Skip to content

Commit 3155f5c

Browse files
committed
for 1-bit lower inner dimension need to go slower path to avoid bug
1 parent 4d6ca4d commit 3155f5c

1 file changed

Lines changed: 1 addition & 1 deletion

File tree

mlx/backend/metal/quantized.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1282,7 +1282,7 @@ void dispatch_qmv(
12821282
const Stream& s,
12831283
const std::string& mode) {
12841284
// It is a qmv with a small inner dimension so route to qmv_quad kernel
1285-
if ((K == 128 || K == 64) && is_power_of_2(bits)) {
1285+
if ((K == 128 || (K == 64 && bits >= 2)) && is_power_of_2(bits)) {
12861286
qmv_quad(x, w, scales, biases, out, group_size, bits, M, N, K, d, s, mode);
12871287
return;
12881288
}

0 commit comments

Comments
 (0)