Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: avoid using Float16 capability in scalar FA

Browse Source
This commit is contained in:
Jeff Bolz
2025-05-08 14:55:52 -05:00
parent 615958f42c
commit 20a6246f29
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+3 -3
View File
@@ -137,7 +137,7 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i
shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
shared vec4 tmpshv4[gl_WorkGroupSize.x];
shared float16_t masksh[Bc][Br];
shared float masksh[Bc][Br];
shared vec4 Qf[Br][D / 4];
void main() {
@@ -296,14 +296,14 @@ void main() {
uint32_t c = (idx + tid) % Bc;
uint32_t r = (idx + tid) / Bc;
if (idx + tid < Bc * Br) {
masksh[c][r] = data_m[(i * Br + r) * m_stride + (j * Bc + c)];
masksh[c][r] = float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]);
}
}
barrier();
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
float mvf = float(masksh[c * cols_per_iter + col_tid][r]);
float mvf = masksh[c * cols_per_iter + col_tid][r];
Sf[r][c] += slope[r]*mvf;
}