mrmd-python

Independent Python runtime for MRMD notebooks. Runs as a daemon process with full GPU memory isolation.

Features

• Independent daemon process - survives if parent dies, variables persist
• GPU memory isolation - kill daemon to release VRAM (critical for vLLM)
• Auto venv detection - uses current venv or VIRTUAL_ENV
• Registry-based discovery - find runtimes via ~/.mrmd/runtimes/
• Full MRP protocol - execute, completions, inspect, variables, streaming

Installation

# With uv
uv pip install mrmd-python

# Or run directly without installing
uvx mrmd-python

Quick Start

# Start a daemon runtime (auto-detects venv)
mrmd-python

# The daemon runs in background. Use the API:
curl http://localhost:PORT/mrp/v1/capabilities

# List running runtimes
mrmd-python --list

# Kill when done (releases GPU memory)
mrmd-python --kill default

CLI Reference

# Start daemon
mrmd-python                     # Start with ID "default"
mrmd-python --id vllm           # Start with custom ID
mrmd-python --venv /path/venv   # Use specific venv
mrmd-python --port 8000         # Use specific port

# Management
mrmd-python --list              # List all running runtimes
mrmd-python --info ID           # Get runtime details
mrmd-python --kill ID           # Kill a runtime
mrmd-python --kill-all          # Kill all runtimes

# Development
mrmd-python --foreground        # Run in foreground (no daemon)

Virtual Environment Detection

When --venv is not specified, mrmd-python auto-detects:

1. Current venv (if running inside one via sys.prefix)
2. VIRTUAL_ENV environment variable
3. Falls back to system Python

API Endpoints

All endpoints at /mrp/v1/:

Architecture

~/.mrmd/
├── runtimes/
│   └── {id}.json    # Registry: pid, port, url, venv, cwd
└── logs/
    └── {id}.log     # Daemon logs

Each runtime is a fully independent process:

• Double-forked daemon (survives parent death)
• Own IPython shell with persistent variables
• HTTP server on auto-assigned port
• Registered in ~/.mrmd/runtimes/ for discovery

GPU Memory Management

For vLLM and other GPU workloads, memory is only released when the process dies:

# Load model in runtime
mrmd-python --id vllm
# ... use the model ...

# Release GPU memory
mrmd-python --kill vllm

Programmatic Usage

from mrmd_python import create_app
import uvicorn

# Create app (daemon_mode=True for use inside daemon)
app = create_app(
    cwd="/path/to/project",
    venv="/path/to/venv",
    daemon_mode=True,
)
uvicorn.run(app, host="localhost", port=8000)

Protocol

See PROTOCOL.md for the full MRP specification.