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

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	ggml/src/ggml-cann/aclnn_ops.cpp
#	ggml/src/ggml-cann/aclnn_ops.h
#	ggml/src/ggml-cann/common.h
#	ggml/src/ggml-cann/ggml-cann.cpp
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo 2025-04-18 12:44:20 +08:00
commit bce519cee7
12 changed files with 130 additions and 34 deletions
Download patch file Download diff file Expand all files Collapse all files

7
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file

@ -5555,7 +5555,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
uint32_t workgroups_y = (uint32_t)neq2;
uint32_t workgroups_z = (uint32_t)neq3;
if (N == 1 && qk_ratio > 1 && is_pow2(qk_ratio) && gqa_ratio <= flash_attention_num_small_rows &&
if (N == 1 && qk_ratio > 1 && gqa_ratio <= flash_attention_num_small_rows &&
qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) {
// grouped query attention - make the N dimension equal to gqa_ratio, reduce
// workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1
@ -5568,8 +5568,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
uint32_t split_kv = KV;
uint32_t split_k = 1;
if (gqa_ratio > 1 && ctx->device->shader_core_count > 0) {
GGML_ASSERT(workgroups_x == 1);
// Try to use split_k when KV is large enough to be worth the overhead
if (workgroups_x == 1 && ctx->device->shader_core_count > 0 && KV >= 512) {
// Try to run two workgroups per SM.
split_k = ctx->device->shader_core_count * 2 / workgroups_y;
if (split_k > 1) {
@ -9285,6 +9285,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case 112:
case 128:
case 256:
case 575: // DeepSeek MLA
break;
default:
return false;