zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-06-24 09:21:16 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

OOB QMV fix (#1579)

Browse Source 
* fix oob access in qmv

* skip more

* fix small case
This commit is contained in:
Alex Barron 2024-11-08 17:59:45 -08:00 committed by GitHub
parent 111fefd5e9
commit a4c47b0276
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 13 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
mlx/backend/metal/kernels/quantized.h
View File

@ -564,18 +564,21 @@ METAL_FUNC void qmv_impl(
static_cast<int>(in_vec_size - k - simd_lid * values_per_thread),
0,
values_per_thread);
U sum =
load_vector_safe<T, U, values_per_thread, bits>(x, x_thread, remaining);
if (remaining > 0) {
U sum = load_vector_safe<T, U, values_per_thread, bits>(
x, x_thread, remaining);
for (int row = 0; out_row + row < out_vec_size; row++) {
const device uint8_t* wl =
(const device uint8_t*)(w + row * in_vec_size_w);
const device T* sl = scales + row * in_vec_size_g;
const device T* bl = biases + row * in_vec_size_g;
for (int row = 0; out_row + row < out_vec_size; row++) {
const device uint8_t* wl =
(const device uint8_t*)(w + row * in_vec_size_w);
const device T* sl = scales + row * in_vec_size_g;
const device T* bl = biases + row * in_vec_size_g;
U s = sl[0];
U b = bl[0];
result[row] += qdot<U, values_per_thread, bits>(wl, x_thread, s, b, sum);
U s = sl[0];
U b = bl[0];
result[row] +=
qdot<U, values_per_thread, bits>(wl, x_thread, s, b, sum);
}
}
for (int row = 0; out_row + row < out_vec_size; row++) {