char-fork/llama.cpp
git clone https://git.t4t.associates/char-fork/llama.cpp
0320ac526
char/aliceai-t5-moe
1// Note: this description is outdated 2// 3// An interface allowing to compute ggml_cgraph with Metal 4// 5// This is a fully functional interface that extends ggml with GPU support for Apple devices. 6// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, etc.) 7// 8// How it works? 9// 10// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this 11// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you 12// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.) 13// 14// You only need to make sure that all memory buffers that you used during the graph creation 15// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is 16// used during the graph evaluation to determine the arguments of the compute kernels. 17// 18// Synchronization between device and host memory (for example for input and output tensors) 19// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions. 20// 21 22#pragma once 23 24#include "ggml.h" 25#include "ggml-backend.h" 26 27#include <stddef.h> 28#include <stdbool.h> 29 30struct ggml_tensor ; 31struct ggml_cgraph ; 32 33#ifdef __cplusplus 34extern "C" { 35#endif 36 37// 38// backend API 39// user-code should use only these functions 40// 41 42// TODO: remove in the future 43GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init (void ); 44 45GGML_BACKEND_API bool ggml_backend_is_metal (ggml_backend_t backend ); 46 47GGML_BACKEND_API void ggml_backend_metal_set_abort_callback (ggml_backend_t backend ,ggml_abort_callback abort_callback ,void * user_data ); 48 49// helper to check if the device supports a specific family 50// ideally, the user code should be doing these checks 51// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf 52GGML_BACKEND_API bool ggml_backend_metal_supports_family (ggml_backend_t backend ,int family ); 53 54// capture all command buffers committed the next time `ggml_backend_graph_compute` is called 55GGML_BACKEND_API void ggml_backend_metal_capture_next_compute (ggml_backend_t backend ); 56 57GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg (void ); 58 59#ifdef __cplusplus 60} 61#endif