root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-07 17:21:46 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

metal : minor fixup in FA kernel

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-11-02 20:41:42 +02:00
parent 42cadc74bd
commit 40e717263e
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-metal.metal
View File

@ -2776,11 +2776,11 @@ kernel void kernel_flash_attn_ext_vec_f16(
const short iv3 = iq3 / rv3; const short iv3 = iq3 / rv3;
// load the queries from shared memory into local memory // load the queries from shared memory into local memory
float4 mq[D4]; float4 mq[D4/NW];
for (short ii = 0; ii < D4; ii += NW) { for (short ii = 0; ii < D4; ii += NW) {
short i = ii + tiisg; short i = ii + tiisg;
mq[i] = (float4) sq4[i]; mq[i/NW] = (float4) sq4[i];
} }
// pointer to the mask // pointer to the mask
@ -2812,7 +2812,7 @@ kernel void kernel_flash_attn_ext_vec_f16(
mk[2] = (float4) pk4[i + 2*(nb11/8)]; mk[2] = (float4) pk4[i + 2*(nb11/8)];
mk[3] = (float4) pk4[i + 3*(nb11/8)]; mk[3] = (float4) pk4[i + 3*(nb11/8)];
mqk += (float4) (mq[i] * mk); mqk += (float4) (mq[i/NW] * mk);
} }
// reduce the results from the threads in the simdgroup // reduce the results from the threads in the simdgroup