4 ms·

TLDR: Quantized versions of Llama 3.2 1B and 3B models with "competitive accuracy" to the original versions (meaning some degraded performance; plots included i

TLDR: Quantized versions of Llama 3.2 1B and 3B models with "competitive accuracy" to the original versions (meaning some degraded performance; plots included in the release notes).

newfocogi 2y ago
Quantization schemes include post-training quantization (PTQ), SpinQuant, and QLoRA.
grahamj 2y ago
Thx, I prefer not to visit meta properties :X They were already pretty small but I guess the smaller the better as long as accuracy doesn't suffer too much.