Why it matters
bitsandbytes made INT8 loading trivial in the transformers ecosystem. Understanding LLM.int8() explains why 'load_in_8bit=True' works well.
Advertisement
The architecture
Identify outlier columns: features with values above threshold (~6).
Split matmul: outliers in FP16, non-outliers in INT8. Sum results.
Advertisement
How it works end to end
Threshold: ~6 by default. Higher threshold = fewer outliers, faster, but more quality loss.
Overhead: FP16 computation for outliers has some cost but is small (few features).
Model support: bitsandbytes integrates with transformers via load_in_8bit=True.