tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-19 22:36:13 -04:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: optimize and refactor MMQ (#8416)

Browse Source 
* CUDA: optimize and refactor MMQ

* explicit q8_1 memory layouts, add documentation
This commit is contained in:
Johannes Gäßler
2024-07-11 16:47:47 +02:00
committed by GitHub
parent a977c11544
commit 808aba3916
5 changed files with 867 additions and 687 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-cuda/mma.cuh
View File

@@ -70,6 +70,10 @@ struct mma_int_A_I16K8 {
}
#endif // defined(INT8_MMA_AVAILABLE)
}
__device__ __forceinline__ void load_low(const int * __restrict__ xs0, const int & stride) {
((mma_int_A_I16K4 *) x)[0].load(xs0, stride);
}
};
struct mma_int_B_J8K4 {