CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-11 10:36:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Support FA with K/V in F32 (#16543)

Browse Source
This commit is contained in:
Jeff Bolz
2025-10-14 08:53:37 -05:00
committed by GitHub
parent 7ea15bb64c
commit 4258e0cfe7
4 changed files with 49 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl
View File

@@ -1,6 +1,18 @@
#include "types.glsl"
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufF32 {
vec4 block;
};
float16_t dequantFuncF32(const in decodeBufF32 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const vec4 v = bl.block;
const uint idx = coordInBlock[1];
const f16vec4 vf16 = f16vec4(v);
return vf16[idx];
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ4_0 {
block_q4_0_packed16 block;
};
@@ -717,4 +729,6 @@ float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords
#define dequantFuncA dequantFuncIQ4_NL
#elif defined(DATA_A_MXFP4)
#define dequantFuncA dequantFuncMXFP4
#elif defined(DATA_A_F32)
#define dequantFuncA dequantFuncF32
#endif

20
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl
View File

@@ -64,13 +64,31 @@ layout (binding = 4) readonly buffer S {float data_s[];};
layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
#if defined(A_TYPE_PACKED16)
#define BINDING_IDX_K 0
#define BINDING_IDX_V 1
#if defined(DATA_A_F32)
layout (binding = 1) readonly buffer K_PACKED {vec4 k_data_packed[];} k_packed;
layout (binding = 2) readonly buffer V_PACKED {vec4 v_data_packed[];} v_packed;
#elif defined(A_TYPE_PACKED16)
layout (binding = 1) readonly buffer K_PACKED16 {A_TYPE_PACKED16 k_data_packed16[];} k_packed;
layout (binding = 2) readonly buffer V_PACKED16 {A_TYPE_PACKED16 v_data_packed16[];} v_packed;
#endif
#if defined(DATA_A_F32)
#undef BLOCK_SIZE
#define BLOCK_SIZE 4
#define BLOCK_BYTE_SIZE 16
vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
// iqs is currently always zero in the flash attention shaders
if (binding_idx == BINDING_IDX_K) {
return k_packed.k_data_packed[a_offset + ib];
} else {
return v_packed.v_data_packed[a_offset + ib];
}
}
#endif
#if defined(DATA_A_Q4_0)
#define BLOCK_BYTE_SIZE 18

7
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View File

@@ -611,9 +611,6 @@ void process_shaders() {
}
for (const auto& tname : type_names) {
if (tname == "f32") {
continue;
}
if (tname == "bf16") continue;
#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
@@ -630,7 +627,7 @@ void process_shaders() {
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"COOPMAT", "1"}}), true, true, false, f16acc);
} else if (tname == "q4_0" || tname == "q8_0") {
} else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), true, true, false, f16acc);
@@ -639,7 +636,7 @@ void process_shaders() {
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, false, f16acc);
} else if (tname == "q4_0" || tname == "q8_0") {
} else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
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"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc);