inference/research
verifiedhigh · 0.88updated 2026-10-03

LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale (Dettmers et al.)

paper

Fields

id
sources/qlab-llmint8-paper
Publisher
arXiv
Kind
paper
Published
2022-08-15
Accessed
2026-10-03
notes
v2. Primary source for the mixed-precision decomposition: outlier feature dimensions into 16-bit, >99.9% of values still in 8-bit.

Notes

v2. Primary source for the mixed-precision decomposition: outlier feature dimensions into 16-bit, >99.9% of values still in 8-bit.

Cross-references

This record references no other records.