firehose> #llmops

Fine-tuning's dominant cost is memory, and memory scales with the number of trainable parameters — so the central move of parameter-efficient fine-tuning (PEFT) is to update only a tiny inserted set of parameters while freezing the rest. PEFT methods split into adapter-based (additive) methods that add new weights and soft-prompt methods that introduce trainable tokens. The most popular is LoRA (low-rank adaptation): for a weight matrix of size n×m, LoRA picks a small rank r and learns two matrices (n×r and r×m); only those are updated, the original weights stay frozen, and at inference the two can be multiplied and merged back into the original weights — so, unlike classic adapters, LoRA adds no inference latency. PEFT is not only parameter-efficient but sample-efficient (hundreds to a few thousand examples versus the millions full fine-tuning can need), and adapter methods let you serve many task variants that share one base model.

Claims


Linked from