Module lora 

Zero-Copy LoRA Multiplexing — context-driven neural adaptation.

Maintains a single quantized base model in memory and streams tiny LoRA (Low-Rank Adaptation) adapters based on context triggers encoded in the NQuin 5th metadata vector. Context switching takes <10 ms; the extra memory footprint is ≤15 MB per cached adapter.

Architecture

Prompt / NQuin
      │
      ▼
 ContextDetector  ──► ContextType (Medical / Legal / Chemical / …)
      │
      ▼
 LoRAAdapterManager (LRU-10 cache)
      ├── adapter_cache: HashMap<ContextType, LoRAAdapter>
      │      LoRAAdapter { lora_a: [rank × n_in], lora_b: [n_out × rank] }
      │
      ▼
 CPU apply: output += B @ (A @ x) * scaling
 GPU apply: lora_apply.wgsl dispatch (additive delta on hidden state)
      │
      ▼
 Modified hidden-state fed into fused_transformer.wgsl

NQuin metadata encoding (bits 60–48)

Bits	Field	Notes
63–60	ContextType	0=General … 5=Technical
59–56	AdapterID	0–15 (4 bits)
55–48	Confidence	0–255 → 0.0–1.0

Re-exports

pub use adapter_manager::LoRAAdapter;

pub use adapter_manager::LoRAAdapterManager;

pub use adapter_manager::LoRAError;

pub use adapter_manager::LoRAMetadata;

pub use adapter_manager::LoRATensor;

pub use context_detector::ContextDetector;

pub use context_detector::ContextType;

Modules

adapter_manager

LoRA adapter loading, caching, and application.

context_detector

Context detection for LoRA adapter selection.

webgpu_lora

GPU-accelerated LoRA delta application via wgpu compute shader.