Method mm

python/tests/test_quantized.py:1113–1114

(sb, x, wq)

        wq, s, b = mx.quantize(w, bits=4, group_size=64)

        def mm(sb, x, wq):
            return mx.quantized_matmul(x, wq, *sb, bits=4, group_size=64).sum()

        params = (s, b)
        dparams = mx.grad(mm)((s, b), x, wq)