CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-05 09:36:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: clamp matmul and FA results to the max finite value (#15652)

Browse Source 
* vulkan: clamp matmul and FA results to the max finite value

* only clamp for fp16
This commit is contained in:
Jeff Bolz
2025-08-31 01:27:57 -05:00
committed by GitHub
parent 4d74393bcc
commit 94e82c7ead
7 changed files with 58 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -334,6 +334,9 @@ void main() {
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
Of[r][d] *= Lfrcp[r];
#if defined(ACC_TYPE_MAX)
Of[r][d] = clamp(Of[r][d], -vec4(ACC_TYPE_MAX), vec4(ACC_TYPE_MAX));
#endif
}
}