The Savings Multiply and So Does the Damage
Last timeMeasuring What You Broke
Compression techniques combine, but not additively: the memory savings multiply and the quality costs more than add. Order decides how much more, and whether you can tell what did it.
Everything in this course has been presented one technique at a time. In
practice a served model has usually had three or four of them applied, and the
interesting questions are about the interaction rather than about any one of
them.
What multiplies
Start with the good news, which is the memory arithmetic.
- memory after the pipeline
- memory before any of it
- fraction left after the first technique
- fraction left after the second
- fraction left after the third
This is worth internalising because it changes which choice is better. Three
techniques each taken to a modest setting beat one technique taken to an
aggressive setting, for the same final size, almost every time. Four-bit
rounding of a half-size distilled model is a far better model than two-bit
rounding of the original, and both land in roughly the same memory.
The lesson stops here
6 more paragraphs to go
You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.
The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.
See the planThe contentsThis is the reading half
Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.
The contents