vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

unpack should be u16, add vim swap to gitignore (about time)

Browse source
This commit is contained in:
Eve 2025-01-05 21:59:43 -05:00
parent 173077180f
commit b4ae7005e6
3 changed files with 13 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

1
.gitignore vendored
View file

@ -18,6 +18,7 @@
*.metallib *.metallib
*.o *.o
*.so *.so
*.swp
*.tmp *.tmp
# IDE / OS # IDE / OS

12
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
View file

@ -47,12 +47,12 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const FLOAT_TYPE dall = FLOAT_TYPE(d.x); const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; uint16_t scale0_u16 = data_a_packed16[ib0 + i].scales[v_im ];
uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; uint16_t scale4_u16 = data_a_packed16[ib0 + i].scales[v_im + 2];
uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; uint16_t scale8_u16 = data_a_packed16[ib0 + i].scales[v_im + 4];
uvec2 scale0 = uvec2(unpack8(scale0_u32)); uvec2 scale0 = uvec2(unpack8(scale0_u16));
uvec2 scale4 = uvec2(unpack8(scale4_u32)); uvec2 scale4 = uvec2(unpack8(scale4_u16));
uvec2 scale8 = uvec2(unpack8(scale8_u32)); uvec2 scale8 = uvec2(unpack8(scale8_u16));
const uint32_t sc0 = ( scale0.x & 0x3f); const uint32_t sc0 = ( scale0.x & 0x3f);
const uint32_t sc1 = ( scale0.y & 0x3f); const uint32_t sc1 = ( scale0.y & 0x3f);

12
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
View file

@ -46,12 +46,12 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const FLOAT_TYPE dall = FLOAT_TYPE(d.x); const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; uint16_t scale0_u16 = data_a_packed16[ib0 + i].scales[v_im ];
uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; uint16_t scale4_u16 = data_a_packed16[ib0 + i].scales[v_im + 2];
uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; uint16_t scale8_u16 = data_a_packed16[ib0 + i].scales[v_im + 4];
uvec2 scale0 = uvec2(unpack8(scale0_u32)); uvec2 scale0 = uvec2(unpack8(scale0_u16));
uvec2 scale4 = uvec2(unpack8(scale4_u32)); uvec2 scale4 = uvec2(unpack8(scale4_u16));
uvec2 scale8 = uvec2(unpack8(scale8_u32)); uvec2 scale8 = uvec2(unpack8(scale8_u16));
const uint32_t sc0 = ( scale0.x & 0x3f); const uint32_t sc0 = ( scale0.x & 0x3f);
const uint32_t sc1 = ( scale0.y & 0x3f); const uint32_t sc1 = ( scale0.y & 0x3f);