@johnhenry/aimatey-native-model-runner

Note: Previously published as [email protected].

Base class for wrapping any local model CLI or binary as an Aimatey backend — llama.cpp's main, whisper.cpp, MLX scripts, custom inference servers driven over stdio. Part of the aimatey monorepo.

This package ships an abstract class, not a ready-made backend. If you want a turnkey local backend, see @johnhenry/aimatey-native-node-llamacpp or @johnhenry/aimatey-native-apple, or use the Ollama/LM Studio backends in @johnhenry/aimatey-backend.

Installation

npm install @johnhenry/aimatey-native-model-runner

Usage

Subclass GenericModelRunnerBackend and implement the four translation hooks:

import { GenericModelRunnerBackend } from '@johnhenry/aimatey-native-model-runner';
import type { IRChatRequest } from '@johnhenry/aimatey-types';

class LlamaCliBackend extends GenericModelRunnerBackend {
  constructor() {
    super({
      command: '/usr/local/bin/llama',
      name: 'llama-cli',
      restartOnCrash: true,
    });
  }

  protected buildCommandArgs(request: IRChatRequest): string[] {
    return ['-m', '/models/model.gguf', '--temp', String(request.parameters?.temperature ?? 0.7)];
  }

  protected formatPrompt(request: IRChatRequest): string {
    return request.messages
      .map((m) => `${m.role}: ${typeof m.content === 'string' ? m.content : ''}`)
      .join('\n');
  }

  protected parseResponse(output: string) {
    return { content: output.trim() };
  }

  protected parseStreamChunk(chunk: string) {
    return { delta: chunk };
  }
}

The base class handles process lifecycle (spawn, health checks, restart on crash), stdio plumbing, and adapting everything to the BackendAdapter interface so your subclass works in a Bridge or Router like any cloud provider.

License

MIT - see LICENSE for details.