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

partially revert changes

Browse source
This commit is contained in:
Johannes Gäßler 2024-05-07 18:36:47 +02:00
parent d9bcb92f75
commit fa81c3a22c
2 changed files with 7 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-cuda/common.cuh
View file

@ -430,7 +430,7 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
GGML_UNUSED(a);
GGML_UNUSED(b);
NO_DEVICE_CODE;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
}
static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {

6
ggml-cuda/fattn.cu
View file

@ -61,6 +61,7 @@ static __global__ void flash_attn_vec_ext_f16(
static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
constexpr int nwarps = D / WARP_SIZE;
const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
__builtin_assume(tid < D);
__shared__ half KQ[ncols*D];
#pragma unroll
@ -106,7 +107,10 @@ static __global__ void flash_attn_vec_ext_f16(
for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
// Calculate KQ tile and keep track of new maximum KQ values:
half kqmax_new[ncols];
memcpy(kqmax_new, kqmax, sizeof(kqmax));
#pragma unroll
for (int j = 0; j < ncols; ++j) {
kqmax_new[j] = kqmax[j];
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {