vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix to allow clblast to work even after blas backend splitoff

Browse source
This commit is contained in:
Concedo 2024-06-17 15:02:55 +08:00
parent b53e760557
commit ba9ef4d01b
20 changed files with 802 additions and 466 deletions
Download patch file Download diff file Expand all files Collapse all files

28
ggml-backend-impl.h
View file

@ -17,13 +17,15 @@ extern "C" {
struct ggml_backend_buffer_type_i {
const char * (*GGML_CALL get_name) (ggml_backend_buffer_type_t buft);
// allocate a buffer of this type
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
// tensor alignment
size_t (*GGML_CALL get_alignment) (ggml_backend_buffer_type_t buft);
// max buffer size that can be allocated
size_t (*GGML_CALL get_max_size) (ggml_backend_buffer_type_t buft);
// data size needed to allocate the tensor, including padding
size_t (*GGML_CALL get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
// 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);
};
@ -92,27 +94,37 @@ extern "C" {
void (*GGML_CALL synchronize)(ggml_backend_t backend);
// compute graph with a plan (not used currently)
// create a new plan for a graph
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);
// update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
void (*GGML_CALL graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
// compute the graph with the plan
enum ggml_status (*GGML_CALL graph_plan_compute)(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
// check if the backend can compute an operation
bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
// check if the backend can use tensors allocated in a buffer type
bool (*GGML_CALL supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
// check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
// these should be expensive operations with large batch sizes that may benefit from running on this backend
// even if the weight has to be copied from the CPU temporarily
bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
// (optional) event synchronization
// create a new event that can record events on this backend instance
ggml_backend_event_t (*GGML_CALL event_new) (ggml_backend_t backend);
void (*GGML_CALL event_free) (ggml_backend_event_t event);
// record an event on the backend instance that created it
void (*GGML_CALL event_record) (ggml_backend_event_t event);
// wait for an event on on a different backend instance
void (*GGML_CALL event_wait) (ggml_backend_t backend, ggml_backend_event_t event);
// block until an event is recorded
void (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
};