char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Andreas KieslingerProgrammatic Dependent Launch (PDL) for more performance on newer NVIDIA GPUs (Hopper+) (#22522)e94722822

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
1.4 KiB37 linesraw
1#include "softcap.cuh"
2
3static __global__ void softcap_f32(const float * x, float * dst, const float scale, const float softcap, const int k) {
4    ggml_cuda_pdl_lc();
5    const int i = blockDim.x*blockIdx.x + threadIdx.x;
6
7    if (i >= k) {
8        return;
9    }
10
11    ggml_cuda_pdl_sync();
12    dst[i] = tanhf(scale * x[i]) * softcap;
13}
14
15static void softcap_f32_cuda(const float * x, float * dst, const float scale, const float softcap, const int k, cudaStream_t stream) {
16    const int num_blocks = (k + CUDA_SOFTCAP_BLOCK_SIZE - 1) / CUDA_SOFTCAP_BLOCK_SIZE;
17    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream);
18    ggml_cuda_kernel_launch(softcap_f32, launch_params, x, dst, scale, softcap, k);
19}
20
21// fused GGML_OP_SCALE + GGML_UNARY_OP_TANH + GGML_OP_SCALE
22void ggml_cuda_op_softcap(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * src) {
23    const ggml_tensor * src0 = src->src[0];
24    const float * src0_d = (const float *)src0->data;
25    float * dst_d = (float *)dst->data;
26    cudaStream_t stream = ctx.stream();
27
28    GGML_ASSERT(src0->type == GGML_TYPE_F32);
29    GGML_ASSERT( dst->type == GGML_TYPE_F32);
30
31    float scale;
32    float softcap;
33    memcpy(&scale,   (float *) src->op_params + 0, sizeof(float));
34    memcpy(&softcap, (float *) dst->op_params + 0, sizeof(float));
35
36    softcap_f32_cuda(src0_d, dst_d, scale, softcap, ggml_nelements(src0), stream);
37}