tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-14 20:29:41 -04:00
Code Issues Packages Projects Releases Wiki Activity

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
7 changed files with 222 additions and 193 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/vulkan-shaders/dequant_q4_0.comp
View File

@@ -18,15 +18,13 @@ void main() {
return;
}
const uint b_idx = 1024*i + 32*ir + 8*il;
const uint q_idx = 8*il;
const uint b_idx = 1024*i + 32*ir + q_idx;
const float d = float(data_a[ib].d);
const float dm = -8.0f * d;
const uint q_idx = 8*il;
[[unroll]] for (uint l = 0; l < 8; ++l) {
data_b[b_idx + l + 0] = D_TYPE(d * (data_a[ib].qs[q_idx + l] & 0xF) + dm);
data_b[b_idx + l + 16] = D_TYPE(d * (data_a[ib].qs[q_idx + l] >> 4) + dm);
data_b[b_idx + l + 0] = D_TYPE(d * ((data_a[ib].qs[q_idx + l] & 0xF) - 8.0f));
data_b[b_idx + l + 16] = D_TYPE(d * ((data_a[ib].qs[q_idx + l] >> 4) - 8.0f));
}
}