happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: Fix data race/hang in scalar/cm1 flash attention (#17887)

This commit is contained in:
Jeff Bolz 2025-12-14 02:00:00 -06:00 committed by GitHub
parent 4722671641
commit 3238b1400c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 6 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
View File

@ -256,6 +256,9 @@ void main() {
barrier(); barrier();
} }
// prevent race on tmpsh
barrier();
// reduce across threads // reduce across threads
[[unroll]] for (uint32_t r = 0; r < Br; ++r) { [[unroll]] for (uint32_t r = 0; r < Br; ++r) {

3
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
View File

@ -302,6 +302,9 @@ void main() {
barrier(); barrier();
} }
// prevent race on tmpsh
barrier();
// reduce across threads // reduce across threads
float rowmaxf[rows_per_thread], eMf[rows_per_thread], Moldf[rows_per_thread]; float rowmaxf[rows_per_thread], eMf[rows_per_thread], Moldf[rows_per_thread];