ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

vulkan: support copy from f32 to q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl (#11166)

Browse source 
* vulkan: support copy from f32 to q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl

Shaders are based on cpy.cu.

* vulkan: support copy from q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl to f32

* ggml: copy q->f32 assumes some contiguity in the destination
This commit is contained in:
Jeff Bolz 2025-01-16 15:47:10 -06:00 • committed by GitHub
parent 466300fe14
commit bd38ddea01
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 446 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

51
ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp Normal file
View file

@ -0,0 +1,51 @@
#version 450
#include "types.comp"
#include "generic_unary_head.comp"
#include "dequant_funcs.comp"
#if defined(DATA_A_IQ4_NL)
// 16 invocations needed for init_iq4nl_shmem
layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
#else
layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
#endif
void main() {
#if defined(DATA_A_IQ4_NL)
init_iq4nl_shmem();
if (gl_LocalInvocationIndex.x != 0) {
return;
}
#endif
const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
if (idx >= p.ne) {
return;
}
uint dst_idx = get_doffset() + dst_idx(idx);
uint src_idx = src0_idx_quant(idx, QUANT_K);
const uint a_offset = 0;
const uint ib = src_idx;
const vec2 dm = get_dm(ib, a_offset);
[[unroll]] for (int j = 0; j < QUANT_K; j += 4) {
vec4 v = dequantize4(ib, j / QUANT_R, a_offset);
v = v * dm.x + vec4(dm.y);
#if QUANT_R == 2
data_d[dst_idx + j/2 + 0] = v[0];
data_d[dst_idx + j/2 + QUANT_K/2 + 0] = v[1];
data_d[dst_idx + j/2 + 1] = v[2];
data_d[dst_idx + j/2 + QUANT_K/2 + 1] = v[3];
#else
data_d[dst_idx + j + 0] = v[0];
data_d[dst_idx + j + 1] = v[1];
data_d[dst_idx + j + 2] = v[2];
data_d[dst_idx + j + 3] = v[3];
#endif
}
}