LoRA adapter loading and hot-swapping in browser

[sauravpanda]

Context

MobileFineTuner (Dec 2025): end-to-end LoRA fine-tuning on mobile phones. LoRA adapters are typically 5-50MB — trivial to download and swap.

Proposal

• Load base model once, swap LoRA adapters for different tasks/personas
• Store adapters in OPFS (~5-50MB each)
• Apply LoRA merge at inference time: `output = base_output + alpha * (B @ A @ input)`
• Enable personalization: users fine-tune locally, adapters never leave device
• Ship a collection of pre-made adapters (code, creative writing, math, etc.)

Impact

• One base model download, infinite specialization via tiny adapters
• Privacy-preserving personalization (adapters stored locally)