Module lora

Module lora 

Source
Expand description

LoRA (Low-Rank Adaptation) implementations for SONA

Two-tier LoRA system:

  • MicroLoRA: Rank 1-2, per-request adaptation (<100μs)
  • BaseLoRA: Rank 4-16, background adaptation (hourly)

Structs§

BaseLoRA
Base LoRA for background adaptation
LoRAEngine
Combined LoRA engine managing both tiers
LoRALayer
Single LoRA layer
MicroLoRA
Micro-LoRA for per-request adaptation

Constants§

OPTIMAL_BATCH_SIZE
Optimal batch size for processing (benchmark-validated)