vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

ggml : fix FA mask dim 2 and 3 (#14505)

Browse source 
* ggml : fix FA mask dim 2 and 3

ggml-ci

* backends : unsupport batched FA in CUDA and Vulkan

ggml-ci

* vulkan : disable FA for mask->ne[2] != 1
This commit is contained in:
Georgi Gerganov 2025-07-03 10:46:57 +03:00 • committed by GitHub
parent d4cdd9c1c3
commit 9067487c44
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 26 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file

@ -10265,6 +10265,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) {
return false;
}
// TODO: support broadcast
// note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14449, but
// the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
if (op->src[0]->ne[3] != 1 || (op->src[3] && op->src[3]->ne[2] != 1)) {
return false;
}
// It's straightforward to support different K/V dequant, but would
// significantly increase the number of pipelines
if (op->src[1]->type != op->src[2]->type) {