Ahmadzei's picture
added 3 more tables for large emb model
5fa1a76
This feature performs a second quantization of the already quantized weights to save an addition 0.4 bits/parameter.