| 1 | #include "llama-impl.h" |
| 2 | |
| 3 | #include "llama-chat.h" |
| 4 | #include "llama-mmap.h" |
| 5 | #include "llama-vocab.h" |
| 6 | #include "llama-model-loader.h" |
| 7 | #include "llama-model-saver.h" |
| 8 | #include "llama-model.h" |
| 9 | |
| 10 | #include "ggml.h" |
| 11 | #include "ggml-backend.h" |
| 12 | |
| 13 | #include <algorithm> |
| 14 | #include <cstddef> |
| 15 | #include <cstdint> |
| 16 | #include <cstdio> |
| 17 | #include <cstring> |
| 18 | #include <ctime> |
| 19 | |
| 20 | #if defined(_MSC_VER) |
| 21 | #pragma warning(disable: 4244 4267) // possible loss of data |
| 22 | #endif |
| 23 | |
| 24 | // |
| 25 | // interface implementation |
| 26 | // |
| 27 | |
| 28 | const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_type) { |
| 29 | switch (flash_attn_type) { |
| 30 | case LLAMA_FLASH_ATTN_TYPE_AUTO: |
| 31 | return "auto" ; |
| 32 | case LLAMA_FLASH_ATTN_TYPE_DISABLED: |
| 33 | return "disabled" ; |
| 34 | case LLAMA_FLASH_ATTN_TYPE_ENABLED: |
| 35 | return "enabled" ; |
| 36 | } |
| 37 | GGML_ABORT("fatal error" ); |
| 38 | } |
| 39 | |
| 40 | struct llama_sampler_chain_params llama_sampler_chain_default_params() { |
| 41 | struct llama_sampler_chain_params result = { |
| 42 | /*.no_perf =*/ true, |
| 43 | }; |
| 44 | |
| 45 | return result; |
| 46 | } |
| 47 | |
| 48 | size_t llama_max_devices(void) { |
| 49 | return 16; |
| 50 | } |
| 51 | |
| 52 | bool llama_supports_mmap(void) { |
| 53 | return llama_mmap::SUPPORTED; |
| 54 | } |
| 55 | |
| 56 | bool llama_supports_mlock(void) { |
| 57 | return llama_mlock::SUPPORTED; |
| 58 | } |
| 59 | |
| 60 | bool llama_supports_gpu_offload(void) { |
| 61 | return ggml_backend_dev_by_type(type: GGML_BACKEND_DEVICE_TYPE_GPU) != nullptr || |
| 62 | ggml_backend_dev_by_type(type: GGML_BACKEND_DEVICE_TYPE_IGPU) != nullptr || |
| 63 | llama_supports_rpc(); |
| 64 | } |
| 65 | |
| 66 | bool llama_supports_rpc(void) { |
| 67 | return ggml_backend_reg_by_name(name: "RPC" ) != nullptr; |
| 68 | } |
| 69 | |
| 70 | void llama_backend_init(void) { |
| 71 | ggml_time_init(); |
| 72 | |
| 73 | // needed to initialize f16 tables |
| 74 | { |
| 75 | struct ggml_init_params params = { .mem_size: 0, NULL, .no_alloc: false }; |
| 76 | struct ggml_context * ctx = ggml_init(params); |
| 77 | ggml_free(ctx); |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | void llama_numa_init(enum ggml_numa_strategy numa) { |
| 82 | if (numa != GGML_NUMA_STRATEGY_DISABLED) { |
| 83 | auto * dev = ggml_backend_dev_by_type(type: GGML_BACKEND_DEVICE_TYPE_CPU); |
| 84 | GGML_ASSERT(dev && "CPU backend is not loaded" ); |
| 85 | auto * reg = ggml_backend_dev_backend_reg(device: dev); |
| 86 | auto * numa_init_fn = (decltype(ggml_numa_init) *) ggml_backend_reg_get_proc_address(reg, name: "ggml_backend_cpu_numa_init" ); |
| 87 | if (numa_init_fn) { |
| 88 | numa_init_fn(numa); |
| 89 | } |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | void llama_backend_free(void) { |
| 94 | ggml_quantize_free(); |
| 95 | } |
| 96 | |
| 97 | int64_t llama_time_us(void) { |
| 98 | return ggml_time_us(); |
| 99 | } |
| 100 | |
| 101 | // Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback |
| 102 | static int llama_model_load(const std::string & fname, std::vector<std::string> & splits, llama_model & model, llama_model_params & params) { |
| 103 | // loading time will be recalculated after the first eval, so |
| 104 | // we take page faults deferred by mmap() into consideration |
| 105 | model.t_load_us = 0; |
| 106 | time_meas tm(model.t_load_us); |
| 107 | |
| 108 | model.t_start_us = tm.t_start_us; |
| 109 | |
| 110 | try { |
| 111 | llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides, params.tensor_buft_overrides); |
| 112 | |
| 113 | ml.print_info(); |
| 114 | |
| 115 | model.hparams.vocab_only = params.vocab_only; |
| 116 | |
| 117 | try { |
| 118 | model.load_arch(ml); |
| 119 | } catch(const std::exception & e) { |
| 120 | throw std::runtime_error("error loading model architecture: " + std::string(e.what())); |
| 121 | } |
| 122 | try { |
| 123 | model.load_hparams(ml); |
| 124 | } catch(const std::exception & e) { |
| 125 | throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what())); |
| 126 | } |
| 127 | if (model.arch == LLM_ARCH_CLIP) { |
| 128 | throw std::runtime_error("CLIP cannot be used as main model, use it with --mmproj instead" ); |
| 129 | } |
| 130 | try { |
| 131 | model.load_vocab(ml); |
| 132 | } catch(const std::exception & e) { |
| 133 | throw std::runtime_error("error loading model vocabulary: " + std::string(e.what())); |
| 134 | } |
| 135 | |
| 136 | model.load_stats(ml); |
| 137 | model.print_info(); |
| 138 | |
| 139 | if (params.vocab_only) { |
| 140 | LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n" , __func__); |
| 141 | return 0; |
| 142 | } |
| 143 | |
| 144 | if (!model.load_tensors(ml)) { |
| 145 | return -2; |
| 146 | } |
| 147 | } catch (const std::exception & err) { |
| 148 | LLAMA_LOG_ERROR("%s: error loading model: %s\n" , __func__, err.what()); |
| 149 | return -1; |
| 150 | } |
| 151 | |
| 152 | return 0; |
| 153 | } |
| 154 | |
| 155 | static struct llama_model * llama_model_load_from_file_impl( |
| 156 | const std::string & path_model, |
| 157 | std::vector<std::string> & splits, |
| 158 | struct llama_model_params params) { |
| 159 | ggml_time_init(); |
| 160 | |
| 161 | if (!params.vocab_only && ggml_backend_reg_count() == 0) { |
| 162 | LLAMA_LOG_ERROR("%s: no backends are loaded. hint: use ggml_backend_load() or ggml_backend_load_all() to load a backend before calling this function\n" , __func__); |
| 163 | return nullptr; |
| 164 | } |
| 165 | |
| 166 | unsigned cur_percentage = 0; |
| 167 | if (params.progress_callback == NULL) { |
| 168 | params.progress_callback_user_data = &cur_percentage; |
| 169 | params.progress_callback = [](float progress, void * ctx) { |
| 170 | unsigned * cur_percentage_p = (unsigned *) ctx; |
| 171 | unsigned percentage = (unsigned) (100 * progress); |
| 172 | while (percentage > *cur_percentage_p) { |
| 173 | *cur_percentage_p = percentage; |
| 174 | LLAMA_LOG_CONT("." ); |
| 175 | if (percentage >= 100) { |
| 176 | LLAMA_LOG_CONT("\n" ); |
| 177 | } |
| 178 | } |
| 179 | return true; |
| 180 | }; |
| 181 | } |
| 182 | |
| 183 | llama_model * model = new llama_model(params); |
| 184 | |
| 185 | // create list of devices to use with this model |
| 186 | if (params.devices) { |
| 187 | for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) { |
| 188 | model->devices.push_back(x: *dev); |
| 189 | } |
| 190 | } else { |
| 191 | // default device selection |
| 192 | |
| 193 | // build list of available devices |
| 194 | std::vector<ggml_backend_dev_t> gpus; |
| 195 | std::vector<ggml_backend_dev_t> igpus; |
| 196 | std::vector<ggml_backend_dev_t> rpc_servers; |
| 197 | |
| 198 | for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { |
| 199 | ggml_backend_dev_t dev = ggml_backend_dev_get(index: i); |
| 200 | switch (ggml_backend_dev_type(device: dev)) { |
| 201 | case GGML_BACKEND_DEVICE_TYPE_CPU: |
| 202 | case GGML_BACKEND_DEVICE_TYPE_ACCEL: |
| 203 | // skip CPU backends since they are handled separately |
| 204 | break; |
| 205 | |
| 206 | case GGML_BACKEND_DEVICE_TYPE_GPU: { |
| 207 | ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(device: dev); |
| 208 | if (ggml_backend_reg_name(reg) == std::string("RPC" )) { |
| 209 | rpc_servers.push_back(x: dev); |
| 210 | } else { |
| 211 | // check if there is already a GPU with the same device id |
| 212 | ggml_backend_dev_props props; |
| 213 | ggml_backend_dev_get_props(device: dev, props: &props); |
| 214 | auto it = std::find_if(first: gpus.begin(), last: gpus.end(), pred: [&props](ggml_backend_dev_t d) { |
| 215 | ggml_backend_dev_props d_props; |
| 216 | ggml_backend_dev_get_props(device: d, props: &d_props); |
| 217 | if (props.device_id && d_props.device_id) { |
| 218 | return strcmp(s1: props.device_id, s2: d_props.device_id) == 0; |
| 219 | } |
| 220 | return false; |
| 221 | }); |
| 222 | |
| 223 | if (it != gpus.end()) { |
| 224 | LLAMA_LOG_INFO("%s: skipping device %s (%s) with id %s - already using device %s (%s) with the same id\n" , |
| 225 | __func__, |
| 226 | ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), |
| 227 | props.device_id ? props.device_id : "unknown id" , |
| 228 | ggml_backend_dev_name(*it), ggml_backend_dev_description(*it)); |
| 229 | } else { |
| 230 | gpus.push_back(x: dev); |
| 231 | } |
| 232 | } |
| 233 | break; |
| 234 | } |
| 235 | |
| 236 | case GGML_BACKEND_DEVICE_TYPE_IGPU: |
| 237 | igpus.push_back(x: dev); |
| 238 | break; |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | // add RPC servers at the front of the list to minimize network transfers |
| 243 | model->devices.insert(position: model->devices.begin(), first: rpc_servers.begin(), last: rpc_servers.end()); |
| 244 | |
| 245 | // add GPUs |
| 246 | model->devices.insert(position: model->devices.end(), first: gpus.begin(), last: gpus.end()); |
| 247 | |
| 248 | // add integrated GPUs only if no other devices were found |
| 249 | if (model->devices.empty()) { |
| 250 | model->devices.insert(position: model->devices.end(), first: igpus.begin(), last: igpus.end()); |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | // if using single GPU mode, remove all except the main GPU |
| 255 | if (params.split_mode == LLAMA_SPLIT_MODE_NONE) { |
| 256 | if (params.main_gpu < 0) { |
| 257 | model->devices.clear(); |
| 258 | } else { |
| 259 | if (params.main_gpu >= (int)model->devices.size()) { |
| 260 | LLAMA_LOG_ERROR("%s: invalid value for main_gpu: %d (available devices: %zu)\n" , __func__, params.main_gpu, model->devices.size()); |
| 261 | llama_model_free(model); |
| 262 | return nullptr; |
| 263 | } |
| 264 | ggml_backend_dev_t main_gpu = model->devices[params.main_gpu]; |
| 265 | model->devices.clear(); |
| 266 | model->devices.push_back(x: main_gpu); |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | for (auto * dev : model->devices) { |
| 271 | ggml_backend_dev_props props; |
| 272 | ggml_backend_dev_get_props(device: dev, props: &props); |
| 273 | LLAMA_LOG_INFO("%s: using device %s (%s) (%s) - %zu MiB free\n" , __func__, |
| 274 | ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), |
| 275 | props.device_id ? props.device_id : "unknown id" , |
| 276 | props.memory_free/1024/1024); |
| 277 | } |
| 278 | |
| 279 | const int status = llama_model_load(fname: path_model, splits, model&: *model, params); |
| 280 | GGML_ASSERT(status <= 0); |
| 281 | if (status < 0) { |
| 282 | if (status == -1) { |
| 283 | LLAMA_LOG_ERROR("%s: failed to load model\n" , __func__); |
| 284 | } else if (status == -2) { |
| 285 | LLAMA_LOG_INFO("%s: cancelled model load\n" , __func__); |
| 286 | } |
| 287 | |
| 288 | llama_model_free(model); |
| 289 | return nullptr; |
| 290 | } |
| 291 | |
| 292 | return model; |
| 293 | } |
| 294 | |
| 295 | // deprecated |
| 296 | struct llama_model * llama_load_model_from_file( |
| 297 | const char * path_model, |
| 298 | struct llama_model_params params) { |
| 299 | return llama_model_load_from_file(path_model, params); |
| 300 | } |
| 301 | |
| 302 | struct llama_model * llama_model_load_from_file( |
| 303 | const char * path_model, |
| 304 | struct llama_model_params params) { |
| 305 | std::vector<std::string> splits = {}; |
| 306 | return llama_model_load_from_file_impl(path_model, splits, params); |
| 307 | } |
| 308 | |
| 309 | struct llama_model * llama_model_load_from_splits( |
| 310 | const char ** paths, |
| 311 | size_t n_paths, |
| 312 | struct llama_model_params params) { |
| 313 | std::vector<std::string> splits; |
| 314 | if (n_paths == 0) { |
| 315 | LLAMA_LOG_ERROR("%s: list of splits is empty\n" , __func__); |
| 316 | return nullptr; |
| 317 | } |
| 318 | splits.reserve(n: n_paths); |
| 319 | for (size_t i = 0; i < n_paths; ++i) { |
| 320 | splits.push_back(x: paths[i]); |
| 321 | } |
| 322 | return llama_model_load_from_file_impl(path_model: splits.front(), splits, params); |
| 323 | } |
| 324 | |
| 325 | void llama_model_save_to_file(const struct llama_model * model, const char * path_model) { |
| 326 | llama_model_saver ms(*model); |
| 327 | ms.add_kv_from_model(); |
| 328 | ms.add_tensors_from_model(); |
| 329 | ms.save(path_model); |
| 330 | } |
| 331 | |
| 332 | // |
| 333 | // chat templates |
| 334 | // |
| 335 | |
| 336 | int32_t llama_chat_apply_template( |
| 337 | const char * tmpl, |
| 338 | const struct llama_chat_message * chat, |
| 339 | size_t n_msg, |
| 340 | bool add_ass, |
| 341 | char * buf, |
| 342 | int32_t length) { |
| 343 | const std::string curr_tmpl(tmpl == nullptr ? "chatml" : tmpl); |
| 344 | |
| 345 | // format the chat to string |
| 346 | std::vector<const llama_chat_message *> chat_vec; |
| 347 | chat_vec.resize(new_size: n_msg); |
| 348 | for (size_t i = 0; i < n_msg; i++) { |
| 349 | chat_vec[i] = &chat[i]; |
| 350 | } |
| 351 | |
| 352 | std::string formatted_chat; |
| 353 | llm_chat_template detected_tmpl = llm_chat_detect_template(tmpl: curr_tmpl); |
| 354 | if (detected_tmpl == LLM_CHAT_TEMPLATE_UNKNOWN) { |
| 355 | return -1; |
| 356 | } |
| 357 | int32_t res = llm_chat_apply_template(tmpl: detected_tmpl, chat: chat_vec, dest&: formatted_chat, add_ass); |
| 358 | if (res < 0) { |
| 359 | return res; |
| 360 | } |
| 361 | if (buf && length > 0) { |
| 362 | strncpy(dest: buf, src: formatted_chat.c_str(), n: length); |
| 363 | } |
| 364 | return res; |
| 365 | } |
| 366 | |
| 367 | // |
| 368 | // model split |
| 369 | // |
| 370 | |
| 371 | int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) { |
| 372 | static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf" ; |
| 373 | if (snprintf(s: split_path, maxlen: maxlen, format: SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) { |
| 374 | return strlen(s: split_path); |
| 375 | } |
| 376 | return 0; |
| 377 | } |
| 378 | |
| 379 | int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count) { |
| 380 | std::string str_split_path(split_path); |
| 381 | char postfix[32]; |
| 382 | snprintf(s: postfix, maxlen: 32, format: "-%05d-of-%05d.gguf" , split_no + 1, split_count); |
| 383 | std::string str_postfix(postfix); |
| 384 | |
| 385 | // check if split_prefix ends with postfix |
| 386 | int size_prefix = str_split_path.size() - str_postfix.size(); |
| 387 | if (size_prefix > 0 && str_split_path.find(str: str_postfix, pos: size_prefix) != std::string::npos) { |
| 388 | snprintf(s: split_prefix, maxlen: std::min(a: (size_t) size_prefix + 1, b: maxlen), format: "%s" , split_path); |
| 389 | return size_prefix; |
| 390 | } |
| 391 | |
| 392 | return 0; |
| 393 | } |
| 394 | |
| 395 | const char * llama_print_system_info(void) { |
| 396 | static std::string s; |
| 397 | s.clear(); // Clear the string, since it's static, otherwise it will accumulate data from previous calls. |
| 398 | |
| 399 | for (size_t i = 0; i < ggml_backend_reg_count(); i++) { |
| 400 | auto * reg = ggml_backend_reg_get(index: i); |
| 401 | auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, name: "ggml_backend_get_features" ); |
| 402 | if (get_features_fn) { |
| 403 | ggml_backend_feature * features = get_features_fn(reg); |
| 404 | s += ggml_backend_reg_name(reg); |
| 405 | s += " : " ; |
| 406 | for (; features->name; features++) { |
| 407 | s += features->name; |
| 408 | s += " = " ; |
| 409 | s += features->value; |
| 410 | s += " | " ; |
| 411 | } |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | return s.c_str(); |
| 416 | } |
| 417 | |
| 418 | |