mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-11-14 18:24:37 -06:00
Vijay Janapa Reddi
61c71626ea
Keep the configurable baseline_precision introduced in 5a60c0dd11 but
default it to "fp32" in solve, candidate, and sweep, so every existing
CompressionModel result, and every value downstream consumers render,
is unchanged. Callers measuring FP16/BF16-served models pass
baseline_precision="fp16" to get INT4 = 4x.
Tests now pin the FP32 default and pass baseline_precision="fp16"
explicitly where they assert ratios against FP16. The docstring, Wall 13
equation, laws-explained, API page, tutorial exercise, DESIGN.md,
tutorial slides, and changelog describe a configurable baseline that
defaults to FP32.