char-fork/llama.cpp

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

Johannes GäßlerCUDA: batched+noncont MMQ, refactor bs>1 MoE code (#13199)e1e8e0991

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
622 B15 linesraw
1#include "common.cuh"
2
3#define CUDA_GET_ROWS_BLOCK_SIZE 256
4#define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256
5
6void get_rows_cuda(
7        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
8        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
9        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
10        size_t nb1, size_t nb2, size_t nb3,
11        cudaStream_t stream);
12
13void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
14
15void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);