A single LLM can be trained to support multiple weight precisions at once, with the 2-bit version significantly more accurate than standard 2-bit quantization.
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
-
Matryoshka Quantization
A single LLM can be trained to support multiple weight precisions at once, with the 2-bit version significantly more accurate than standard 2-bit quantization.