llama.cpp/ggml-backend-impl.h

117 lines
5.2 KiB
C
Raw Normal View History

#pragma once
// ggml-backend internal header
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
#endif
//
// Backend buffer
//
// buffer type
typedef void * ggml_backend_buffer_type_context_t;
struct ggml_backend_buffer_type_i {
const char * (*GGML_CALL get_name) (ggml_backend_buffer_type_t buft);
ggml_backend_buffer_t (*GGML_CALL alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
size_t (*GGML_CALL get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment
size_t (*GGML_CALL get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
bool (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
// check if tensor data is in host memory
// should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
bool (*GGML_CALL is_host) (ggml_backend_buffer_type_t buft);
};
struct ggml_backend_buffer_type {
struct ggml_backend_buffer_type_i iface;
ggml_backend_buffer_type_context_t context;
};
// buffer
typedef void * ggml_backend_buffer_context_t;
struct ggml_backend_buffer_i {
const char * (*GGML_CALL get_name) (ggml_backend_buffer_t buffer);
void (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
void * (*GGML_CALL get_base) (ggml_backend_buffer_t buffer);
void (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
void (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
void (*GGML_CALL clear) (ggml_backend_buffer_t buffer, uint8_t value);
void (*GGML_CALL reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
};
struct ggml_backend_buffer {
struct ggml_backend_buffer_i iface;
ggml_backend_buffer_type_t buft;
ggml_backend_buffer_context_t context;
size_t size;
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 19:07:38 +00:00
enum ggml_backend_buffer_usage usage;
};
GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
ggml_backend_buffer_type_t buft,
struct ggml_backend_buffer_i iface,
ggml_backend_buffer_context_t context,
size_t size);
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 19:07:38 +00:00
// do not use directly, use ggml_backend_tensor_copy instead
bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
//
// Backend
//
typedef void * ggml_backend_context_t;
struct ggml_backend_i {
const char * (*GGML_CALL get_name)(ggml_backend_t backend);
void (*GGML_CALL free)(ggml_backend_t backend);
// buffer allocation
ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 19:07:38 +00:00
// (optional) asynchronous tensor data access
void (*GGML_CALL set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 19:07:38 +00:00
// (optional) complete all pending operations
void (*GGML_CALL synchronize)(ggml_backend_t backend);
// compute graph with a plan
ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
void (*GGML_CALL graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
void (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 19:07:38 +00:00
// compute graph without a plan (async)
bool (*GGML_CALL graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
// check if the backend supports an operation
bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
};
struct ggml_backend {
struct ggml_backend_i iface;
ggml_backend_context_t context;
};
//
// Backend registry
//
typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
#ifdef __cplusplus
}
#endif