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

Merge commit '9baf9ef304' into concedo_experimental

Browse source
This commit is contained in:
Concedo 2023-07-30 18:18:23 +08:00
commit 82d0695f0f
1 changed files with 4 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml-cuda.cu
View file

@ -3521,13 +3521,12 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
if (split) {
// src0 = weight matrix is saved as a transposed matrix for better memory layout.
// dst is NOT transposed.
// The outputs of cuBLAS matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
// The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
// Instead they need to be copied to the correct slice in ne0 = dst row index.
// If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
for (int64_t j = 0; j < ne1; ++j) {
float * dhf_dst_i = (float *) ((char *) dst_off_device + (j*ne0 + i01_low)*sizeof(float) + i02*nb2 + i03*nb3);
CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_ddf_i + j*i01_diff, i01_diff*sizeof(float), kind, cudaStream_main));
}
float * dhf_dst_i = (float *) ((char *) dst_off_device + i01_low*sizeof(float) + i02*nb2 + i03*nb3);
CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float), dst_ddf_i, i01_diff*sizeof(float),
i01_diff*sizeof(float), ne1, kind, cudaStream_main));
} else {
float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_ddf_i, dst_stride*sizeof(float), kind, cudaStream_main));