Skip to content

Fix incorrect nvfp4 quantized_matmul through the split-K path #3854

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
metascroy wants to merge 1 commit into
base: main
Choose a base branch
from
+9 −5
Open

Fix incorrect nvfp4 quantized_matmul through the split-K path #3854

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
14 changes: 9 additions & 5 deletions mlx/backend/metal/quantized.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -884,11 +884,15 @@ void qmm_splitk(
int current_tgs = n_tiles * m_tiles;
int split_k = std::max(1, 512 / current_tgs);

// Cap split_k by the number of quantization groups
split_k = std::min(split_k, K / group_size);

// Ensure K divides evenly by split_k * group_size
while (split_k > 1 && (K % (split_k * group_size) != 0)) {
// Each K partition must be a whole number of BK-wide (32) K-tiles as well as
// whole quantization groups. The qmm_t_splitk kernels tile K by BK=32 and do
// not bound the K dimension, so a partition smaller than BK (e.g. nvfp4's
// group_size=16) would over-read into the next group's weights/scales.
int k_align = group_size > 32 ? group_size : 32;
split_k = std::min(split_k, K / k_align);

// Ensure K divides evenly by split_k * k_align
while (split_k > 1 && (K % (split_k * k_align) != 0)) {
split_k--;
}
if (split_k <= 1) {
Expand Down