An input-adaptive 8-bit quantization scheme estimates activation ranges via a lightweight probabilistic surrogate, achieving near-dynamic accuracy with static-like memory overhead.
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
-
A probabilistic framework for dynamic quantization
An input-adaptive 8-bit quantization scheme estimates activation ranges via a lightweight probabilistic surrogate, achieving near-dynamic accuracy with static-like memory overhead.