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

21
ggml-cuda.h
View file

@ -17,29 +17,17 @@ extern "C" {
#define GGML_CUDA_MAX_DEVICES 16
// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
GGML_API GGML_CALL void ggml_init_cublas(void);
// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
GGML_API GGML_CALL bool ggml_cublas_loaded(void);
GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
GGML_API GGML_CALL void ggml_cuda_host_free(void * ptr);
GGML_API GGML_CALL bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API GGML_CALL bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
GGML_API GGML_CALL int ggml_cuda_get_device_count(void);
GGML_API GGML_CALL void ggml_cuda_get_device_description(int device, char * description, size_t description_size);
// backend API
GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
// device buffer
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
// split tensor buffer that splits matrices by rows across multiple devices
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
@ -47,6 +35,9 @@ GGML_API GGML_CALL int ggml_backend_cuda_get_device_count(void);
GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
#ifdef __cplusplus
}
#endif