Skip to content

metal : fix F32 accumulation in FA vec kernel #10232

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 1 commit into from
Nov 9, 2024
Merged

metal : fix F32 accumulation in FA vec kernel #10232

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
13 changes: 7 additions & 6 deletions ggml/src/ggml-metal.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3450,7 +3450,7 @@ kernel void kernel_flash_attn_ext_vec(
{
// each simdgroup processes 1 query and 4 keys
for (short cc = 0; cc < C/4; ++cc) {
qk_t mqk = 0.0;
qk_t mqka[4] = { 0.0, 0.0, 0.0, 0.0 };

device const kd4x4_t * pk = (device const kd4x4_t *) ((device const char *) k + ((ic + 4*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));

Expand All @@ -3461,13 +3461,14 @@ kernel void kernel_flash_attn_ext_vec(
k4x4_t mk;
deq_k(pk + i/nl_k, i%nl_k, mk);

mqk +=
dot(mq[ii/NL][0], mk[0]) +
dot(mq[ii/NL][1], mk[1]) +
dot(mq[ii/NL][2], mk[2]) +
dot(mq[ii/NL][3], mk[3]);
mqka[0] += dot(mq[ii/NL][0], mk[0]);
mqka[1] += dot(mq[ii/NL][1], mk[1]);
mqka[2] += dot(mq[ii/NL][2], mk[2]);
mqka[3] += dot(mq[ii/NL][3], mk[3]);
}

qk_t mqk = mqka[0] + mqka[1] + mqka[2] + mqka[3];

// simdgroup reduce
// [ 0 .. 7] -> [ 0]
// [ 8 .. 15] -> [ 8]
Expand Down
Loading