Researchers have introduced LoCA (Local Credit Assignment), a novel two-stage method for parameter-efficient fine-tuning of large language models. This approach aims to reduce computational requirements by replacing repeated end-to-end backpropagation with a single calibration pass. LoCA achieves this by fitting local maps within transformer blocks to correct prediction errors, which are then used for forward-only tuning. Evaluations on Qwen2.5 and SmolLM2-1.7B models show that LoCA can lead to lower cross-entropy, reduced GPU peak usage, and decreased CPU memory and per-pass time compared to LoRA. AI
IMPACT LoCA's efficiency gains could enable broader adoption of LLM fine-tuning on less powerful hardware.
RANK_REASON The cluster contains an academic paper detailing a new method for LLM tuning.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →