mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-12-27 20:04:35 +00:00
4760e7cc0b
* sync : ggml (backend v2) (wip) * sync : migrate examples and llama.cpp to dynamic graphs (wip) * sync : update tests + fix max op params to 64 ggml-ci * sync : ggml-cuda ggml-ci * llama : fix save/load state context size ggml-ci * sync : try to fix build on tvOS * sync : pass custom graph sizes in training examples * sync : update graph copies to new ggml API * sync : update sync-ggml.sh with new files * scripts : fix header in sync script * train : fix context size calculations * llama : increase inference graph size up to 4096 nodes * train : allocate grads for backward graphs * train : allocate grads for gb_tmp
86 lines
3.4 KiB
C
86 lines
3.4 KiB
C
#pragma once
|
|
|
|
#include "ggml.h"
|
|
|
|
#ifdef __cplusplus
|
|
extern "C" {
|
|
#endif
|
|
|
|
struct ggml_backend;
|
|
struct ggml_backend_buffer;
|
|
|
|
//
|
|
// Legacy API
|
|
//
|
|
|
|
typedef struct ggml_allocr * ggml_allocr_t;
|
|
|
|
// initialize allocator for use with CPU backend only
|
|
GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment);
|
|
GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment);
|
|
|
|
// initialize allocator for use with ggml-backend
|
|
GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
|
|
GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
|
|
GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend);
|
|
|
|
GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc);
|
|
|
|
// tell the allocator to parse nodes following the order described in the list
|
|
// you should call this if your graph are optimized to execute out-of-order
|
|
GGML_API void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n);
|
|
|
|
GGML_API void ggml_allocr_free (ggml_allocr_t alloc);
|
|
GGML_API bool ggml_allocr_is_measure (ggml_allocr_t alloc);
|
|
GGML_API void ggml_allocr_reset (ggml_allocr_t alloc);
|
|
GGML_API void ggml_allocr_alloc (ggml_allocr_t alloc, struct ggml_tensor * tensor);
|
|
GGML_API size_t ggml_allocr_max_size (ggml_allocr_t alloc);
|
|
|
|
GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph);
|
|
|
|
//
|
|
// ggml-backend v2 API
|
|
//
|
|
|
|
// Seperate tensor and graph allocator objects
|
|
// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
|
|
// The original API is kept as a wrapper around the new API
|
|
|
|
// Tensor allocator
|
|
typedef struct ggml_tallocr * ggml_tallocr_t;
|
|
|
|
GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
|
|
GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
|
|
GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
|
|
GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
|
|
GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
|
|
|
|
GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);
|
|
|
|
GGML_API void ggml_tallocr_free (ggml_tallocr_t talloc);
|
|
GGML_API bool ggml_tallocr_is_measure (ggml_tallocr_t talloc);
|
|
GGML_API void ggml_tallocr_reset (ggml_tallocr_t talloc);
|
|
GGML_API void ggml_tallocr_alloc (ggml_tallocr_t talloc, struct ggml_tensor * tensor);
|
|
GGML_API size_t ggml_tallocr_max_size (ggml_tallocr_t talloc);
|
|
|
|
|
|
// Graph allocator
|
|
typedef struct ggml_gallocr * ggml_gallocr_t;
|
|
|
|
GGML_API ggml_gallocr_t ggml_gallocr_new(void);
|
|
GGML_API void ggml_gallocr_free(ggml_gallocr_t galloc);
|
|
|
|
GGML_API void ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n);
|
|
GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph);
|
|
|
|
// Allocate tensors from the allocators given by the hash table
|
|
GGML_API void ggml_gallocr_alloc_graph_n(
|
|
ggml_gallocr_t galloc,
|
|
struct ggml_cgraph * graph,
|
|
struct ggml_hash_set hash_set,
|
|
ggml_tallocr_t * hash_node_talloc);
|
|
|
|
#ifdef __cplusplus
|
|
}
|
|
#endif
|