Struct LoRAAdapterManager 

pub struct LoRAAdapterManager {
    pub detector: ContextDetector,
    pub expected_n_in: Option<usize>,
    pub expected_n_out: Option<usize>,
    /* private fields */
}

Manages LoRA adapter loading, caching, and active-adapter switching.

The manager maintains:

• A filesystem directory for .lora files
• An LRU cache of up to 10 loaded adapters (≈150 MB at 15 MB each)
• The currently active adapter (fast path for repeated calls)

Fields

detector: ContextDetector

expected_n_in: Option<usize>

Expected embedding dimension — used for dim-check at apply time.

expected_n_out: Option<usize>

Expected hidden/output dimension — used for dim-check at apply time.

Implementations

impl LoRAAdapterManager

pub fn new(adapter_dir: impl Into<PathBuf>) -> Self

Create a manager pointing at adapter_dir.

The directory does not need to exist at construction time; adapters are loaded lazily on first switch_to.

pub fn default_path() -> PathBuf

Standard location: ~/.qualia/lora_adapters/

pub fn set_expected_dims(&mut self, n_in: usize, n_out: usize)

Hint the expected input/output dimensions so that mis-shaped adapters are rejected early rather than causing panic in apply_cpu.

pub fn detect_context(&self, prompt: &str) -> (ContextType, f32)

Detect ContextType from prompt text. LoRA context is no longer encoded in NQuin metadata bits (§4.1 migration); use active_adapter_for_hash() to look up the adapter for a specific quin.

pub fn active_adapter_for_hash(&self, content_hash: u64) -> Option<u64>

Look up the active adapter ID for a given NQuin content-hash. Returns None if no adapter has been associated with that quin.

pub fn set_adapter_for_hash(&mut self, content_hash: u64, adapter_id: u64)

Associate an adapter ID with a NQuin content-hash in the side-table.

pub fn clear_hash_associations(&mut self)

Remove all side-table associations (e.g. at end of an inference batch).

pub fn switch_to(&mut self, target: ContextType) -> Result<bool, LoRAError>

Ensure the adapter for target is loaded and set as active.

Returns Ok(true) if a switch occurred, Ok(false) if already active.

pub fn auto_switch( &mut self, prompt: &str, threshold: f32, ) -> (ContextType, f32, bool)

Detect context from prompt and switch adapter if confidence exceeds threshold.

Returns (context_type, confidence, switched).

pub fn apply_active( &self, input: &[f32], output: &mut [f32], ) -> Result<(), LoRAError>

Apply the active adapter’s delta to output (if any adapter is loaded).

input is the current hidden-state vector before the LoRA correction. output is incremented in-place: output += B @ (A @ input) * scaling.

pub fn active(&self) -> Option<&LoRAAdapter>

Return a reference to the active adapter, if any.

pub fn deactivate(&mut self)

Clear the active adapter (revert to base model behaviour).

pub fn available_adapters(&self) -> Vec<ContextType>

Check which adapters exist on disk (for UI / resource catalog).

pub fn save_adapter( &self, ctx: ContextType, adapter: &LoRAAdapter, ) -> Result<PathBuf, LoRAError>

Persist an adapter to disk in the .lora binary format.

pub fn build_synthetic( ctx: ContextType, rank: u32, alpha: f32, n_in: usize, n_out: usize, adapter_id: u8, ) -> LoRAAdapter

Build a synthetic adapter populated with Kaiming-uniform A / zero B weights. Used for fine-tuning initialisation and test harness.