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

Basic Vulkan Multi-GPU implementation (#5321)

Browse source 
* Initial Vulkan multi-gpu implementation

Move most global variables into backend context

* Add names to backend device functions

* Add further missing cleanup code

* Reduce code duplication in tensor split layer assignment

* generalize LLAMA_SPLIT_LAYER for all backends, do not expose device count and memory in llama.h

* Only do device info print in the beginning and initialize one backend for cpu assist

Add missing cleanup code

* Rework backend memory management to make sure devices and buffers get properly allocated and freed

* Rename cpu assist free function

---------

Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
0cc4m 2024-02-07 07:54:50 +01:00 • committed by GitHub
parent ed0bf32290
commit ee1628bdfe
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 1587 additions and 1166 deletions
Download patch file Download diff file Expand all files Collapse all files

23
ggml-vulkan.h
View file

@ -8,24 +8,29 @@ extern "C" {
#endif
#define GGML_VK_NAME "Vulkan"
#define GGML_VK_MAX_DEVICES 16
GGML_API void ggml_vk_init(void);
GGML_API void ggml_vk_init_cpu_assist(void);
GGML_API void ggml_vk_preallocate_buffers_graph(struct ggml_tensor * node);
GGML_API void ggml_vk_preallocate_buffers(void);
GGML_API void ggml_vk_build_graph(struct ggml_tensor * node, bool last_node);
GGML_API bool ggml_vk_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
GGML_API void ggml_vk_preallocate_buffers_graph_cpu_assist(struct ggml_tensor * node);
GGML_API void ggml_vk_preallocate_buffers_cpu_assist(void);
GGML_API void ggml_vk_build_graph_cpu_assist(struct ggml_tensor * node, bool last_node);
GGML_API bool ggml_vk_compute_forward_cpu_assist(struct ggml_compute_params * params, struct ggml_tensor * tensor);
#ifdef GGML_VULKAN_CHECK_RESULTS
void ggml_vk_check_results_1(struct ggml_compute_params * params, struct ggml_tensor * tensor);
void ggml_vk_check_results_1_cpu_assist(struct ggml_compute_params * params, struct ggml_tensor * tensor);
#endif
GGML_API void ggml_vk_graph_cleanup(void);
GGML_API void ggml_vk_graph_cleanup_cpu_assist(void);
GGML_API void ggml_vk_free_cpu_assist(void);
// backend API
GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(void);
GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);
GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
GGML_API GGML_CALL int ggml_backend_vk_get_device_count(void);
GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(void);
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
// 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_vk_host_buffer_type(void);