#pragma once

#include "ggml.h"
#include "ggml-backend.h"

#ifdef  __cplusplus
extern "C" {
#endif

#ifdef GGML_USE_HIP
#define GGML_CUDA_NAME "ROCm"
#define GGML_CUBLAS_NAME "hipBLAS"
#elif defined(GGML_USE_MUSA)
#define GGML_CUDA_NAME "MUSA"
#define GGML_CUBLAS_NAME "muBLAS"
#else
#define GGML_CUDA_NAME "CUDA"
#define GGML_CUBLAS_NAME "cuBLAS"
#endif
#define GGML_CUDA_MAX_DEVICES       16

// backend API
GGML_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device);

GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
GGML_BACKEND_API void ggml_backend_cuda_trim_pools(ggml_backend_t backend);
// Sets the CUDA scheduling priority used for this backend instance's lazily
// created streams (lower = higher priority, 0 = default). Scoped to the
// instance; already-created streams keep their priority.
GGML_BACKEND_API void ggml_backend_cuda_set_stream_priority(ggml_backend_t backend, int priority);
// Opts this backend instance into folding the bf16-reference rounding
// (ggml_round_bf16) that directly follows RMS_NORM*MUL, ADD or SwiGLU into the
// producing kernel. Off by default; results are bit-identical either way.
GGML_BACKEND_API void ggml_backend_cuda_set_round_bf16_fusion(ggml_backend_t backend, bool enabled);
// Returns the backend's current compute CUDA stream (cudaStream_t) so host
// code can enqueue its own kernels/copies ordered with graph computes.
GGML_BACKEND_API void * ggml_backend_cuda_get_stream(ggml_backend_t backend);
GGML_BACKEND_API void ggml_backend_cuda_clear_graph(ggml_backend_t backend, const struct ggml_cgraph * graph);

// device buffer
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);

// conduct allreduce operation between devices
GGML_BACKEND_API bool ggml_backend_cuda_allreduce_tensor(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends);

// split tensor buffer that splits matrices by rows across multiple devices
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);

// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);

GGML_BACKEND_API int  ggml_backend_cuda_get_device_count(void);
GGML_BACKEND_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
GGML_BACKEND_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);

GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
GGML_BACKEND_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);

GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void);

#ifdef  __cplusplus
}
#endif