Skip to content

metal : use FA-vec kernel up to batch size 20 #13496

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

Merged
merged 3 commits into from
May 13, 2025
Merged

metal : use FA-vec kernel up to batch size 20 #13496

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f078c79
batched-bench : fix pp batch contents
ggerganov May 13, 2025
fdfc7de
metal : optimize multi-sequence FA vec kernel
ggerganov May 13, 2025
78d7022
metal : use FA-vec kernel up to batch size 20
ggerganov May 13, 2025
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
2 changes: 1 addition & 1 deletion ggml/src/ggml-metal/ggml-metal.m
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4358,7 +4358,7 @@ static bool ggml_metal_encode_node(
// TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
// for now avoiding mainly to keep the number of templates/kernels a bit lower
// these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
if (ne01 >= 20 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
switch (src1->type) {
case GGML_TYPE_F16:
{
Expand Down
5 changes: 5 additions & 0 deletions ggml/src/ggml-metal/ggml-metal.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3887,6 +3887,11 @@ kernel void kernel_flash_attn_ext_vec(
sm[tiisg] = pm[ic + tiisg];
}

// skip -INF blocks
if (simd_max(sm[tiisg]) == -INFINITY) {
continue;
}

// Q*K^T
{
// each simdgroup processes 1 query and NE (NW/NL) head elements
Expand Down
4 changes: 2 additions & 2 deletions tools/batched-bench/batched-bench.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,8 +123,8 @@ int main(int argc, char ** argv) {

common_batch_clear(batch);

for (int i = 0; i < pp; ++i) {
for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
for (int i = 0; i < pp; ++i) {
common_batch_add(batch, 0, i, { j }, false);
}
}
Expand Down
Loading