EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

vulkan: matmul dequantization improvements (#12015)

Browse Source 
* faster dequant for old quants

* dont use unpack for iq4_nl

* vec2 unpack for q8
This commit is contained in:
Eve
2025-02-28 07:20:08 +00:00
committed by GitHub
parent 581650b7ca
commit fbeda9002d
5 changed files with 93 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
View File

@@ -92,7 +92,7 @@ float16_t dequantFuncQ8_0(const in decodeBufQ8_0 bl, const in uint blockCoords[2
const uint iqs = idx;
// Load 16b and select the byte for this element
int32_t qs = unpack8(int32_t(bl.block.qs[(iqs & 0x1E) >> 1]))[iqs & 1];
int32_t qs = unpack8(bl.block.qs[(iqs & 0x1E) >> 1])[iqs & 1];
float16_t ret = float16_t(qs) * d;
return ret;
}