ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-05 03:20:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

softcap: CUDA

Browse Source 
~1% speedup for Gemma2-9b
This commit is contained in:
Iwan Kawrakow
2024-08-02 06:34:29 +03:00
parent a8a0c692d4
commit e49ce89901
2 changed files with 7 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
ggml/src/ggml-cuda.cu
View File

@@ -2869,6 +2869,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
case GGML_OP_DIV:
case GGML_OP_RMS_NORM:
case GGML_OP_SCALE:
case GGML_OP_SOFTCAP:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_CLAMP:

11
ggml/src/ggml-cuda/softcap.cu
View File

@@ -1,18 +1,19 @@
#include "softcap.cuh"
static __global__ void softcap_f32(const float * x, float * dst, const float * scales, const int k) {
static __global__ void softcap_f32(const float * x, float * dst, float s_before, float s_after, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = scales[1] * tanh(scales[0]*x[i]);
float xi = s_before*x[i];
dst[i] = s_after * tanh(xi);
}
static void softcap_f32_cuda(const float * x, float * dst, const float * scales, const int k, cudaStream_t stream) {
static void softcap_f32_cuda(const float * x, float * dst, float s_before, float s_after, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_SOFTCAP_BLOCK_SIZE - 1) / CUDA_SOFTCAP_BLOCK_SIZE;
softcap_f32<<<num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream>>>(x, dst, scales, k);
softcap_f32<<<num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream>>>(x, dst, s_before, s_after, k);
}
void ggml_cuda_op_softcap(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -27,5 +28,5 @@ void ggml_cuda_op_softcap(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
float scales[2];
memcpy(scales, dst->op_params, sizeof(scales));
softcap_f32_cuda(src0_d, dst_d, scales, ggml_nelements(src0), stream);
softcap_f32_cuda(src0_d, dst_d, scales[0], scales[1], ggml_nelements(src0), stream);
}