ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-04-28 06:49:42 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

metal : fix floating-point range of attention scores in FA kernels (llama/13090)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2025-04-24 10:38:30 +03:00
parent cf3eb291ab
commit 01e1600edd
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/ggml-metal.metal
View File

@ -3192,7 +3192,7 @@ kernel void kernel_flash_attn_ext(
{
float S[Q] = { [0 ... Q-1] = 0.0f };
float M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 };
float M[Q] = { [0 ... Q-1] = -__FLT_MAX__/2 };
// thread indices inside the simdgroup
// TODO: see if we can utilize quad-group functions for better performance
@ -3452,7 +3452,7 @@ kernel void kernel_flash_attn_ext(
// reduce the warps sequentially
for (ushort sg = 1; sg < nsg; ++sg) {
float S = { 0.0f };
float M = { -__FLT16_MAX__/2 };
float M = { -__FLT_MAX__/2 };
threadgroup_barrier(mem_flags::mem_threadgroup);
@ -3699,7 +3699,7 @@ kernel void kernel_flash_attn_ext_vec(
{
float S = 0.0f;
float M = -__FLT16_MAX__/2;
float M = -__FLT_MAX__/2;
// thread indices inside the simdgroup
const short tx = tiisg%NL;