mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 33 Packages Projects Releases Wiki Activity

vulkan: Flash Attention DP4A shader for quantized KV cache (#20797)

Browse Source 
* use integer dot product for quantized KV flash attention

* small improvements

* fix SHMEM_STAGING indexing

* add missing KV type quants

* fixes

* add supported quants to FA tests

* readd fast paths for <8bit quants

* fix mmq gate and shmem checks
This commit is contained in:
Ruben Ortlam
2026-04-13 14:21:31 +02:00
committed by GitHub
parent aa00911d12
commit 75f3bc94e6
8 changed files with 431 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+11 -4
View File
@@ -406,8 +406,8 @@ std::map<std::string, std::string> merge_maps(const std::map<std::string, std::s
}
static std::vector<std::future<void>> compiles;
void string_to_spv(std::string name, const std::string& source, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) {
name = name + (f16acc ? "_f16acc" : "") + (coopmat ? "_cm1" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32"));
void string_to_spv(std::string name, const std::string& source, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false, const std::string& suffix = "") {
name = name + (f16acc ? "_f16acc" : "") + (coopmat ? "_cm1" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32")) + suffix;
std::string out_path = join_paths(output_dir, name + ".spv");
if (input_filepath == "") {
@@ -625,15 +625,16 @@ void process_shaders() {
for (const bool& fp16 : {false, true}) {
std::map<std::string, std::string> base_dict;
if (fp16) {
base_dict = {{"FLOAT_TYPE", "float16_t"}, {"FLOAT_TYPEV4", "f16vec4"}, {"FLOAT16", "1"}, {"FLOAT_TYPE_MAX", "float16_t(65504.0)"}};
base_dict = {{"FLOAT_TYPE", "float16_t"}, {"FLOAT_TYPEV2", "f16vec2"}, {"FLOAT_TYPEV4", "f16vec4"}, {"FLOAT16", "1"}, {"FLOAT_TYPE_MAX", "float16_t(65504.0)"}};
} else {
base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV4", "vec4"}};
base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"FLOAT_TYPEV4", "vec4"}};
}
// flash attention
for (const bool& f16acc : {false, true}) {
std::map<std::string, std::string> fa_base_dict = base_dict;
fa_base_dict["ACC_TYPE"] = fp16 && f16acc ? "float16_t" : "float";
fa_base_dict["ACC_TYPEV2"] = fp16 && f16acc ? "f16vec2" : "vec2";
fa_base_dict["ACC_TYPEV4"] = fp16 && f16acc ? "f16vec4" : "vec4";
if (fp16 && f16acc) {
fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
@@ -672,6 +673,12 @@ void process_shaders() {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, false, false, f16acc);
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
if (tname != "f32") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }, {"MMQ", "1"}}), fp16, false, false, f16acc, "_int8");
}
#endif
}
}
}