selection) basilica up h100 basilica up a100 # Specify cloud type basilica up --compute secure-cloud basilica up --compute community-cloud # Detached mode (don't auto-connect) basilica up -d # Additional options basilica up --gpu-count 4 basilica up --country US basilica up --no-ssh # Faster startup without SSH ``` ### Managing Rentals ```bash # List active rentals basilica ps # Show rental history basilica ps --history # Filter by cloud type basilica ps --compute secure-cloud # Check specific rental basilica status <rental-id> # Terminate rental basilica down <rental-id> # Terminate all rentals basilica down --all # Restart container basilica restart <rental-id> ``` ### Executing Code ```bash # Execute command on rental basilica exec "python train.py" --target <rental-id> # If only one active rental, --target is optional basilica exec "nvidia-smi" basilica exec "pip install -r requirements.txt" ``` ### File Transfer ```bash # Copy file to rental basilica cp train.py <rental-id>:/workspace/ # Copy directory to rental basilica cp ./project/ <rental-id>:/workspace/project/ # Download file from rental basilica cp <rental-id>:/workspace/model.pth ./ # Download directory basilica cp <rental-id>:/workspace/checkpoints/ ./checkpoints/ ``` ### SSH Access ```bash # SSH into instance basilica ssh <rental-id> # Port forwarding (e.g., Jupyter) basilica ssh <rental-id> -L 8888:localhost:8888 # Remote port forwarding basilica ssh <rental-id> -R 9999:localhost:9999 ``` ### Logs ```bash # View logs basilica logs <rental-id> # Follow logs in real-time basilica logs <rental-id> --follow # Tail last N lines basilica logs <rental-id> --tail 100 ``` ### SSH Key Management ```bash # Add SSH key basilica ssh-keys add # Add with specific file basilica ssh-keys add --file ~/.ssh/id_rsa.pub # List registered keys basilica ssh-keys list # Delete key basilica ssh-keys delete ``` ### API Token Management ```bash # Create API token basilica tokens create <name> # List tokens basilica tokens list # Revoke token basilica tokens revoke <name> ``` ### Funding ```bash # Show deposit address basilica fund # List deposit history basilica fund list --limit 100 ``` ## Common Workflows ### PyTorch Training (Claude Code) ```bash # 1. List available GPUs using workaround script python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py list --compute secure-cloud # 2. Rent GPU by selection number python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py rent --select 1 # Note the rental ID from output # 3. Copy training files basilica cp train.py <rental-id>:/workspace/ basilica cp requirements.txt <rental-id>:/workspace/ # 4. Install dependencies and run training basilica exec "pip install -r /workspace/requirements.txt" --target <rental-id> basilica exec "python /workspace/train.py --epochs 10" --target <rental-id> # 5. Download results basilica cp <rental-id>:/workspace/model.pth ./ # 6. Terminate when done basilica down <rental-id> ``` ### PyTorch Training (Terminal with TTY) ```bash # 1. Start GPU rental (interactive selection) basilica up h100 --compute secure-cloud -d # Note the rental ID # 2. Copy training files basilica cp train.py <rental-id>:/workspace/ # 3. Run training basilica exec "python /workspace/train.py" --target <rental-id> # 4. Download results and cleanup basilica cp <rental-id>:/workspace/model.pth ./ basilica down <rental-id> ``` ### Jupyter Notebook ```bash # 1. Start rental and SSH with port forward basilica up h100 -d basilica ssh <rental-id> -L 8888:localhost:8888 # 2. In SSH session, start Jupyter jupyter lab --ip=0.0.0.0 --port=8888 --no-browser # 3. Open http://localhost:8888 in browser ``` ### Check Costs ```bash # Check current balance basilica balance # View rental history with costs basilica ps --history # Get deposit address if needed basilica fund ``` ## Troubleshooting ### "Selection failed: not a terminal" This occurs when running `basilica up` in Claude Code. Use the workaround script: ```bash python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py list python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py rent --select <number> ``` ### "Not logged in" ```bash basilica login # Or for non-browser environments: basilica login --device-code ``` ### "Insufficient balance" ```bash basilica balance # Check current balance basilica fund # Get deposit address ``` ### "No GPUs available" ```bash basilica ls # Check different GPU types basilica ls --compute community-cloud # Try community cloud ``` ### "SSH key not registered" ```bash basilica ssh-keys add ``` ### "Connection timeout" ```bash basilica status <rental-id> # Check if still running basilica logs <rental-id> # Check for errors ``` ## GPU Selection Guide | Use Case | Recommended GPU | Typical Price | |----------|-----------------|---------------| | Small models, fine-tuning | A100 (1x) | $1-2/hr | | Medium models | H100 (1x) | $2-3/hr | | Large models | 4-8x A100/H100 | $5-20/hr | | Inference testing | Any 1x GPU | $1-3/hr | ## Resources - **Basilica Homepage**: https://basilica.ai - **CLI Help**: `basilica help <command>` - **Version**: Check with `basilica --version` ## Scripts Reference ### basilica_up.py Non-interactive GPU rental script for Claude Code and other non-TTY environments. ```bash # Show help python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py --help # List offerings python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py list [--gpu-type TYPE] [--compute CLOUD] # Rent by selection python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py rent --select NUMBER # Rent by offering ID python ~/.claude/skills/basilica-cli-helper/scripts/basilica_up.py rent OFFERING_ID ``` The script caches offering data to `/tmp/basilica_offerings.json` for the `--select` option.