Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-04 02:37:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: fix FA dequant path engagement (#28190)

Browse Source 
Skip the nb[3] check when ne[3] == 1, the shader never reads it for a
single stream. Cache views carry the full-buffer stride there, so the old
check reduced to n_kv == kv_size and the path only engaged with the
cache full.
This commit is contained in:
Nathan Wilson
2026-09-03 10:40:34 +02:00
committed by GitHub
parent 0df974d777
commit c7bda030e7
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+1 -1
View File
@@ -10966,7 +10966,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
return t->nb[0] == ggml_type_size(t->type) &&
t->nb[2] == ggml_row_size(t->type, t->ne[0]) &&
t->nb[1] == t->nb[2] * t->ne[2] &&
t->nb[3] == t->nb[1] * t->ne[1];
(t->ne[3] == 1 || t->nb[3] == t->nb[1] * t->ne[1]);
};
const bool k_quant = k->type != GGML_TYPE_F16 && k->type != GGML_TYPE_BF16 && k->type != GGML_TYPE_F32;
const bool v_quant = v->type != GGML_TYPE_F16 && v->type != GGML_TYPE_BF16 && v->type != GGML_TYPE_F32;