Skip to content

metal : use F32 prec for K*Q in vec FA #9595

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
Sep 23, 2024
Merged

metal : use F32 prec for K*Q in vec FA #9595

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: 7 additions & 7 deletions ggml/src/ggml-metal.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2631,11 +2631,11 @@ kernel void kernel_flash_attn_ext_vec_f16(
const short iv3 = iq3 / rv3;

// load the queries from shared memory into local memory
half4 mq[D4];
float4 mq[D4];

for (short ii = 0; ii < D4; ii += NW) {
short i = ii + tiisg;
mq[i] = sq4[i];
mq[i] = (float4) sq4[i];
}

// pointer to the mask
Expand All @@ -2661,11 +2661,11 @@ kernel void kernel_flash_attn_ext_vec_f16(
for (short ii = 0; ii < D4; ii += NW) {
const short i = ii + tiisg;

half4x4 mk;
mk[0] = pk4[i + 0*(nb11/8)];
mk[1] = pk4[i + 1*(nb11/8)];
mk[2] = pk4[i + 2*(nb11/8)];
mk[3] = pk4[i + 3*(nb11/8)];
float4x4 mk;
mk[0] = (float4) pk4[i + 0*(nb11/8)];
mk[1] = (float4) pk4[i + 1*(nb11/8)];
mk[2] = (float4) pk4[i + 2*(nb11/8)];
mk[3] = (float4) pk4[i + 3*(nb11/8)];

mqk += (float4) (mq[i] * mk);
}
Expand Down
Loading