Skip to content

Raise qmv batch limit for large matrices on M5-class GPUs #3791

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
pierre427 wants to merge 1 commit into
base: main
Choose a base branch
from
+15 −0
Open

Raise qmv batch limit for large matrices on M5-class GPUs #3791

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 15 additions & 0 deletions mlx/backend/metal/quantized.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,6 +84,21 @@ inline array ensure_row_contiguous_matrix(
inline int get_qmv_batch_limit(int D, int O, metal::Device& d) {
auto arch_size = d.get_architecture().back();
auto arch_gen = d.get_architecture_gen();
// M5-generation GPUs run the qmv_wide kernels efficiently at larger row
// counts than the generic fallback assumes. Measured on M5 Max (g17s),
// 4-bit and 8-bit affine weights (group_size 64), D=8192 with
// O in {1024, 8192, 28672, 250112}: qmv_wide is ~18% faster than the qmm
// path at M=12 and still ahead at M=16. This mostly benefits speculative
// decoding verification (see #3553 for the small-M step it mitigates).
if (arch_gen >= 17 && arch_size != 'd') {
if (D <= 2048 && O <= 2048) {
return 18;
} else if (D <= 4096 && O <= 4096) {
return 12;
} else {
return 16;
}
}
if (arch_gen == 13 || arch_gen == 14) {
switch (arch_size) {
case 'd':
Expand Down