ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

ggml : add epsilon as a parameter for group_norm (#8818)

Browse source 
Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
This commit is contained in:
Molly Sophia 2024-08-06 15:26:46 +08:00 • committed by GitHub
parent cdd1889de6
commit 2d5dd7bb3f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 38 additions and 24 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml/src/ggml-cuda/norm.cu
View file

@ -142,8 +142,7 @@ static void norm_f32_cuda(const float * x, float * dst, const int ncols, const i
}
}
static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
static const float eps = 1e-6f;
static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
if (group_size < 1024) {
const dim3 block_dims(WARP_SIZE, 1, 1);
group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
@ -196,8 +195,12 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
GGML_ASSERT( dst->type == GGML_TYPE_F32);
int num_groups = dst->op_params[0];
float eps;
memcpy(&eps, dst->op_params + 1, sizeof(float));
int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], group_size, ggml_nelements(src0), stream);
group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
}
void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {