puja93

HI, i've been trying to reload a 4-bit quantized llama checkpoint. For any mlp or attention layer, it expands to 5 new layers (bias, g_idx, qweight, qzeros, scales. Each with...