RunPod CLI Dashboard

CLI tool for managing RunPod GPU pods — create, connect, run commands, and open services in one step.

Installation

uv add git+https://github.com/Butanium/runpod-cli-dashboard

# Or with uv (development)
uv sync

This installs the runpod-cli command.

Quick start

# First run — interactive setup (username, API key, git config, HF token)
runpod-cli config

# Launch a dashboard pod (default: A40 GPU)
runpod-cli

# Launch a vLLM server with LoRA support
runpod-cli task=vllm gpu=a100 app_port=8000 task.model=meta-llama/Llama-3.1-8B-Instruct

# Stop a pod (keeps it for later reuse)
runpod-cli pause

# Destroy a pod
runpod-cli destroy

Commands

Command	Description
runpod-cli	Launch/connect to a pod (default task: dashboard)
runpod-cli config	Create or update global user configuration
runpod-cli config --local	Create or update local config (.runpod/config.yaml)
runpod-cli pause / stop	Stop the latest pod (can be resumed later)
runpod-cli destroy	Terminate and delete the latest pod

Configuration

User config

Created interactively on first run or via runpod-cli config:

• username (required) — prefixes pod names for identification
• api_key — RunPod API key (or set RUNPOD_API_KEY env var)
• git_name / git_email — for git on pods
• hf_token — HuggingFace token for private models

Stored in ~/.config/runpod-cli/config.yaml (global). Use runpod-cli config --local to save per-project overrides in .runpod/config.yaml — local values override global.

Hydra overrides

The CLI uses Hydra for configuration. Override any setting on the command line:

# Different GPU
runpod-cli gpu=h100

# Multiple overrides
runpod-cli gpu=a100 storage=large open_ide=false stream_output=false

# Show resolved config without running
runpod-cli --cfg job

Built-in config options

GPU (gpu=): a40 (default), a100, h100, 2a100, 4a40

Storage (storage=): default (100GB vol / 200GB disk), large (500GB / 500GB)

Task (task=): dashboard (default), testing, vllm

IDE (ide=): cursor (default), vscode

Custom config overrides

Add your own configs at two levels — both are auto-discovered, no extra flags needed:

Global (~/.config/runpod-cli/configs/) — available everywhere:

~/.config/runpod-cli/configs/
  task/
    my_vllm.yaml       # runpod-cli task=my_vllm
  gpu/
    custom.yaml         # runpod-cli gpu=custom

Local (.runpod/ in current directory) — per-project, takes precedence over global:

your-project/
  .runpod/
    task/
      experiment.yaml   # runpod-cli task=experiment

Precedence: local .runpod/ > global ~/.config/runpod-cli/configs/ > built-in defaults.

vLLM server

Launch a vLLM OpenAI-compatible server with dynamic LoRA support:

runpod-cli task=vllm gpu=a100 app_port=8000 task.model=meta-llama/Llama-3.1-8B-Instruct

This uses RunPod's official vLLM template (vllm/vllm-openai:latest). The server starts as the container's main process — no SSH/tmux needed.

Endpoints available at http://<pod-ip>:<port>/v1/:

• /v1/chat/completions
• /v1/completions
• /v1/models
• /v1/load_lora_adapter / /v1/unload_lora_adapter (dynamic LoRA)

Override vLLM settings:

# Custom LoRA config
runpod-cli task=vllm gpu=h100 app_port=8000 \
  task.model=my-org/my-model \
  task.max_loras=8 \
  task.max_lora_rank=128 \
  task.gpu_memory_utilization=0.9

How it works

1. Creates (or resumes) a RunPod GPU pod
2. For remote_command tasks: SSHes in, runs the command in a tmux session
3. For docker_args tasks (vLLM): the container starts the server directly
4. Updates ~/.ssh/config for easy SSH access (ssh <pod-name>)
5. Opens the service URL in your browser