[Python] Can I decrease the numeric precision of scikit-learn's predict_proba() method?

The question behind the question is that I am looking to decrease the runtime of and computational resources expended by my scikit-learn model when using it in production.

When I run predict_proba(), it generates an array of probabilities. Is there a way to have the method use less resources by not spitting out the result to 16 decimals of precision, e.g., but in 8 decimals? The comparison I'm thinking of is quantizing LLMs.

Rounding afterward does not accomplish this, since predict_proba(), where I'm facing a bottleneck, has already ran.

Continue reading...