llama.cpp/ggml-backend-impl.h
slaren f30ea47a87
llama : add pipeline parallelism support (#6017)
* llama : add pipeline parallelism support for batch processing with multiple CUDA GPUs

ggml-ci

* server : add -ub, --ubatch-size parameter

* fix server embedding test

* llama : fix Mamba inference for pipeline parallelism

Tested to work correctly with both `main` and `parallel` examples.

* llama : limit max batch size to n_batch

* add LLAMA_SCHED_MAX_COPIES to configure the number of input copies for pipeline parallelism
default increase to 4 (from 2)

changing this value may improve performance for some systems, but increases memory usage

* fix hip build

* fix sycl build (disable cpy_tensor_async)

* fix hip build

* llama : limit n_batch and n_ubatch to n_ctx during context creation

* llama : fix norm backend

* batched-bench : sync after decode

* swiftui : sync after decode

* ggml : allow ggml_get_rows to use multiple threads if they are available

* check n_ubatch >= n_tokens with non-casual attention

* llama : do not limit n_batch to n_ctx with non-casual attn

* server : construct batch with size of llama_n_batch

* ggml_backend_cpu_graph_compute : fix return value when alloc fails

* llama : better n_batch and n_ubatch comment

* fix merge

* small fix

* reduce default n_batch to 2048

---------

Co-authored-by: Francis Couture-Harpin <git@compilade.net>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-03-13 18:54:21 +01:00

137 lines
6.5 KiB
C

#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_max_size) (ggml_backend_buffer_type_t buft); // allocation max size
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;
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);
// 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);
// buffer that contains a collection of buffers
GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
//
// 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);
// (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_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
// (optional) complete all pending operations
void (*GGML_CALL synchronize)(ggml_backend_t backend);
// compute graph with a plan (not used currently)
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);
// compute graph with a plan
enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
// compute graph without a plan (async)
enum ggml_status (*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);
// (optional) event synchronization
ggml_backend_event_t (*GGML_CALL event_new) (ggml_backend_t backend);
void (*GGML_CALL event_free) (ggml_backend_event_t event);
void (*GGML_CALL event_record) (ggml_backend_event_t event);
void (*GGML_CALL event_wait) (ggml_backend_t backend, ggml_backend_event_t event);
void (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
};
struct ggml_backend {
ggml_guid_t guid;
struct ggml_backend_i iface;
ggml_backend_context_t context;
};
struct ggml_backend_event {
ggml_backend_t backend;
void * 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