BitMoD quantizes LLM weights to 3-4 bits using per-group custom floating-point formats with a repurposed zero value, and accelerates them with a bit-serial PE and dequantization unit.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
BitMoD: Bit-serial Mixture-of-Datatype LLM Acceleration
BitMoD quantizes LLM weights to 3-4 bits using per-group custom floating-point formats with a repurposed zero value, and accelerates them with a bit-serial PE and dequantization unit.