Struct LoraBank 

pub struct LoraBank {
    pub slot: Option<SlotEmbed>,
    pub meta: HashMap<String, String>,
    /* private fields */
}

Every branch an adapter file carries, keyed by the projection it belongs to — the container’s name with the dit. prefix dropped, which is what the adapters use after their own diffusion_model..

Fields

slot: Option<SlotEmbed>

meta: HashMap<String, String>

Implementations

impl LoraBank

pub fn load(path: &Path, strength: f32) -> Result<LoraBank, String>

Read an adapter. scale multiplies every branch — the strength dial.

alpha is honoured when the file records it: the trained convention is scale = strength · alpha / rank, and an adapter that ships neither alpha nor lora_alpha is taken at strength as-is, which is what the diffusers loaders do for a file whose A/B are already scaled.

pub fn len(&self) -> usize

pub fn is_empty(&self) -> bool

pub fn rank(&self) -> usize

The rank the file was trained at, for the log line.

pub fn keys(&self) -> Vec<&str>

Every projection this adapter names. The caller binds what its container has and reports the rest: an adapter that trained a module we fold away (adaLN, on a curve-form pack) must say so rather than render as though it had been applied.

pub fn branch_for( &self, name: &str, out: usize, inn: usize, ) -> Result<Option<LoraBranch>, String>

The branch for a projection of a KNOWN shape. A branch whose matrices do not match the panel it will be handed belongs to another model — add writes n·out floats through a raw pointer, so a mismatch is a buffer overrun, not a bad picture. Refuse it by name instead.

pub fn branch(&self, name: &str) -> Option<LoraBranch>

The branch for a container tensor name, if this adapter carries one. name is the projection without .weight — dit.transformer_blocks.0.attn1.to_q.

pub fn check_reference_convention(&self) -> Result<(), String>

How the adapter wants reference tokens placed. Both are refused rather than approximated when they are not what this implementation does.