vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

backend : offload large batches to GPU (#6083)

Browse source 
* backend : offload large batches to GPU

* fix hip

* code cleanup

* fix CUDA split buffers

* Update ggml-backend-impl.h

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* cuda : fix memset without set_device

* imatrix : remove sched affix from weight names

* sched : add a new split if the current one has too many inputs
reduce max inputs per split
more cleanup

* update backends

ggml-ci

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
slaren 2024-03-18 11:03:04 +01:00 • committed by GitHub
parent 496bc79bc2
commit 2bf8d0f7c4
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
14 changed files with 349 additions and 396 deletions
Download patch file Download diff file Expand all files Collapse all files

5
ggml-backend-impl.h
View file

@ -103,6 +103,11 @@ extern "C" {
// check if the backend supports an operation
bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
// 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
ggml_backend_event_t (*GGML_CALL event_new) (ggml_backend_t backend);
void (*GGML_CALL event_free) (ggml_backend_event_t event);