vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 17:44:38 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

vulkan: scalar flash attention implementation (#13324)

Browse source 
* vulkan: scalar flash attention implementation

* vulkan: always use fp32 for scalar flash attention

* vulkan: use vector loads in scalar flash attention shader

* vulkan: remove PV matrix, helps with register usage

* vulkan: reduce register usage in scalar FA, but perf may be slightly worse

* vulkan: load each Q value once. optimize O reduction. more tuning

* vulkan: support q4_0/q8_0 KV in scalar FA

* CI: increase timeout to accommodate newly-supported tests

* vulkan: for scalar FA, select between 1 and 8 rows

* vulkan: avoid using Float16 capability in scalar FA
This commit is contained in:
Jeff Bolz 2025-05-09 23:07:07 -07:00 • committed by GitHub
parent 7c28a74e07
commit dc1d2adfc0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 646 additions and 94 deletions
Download patch file Download diff file Expand all files Collapse all files

12
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View file

@ -421,7 +421,6 @@ void process_shaders() {
#endif
}
#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
// flash attention
for (const auto& f16acc : {false, true}) {
std::string acctype = f16acc ? "float16_t" : "float";
@ -432,6 +431,7 @@ void process_shaders() {
}
if (tname == "bf16") continue;
#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, true, f16acc);
@ -440,9 +440,17 @@ void process_shaders() {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
}
#endif
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, false, f16acc);
} else if (tname == "q4_0" || tname == "q8_0") {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc);
}
}
}
#endif
for (const auto& tname : type_names) {
// mul mat vec