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

Vulkan IQ4_NL Support (#8613)

Browse source 
* Fix Vulkan matmul tests compile errors

* Add Vulkan IQ4_NL support

* Fix Vulkan DeepSeek-Coder-V2-Lite MoE support
This commit is contained in:
0cc4m 2024-07-23 10:56:49 +02:00 • committed by GitHub
parent 46e47417aa
commit 751fcfc6c3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 222 additions and 193 deletions
Download patch file Download diff file Expand all files Collapse all files

8
ggml/src/vulkan-shaders/dequant_funcs.comp
View file

@ -58,3 +58,11 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
}
#endif
#if defined(DATA_A_IQ4_NL)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const float d = float(data_a[a_offset + ib].d);
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
}
#endif