Implementation:LMCache LMCache Runtime Plugin Launcher

Knowledge Sources	LMCache
Domains	Plugin System, Process Management
Last Updated	2026-02-09 00:00 GMT

Overview

RuntimePluginLauncher discovers and launches external plugin scripts (Python or Bash) as subprocess alongside the LMCache engine.

Description

This class manages the lifecycle of runtime plugins defined in the engine configuration. It scans configured file paths or directories for .py and .sh files, applies role-based and worker-ID-based filtering using filename conventions (e.g., WORKER_0_myplugin.py), determines the interpreter from shebang lines or file extensions, and launches each plugin as a subprocess with environment variables containing the role, config JSON, worker count, and worker ID. A daemon thread captures each plugin's stdout and logs it. Plugins are automatically terminated on process exit via an atexit handler. The launcher gracefully handles missing interpreters by trying fallback options and skipping unsupported file types.

Usage

Use this launcher when the LMCache configuration specifies runtime_plugin_locations. It is instantiated during engine initialization and automatically launches all matching plugins. Call stop_plugins() to terminate all running plugin processes.

Code Reference

Source Location

• Repository: LMCache
• File: lmcache/v1/plugin/runtime_plugin_launcher.py
• Lines: 1-184

Signature

class RuntimePluginLauncher:
    def __init__(self, config, role: str, worker_count: int, worker_id: int): ...

    def launch_plugins(self): ...
    def stop_plugins(self): ...

Import

from lmcache.v1.plugin.runtime_plugin_launcher import RuntimePluginLauncher

I/O Contract

Inputs

Name	Type	Required	Description
config	LMCacheEngineConfig	Yes	Engine configuration containing runtime_plugin_locations
role	str	Yes	Current instance role (e.g., "worker", "scheduler")
worker_count	int	Yes	Total number of workers in the deployment
worker_id	int	Yes	ID of the current worker, used for plugin filtering

Outputs

Name	Type	Description
plugin_processes	list[subprocess.Popen]	List of launched plugin subprocess handles

Plugin Naming Convention

Plugins are filtered by filename parts separated by underscores:

• First part: role filter (e.g., WORKER, SCHEDULER, ALL)
• Second part (optional, if numeric): worker ID filter
• Example: WORKER_0_my_plugin.py runs only on worker 0 with role "worker"

Environment Variables

Plugins receive the following environment variables:

Variable	Description
LMCACHE_RUNTIME_PLUGIN_ROLE	Current instance role
LMCACHE_RUNTIME_PLUGIN_CONFIG	JSON-serialized engine configuration
LMCACHE_RUNTIME_PLUGIN_WORKER_COUNT	Total worker count
LMCACHE_RUNTIME_PLUGIN_WORKER_ID	Current worker ID

Usage Examples

from lmcache.v1.plugin.runtime_plugin_launcher import RuntimePluginLauncher

launcher = RuntimePluginLauncher(
    config=config,
    role="worker",
    worker_count=4,
    worker_id=0,
)

# Launch all configured plugins
launcher.launch_plugins()

# Later, stop all plugins
launcher.stop_plugins()