LoRA Weight Save/Load

Persists LoRA adapter weights to a simple binary format.

File Format

Header (24 bytes):
  magic    : u32 = 0x4C4F5241 ("LORA")
  version  : u32 = 1
  rank     : u32
  alpha    : f32
  numLayers: u32
  reserved : u32 = 0

Per layer (2 * (A_size + B_size) bytes):
  Q_A data : f32[rank * dim]
  Q_B data : f32[dim * rank]
  V_A data : f32[rank * dim]
  V_B data : f32[kvDim * rank]

def Hesper.LoRA.IO.saveAdapter (device : WebGPU.Device) (adapter : Adapter) (path : String) : IO Unit

Save LoRA adapter weights to a binary file

Equations

• One or more equations did not get rendered due to their size.

def Hesper.LoRA.IO.loadAdapter (device : WebGPU.Device) (path : String) (dim kvDim : Nat) : IO Adapter

Load LoRA adapter weights from a binary file

Equations

• One or more equations did not get rendered due to their size.