Method fn

python/tests/test_quantized.py:263–266

(x)

            w_q, scales, biases = mx.quantize(w, group_size, bits)

            def fn(x):
                return mx.quantized_matmul(
                    x, w_q, scales, biases, transposed, group_size, bits
                )

            _, vjp_out = mx.vjp(fn, primals=(x,), cotangents=(c,))