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

vulkan: small dequantization improvements (#18380)

Browse Source 
* iq4_xs

* quants
This commit is contained in:
Eve
2025-12-26 18:12:11 +01:00
committed by GitHub
parent b96b82fc85
commit cb999704fb
3 changed files with 42 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl
+1 -7
View File
@@ -401,13 +401,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint sl = (data_a[a_offset + ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
const uint sh = (data_a[a_offset + ib].scales_h >> (2 * ib32)) & 3;
const uint qshift = (iqs & 16) >> 2;
u8vec4 qs = u8vec4(
data_a[a_offset + ib].qs[iq + 0],
data_a[a_offset + ib].qs[iq + 1],
data_a[a_offset + ib].qs[iq + 2],
data_a[a_offset + ib].qs[iq + 3]
);
qs = (qs >> qshift) & uint8_t(0xF);
const u8vec4 qs = unpack8((data_a_packed32[a_offset + ib].qs[iq/4] >> qshift) & 0x0F0F0F0F);
const float dl = float(int(sl | (sh << 4)) - 32);
return dl * vec4(