CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-29 08:41:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: workaround FA compile failures on macos (#13517)

Browse Source
This commit is contained in:
Jeff Bolz
2025-05-14 13:15:50 +09:00
committed by GitHub
parent e5c834f718
commit ab3971f2a0
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
View File

@@ -12,6 +12,7 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
layout (constant_id = 1) const uint32_t Br = 1; layout (constant_id = 1) const uint32_t Br = 1;
layout (constant_id = 2) const uint32_t Bc = 32; layout (constant_id = 2) const uint32_t Bc = 32;
layout (constant_id = 3) const uint32_t D = 32; layout (constant_id = 3) const uint32_t D = 32;
@@ -19,7 +20,7 @@ layout (constant_id = 3) const uint32_t D = 32;
layout (constant_id = 5) const uint32_t D_split = 16; layout (constant_id = 5) const uint32_t D_split = 16;
const uint32_t D_per_thread = D / D_split; const uint32_t D_per_thread = D / D_split;
const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split; const uint32_t cols_per_iter = WorkGroupSize / D_split;
const uint32_t cols_per_thread = Bc / cols_per_iter; const uint32_t cols_per_thread = Bc / cols_per_iter;
layout (push_constant) uniform parameter { layout (push_constant) uniform parameter {
@@ -134,8 +135,8 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i
return ACC_TYPE(pow(base, ACC_TYPE(exph))); return ACC_TYPE(pow(base, ACC_TYPE(exph)));
} }
shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x]; shared FLOAT_TYPE tmpsh[WorkGroupSize];
shared vec4 tmpshv4[gl_WorkGroupSize.x]; shared vec4 tmpshv4[WorkGroupSize];
shared float masksh[Bc][Br]; shared float masksh[Bc][Br];
shared vec4 Qf[Br][D / 4]; shared vec4 Qf[Br][D / 4];