Build the quantization config object: a {group_size, bits, mode} default plus one entry per module in perLayer. A module mapped to false is serialized as false (mlx's "not quantized" convention); otherwise it becomes {bits, group_size}.
quantization
{group_size, bits, mode}
perLayer
false
{bits, group_size}
Build the
quantizationconfig object: a{group_size, bits, mode}default plus one entry per module inperLayer. A module mapped tofalseis serialized asfalse(mlx's "not quantized" convention); otherwise it becomes{bits, group_size}.