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

CUDA: use tensor cores for MMQ (#7676)

Browse source 
* CUDA: int8 tensor cores for MMQ (legacy quants)

* fix out-of-bounds writes

* __builtin_assume -> GGML_CUDA_ASSUME

* fix writeback returning too early
This commit is contained in:
Johannes Gäßler 2024-06-10 11:45:13 +02:00 • committed by GitHub
parent af4ae502dd
commit 1f0dabda8d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 550 additions and 55 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml-cuda/fattn-wmma-f16.cuh
View file

@ -1,9 +1,9 @@
#include "common.cuh"
#include "fattn-common.cuh"
#if FP16_MMA_AVAILABLE
#ifdef FP16_MMA_AVAILABLE
#include <mma.h>
#endif
#endif // FP16_MMA_AVAILABLE
// D == head size, VKQ_stride == num VKQ rows calculated in parallel:
template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t>
@ -45,7 +45,7 @@ static __global__ void flash_attn_ext_f16(
const int ne1,
const int ne2,
const int ne3) {
#if FP16_MMA_AVAILABLE
#ifdef FP16_MMA_AVAILABLE
//In this kernel Q, K, V are matrices while i, j, k are matrix indices.
const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on.