| 1 | #pragma once |
| 2 | |
| 3 | #include "ggml.h" |
| 4 | #include "ggml-backend.h" |
| 5 | |
| 6 | #ifdef __cplusplus |
| 7 | extern "C" { |
| 8 | #endif |
| 9 | |
| 10 | #ifdef GGML_USE_HIP |
| 11 | #define GGML_CUDA_NAME "ROCm" |
| 12 | #define GGML_CUBLAS_NAME "hipBLAS" |
| 13 | #elif defined(GGML_USE_MUSA) |
| 14 | #define GGML_CUDA_NAME "MUSA" |
| 15 | #define GGML_CUBLAS_NAME "muBLAS" |
| 16 | #else |
| 17 | #define GGML_CUDA_NAME "CUDA" |
| 18 | #define GGML_CUBLAS_NAME "cuBLAS" |
| 19 | #endif |
| 20 | #define GGML_CUDA_MAX_DEVICES 16 |
| 21 | |
| 22 | // backend API |
| 23 | GGML_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device); |
| 24 | |
| 25 | GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend); |
| 26 | |
| 27 | // device buffer |
| 28 | GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device); |
| 29 | |
| 30 | // split tensor buffer that splits matrices by rows across multiple devices |
| 31 | GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split); |
| 32 | |
| 33 | // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU |
| 34 | GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void); |
| 35 | |
| 36 | GGML_BACKEND_API int ggml_backend_cuda_get_device_count(void); |
| 37 | GGML_BACKEND_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size); |
| 38 | GGML_BACKEND_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total); |
| 39 | |
| 40 | GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size); |
| 41 | GGML_BACKEND_API void ggml_backend_cuda_unregister_host_buffer(void * buffer); |
| 42 | |
| 43 | GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void); |
| 44 | |
| 45 | #ifdef __cplusplus |
| 46 | } |
| 47 | #endif |
| 48 | |