New method squeezes LLM weights by 10x with minimal accuracy loss

Researchers have developed Squeeze10-LLM, a novel post-training quantization framework designed to significantly reduce the size of large language models. This method achieves an average of 1.6 bits per weight by quantizing 80% of weights to 1 bit and 20% to 4 bits, effectively compressing models by a factor of 10. Key innovations include Post-Binarization Activation Robustness (PBAR) and Full Information Activation Supervision (FIAS) to mitigate performance degradation. Experiments on LLaMA and LLaMA2 models demonstrated Squeeze10-LLM's superior performance in sub-2bit weight-only quantization, improving accuracy on zero-shot classification tasks. AI

IMPACT Enables deployment of larger models on resource-constrained devices, potentially accelerating AI accessibility.

RANK_REASON The cluster contains a research paper detailing a new method for LLM quantization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Full Information Activation Supervision (FIAS)
• Hugging Face
• LLaMA
• LLMs
• Post-Binarization Activation Robustness (PBAR)
• Squeeze10-LLM
• Yangyang Ren

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →