Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: make FA mask/softcap enables spec constants (#19309)

Browse Source 
* vulkan: make FA mask/softcap enables spec constants

* don't specialize for sinks

* bump timeout a little bit
This commit is contained in:
Jeff Bolz
2026-02-06 08:49:58 +01:00
committed by GitHub
parent 7fcf1ef45d
commit f9bd518a6b
6 changed files with 45 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+3 -3
View File
@@ -160,7 +160,7 @@ void main() {
mask_cache[idx] = f16vec4(0);
}
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
if (MASK_ENABLE) {
if (USE_MASK_OPT && mask_opt_idx != j / 16) {
mask_opt_idx = j / 16;
@@ -303,7 +303,7 @@ void main() {
coopMatStore(SfMat, sfsh, coord, sfshstride, gl_CooperativeMatrixLayoutRowMajor);
barrier();
if (p.logit_softcap != 0.0f) {
if (LOGIT_SOFTCAP) {
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
uint32_t c = (idx + tid) / (Br / 4);
uint32_t r = (idx + tid) % (Br / 4);
@@ -314,7 +314,7 @@ void main() {
barrier();
}
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
if (MASK_ENABLE) {
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
uint32_t c = (idx + tid) / (Br / 4);
uint32_t r = (idx + tid) % (Br / 4);