Add LoRA adapter support via Flare engine

[sauravpanda]

Summary

Expose Flare's LoRA adapter loading/merging through BrowserAI's API, enabling task-specific model specialization with tiny (~5-50MB) adapter files.

Use case

const ai = new BrowserAI({ engine: 'flare' });
await ai.loadModel('llama-3.2-1b-flare');

// Load a code-specialized LoRA adapter
await ai.loadAdapter('code-assistant-lora', {
    url: 'https://huggingface.co/.../adapter.safetensors',
    alpha: 16
});

// Now the model is specialized for code
const response = await ai.generateText('Write a Python function...');

Flare API

Flare already supports LoRA:

• FlareEngine.merge_lora(adapter_bytes) — load SafeTensors + merge at load time
• FlareEngine.merge_lora_with_alpha(adapter_bytes, alpha) — with custom alpha

BrowserAI API additions

interface BrowserAI {
    loadAdapter(id: string, options?: { url: string, alpha?: number }): Promise<void>;
    // Note: unmerge requires reloading the base model
}

Adapter registry

Pre-registered adapters (community contributions welcome):

• Code assistant LoRA
• Creative writing LoRA
• Instruction-following LoRA
• Math reasoning LoRA

Depends on

• Implement FlareEngine adapter for BrowserAI engine interface #295 (FlareEngine adapter)

Related

• Parent: Add Flare as a third inference engine backend #293

[sauravpanda]

Completed in PR #301 — BrowserAI.loadAdapter() delegates to FlareEngineWrapper.loadAdapter() which accepts SafeTensors adapter URLs. See README for usage example.