6 ms·
TIL: > Quantization-Aware Training (QAT) [...] allows preserving similar quality to bfloat16 while dramatically reducing the memory requirements to load the mo
by me_bx 4mo ago
TIL:
> Quantization-Aware Training (QAT) [...] allows preserving similar quality to bfloat16 while dramatically reducing the memory requirements to load the model