wylab/llama.cpp
0
0
Fork 1
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-14 18:02:52 +02:00
Code Issues Packages Projects Releases Wiki Activity

simplify logic and make variable names more consistent

Browse Source
This commit is contained in:
Ruben Ortlam
2026-09-14 16:27:25 +02:00
parent 8d9b18117c
commit de18aa930b
1 changed files with 6 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+6 -7
View File
@@ -414,19 +414,19 @@ void main() {
uint32_t c = (idx + tid) / (HSV_pad / 4);
if (idx + gl_WorkGroupSize.x <= Bc * HSV_pad / 4 || c < Bc) {
FLOAT_TYPEV4 V_Tf = FLOAT_TYPEV4(0);
uint32_t vcol;
bool kv_active = fa_kv_index(j * Bc + c, vcol);
uint32_t v_row;
bool kv_active = fa_kv_index(j * Bc + c, v_row);
if (kv_active && (HSV == HSV_pad || d < HSV / 4)) {
#if !defined(BFLOAT16)
if (USE_DECODE_V) {
uint coord = vcol * v_stride * BLOCK_SIZE_V + 4 * d;
uint coord = v_row * v_stride * BLOCK_SIZE_V + 4 * d;
uint ib = coord / BLOCK_SIZE_V;
uint iqs = (coord % BLOCK_SIZE_V);
V_Tf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
} else
#endif
{
V_Tf = FLOAT_TYPEV4(data_vv4[v_offset / 4 + vcol * v_stride / 4 + d]);
V_Tf = FLOAT_TYPEV4(data_vv4[v_offset / 4 + v_row * v_stride / 4 + d]);
}
}
@@ -463,9 +463,8 @@ void main() {
const uint row = idx / v_cols;
const uint col = idx % v_cols;
uint32_t vcol;
bool kv_active = fa_kv_index(j * Bc + row, vcol);
const uint v_row = USE_SPARSE ? vcol : (j * Bc + row);
uint32_t v_row;
bool kv_active = fa_kv_index(j * Bc + row, v_row);
const uint v_col = hsv_tile * MatBc * row_split + col * 4;
const uint coord = v_row * v_stride * BLOCK_SIZE_V + v_col;