Struct PeftAdapterConfig 

pub struct PeftAdapterConfig {
    pub r: usize,
    pub lora_alpha: f64,
    pub target_modules: Vec<String>,
    pub target_modules_pattern: Option<String>,
    pub use_rslora: bool,
    pub layers_to_transform: Option<Vec<usize>>,
    pub trainable_token_indices: Vec<u32>,
    pub modules_to_save: Vec<String>,
    pub lora_embedding: bool,
}

2026-09-26: The fields of a PEFT adapter_config.json that the engine uses, as validated by parse_peft_adapter_config. lora_dropout, a training-only setting, is ignored.

Fields

r: usize

2026-09-26: LoRA rank, greater than 0.

lora_alpha: f64

2026-09-26: LoRA alpha, finite and positive; the JSON may hold an integer or a float.

target_modules: Vec<String>

2026-09-26: The target_modules list as written: bare module names or full paths, each validated on its last .-segment. Per-layer matching is the weight loader’s.

target_modules_pattern: Option<String>

2026-09-26: PEFT’s regex form of target_modules (a JSON string other than all-linear), kept as written and not expanded. It skips the name-level allow-list, but every adapter tensor is still checked by classify_key, which accepts only the projections it names.

use_rslora: bool

2026-09-26: rsLoRA: scaling is alpha/sqrt(r) instead of alpha/r. The config must state it; an absent key is rejected rather than read as false.

layers_to_transform: Option<Vec<usize>>

2026-09-26: The layers_to_transform list, if present. Nothing reads it; the weight loader decides which layers receive deltas.

trainable_token_indices: Vec<u32>

2026-09-26: Token ids for the trainable-token overlay, unique and ascending, from the trainable_token_indices list or the list shared by its embed_tokens and lm_head entries. Empty means no overlay.

modules_to_save: Vec<String>

2026-09-26: The accepted modules_to_save leaves, embed_tokens and lm_head, sorted and deduplicated. Any other leaf is a REJECT(modules_to_save). Empty means no full-module overlay.

lora_embedding: bool

2026-09-26: Always false from this parser. Low-rank embedding LoRA (lora_embedding_A/B) is detected from tensor names by the overlay loader, which rejects it.

Implementations

impl PeftAdapterConfig

pub fn scaling(&self) -> f32

2026-09-26: The delta scale, y += scaling() * (x @ Aᵀ) @ Bᵀ: alpha/r, or alpha/sqrt(r) with use_rslora.

Trait Implementations

impl Clone for PeftAdapterConfig

fn clone(&self) -> PeftAdapterConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PeftAdapterConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.