| 1 | #pragma once |
| 2 | |
| 3 | #include "llama.h" |
| 4 | |
| 5 | #include "llama-impl.h" |
| 6 | #include "llama-arch.h" |
| 7 | #include "llama-mmap.h" |
| 8 | |
| 9 | #include "ggml-cpp.h" |
| 10 | |
| 11 | #include <cstddef> |
| 12 | #include <map> |
| 13 | #include <stdexcept> |
| 14 | #include <unordered_map> |
| 15 | |
| 16 | using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>; |
| 17 | |
| 18 | enum llama_fver { |
| 19 | GGUF_FILE_VERSION_V1 = 1, |
| 20 | GGUF_FILE_VERSION_V2 = 2, |
| 21 | GGUF_FILE_VERSION_V3 = 3, |
| 22 | }; |
| 23 | |
| 24 | const char * llama_file_version_name(llama_fver version); |
| 25 | |
| 26 | struct llama_model_loader { |
| 27 | // Holds information on a model weight |
| 28 | struct llama_tensor_weight { |
| 29 | uint16_t idx; // source file index |
| 30 | size_t offs; // tensor data offset in the original file |
| 31 | |
| 32 | ggml_tensor * tensor; |
| 33 | |
| 34 | llama_tensor_weight(const llama_file * file, uint16_t idx, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) { |
| 35 | const int tensor_idx = gguf_find_tensor(ctx: gguf_ctx, name: ggml_get_name(tensor)); |
| 36 | if (tensor_idx < 0) { |
| 37 | throw std::runtime_error(format(fmt: "tensor '%s' not found in the model" , ggml_get_name(tensor))); |
| 38 | } |
| 39 | |
| 40 | offs = gguf_get_data_offset(ctx: gguf_ctx) + gguf_get_tensor_offset(ctx: gguf_ctx, tensor_id: tensor_idx); |
| 41 | if (offs + ggml_nbytes(tensor) < offs || offs + ggml_nbytes(tensor) > file->size()) { |
| 42 | throw std::runtime_error(format(fmt: "tensor '%s' data is not within the file bounds, model is corrupted or incomplete" , ggml_get_name(tensor))); |
| 43 | } |
| 44 | } |
| 45 | }; |
| 46 | |
| 47 | // custom comparator to sort weights more nicely by layer |
| 48 | struct weight_name_comparer { |
| 49 | bool operator()(const std::string & a, const std::string & b) const { |
| 50 | int a_layer = -1; |
| 51 | int b_layer = -1; |
| 52 | sscanf(s: a.c_str(), format: "blk.%d." , &a_layer); |
| 53 | sscanf(s: b.c_str(), format: "blk.%d." , &b_layer); |
| 54 | if (a_layer != b_layer) { |
| 55 | return a_layer < b_layer; |
| 56 | } |
| 57 | return a < b; |
| 58 | } |
| 59 | }; |
| 60 | |
| 61 | static const int TENSOR_NOT_REQUIRED = 1 << 0; |
| 62 | static const int TENSOR_DUPLICATED = 1 << 1; |
| 63 | static const int TENSOR_SKIP = 1 << 2; |
| 64 | |
| 65 | int n_kv = 0; |
| 66 | int n_tensors = 0; |
| 67 | int n_created = 0; |
| 68 | |
| 69 | uint64_t n_elements = 0; |
| 70 | size_t n_bytes = 0; |
| 71 | |
| 72 | bool use_mmap = false; |
| 73 | bool check_tensors; |
| 74 | |
| 75 | llama_files files; |
| 76 | llama_ftype ftype; |
| 77 | llama_fver fver; |
| 78 | |
| 79 | llama_mmaps mappings; |
| 80 | |
| 81 | std::map<std::string, llama_tensor_weight, weight_name_comparer> weights_map; |
| 82 | std::unordered_map<std::string, llama_model_kv_override> kv_overrides; |
| 83 | const llama_model_tensor_buft_override * tensor_buft_overrides; |
| 84 | |
| 85 | gguf_context_ptr meta; |
| 86 | std::vector<ggml_context_ptr> contexts; |
| 87 | |
| 88 | std::string arch_name; |
| 89 | LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN); |
| 90 | |
| 91 | size_t size_done = 0; |
| 92 | size_t size_data = 0; |
| 93 | std::vector<std::pair<size_t, size_t>> mmaps_used; |
| 94 | |
| 95 | llama_model_loader( |
| 96 | const std::string & fname, |
| 97 | std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme |
| 98 | bool use_mmap, |
| 99 | bool check_tensors, |
| 100 | const llama_model_kv_override * param_overrides_p, |
| 101 | const llama_model_tensor_buft_override * param_tensor_buft_overrides_p); |
| 102 | |
| 103 | template<typename T> |
| 104 | typename std::enable_if<std::is_integral<T>::value, bool>::type |
| 105 | get_arr_n(const std::string & key, T & result, bool required = true); |
| 106 | |
| 107 | template<typename T> |
| 108 | typename std::enable_if<std::is_integral<T>::value, bool>::type |
| 109 | get_arr_n(enum llm_kv kid, T & result, bool required = true); |
| 110 | |
| 111 | template<typename T> |
| 112 | bool get_arr(const std::string & key, std::vector<T> & result, bool required = true); |
| 113 | |
| 114 | template<typename T, size_t N_MAX> |
| 115 | bool get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required = true); |
| 116 | |
| 117 | template<typename T> |
| 118 | bool get_arr(enum llm_kv kid, T & result, bool required = true); |
| 119 | |
| 120 | template<typename T> |
| 121 | bool get_key(const std::string & key, T & result, bool required = true); |
| 122 | |
| 123 | template<typename T> |
| 124 | bool get_key(enum llm_kv kid, T & result, bool required = true); |
| 125 | |
| 126 | template<typename T, size_t N_MAX> |
| 127 | bool get_key_or_arr(const std::string & key, std::array<T, N_MAX> & result, uint32_t n, bool required = true); |
| 128 | |
| 129 | template<typename T> |
| 130 | bool get_key_or_arr(enum llm_kv kid, T & result, uint32_t n, bool required = true); |
| 131 | |
| 132 | std::string get_arch_name() const; |
| 133 | |
| 134 | enum llm_arch get_arch() const; |
| 135 | |
| 136 | const llama_tensor_weight * get_weight(const char * name) const; |
| 137 | |
| 138 | const llama_tensor_weight & require_weight(const char * name) const; |
| 139 | |
| 140 | struct ggml_tensor * get_tensor_meta(const char * name) const; |
| 141 | |
| 142 | struct ggml_tensor * require_tensor_meta(const std::string & name) const; |
| 143 | |
| 144 | const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const; |
| 145 | |
| 146 | struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::initializer_list<int64_t> & ne, int flags = 0); |
| 147 | |
| 148 | struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::initializer_list<int64_t> & ne, size_t offset, bool required = true); |
| 149 | |
| 150 | void done_getting_tensors() const; |
| 151 | |
| 152 | void init_mappings(bool prefetch = true, llama_mlocks * mlock_mmaps = nullptr); |
| 153 | |
| 154 | void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const; |
| 155 | |
| 156 | // for backwards compatibility, does not support ggml-backend |
| 157 | void load_data_for(struct ggml_tensor * cur) const; |
| 158 | |
| 159 | // Returns false if cancelled by progress_callback |
| 160 | bool load_all_data( |
| 161 | struct ggml_context * ctx, |
| 162 | llama_buf_map & bufs, |
| 163 | llama_mlocks * lmlocks, |
| 164 | llama_progress_callback progress_callback, |
| 165 | void * progress_callback_user_data); |
| 166 | |
| 167 | std::string ftype_name() const; |
| 168 | |
| 169 | void print_info() const; |
| 170 | }; |
| 171 | |