char-fork/llama.cpp

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

Aman GuptaCUDA: use mmvq for mul-mat-id for small batch sizes (#18958)8bece2eb2

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
826 B14 linesraw
1#include "common.cuh"
2
3#define MMVF_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVF kernels.
4
5void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
6    const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
7
8void ggml_cuda_op_mul_mat_vec_f(
9    ggml_backend_cuda_context & ctx,
10    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
11    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
12    const int64_t src1_padded_row_size, cudaStream_t stream);
13
14bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11);