NestQuant, a lattice-based post-training quantizer, cuts the 4-bit perplexity gap for Llama-3-8B by more than 55% versus state-of-the-art methods.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
NestQuant: Nested Lattice Quantization for Matrix Products and LLMs
NestQuant, a lattice-based post-training quantizer, cuts the 4-bit perplexity gap for Llama-3-8B by more than 55% versus state-of-the-art methods.