Eligibility predicate (port of mlx-lm's default class_predicate): a module
is quantizable iff it has a 2D .weight whose last dim is divisible by the
group size. This captures Linear projections, the token embedding, and the
(untied) lm_head, and excludes norms/biases/1D tensors.
Eligibility predicate (port of mlx-lm's default class_predicate): a module is quantizable iff it has a 2D
.weightwhose last dim is divisible by the group size. This captures Linear projections, the token embedding, and the (untied) lm_head, and excludes norms/biases/1D tensors.