Brunobkr/llama.cpp_AlgMor24_github

1#pragma once2 3#include "ggml.h"4#include "llama.h"5 6#include <vector>7 8enum common_params_fit_status {9    COMMON_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit10    COMMON_PARAMS_FIT_STATUS_FAILURE = 1, // could not find allocations that are projected to fit11    COMMON_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path12};13 14// fits mparams and cparams to free device memory (assumes system memory is unlimited)15//   - returns true if the parameters could be successfully modified to fit device memory16//   - this function is NOT thread safe because it modifies the global llama logger state17//   - only parameters that have the same value as in llama_default_model_params are modified18//     with the exception of the context size which is modified if and only if equal to 019common_params_fit_status common_fit_params(20                         const char * path_model,21                 llama_model_params * mparams,22               llama_context_params * cparams,23                              float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements24   llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements25                             size_t * margins,               // margins of memory to leave per device in bytes26                           uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use27                     ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log28 29// print estimated memory to stdout30void common_fit_print(31                         const char * path_model,32                 llama_model_params * mparams,33               llama_context_params * cparams);34 35void common_memory_breakdown_print(const llama_context * ctx);36 37struct common_device_memory_data {38    int64_t total;39    int64_t free;40    size_t  model;41    size_t  context;42    size_t  compute;43};44 45using common_device_memory_data_vec = std::vector<common_device_memory_data>;46 47// Load a model + context with no_alloc and return the per-device memory breakdown.48common_device_memory_data_vec common_get_device_memory_data(49                         const char * path_model,50           const llama_model_params * mparams,51         const llama_context_params * cparams,52    std::vector<ggml_backend_dev_t> & devs,53                           uint32_t & hp_ngl,54                           uint32_t & hp_n_ctx_train,55                           uint32_t & hp_n_expert,56                     ggml_log_level   log_level);57