Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-17 08:19:49 +02:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: always use fp32 for scalar flash attention

Browse Source
This commit is contained in:
Jeff Bolz
2025-05-06 23:08:39 -05:00
parent 005756a2a9
commit 3a8d954e0c
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+2 -2
View File
@@ -5726,9 +5726,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
assert(k->type == v->type);
vk_pipeline *pipelines;
// XXX TODO other backends may be changing accumulator precision to default to f32 soon
bool f32acc = dst->op_params[3] == GGML_PREC_F32;
bool scalar = !ctx->device->coopmat2;
// XXX TODO other backends may be changing accumulator precision to default to f32 soon
bool f32acc = scalar || dst->op_params[3] == GGML_PREC_F32;
bool small_rows = N <= get_fa_num_small_rows(scalar);
if (scalar) {