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

Undo mess

Browse source 
---------

Co-authored-by: ardfork <134447697+ardfork@users.noreply.github.com>
This commit is contained in:
Henri Vasserman 2023-08-11 09:30:28 +03:00
parent 42e055d9d6
commit e6b6ae55f4
No known key found for this signature in database
GPG key ID: 2995FC0F58B1A986
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml-cuda.cu
View file

@ -1663,8 +1663,8 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp
#else #else
const float2 dm8f = __half22float2(dm8); const float2 dm8f = __half22float2(dm8);
const float2 ds8f = __half22float2(ds8); const float2 ds8f = __half22float2(ds8);
const float d8d8 = __low2float(dm8) * __low2float(ds8); const float d8d8 = dm8f.x * ds8f.x;
const float m8s8 = __high2float(dm8) * __high2float(ds8); const float m8s8 = dm8f.y * ds8f.y;
#endif // GGML_CUDA_F16 #endif // GGML_CUDA_F16
// scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it