nwpie/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2026-02-05 13:53:23 +02:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: work around Intel fp16 bug in mmq (#18814)

Browse Source
This commit is contained in:
Ruben Ortlam
2026-01-14 09:41:23 +01:00
committed by GitHub
parent 7d587e5544
commit 635ef78ec5
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
View File

@@ -264,7 +264,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
const i8vec2 scales = i8vec2(unpack8(uint32_t(((data_a_packed16[ib_k].scales[(is % 8 ) / 2] >> (4 * (is / 8))) & 0x0F0F) |
(((data_a_packed16[ib_k].scales[(8 + (is % 4)) / 2] >> (2 * (is / 4))) & 0x0303) << 4))).xy); // vec4 used due to #12147
buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales - 32);
buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales - 32));
}
}
@@ -334,7 +334,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
(data_a[ib_k].scales[is+4] >> 4) | ((data_a[ib_k].scales[is ] & 0xC0) >> 2));
}
buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm) * FLOAT_TYPE_VEC2(scale_dm);
buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(vec2(data_a_packed32[ib_k].dm) * vec2(scale_dm));
}
}
@@ -385,7 +385,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
const uint is = iqs_k / 4;
const i8vec2 scales = unpack8(int32_t(data_a_packed16[ib_k].scales[is / 2])).xy;
buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales);
buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales));
}
}