| 1 | #include <algorithm> |
|---|---|
| 2 | #include <array> |
| 3 | #include <cassert> |
| 4 | #include <chrono> |
| 5 | #include <cinttypes> |
| 6 | #include <clocale> |
| 7 | #include <cmath> |
| 8 | #include <cstdio> |
| 9 | #include <cstdlib> |
| 10 | #include <cstring> |
| 11 | #include <ctime> |
| 12 | #include <iterator> |
| 13 | #include <map> |
| 14 | #include <numeric> |
| 15 | #include <regex> |
| 16 | #include <sstream> |
| 17 | #include <string> |
| 18 | #include <thread> |
| 19 | #include <vector> |
| 20 | #include <unordered_set> |
| 21 | |
| 22 | #include "common.h" |
| 23 | #include "ggml.h" |
| 24 | #include "llama.h" |
| 25 | |
| 26 | #ifdef _WIN32 |
| 27 | # define WIN32_LEAN_AND_MEAN |
| 28 | # ifndef NOMINMAX |
| 29 | # define NOMINMAX |
| 30 | # endif |
| 31 | # include <windows.h> |
| 32 | #endif |
| 33 | |
| 34 | // utils |
| 35 | static uint64_t get_time_ns() { |
| 36 | using clock = std::chrono::high_resolution_clock; |
| 37 | return std::chrono::nanoseconds(clock::now().time_since_epoch()).count(); |
| 38 | } |
| 39 | |
| 40 | static bool tensor_buft_override_equal(const llama_model_tensor_buft_override& a, const llama_model_tensor_buft_override& b) { |
| 41 | if (a.pattern != b.pattern) { |
| 42 | // cString comparison that may be null |
| 43 | if (a.pattern == nullptr || b.pattern == nullptr) { |
| 44 | return false; |
| 45 | } |
| 46 | if (strcmp(s1: a.pattern, s2: b.pattern) != 0) { |
| 47 | return false; |
| 48 | } |
| 49 | } |
| 50 | if (a.buft != b.buft) { |
| 51 | return false; |
| 52 | } |
| 53 | return true; |
| 54 | } |
| 55 | |
| 56 | static bool vec_tensor_buft_override_equal(const std::vector<llama_model_tensor_buft_override>& a, const std::vector<llama_model_tensor_buft_override>& b) { |
| 57 | if (a.size() != b.size()) { |
| 58 | return false; |
| 59 | } |
| 60 | for (size_t i = 0; i < a.size(); i++) { |
| 61 | if (!tensor_buft_override_equal(a: a[i], b: b[i])) { |
| 62 | return false; |
| 63 | } |
| 64 | } |
| 65 | return true; |
| 66 | } |
| 67 | |
| 68 | static bool vec_vec_tensor_buft_override_equal(const std::vector<std::vector<llama_model_tensor_buft_override>>& a, const std::vector<std::vector<llama_model_tensor_buft_override>>& b) { |
| 69 | if (a.size() != b.size()) { |
| 70 | return false; |
| 71 | } |
| 72 | for (size_t i = 0; i < a.size(); i++) { |
| 73 | if (!vec_tensor_buft_override_equal(a: a[i], b: b[i])) { |
| 74 | return false; |
| 75 | } |
| 76 | } |
| 77 | return true; |
| 78 | } |
| 79 | |
| 80 | template <class T> static std::string join(const std::vector<T> & values, const std::string & delim) { |
| 81 | std::ostringstream str; |
| 82 | for (size_t i = 0; i < values.size(); i++) { |
| 83 | str << values[i]; |
| 84 | if (i < values.size() - 1) { |
| 85 | str << delim; |
| 86 | } |
| 87 | } |
| 88 | return str.str(); |
| 89 | } |
| 90 | |
| 91 | template <typename T, typename F> static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) { |
| 92 | std::vector<std::string> str_values; |
| 93 | std::transform(values.begin(), values.end(), std::back_inserter(x&: str_values), f); |
| 94 | return str_values; |
| 95 | } |
| 96 | |
| 97 | template <typename T> static T avg(const std::vector<T> & v) { |
| 98 | if (v.empty()) { |
| 99 | return 0; |
| 100 | } |
| 101 | T sum = std::accumulate(v.begin(), v.end(), T(0)); |
| 102 | return sum / (T) v.size(); |
| 103 | } |
| 104 | |
| 105 | template <typename T> static T stdev(const std::vector<T> & v) { |
| 106 | if (v.size() <= 1) { |
| 107 | return 0; |
| 108 | } |
| 109 | T mean = avg(v); |
| 110 | T sq_sum = std::inner_product(v.begin(), v.end(), v.begin(), T(0)); |
| 111 | T stdev = std::sqrt(sq_sum / (T) (v.size() - 1) - mean * mean * (T) v.size() / (T) (v.size() - 1)); |
| 112 | return stdev; |
| 113 | } |
| 114 | |
| 115 | static std::string get_cpu_info() { |
| 116 | std::vector<std::string> cpu_list; |
| 117 | for (size_t i = 0; i < ggml_backend_dev_count(); i++) { |
| 118 | auto * dev = ggml_backend_dev_get(index: i); |
| 119 | auto dev_type = ggml_backend_dev_type(device: dev); |
| 120 | if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU || dev_type == GGML_BACKEND_DEVICE_TYPE_ACCEL) { |
| 121 | cpu_list.push_back(x: ggml_backend_dev_description(device: dev)); |
| 122 | } |
| 123 | } |
| 124 | return join(values: cpu_list, delim: ", "); |
| 125 | } |
| 126 | |
| 127 | static std::string get_gpu_info() { |
| 128 | std::vector<std::string> gpu_list; |
| 129 | for (size_t i = 0; i < ggml_backend_dev_count(); i++) { |
| 130 | auto * dev = ggml_backend_dev_get(index: i); |
| 131 | auto dev_type = ggml_backend_dev_type(device: dev); |
| 132 | if (dev_type == GGML_BACKEND_DEVICE_TYPE_GPU || dev_type == GGML_BACKEND_DEVICE_TYPE_IGPU) { |
| 133 | gpu_list.push_back(x: ggml_backend_dev_description(device: dev)); |
| 134 | } |
| 135 | } |
| 136 | return join(values: gpu_list, delim: ", "); |
| 137 | } |
| 138 | |
| 139 | static std::vector<ggml_backend_dev_t> parse_devices_arg(const std::string & value) { |
| 140 | std::vector<ggml_backend_dev_t> devices; |
| 141 | std::string trimmed = string_strip(str: value); |
| 142 | if (trimmed.empty()) { |
| 143 | throw std::invalid_argument("no devices specified"); |
| 144 | } |
| 145 | if (trimmed == "auto") { |
| 146 | return devices; |
| 147 | } |
| 148 | |
| 149 | auto dev_names = string_split<std::string>(input: trimmed, separator: '/'); |
| 150 | if (dev_names.size() == 1 && string_strip(str: dev_names[0]) == "none") { |
| 151 | devices.push_back(x: nullptr); |
| 152 | return devices; |
| 153 | } |
| 154 | |
| 155 | for (auto & name : dev_names) { |
| 156 | std::string dev_name = string_strip(str: name); |
| 157 | if (dev_name.empty()) { |
| 158 | throw std::invalid_argument("invalid device specification"); |
| 159 | } |
| 160 | auto * dev = ggml_backend_dev_by_name(name: dev_name.c_str()); |
| 161 | if (!dev || ggml_backend_dev_type(device: dev) == GGML_BACKEND_DEVICE_TYPE_CPU) { |
| 162 | throw std::invalid_argument(string_format(fmt: "invalid device: %s", dev_name.c_str())); |
| 163 | } |
| 164 | devices.push_back(x: dev); |
| 165 | } |
| 166 | |
| 167 | devices.push_back(x: nullptr); |
| 168 | return devices; |
| 169 | } |
| 170 | |
| 171 | static void register_rpc_server_list(const std::string & servers) { |
| 172 | auto rpc_servers = string_split<std::string>(input: servers, separator: ','); |
| 173 | if (rpc_servers.empty()) { |
| 174 | throw std::invalid_argument("no RPC servers specified"); |
| 175 | } |
| 176 | |
| 177 | auto * rpc_reg = ggml_backend_reg_by_name(name: "RPC"); |
| 178 | if (!rpc_reg) { |
| 179 | throw std::invalid_argument("failed to find RPC backend"); |
| 180 | } |
| 181 | |
| 182 | using add_rpc_server_fn = ggml_backend_reg_t (*)(const char * endpoint); |
| 183 | auto * ggml_backend_rpc_add_server_fn = (add_rpc_server_fn) ggml_backend_reg_get_proc_address(reg: rpc_reg, name: "ggml_backend_rpc_add_server"); |
| 184 | if (!ggml_backend_rpc_add_server_fn) { |
| 185 | throw std::invalid_argument("failed to find RPC add server function"); |
| 186 | } |
| 187 | for (const auto & server : rpc_servers) { |
| 188 | auto reg = ggml_backend_rpc_add_server_fn(server.c_str()); |
| 189 | ggml_backend_register(reg); |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | static std::string devices_to_string(const std::vector<ggml_backend_dev_t> & devices) { |
| 194 | if (devices.empty()) { |
| 195 | return "auto"; |
| 196 | } |
| 197 | |
| 198 | if (devices.size() == 1 && devices[0] == nullptr) { |
| 199 | return "none"; |
| 200 | } |
| 201 | |
| 202 | std::vector<std::string> names; |
| 203 | for (auto * dev : devices) { |
| 204 | if (dev == nullptr) { |
| 205 | break; |
| 206 | } |
| 207 | names.push_back(x: ggml_backend_dev_name(device: dev)); |
| 208 | } |
| 209 | |
| 210 | return join(values: names, delim: "/"); |
| 211 | } |
| 212 | |
| 213 | // command line params |
| 214 | enum output_formats { NONE, CSV, JSON, JSONL, MARKDOWN, SQL }; |
| 215 | |
| 216 | static const char * output_format_str(output_formats format) { |
| 217 | switch (format) { |
| 218 | case NONE: |
| 219 | return "none"; |
| 220 | case CSV: |
| 221 | return "csv"; |
| 222 | case JSON: |
| 223 | return "json"; |
| 224 | case JSONL: |
| 225 | return "jsonl"; |
| 226 | case MARKDOWN: |
| 227 | return "md"; |
| 228 | case SQL: |
| 229 | return "sql"; |
| 230 | default: |
| 231 | GGML_ABORT("invalid output format"); |
| 232 | } |
| 233 | } |
| 234 | |
| 235 | static bool output_format_from_str(const std::string & s, output_formats & format) { |
| 236 | if (s == "none") { |
| 237 | format = NONE; |
| 238 | } else if (s == "csv") { |
| 239 | format = CSV; |
| 240 | } else if (s == "json") { |
| 241 | format = JSON; |
| 242 | } else if (s == "jsonl") { |
| 243 | format = JSONL; |
| 244 | } else if (s == "md") { |
| 245 | format = MARKDOWN; |
| 246 | } else if (s == "sql") { |
| 247 | format = SQL; |
| 248 | } else { |
| 249 | return false; |
| 250 | } |
| 251 | return true; |
| 252 | } |
| 253 | |
| 254 | static const char * split_mode_str(llama_split_mode mode) { |
| 255 | switch (mode) { |
| 256 | case LLAMA_SPLIT_MODE_NONE: |
| 257 | return "none"; |
| 258 | case LLAMA_SPLIT_MODE_LAYER: |
| 259 | return "layer"; |
| 260 | case LLAMA_SPLIT_MODE_ROW: |
| 261 | return "row"; |
| 262 | default: |
| 263 | GGML_ABORT("invalid split mode"); |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | static std::string pair_str(const std::pair<int, int> & p) { |
| 268 | static char buf[32]; |
| 269 | snprintf(s: buf, maxlen: sizeof(buf), format: "%d,%d", p.first, p.second); |
| 270 | return buf; |
| 271 | } |
| 272 | |
| 273 | static std::vector<int> parse_int_range(const std::string & s) { |
| 274 | // first[-last[(+|*)step]] |
| 275 | std::regex range_regex(R"(^(\d+)(?:-(\d+)(?:([\+|\*])(\d+))?)?(?:,|$))"); |
| 276 | |
| 277 | std::smatch match; |
| 278 | std::string::const_iterator search_start(s.cbegin()); |
| 279 | std::vector<int> result; |
| 280 | while (std::regex_search(s: search_start, e: s.cend(), m&: match, re: range_regex)) { |
| 281 | int first = std::stoi(str: match[1]); |
| 282 | int last = match[2].matched ? std::stoi(str: match[2]) : first; |
| 283 | char op = match[3].matched ? match[3].str()[0] : '+'; |
| 284 | int step = match[4].matched ? std::stoi(str: match[4]) : 1; |
| 285 | |
| 286 | for (int i = first; i <= last;) { |
| 287 | result.push_back(x: i); |
| 288 | |
| 289 | int prev_i = i; |
| 290 | |
| 291 | if (op == '+') { |
| 292 | i += step; |
| 293 | } else if (op == '*') { |
| 294 | i *= step; |
| 295 | } else { |
| 296 | throw std::invalid_argument("invalid range format"); |
| 297 | } |
| 298 | |
| 299 | if (i <= prev_i) { |
| 300 | throw std::invalid_argument("invalid range"); |
| 301 | } |
| 302 | } |
| 303 | search_start = match.suffix().first; |
| 304 | } |
| 305 | |
| 306 | if (search_start != s.cend()) { |
| 307 | throw std::invalid_argument("invalid range format"); |
| 308 | } |
| 309 | |
| 310 | return result; |
| 311 | } |
| 312 | |
| 313 | struct cmd_params { |
| 314 | std::vector<std::string> model; |
| 315 | std::vector<int> n_prompt; |
| 316 | std::vector<int> n_gen; |
| 317 | std::vector<std::pair<int, int>> n_pg; |
| 318 | std::vector<int> n_depth; |
| 319 | std::vector<int> n_batch; |
| 320 | std::vector<int> n_ubatch; |
| 321 | std::vector<ggml_type> type_k; |
| 322 | std::vector<ggml_type> type_v; |
| 323 | std::vector<int> n_threads; |
| 324 | std::vector<std::string> cpu_mask; |
| 325 | std::vector<bool> cpu_strict; |
| 326 | std::vector<int> poll; |
| 327 | std::vector<int> n_gpu_layers; |
| 328 | std::vector<int> n_cpu_moe; |
| 329 | std::vector<llama_split_mode> split_mode; |
| 330 | std::vector<int> main_gpu; |
| 331 | std::vector<bool> no_kv_offload; |
| 332 | std::vector<bool> flash_attn; |
| 333 | std::vector<std::vector<ggml_backend_dev_t>> devices; |
| 334 | std::vector<std::vector<float>> tensor_split; |
| 335 | std::vector<std::vector<llama_model_tensor_buft_override>> tensor_buft_overrides; |
| 336 | std::vector<bool> use_mmap; |
| 337 | std::vector<bool> embeddings; |
| 338 | std::vector<bool> no_op_offload; |
| 339 | std::vector<bool> no_host; |
| 340 | ggml_numa_strategy numa; |
| 341 | int reps; |
| 342 | ggml_sched_priority prio; |
| 343 | int delay; |
| 344 | bool verbose; |
| 345 | bool progress; |
| 346 | bool no_warmup; |
| 347 | output_formats output_format; |
| 348 | output_formats output_format_stderr; |
| 349 | }; |
| 350 | |
| 351 | static const cmd_params cmd_params_defaults = { |
| 352 | /* model */ { "models/7B/ggml-model-q4_0.gguf"}, |
| 353 | /* n_prompt */ { 512 }, |
| 354 | /* n_gen */ { 128 }, |
| 355 | /* n_pg */ {}, |
| 356 | /* n_depth */ { 0 }, |
| 357 | /* n_batch */ { 2048 }, |
| 358 | /* n_ubatch */ { 512 }, |
| 359 | /* type_k */ { GGML_TYPE_F16 }, |
| 360 | /* type_v */ { GGML_TYPE_F16 }, |
| 361 | /* n_threads */ { cpu_get_num_math() }, |
| 362 | /* cpu_mask */ { "0x0"}, |
| 363 | /* cpu_strict */ { false }, |
| 364 | /* poll */ { 50 }, |
| 365 | /* n_gpu_layers */ { 99 }, |
| 366 | /* n_cpu_moe */ { 0 }, |
| 367 | /* split_mode */ { LLAMA_SPLIT_MODE_LAYER }, |
| 368 | /* main_gpu */ { 0 }, |
| 369 | /* no_kv_offload */ { false }, |
| 370 | /* flash_attn */ { false }, |
| 371 | /* devices */ { {} }, |
| 372 | /* tensor_split */ { std::vector<float>(llama_max_devices(), 0.0f) }, |
| 373 | /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{ { .pattern: nullptr, .buft: nullptr } } }, |
| 374 | /* use_mmap */ { true }, |
| 375 | /* embeddings */ { false }, |
| 376 | /* no_op_offload */ { false }, |
| 377 | /* no_host */ { false }, |
| 378 | /* numa */ GGML_NUMA_STRATEGY_DISABLED, |
| 379 | /* reps */ 5, |
| 380 | /* prio */ GGML_SCHED_PRIO_NORMAL, |
| 381 | /* delay */ 0, |
| 382 | /* verbose */ false, |
| 383 | /* progress */ false, |
| 384 | /* no_warmup */ false, |
| 385 | /* output_format */ MARKDOWN, |
| 386 | /* output_format_stderr */ NONE, |
| 387 | }; |
| 388 | |
| 389 | static void print_usage(int /* argc */, char ** argv) { |
| 390 | printf(format: "usage: %s [options]\n", argv[0]); |
| 391 | printf(format: "\n"); |
| 392 | printf(format: "options:\n"); |
| 393 | printf(format: " -h, --help\n"); |
| 394 | printf(format: " --numa <distribute|isolate|numactl> numa mode (default: disabled)\n"); |
| 395 | printf(format: " -r, --repetitions <n> number of times to repeat each test (default: %d)\n", |
| 396 | cmd_params_defaults.reps); |
| 397 | printf(format: " --prio <-1|0|1|2|3> process/thread priority (default: %d)\n", |
| 398 | cmd_params_defaults.prio); |
| 399 | printf(format: " --delay <0...N> (seconds) delay between each test (default: %d)\n", |
| 400 | cmd_params_defaults.delay); |
| 401 | printf(format: " -o, --output <csv|json|jsonl|md|sql> output format printed to stdout (default: %s)\n", |
| 402 | output_format_str(format: cmd_params_defaults.output_format)); |
| 403 | printf(format: " -oe, --output-err <csv|json|jsonl|md|sql> output format printed to stderr (default: %s)\n", |
| 404 | output_format_str(format: cmd_params_defaults.output_format_stderr)); |
| 405 | printf(format: " --list-devices list available devices and exit\n"); |
| 406 | printf(format: " -v, --verbose verbose output\n"); |
| 407 | printf(format: " --progress print test progress indicators\n"); |
| 408 | printf(format: " --no-warmup skip warmup runs before benchmarking\n"); |
| 409 | if (llama_supports_rpc()) { |
| 410 | printf(format: " -rpc, --rpc <rpc_servers> register RPC devices (comma separated)\n"); |
| 411 | } |
| 412 | printf(format: "\n"); |
| 413 | printf(format: "test parameters:\n"); |
| 414 | printf(format: " -m, --model <filename> (default: %s)\n", join(values: cmd_params_defaults.model, delim: ",").c_str()); |
| 415 | printf(format: " -p, --n-prompt <n> (default: %s)\n", |
| 416 | join(values: cmd_params_defaults.n_prompt, delim: ",").c_str()); |
| 417 | printf(format: " -n, --n-gen <n> (default: %s)\n", join(values: cmd_params_defaults.n_gen, delim: ",").c_str()); |
| 418 | printf(format: " -pg <pp,tg> (default: %s)\n", |
| 419 | join(values: transform_to_str(values: cmd_params_defaults.n_pg, f: pair_str), delim: ",").c_str()); |
| 420 | printf(format: " -d, --n-depth <n> (default: %s)\n", |
| 421 | join(values: cmd_params_defaults.n_depth, delim: ",").c_str()); |
| 422 | printf(format: " -b, --batch-size <n> (default: %s)\n", |
| 423 | join(values: cmd_params_defaults.n_batch, delim: ",").c_str()); |
| 424 | printf(format: " -ub, --ubatch-size <n> (default: %s)\n", |
| 425 | join(values: cmd_params_defaults.n_ubatch, delim: ",").c_str()); |
| 426 | printf(format: " -ctk, --cache-type-k <t> (default: %s)\n", |
| 427 | join(values: transform_to_str(values: cmd_params_defaults.type_k, f: ggml_type_name), delim: ",").c_str()); |
| 428 | printf(format: " -ctv, --cache-type-v <t> (default: %s)\n", |
| 429 | join(values: transform_to_str(values: cmd_params_defaults.type_v, f: ggml_type_name), delim: ",").c_str()); |
| 430 | printf(format: " -t, --threads <n> (default: %s)\n", |
| 431 | join(values: cmd_params_defaults.n_threads, delim: ",").c_str()); |
| 432 | printf(format: " -C, --cpu-mask <hex,hex> (default: %s)\n", |
| 433 | join(values: cmd_params_defaults.cpu_mask, delim: ",").c_str()); |
| 434 | printf(format: " --cpu-strict <0|1> (default: %s)\n", |
| 435 | join(values: cmd_params_defaults.cpu_strict, delim: ",").c_str()); |
| 436 | printf(format: " --poll <0...100> (default: %s)\n", join(values: cmd_params_defaults.poll, delim: ",").c_str()); |
| 437 | printf(format: " -ngl, --n-gpu-layers <n> (default: %s)\n", |
| 438 | join(values: cmd_params_defaults.n_gpu_layers, delim: ",").c_str()); |
| 439 | printf(format: " -ncmoe, --n-cpu-moe <n> (default: %s)\n", |
| 440 | join(values: cmd_params_defaults.n_cpu_moe, delim: ",").c_str()); |
| 441 | printf(format: " -sm, --split-mode <none|layer|row> (default: %s)\n", |
| 442 | join(values: transform_to_str(values: cmd_params_defaults.split_mode, f: split_mode_str), delim: ",").c_str()); |
| 443 | printf(format: " -mg, --main-gpu <i> (default: %s)\n", |
| 444 | join(values: cmd_params_defaults.main_gpu, delim: ",").c_str()); |
| 445 | printf(format: " -nkvo, --no-kv-offload <0|1> (default: %s)\n", |
| 446 | join(values: cmd_params_defaults.no_kv_offload, delim: ",").c_str()); |
| 447 | printf(format: " -fa, --flash-attn <0|1> (default: %s)\n", |
| 448 | join(values: cmd_params_defaults.flash_attn, delim: ",").c_str()); |
| 449 | printf(format: " -dev, --device <dev0/dev1/...> (default: auto)\n"); |
| 450 | printf(format: " -mmp, --mmap <0|1> (default: %s)\n", |
| 451 | join(values: cmd_params_defaults.use_mmap, delim: ",").c_str()); |
| 452 | printf(format: " -embd, --embeddings <0|1> (default: %s)\n", |
| 453 | join(values: cmd_params_defaults.embeddings, delim: ",").c_str()); |
| 454 | printf(format: " -ts, --tensor-split <ts0/ts1/..> (default: 0)\n"); |
| 455 | printf(format: " -ot --override-tensor <tensor name pattern>=<buffer type>;...\n"); |
| 456 | printf(format: " (default: disabled)\n"); |
| 457 | printf(format: " -nopo, --no-op-offload <0|1> (default: 0)\n"); |
| 458 | printf(format: " --no-host <0|1> (default: %s)\n", |
| 459 | join(values: cmd_params_defaults.no_host, delim: ",").c_str()); |
| 460 | printf(format: "\n"); |
| 461 | printf( |
| 462 | format: "Multiple values can be given for each parameter by separating them with ','\n" |
| 463 | "or by specifying the parameter multiple times. Ranges can be given as\n" |
| 464 | "'first-last' or 'first-last+step' or 'first-last*mult'.\n"); |
| 465 | } |
| 466 | |
| 467 | static ggml_type ggml_type_from_name(const std::string & s) { |
| 468 | if (s == "f16") { |
| 469 | return GGML_TYPE_F16; |
| 470 | } |
| 471 | if (s == "bf16") { |
| 472 | return GGML_TYPE_BF16; |
| 473 | } |
| 474 | if (s == "q8_0") { |
| 475 | return GGML_TYPE_Q8_0; |
| 476 | } |
| 477 | if (s == "q4_0") { |
| 478 | return GGML_TYPE_Q4_0; |
| 479 | } |
| 480 | if (s == "q4_1") { |
| 481 | return GGML_TYPE_Q4_1; |
| 482 | } |
| 483 | if (s == "q5_0") { |
| 484 | return GGML_TYPE_Q5_0; |
| 485 | } |
| 486 | if (s == "q5_1") { |
| 487 | return GGML_TYPE_Q5_1; |
| 488 | } |
| 489 | if (s == "iq4_nl") { |
| 490 | return GGML_TYPE_IQ4_NL; |
| 491 | } |
| 492 | |
| 493 | return GGML_TYPE_COUNT; |
| 494 | } |
| 495 | |
| 496 | static cmd_params parse_cmd_params(int argc, char ** argv) { |
| 497 | cmd_params params; |
| 498 | std::string arg; |
| 499 | bool invalid_param = false; |
| 500 | const std::string arg_prefix = "--"; |
| 501 | const char split_delim = ','; |
| 502 | |
| 503 | params.verbose = cmd_params_defaults.verbose; |
| 504 | params.output_format = cmd_params_defaults.output_format; |
| 505 | params.output_format_stderr = cmd_params_defaults.output_format_stderr; |
| 506 | params.reps = cmd_params_defaults.reps; |
| 507 | params.numa = cmd_params_defaults.numa; |
| 508 | params.prio = cmd_params_defaults.prio; |
| 509 | params.delay = cmd_params_defaults.delay; |
| 510 | params.progress = cmd_params_defaults.progress; |
| 511 | params.no_warmup = cmd_params_defaults.no_warmup; |
| 512 | |
| 513 | for (int i = 1; i < argc; i++) { |
| 514 | arg = argv[i]; |
| 515 | if (arg.compare(pos: 0, n: arg_prefix.size(), str: arg_prefix) == 0) { |
| 516 | std::replace(first: arg.begin(), last: arg.end(), old_value: '_', new_value: '-'); |
| 517 | } |
| 518 | |
| 519 | try { |
| 520 | if (arg == "-h"|| arg == "--help") { |
| 521 | print_usage(argc, argv); |
| 522 | exit(status: 0); |
| 523 | } else if (arg == "-m"|| arg == "--model") { |
| 524 | if (++i >= argc) { |
| 525 | invalid_param = true; |
| 526 | break; |
| 527 | } |
| 528 | auto p = string_split<std::string>(input: argv[i], separator: split_delim); |
| 529 | params.model.insert(position: params.model.end(), first: p.begin(), last: p.end()); |
| 530 | } else if (arg == "-p"|| arg == "--n-prompt") { |
| 531 | if (++i >= argc) { |
| 532 | invalid_param = true; |
| 533 | break; |
| 534 | } |
| 535 | auto p = parse_int_range(s: argv[i]); |
| 536 | params.n_prompt.insert(position: params.n_prompt.end(), first: p.begin(), last: p.end()); |
| 537 | } else if (arg == "-n"|| arg == "--n-gen") { |
| 538 | if (++i >= argc) { |
| 539 | invalid_param = true; |
| 540 | break; |
| 541 | } |
| 542 | auto p = parse_int_range(s: argv[i]); |
| 543 | params.n_gen.insert(position: params.n_gen.end(), first: p.begin(), last: p.end()); |
| 544 | } else if (arg == "-pg") { |
| 545 | if (++i >= argc) { |
| 546 | invalid_param = true; |
| 547 | break; |
| 548 | } |
| 549 | auto p = string_split<std::string>(input: argv[i], separator: ','); |
| 550 | if (p.size() != 2) { |
| 551 | invalid_param = true; |
| 552 | break; |
| 553 | } |
| 554 | params.n_pg.push_back(x: { std::stoi(str: p[0]), std::stoi(str: p[1]) }); |
| 555 | } else if (arg == "-d"|| arg == "--n-depth") { |
| 556 | if (++i >= argc) { |
| 557 | invalid_param = true; |
| 558 | break; |
| 559 | } |
| 560 | auto p = parse_int_range(s: argv[i]); |
| 561 | params.n_depth.insert(position: params.n_depth.end(), first: p.begin(), last: p.end()); |
| 562 | } else if (arg == "-b"|| arg == "--batch-size") { |
| 563 | if (++i >= argc) { |
| 564 | invalid_param = true; |
| 565 | break; |
| 566 | } |
| 567 | auto p = parse_int_range(s: argv[i]); |
| 568 | params.n_batch.insert(position: params.n_batch.end(), first: p.begin(), last: p.end()); |
| 569 | } else if (arg == "-ub"|| arg == "--ubatch-size") { |
| 570 | if (++i >= argc) { |
| 571 | invalid_param = true; |
| 572 | break; |
| 573 | } |
| 574 | auto p = parse_int_range(s: argv[i]); |
| 575 | params.n_ubatch.insert(position: params.n_ubatch.end(), first: p.begin(), last: p.end()); |
| 576 | } else if (arg == "-ctk"|| arg == "--cache-type-k") { |
| 577 | if (++i >= argc) { |
| 578 | invalid_param = true; |
| 579 | break; |
| 580 | } |
| 581 | auto p = string_split<std::string>(input: argv[i], separator: split_delim); |
| 582 | |
| 583 | std::vector<ggml_type> types; |
| 584 | for (const auto & t : p) { |
| 585 | ggml_type gt = ggml_type_from_name(s: t); |
| 586 | if (gt == GGML_TYPE_COUNT) { |
| 587 | invalid_param = true; |
| 588 | break; |
| 589 | } |
| 590 | types.push_back(x: gt); |
| 591 | } |
| 592 | if (invalid_param) { |
| 593 | break; |
| 594 | } |
| 595 | params.type_k.insert(position: params.type_k.end(), first: types.begin(), last: types.end()); |
| 596 | } else if (arg == "-ctv"|| arg == "--cache-type-v") { |
| 597 | if (++i >= argc) { |
| 598 | invalid_param = true; |
| 599 | break; |
| 600 | } |
| 601 | auto p = string_split<std::string>(input: argv[i], separator: split_delim); |
| 602 | |
| 603 | std::vector<ggml_type> types; |
| 604 | for (const auto & t : p) { |
| 605 | ggml_type gt = ggml_type_from_name(s: t); |
| 606 | if (gt == GGML_TYPE_COUNT) { |
| 607 | invalid_param = true; |
| 608 | break; |
| 609 | } |
| 610 | types.push_back(x: gt); |
| 611 | } |
| 612 | if (invalid_param) { |
| 613 | break; |
| 614 | } |
| 615 | params.type_v.insert(position: params.type_v.end(), first: types.begin(), last: types.end()); |
| 616 | } else if (arg == "-dev"|| arg == "--device") { |
| 617 | if (++i >= argc) { |
| 618 | invalid_param = true; |
| 619 | break; |
| 620 | } |
| 621 | auto combos = string_split<std::string>(input: argv[i], separator: split_delim); |
| 622 | for (const auto & combo : combos) { |
| 623 | try { |
| 624 | params.devices.push_back(x: parse_devices_arg(value: combo)); |
| 625 | } catch (const std::exception & e) { |
| 626 | fprintf(stderr, format: "error: %s\n", e.what()); |
| 627 | invalid_param = true; |
| 628 | break; |
| 629 | } |
| 630 | } |
| 631 | if (invalid_param) { |
| 632 | break; |
| 633 | } |
| 634 | } else if (arg == "--list-devices") { |
| 635 | std::vector<ggml_backend_dev_t> devices; |
| 636 | for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { |
| 637 | auto * dev = ggml_backend_dev_get(index: i); |
| 638 | if (ggml_backend_dev_type(device: dev) != GGML_BACKEND_DEVICE_TYPE_CPU) { |
| 639 | devices.push_back(x: dev); |
| 640 | } |
| 641 | } |
| 642 | printf(format: "Available devices:\n"); |
| 643 | if (devices.empty()) { |
| 644 | printf(format: " (none)\n"); |
| 645 | } |
| 646 | for (auto * dev : devices) { |
| 647 | size_t free, total; |
| 648 | ggml_backend_dev_memory(device: dev, free: &free, total: &total); |
| 649 | printf(format: " %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(device: dev), ggml_backend_dev_description(device: dev), total / 1024 / 1024, free / 1024 / 1024); |
| 650 | } |
| 651 | exit(status: 0); |
| 652 | } else if (arg == "-t"|| arg == "--threads") { |
| 653 | if (++i >= argc) { |
| 654 | invalid_param = true; |
| 655 | break; |
| 656 | } |
| 657 | auto p = parse_int_range(s: argv[i]); |
| 658 | params.n_threads.insert(position: params.n_threads.end(), first: p.begin(), last: p.end()); |
| 659 | } else if (arg == "-C"|| arg == "--cpu-mask") { |
| 660 | if (++i >= argc) { |
| 661 | invalid_param = true; |
| 662 | break; |
| 663 | } |
| 664 | auto p = string_split<std::string>(input: argv[i], separator: split_delim); |
| 665 | params.cpu_mask.insert(position: params.cpu_mask.end(), first: p.begin(), last: p.end()); |
| 666 | } else if (arg == "--cpu-strict") { |
| 667 | if (++i >= argc) { |
| 668 | invalid_param = true; |
| 669 | break; |
| 670 | } |
| 671 | auto p = string_split<bool>(str: argv[i], delim: split_delim); |
| 672 | params.cpu_strict.insert(position: params.cpu_strict.end(), first: p.begin(), last: p.end()); |
| 673 | } else if (arg == "--poll") { |
| 674 | if (++i >= argc) { |
| 675 | invalid_param = true; |
| 676 | break; |
| 677 | } |
| 678 | auto p = parse_int_range(s: argv[i]); |
| 679 | params.poll.insert(position: params.poll.end(), first: p.begin(), last: p.end()); |
| 680 | } else if (arg == "-ngl"|| arg == "--n-gpu-layers") { |
| 681 | if (++i >= argc) { |
| 682 | invalid_param = true; |
| 683 | break; |
| 684 | } |
| 685 | auto p = parse_int_range(s: argv[i]); |
| 686 | params.n_gpu_layers.insert(position: params.n_gpu_layers.end(), first: p.begin(), last: p.end()); |
| 687 | } else if (arg == "-ncmoe"|| arg == "--n-cpu-moe") { |
| 688 | if (++i >= argc) { |
| 689 | invalid_param = true; |
| 690 | break; |
| 691 | } |
| 692 | auto p = parse_int_range(s: argv[i]); |
| 693 | params.n_cpu_moe.insert(position: params.n_cpu_moe.end(), first: p.begin(), last: p.end()); |
| 694 | } else if (llama_supports_rpc() && (arg == "-rpc"|| arg == "--rpc")) { |
| 695 | if (++i >= argc) { |
| 696 | invalid_param = true; |
| 697 | break; |
| 698 | } |
| 699 | try { |
| 700 | register_rpc_server_list(servers: argv[i]); |
| 701 | } catch (const std::exception & e) { |
| 702 | fprintf(stderr, format: "error: %s\n", e.what()); |
| 703 | invalid_param = true; |
| 704 | break; |
| 705 | } |
| 706 | } else if (arg == "-sm"|| arg == "--split-mode") { |
| 707 | if (++i >= argc) { |
| 708 | invalid_param = true; |
| 709 | break; |
| 710 | } |
| 711 | auto p = string_split<std::string>(input: argv[i], separator: split_delim); |
| 712 | |
| 713 | std::vector<llama_split_mode> modes; |
| 714 | for (const auto & m : p) { |
| 715 | llama_split_mode mode; |
| 716 | if (m == "none") { |
| 717 | mode = LLAMA_SPLIT_MODE_NONE; |
| 718 | } else if (m == "layer") { |
| 719 | mode = LLAMA_SPLIT_MODE_LAYER; |
| 720 | } else if (m == "row") { |
| 721 | mode = LLAMA_SPLIT_MODE_ROW; |
| 722 | } else { |
| 723 | invalid_param = true; |
| 724 | break; |
| 725 | } |
| 726 | modes.push_back(x: mode); |
| 727 | } |
| 728 | if (invalid_param) { |
| 729 | break; |
| 730 | } |
| 731 | params.split_mode.insert(position: params.split_mode.end(), first: modes.begin(), last: modes.end()); |
| 732 | } else if (arg == "-mg"|| arg == "--main-gpu") { |
| 733 | if (++i >= argc) { |
| 734 | invalid_param = true; |
| 735 | break; |
| 736 | } |
| 737 | params.main_gpu = parse_int_range(s: argv[i]); |
| 738 | } else if (arg == "-nkvo"|| arg == "--no-kv-offload") { |
| 739 | if (++i >= argc) { |
| 740 | invalid_param = true; |
| 741 | break; |
| 742 | } |
| 743 | auto p = string_split<bool>(str: argv[i], delim: split_delim); |
| 744 | params.no_kv_offload.insert(position: params.no_kv_offload.end(), first: p.begin(), last: p.end()); |
| 745 | } else if (arg == "--numa") { |
| 746 | if (++i >= argc) { |
| 747 | invalid_param = true; |
| 748 | break; |
| 749 | } |
| 750 | std::string value(argv[i]); |
| 751 | if (value == "distribute"|| value == "") { |
| 752 | params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; |
| 753 | } else if (value == "isolate") { |
| 754 | params.numa = GGML_NUMA_STRATEGY_ISOLATE; |
| 755 | } else if (value == "numactl") { |
| 756 | params.numa = GGML_NUMA_STRATEGY_NUMACTL; |
| 757 | } else { |
| 758 | invalid_param = true; |
| 759 | break; |
| 760 | } |
| 761 | } else if (arg == "-fa"|| arg == "--flash-attn") { |
| 762 | if (++i >= argc) { |
| 763 | invalid_param = true; |
| 764 | break; |
| 765 | } |
| 766 | auto p = string_split<bool>(str: argv[i], delim: split_delim); |
| 767 | params.flash_attn.insert(position: params.flash_attn.end(), first: p.begin(), last: p.end()); |
| 768 | } else if (arg == "-mmp"|| arg == "--mmap") { |
| 769 | if (++i >= argc) { |
| 770 | invalid_param = true; |
| 771 | break; |
| 772 | } |
| 773 | auto p = string_split<bool>(str: argv[i], delim: split_delim); |
| 774 | params.use_mmap.insert(position: params.use_mmap.end(), first: p.begin(), last: p.end()); |
| 775 | } else if (arg == "-embd"|| arg == "--embeddings") { |
| 776 | if (++i >= argc) { |
| 777 | invalid_param = true; |
| 778 | break; |
| 779 | } |
| 780 | auto p = string_split<bool>(str: argv[i], delim: split_delim); |
| 781 | params.embeddings.insert(position: params.embeddings.end(), first: p.begin(), last: p.end()); |
| 782 | } else if (arg == "-nopo"|| arg == "--no-op-offload") { |
| 783 | if (++i >= argc) { |
| 784 | invalid_param = true; |
| 785 | break; |
| 786 | } |
| 787 | auto p = string_split<bool>(str: argv[i], delim: split_delim); |
| 788 | params.no_op_offload.insert(position: params.no_op_offload.end(), first: p.begin(), last: p.end()); |
| 789 | } else if (arg == "--no-host") { |
| 790 | if (++i >= argc) { |
| 791 | invalid_param = true; |
| 792 | break; |
| 793 | } |
| 794 | auto p = string_split<bool>(str: argv[i], delim: split_delim); |
| 795 | params.no_host.insert(position: params.no_host.end(), first: p.begin(), last: p.end()); |
| 796 | } else if (arg == "-ts"|| arg == "--tensor-split") { |
| 797 | if (++i >= argc) { |
| 798 | invalid_param = true; |
| 799 | break; |
| 800 | } |
| 801 | for (auto ts : string_split<std::string>(input: argv[i], separator: split_delim)) { |
| 802 | // split string by ; and / |
| 803 | const std::regex regex{ R"([;/]+)"}; |
| 804 | std::sregex_token_iterator it{ ts.begin(), ts.end(), regex, -1 }; |
| 805 | std::vector<std::string> split_arg{ it, {} }; |
| 806 | GGML_ASSERT(split_arg.size() <= llama_max_devices()); |
| 807 | |
| 808 | std::vector<float> tensor_split(llama_max_devices()); |
| 809 | for (size_t i = 0; i < llama_max_devices(); ++i) { |
| 810 | if (i < split_arg.size()) { |
| 811 | tensor_split[i] = std::stof(str: split_arg[i]); |
| 812 | } else { |
| 813 | tensor_split[i] = 0.0f; |
| 814 | } |
| 815 | } |
| 816 | params.tensor_split.push_back(x: tensor_split); |
| 817 | } |
| 818 | } else if (arg == "-ot"|| arg == "--override-tensor") { |
| 819 | if (++i >= argc) { |
| 820 | invalid_param = true; |
| 821 | break; |
| 822 | } |
| 823 | auto * value = argv[i]; |
| 824 | /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list; |
| 825 | if (buft_list.empty()) { |
| 826 | // enumerate all the devices and add their buffer types to the list |
| 827 | for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { |
| 828 | auto * dev = ggml_backend_dev_get(index: i); |
| 829 | auto * buft = ggml_backend_dev_buffer_type(device: dev); |
| 830 | if (buft) { |
| 831 | buft_list[ggml_backend_buft_name(buft)] = buft; |
| 832 | } |
| 833 | } |
| 834 | } |
| 835 | auto override_group_span_len = std::strcspn(s: value, reject: ","); |
| 836 | bool last_group = false; |
| 837 | do { |
| 838 | if (override_group_span_len == 0) { |
| 839 | // Adds an empty override-tensors for an empty span |
| 840 | params.tensor_buft_overrides.push_back(x: {{}}); |
| 841 | if (value[override_group_span_len] == '\0') { |
| 842 | value = &value[override_group_span_len]; |
| 843 | last_group = true; |
| 844 | } else { |
| 845 | value = &value[override_group_span_len + 1]; |
| 846 | override_group_span_len = std::strcspn(s: value, reject: ","); |
| 847 | } |
| 848 | continue; |
| 849 | } |
| 850 | // Stamps null terminators into the argv |
| 851 | // value for this option to avoid the |
| 852 | // memory leak present in the implementation |
| 853 | // over in arg.cpp. Acceptable because we |
| 854 | // only parse these args once in this program. |
| 855 | auto * override_group = value; |
| 856 | if (value[override_group_span_len] == '\0') { |
| 857 | value = &value[override_group_span_len]; |
| 858 | last_group = true; |
| 859 | } else { |
| 860 | value[override_group_span_len] = '\0'; |
| 861 | value = &value[override_group_span_len + 1]; |
| 862 | } |
| 863 | std::vector<llama_model_tensor_buft_override> group_tensor_buft_overrides{}; |
| 864 | auto override_span_len = std::strcspn(s: override_group, reject: ";"); |
| 865 | while (override_span_len > 0) { |
| 866 | auto * override = override_group; |
| 867 | if (override_group[override_span_len] != '\0') { |
| 868 | override_group[override_span_len] = '\0'; |
| 869 | override_group = &override_group[override_span_len + 1]; |
| 870 | } else { |
| 871 | override_group = &override_group[override_span_len]; |
| 872 | } |
| 873 | auto tensor_name_span_len = std::strcspn(s: override, reject: "="); |
| 874 | if (tensor_name_span_len >= override_span_len) { |
| 875 | invalid_param = true; |
| 876 | break; |
| 877 | } |
| 878 | override[tensor_name_span_len] = '\0'; |
| 879 | auto * tensor_name = override; |
| 880 | auto * buffer_type = &override[tensor_name_span_len + 1]; |
| 881 | if (buft_list.find(x: buffer_type) == buft_list.end()) { |
| 882 | printf(format: "error: unrecognized buffer type '%s'\n", buffer_type); |
| 883 | printf(format: "Available buffer types:\n"); |
| 884 | for (const auto & it : buft_list) { |
| 885 | printf(format: " %s\n", ggml_backend_buft_name(buft: it.second)); |
| 886 | } |
| 887 | invalid_param = true; |
| 888 | break; |
| 889 | } |
| 890 | group_tensor_buft_overrides.push_back(x: {.pattern: tensor_name, .buft: buft_list.at(k: buffer_type)}); |
| 891 | override_span_len = std::strcspn(s: override_group, reject: ";"); |
| 892 | } |
| 893 | if (invalid_param) { |
| 894 | break; |
| 895 | } |
| 896 | group_tensor_buft_overrides.push_back(x: {.pattern: nullptr,.buft: nullptr}); |
| 897 | params.tensor_buft_overrides.push_back(x: group_tensor_buft_overrides); |
| 898 | override_group_span_len = std::strcspn(s: value, reject: ","); |
| 899 | } while (!last_group); |
| 900 | } else if (arg == "-r"|| arg == "--repetitions") { |
| 901 | if (++i >= argc) { |
| 902 | invalid_param = true; |
| 903 | break; |
| 904 | } |
| 905 | params.reps = std::stoi(str: argv[i]); |
| 906 | } else if (arg == "--prio") { |
| 907 | if (++i >= argc) { |
| 908 | invalid_param = true; |
| 909 | break; |
| 910 | } |
| 911 | params.prio = (enum ggml_sched_priority) std::stoi(str: argv[i]); |
| 912 | } else if (arg == "--delay") { |
| 913 | if (++i >= argc) { |
| 914 | invalid_param = true; |
| 915 | break; |
| 916 | } |
| 917 | params.delay = std::stoi(str: argv[i]); |
| 918 | } else if (arg == "-o"|| arg == "--output") { |
| 919 | if (++i >= argc) { |
| 920 | invalid_param = true; |
| 921 | break; |
| 922 | } |
| 923 | invalid_param = !output_format_from_str(s: argv[i], format&: params.output_format); |
| 924 | } else if (arg == "-oe"|| arg == "--output-err") { |
| 925 | if (++i >= argc) { |
| 926 | invalid_param = true; |
| 927 | break; |
| 928 | } |
| 929 | invalid_param = !output_format_from_str(s: argv[i], format&: params.output_format_stderr); |
| 930 | } else if (arg == "-v"|| arg == "--verbose") { |
| 931 | params.verbose = true; |
| 932 | } else if (arg == "--progress") { |
| 933 | params.progress = true; |
| 934 | } else if (arg == "--no-warmup") { |
| 935 | params.no_warmup = true; |
| 936 | } else { |
| 937 | invalid_param = true; |
| 938 | break; |
| 939 | } |
| 940 | } catch (const std::exception & e) { |
| 941 | fprintf(stderr, format: "error: %s\n", e.what()); |
| 942 | invalid_param = true; |
| 943 | break; |
| 944 | } |
| 945 | } |
| 946 | |
| 947 | if (invalid_param) { |
| 948 | fprintf(stderr, format: "error: invalid parameter for argument: %s\n", arg.c_str()); |
| 949 | print_usage(argc, argv); |
| 950 | exit(status: 1); |
| 951 | } |
| 952 | |
| 953 | // set defaults |
| 954 | if (params.model.empty()) { |
| 955 | params.model = cmd_params_defaults.model; |
| 956 | } |
| 957 | if (params.n_prompt.empty()) { |
| 958 | params.n_prompt = cmd_params_defaults.n_prompt; |
| 959 | } |
| 960 | if (params.n_gen.empty()) { |
| 961 | params.n_gen = cmd_params_defaults.n_gen; |
| 962 | } |
| 963 | if (params.n_pg.empty()) { |
| 964 | params.n_pg = cmd_params_defaults.n_pg; |
| 965 | } |
| 966 | if (params.n_depth.empty()) { |
| 967 | params.n_depth = cmd_params_defaults.n_depth; |
| 968 | } |
| 969 | if (params.n_batch.empty()) { |
| 970 | params.n_batch = cmd_params_defaults.n_batch; |
| 971 | } |
| 972 | if (params.n_ubatch.empty()) { |
| 973 | params.n_ubatch = cmd_params_defaults.n_ubatch; |
| 974 | } |
| 975 | if (params.type_k.empty()) { |
| 976 | params.type_k = cmd_params_defaults.type_k; |
| 977 | } |
| 978 | if (params.type_v.empty()) { |
| 979 | params.type_v = cmd_params_defaults.type_v; |
| 980 | } |
| 981 | if (params.n_gpu_layers.empty()) { |
| 982 | params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; |
| 983 | } |
| 984 | if (params.n_cpu_moe.empty()) { |
| 985 | params.n_cpu_moe = cmd_params_defaults.n_cpu_moe; |
| 986 | } |
| 987 | if (params.split_mode.empty()) { |
| 988 | params.split_mode = cmd_params_defaults.split_mode; |
| 989 | } |
| 990 | if (params.main_gpu.empty()) { |
| 991 | params.main_gpu = cmd_params_defaults.main_gpu; |
| 992 | } |
| 993 | if (params.no_kv_offload.empty()) { |
| 994 | params.no_kv_offload = cmd_params_defaults.no_kv_offload; |
| 995 | } |
| 996 | if (params.flash_attn.empty()) { |
| 997 | params.flash_attn = cmd_params_defaults.flash_attn; |
| 998 | } |
| 999 | if (params.devices.empty()) { |
| 1000 | params.devices = cmd_params_defaults.devices; |
| 1001 | } |
| 1002 | if (params.tensor_split.empty()) { |
| 1003 | params.tensor_split = cmd_params_defaults.tensor_split; |
| 1004 | } |
| 1005 | if (params.tensor_buft_overrides.empty()) { |
| 1006 | params.tensor_buft_overrides = cmd_params_defaults.tensor_buft_overrides; |
| 1007 | } |
| 1008 | if (params.use_mmap.empty()) { |
| 1009 | params.use_mmap = cmd_params_defaults.use_mmap; |
| 1010 | } |
| 1011 | if (params.embeddings.empty()) { |
| 1012 | params.embeddings = cmd_params_defaults.embeddings; |
| 1013 | } |
| 1014 | if (params.no_op_offload.empty()) { |
| 1015 | params.no_op_offload = cmd_params_defaults.no_op_offload; |
| 1016 | } |
| 1017 | if (params.no_host.empty()) { |
| 1018 | params.no_host = cmd_params_defaults.no_host; |
| 1019 | } |
| 1020 | if (params.n_threads.empty()) { |
| 1021 | params.n_threads = cmd_params_defaults.n_threads; |
| 1022 | } |
| 1023 | if (params.cpu_mask.empty()) { |
| 1024 | params.cpu_mask = cmd_params_defaults.cpu_mask; |
| 1025 | } |
| 1026 | if (params.cpu_strict.empty()) { |
| 1027 | params.cpu_strict = cmd_params_defaults.cpu_strict; |
| 1028 | } |
| 1029 | if (params.poll.empty()) { |
| 1030 | params.poll = cmd_params_defaults.poll; |
| 1031 | } |
| 1032 | |
| 1033 | return params; |
| 1034 | } |
| 1035 | |
| 1036 | struct cmd_params_instance { |
| 1037 | std::string model; |
| 1038 | int n_prompt; |
| 1039 | int n_gen; |
| 1040 | int n_depth; |
| 1041 | int n_batch; |
| 1042 | int n_ubatch; |
| 1043 | ggml_type type_k; |
| 1044 | ggml_type type_v; |
| 1045 | int n_threads; |
| 1046 | std::string cpu_mask; |
| 1047 | bool cpu_strict; |
| 1048 | int poll; |
| 1049 | int n_gpu_layers; |
| 1050 | int n_cpu_moe; |
| 1051 | llama_split_mode split_mode; |
| 1052 | int main_gpu; |
| 1053 | bool no_kv_offload; |
| 1054 | bool flash_attn; |
| 1055 | std::vector<ggml_backend_dev_t> devices; |
| 1056 | std::vector<float> tensor_split; |
| 1057 | std::vector<llama_model_tensor_buft_override> tensor_buft_overrides; |
| 1058 | bool use_mmap; |
| 1059 | bool embeddings; |
| 1060 | bool no_op_offload; |
| 1061 | bool no_host; |
| 1062 | |
| 1063 | llama_model_params to_llama_mparams() const { |
| 1064 | llama_model_params mparams = llama_model_default_params(); |
| 1065 | |
| 1066 | mparams.n_gpu_layers = n_gpu_layers; |
| 1067 | if (!devices.empty()) { |
| 1068 | mparams.devices = const_cast<ggml_backend_dev_t *>(devices.data()); |
| 1069 | } |
| 1070 | mparams.split_mode = split_mode; |
| 1071 | mparams.main_gpu = main_gpu; |
| 1072 | mparams.tensor_split = tensor_split.data(); |
| 1073 | mparams.use_mmap = use_mmap; |
| 1074 | mparams.no_host = no_host; |
| 1075 | |
| 1076 | if (n_cpu_moe <= 0) { |
| 1077 | if (tensor_buft_overrides.empty()) { |
| 1078 | mparams.tensor_buft_overrides = nullptr; |
| 1079 | } else { |
| 1080 | GGML_ASSERT(tensor_buft_overrides.back().pattern == nullptr && |
| 1081 | "Tensor buffer overrides not terminated with empty pattern"); |
| 1082 | mparams.tensor_buft_overrides = tensor_buft_overrides.data(); |
| 1083 | } |
| 1084 | } else { |
| 1085 | static std::vector<llama_model_tensor_buft_override> merged; |
| 1086 | static std::vector<std::string> patterns; |
| 1087 | |
| 1088 | merged.clear(); |
| 1089 | patterns.clear(); |
| 1090 | |
| 1091 | auto first = tensor_buft_overrides.begin(); |
| 1092 | auto last = tensor_buft_overrides.end(); |
| 1093 | if (first != last && (last - 1)->pattern == nullptr) { |
| 1094 | --last; |
| 1095 | } |
| 1096 | merged.insert(position: merged.end(), first: first, last: last); |
| 1097 | |
| 1098 | patterns.reserve(n: (size_t) n_cpu_moe); |
| 1099 | merged.reserve(n: merged.size() + (size_t) n_cpu_moe + 1); |
| 1100 | |
| 1101 | for (int i = 0; i < n_cpu_moe; ++i) { |
| 1102 | patterns.push_back(x: llm_ffn_exps_block_regex(idx: i)); |
| 1103 | merged.push_back(x: { .pattern: patterns.back().c_str(), |
| 1104 | .buft: ggml_backend_cpu_buffer_type() }); |
| 1105 | } |
| 1106 | |
| 1107 | merged.push_back(x: { .pattern: nullptr, .buft: nullptr }); |
| 1108 | |
| 1109 | mparams.tensor_buft_overrides = merged.data(); |
| 1110 | } |
| 1111 | |
| 1112 | return mparams; |
| 1113 | } |
| 1114 | |
| 1115 | bool equal_mparams(const cmd_params_instance & other) const { |
| 1116 | return model == other.model && n_gpu_layers == other.n_gpu_layers && n_cpu_moe == other.n_cpu_moe && |
| 1117 | split_mode == other.split_mode && |
| 1118 | main_gpu == other.main_gpu && use_mmap == other.use_mmap && tensor_split == other.tensor_split && |
| 1119 | devices == other.devices && |
| 1120 | no_host == other.no_host && |
| 1121 | vec_tensor_buft_override_equal(a: tensor_buft_overrides, b: other.tensor_buft_overrides); |
| 1122 | } |
| 1123 | |
| 1124 | llama_context_params to_llama_cparams() const { |
| 1125 | llama_context_params cparams = llama_context_default_params(); |
| 1126 | |
| 1127 | cparams.n_ctx = n_prompt + n_gen + n_depth; |
| 1128 | cparams.n_batch = n_batch; |
| 1129 | cparams.n_ubatch = n_ubatch; |
| 1130 | cparams.type_k = type_k; |
| 1131 | cparams.type_v = type_v; |
| 1132 | cparams.offload_kqv = !no_kv_offload; |
| 1133 | cparams.flash_attn_type = flash_attn ? LLAMA_FLASH_ATTN_TYPE_ENABLED : LLAMA_FLASH_ATTN_TYPE_DISABLED; |
| 1134 | cparams.embeddings = embeddings; |
| 1135 | cparams.op_offload = !no_op_offload; |
| 1136 | cparams.swa_full = false; |
| 1137 | |
| 1138 | return cparams; |
| 1139 | } |
| 1140 | }; |
| 1141 | |
| 1142 | static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) { |
| 1143 | std::vector<cmd_params_instance> instances; |
| 1144 | |
| 1145 | // this ordering minimizes the number of times that each model needs to be reloaded |
| 1146 | // clang-format off |
| 1147 | for (const auto & m : params.model) |
| 1148 | for (const auto & nl : params.n_gpu_layers) |
| 1149 | for (const auto & ncmoe : params.n_cpu_moe) |
| 1150 | for (const auto & sm : params.split_mode) |
| 1151 | for (const auto & mg : params.main_gpu) |
| 1152 | for (const auto & devs : params.devices) |
| 1153 | for (const auto & ts : params.tensor_split) |
| 1154 | for (const auto & ot : params.tensor_buft_overrides) |
| 1155 | for (const auto & mmp : params.use_mmap) |
| 1156 | for (const auto & noh : params.no_host) |
| 1157 | for (const auto & embd : params.embeddings) |
| 1158 | for (const auto & nopo : params.no_op_offload) |
| 1159 | for (const auto & nb : params.n_batch) |
| 1160 | for (const auto & nub : params.n_ubatch) |
| 1161 | for (const auto & tk : params.type_k) |
| 1162 | for (const auto & tv : params.type_v) |
| 1163 | for (const auto & nkvo : params.no_kv_offload) |
| 1164 | for (const auto & fa : params.flash_attn) |
| 1165 | for (const auto & nt : params.n_threads) |
| 1166 | for (const auto & cm : params.cpu_mask) |
| 1167 | for (const auto & cs : params.cpu_strict) |
| 1168 | for (const auto & nd : params.n_depth) |
| 1169 | for (const auto & pl : params.poll) { |
| 1170 | for (const auto & n_prompt : params.n_prompt) { |
| 1171 | if (n_prompt == 0) { |
| 1172 | continue; |
| 1173 | } |
| 1174 | cmd_params_instance instance = { |
| 1175 | /* .model = */ m, |
| 1176 | /* .n_prompt = */ n_prompt, |
| 1177 | /* .n_gen = */ 0, |
| 1178 | /* .n_depth = */ nd, |
| 1179 | /* .n_batch = */ nb, |
| 1180 | /* .n_ubatch = */ nub, |
| 1181 | /* .type_k = */ tk, |
| 1182 | /* .type_v = */ tv, |
| 1183 | /* .n_threads = */ nt, |
| 1184 | /* .cpu_mask = */ cm, |
| 1185 | /* .cpu_strict = */ cs, |
| 1186 | /* .poll = */ pl, |
| 1187 | /* .n_gpu_layers = */ nl, |
| 1188 | /* .n_cpu_moe = */ ncmoe, |
| 1189 | /* .split_mode = */ sm, |
| 1190 | /* .main_gpu = */ mg, |
| 1191 | /* .no_kv_offload= */ nkvo, |
| 1192 | /* .flash_attn = */ fa, |
| 1193 | /* .devices = */ devs, |
| 1194 | /* .tensor_split = */ ts, |
| 1195 | /* .tensor_buft_overrides = */ ot, |
| 1196 | /* .use_mmap = */ mmp, |
| 1197 | /* .embeddings = */ embd, |
| 1198 | /* .no_op_offload= */ nopo, |
| 1199 | /* .no_host = */ noh, |
| 1200 | }; |
| 1201 | instances.push_back(x: instance); |
| 1202 | } |
| 1203 | |
| 1204 | for (const auto & n_gen : params.n_gen) { |
| 1205 | if (n_gen == 0) { |
| 1206 | continue; |
| 1207 | } |
| 1208 | cmd_params_instance instance = { |
| 1209 | /* .model = */ m, |
| 1210 | /* .n_prompt = */ 0, |
| 1211 | /* .n_gen = */ n_gen, |
| 1212 | /* .n_depth = */ nd, |
| 1213 | /* .n_batch = */ nb, |
| 1214 | /* .n_ubatch = */ nub, |
| 1215 | /* .type_k = */ tk, |
| 1216 | /* .type_v = */ tv, |
| 1217 | /* .n_threads = */ nt, |
| 1218 | /* .cpu_mask = */ cm, |
| 1219 | /* .cpu_strict = */ cs, |
| 1220 | /* .poll = */ pl, |
| 1221 | /* .n_gpu_layers = */ nl, |
| 1222 | /* .n_cpu_moe = */ ncmoe, |
| 1223 | /* .split_mode = */ sm, |
| 1224 | /* .main_gpu = */ mg, |
| 1225 | /* .no_kv_offload= */ nkvo, |
| 1226 | /* .flash_attn = */ fa, |
| 1227 | /* .devices = */ devs, |
| 1228 | /* .tensor_split = */ ts, |
| 1229 | /* .tensor_buft_overrides = */ ot, |
| 1230 | /* .use_mmap = */ mmp, |
| 1231 | /* .embeddings = */ embd, |
| 1232 | /* .no_op_offload= */ nopo, |
| 1233 | /* .no_host = */ noh, |
| 1234 | }; |
| 1235 | instances.push_back(x: instance); |
| 1236 | } |
| 1237 | |
| 1238 | for (const auto & n_pg : params.n_pg) { |
| 1239 | if (n_pg.first == 0 && n_pg.second == 0) { |
| 1240 | continue; |
| 1241 | } |
| 1242 | cmd_params_instance instance = { |
| 1243 | /* .model = */ m, |
| 1244 | /* .n_prompt = */ n_pg.first, |
| 1245 | /* .n_gen = */ n_pg.second, |
| 1246 | /* .n_depth = */ nd, |
| 1247 | /* .n_batch = */ nb, |
| 1248 | /* .n_ubatch = */ nub, |
| 1249 | /* .type_k = */ tk, |
| 1250 | /* .type_v = */ tv, |
| 1251 | /* .n_threads = */ nt, |
| 1252 | /* .cpu_mask = */ cm, |
| 1253 | /* .cpu_strict = */ cs, |
| 1254 | /* .poll = */ pl, |
| 1255 | /* .n_gpu_layers = */ nl, |
| 1256 | /* .n_cpu_moe = */ ncmoe, |
| 1257 | /* .split_mode = */ sm, |
| 1258 | /* .main_gpu = */ mg, |
| 1259 | /* .no_kv_offload= */ nkvo, |
| 1260 | /* .flash_attn = */ fa, |
| 1261 | /* .devices = */ devs, |
| 1262 | /* .tensor_split = */ ts, |
| 1263 | /* .tensor_buft_overrides = */ ot, |
| 1264 | /* .use_mmap = */ mmp, |
| 1265 | /* .embeddings = */ embd, |
| 1266 | /* .no_op_offload= */ nopo, |
| 1267 | /* .no_host = */ noh, |
| 1268 | }; |
| 1269 | instances.push_back(x: instance); |
| 1270 | } |
| 1271 | } |
| 1272 | // clang-format on |
| 1273 | |
| 1274 | return instances; |
| 1275 | } |
| 1276 | |
| 1277 | struct test { |
| 1278 | static const std::string build_commit; |
| 1279 | static const int build_number; |
| 1280 | const std::string cpu_info; |
| 1281 | const std::string gpu_info; |
| 1282 | std::string model_filename; |
| 1283 | std::string model_type; |
| 1284 | uint64_t model_size; |
| 1285 | uint64_t model_n_params; |
| 1286 | int n_batch; |
| 1287 | int n_ubatch; |
| 1288 | int n_threads; |
| 1289 | std::string cpu_mask; |
| 1290 | bool cpu_strict; |
| 1291 | int poll; |
| 1292 | ggml_type type_k; |
| 1293 | ggml_type type_v; |
| 1294 | int n_gpu_layers; |
| 1295 | int n_cpu_moe; |
| 1296 | llama_split_mode split_mode; |
| 1297 | int main_gpu; |
| 1298 | bool no_kv_offload; |
| 1299 | bool flash_attn; |
| 1300 | std::vector<ggml_backend_dev_t> devices; |
| 1301 | std::vector<float> tensor_split; |
| 1302 | std::vector<llama_model_tensor_buft_override> tensor_buft_overrides; |
| 1303 | bool use_mmap; |
| 1304 | bool embeddings; |
| 1305 | bool no_op_offload; |
| 1306 | bool no_host; |
| 1307 | int n_prompt; |
| 1308 | int n_gen; |
| 1309 | int n_depth; |
| 1310 | std::string test_time; |
| 1311 | std::vector<uint64_t> samples_ns; |
| 1312 | |
| 1313 | test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) : |
| 1314 | cpu_info(get_cpu_info()), |
| 1315 | gpu_info(get_gpu_info()) { |
| 1316 | |
| 1317 | model_filename = inst.model; |
| 1318 | char buf[128]; |
| 1319 | llama_model_desc(model: lmodel, buf, buf_size: sizeof(buf)); |
| 1320 | model_type = buf; |
| 1321 | model_size = llama_model_size(model: lmodel); |
| 1322 | model_n_params = llama_model_n_params(model: lmodel); |
| 1323 | n_batch = inst.n_batch; |
| 1324 | n_ubatch = inst.n_ubatch; |
| 1325 | n_threads = inst.n_threads; |
| 1326 | cpu_mask = inst.cpu_mask; |
| 1327 | cpu_strict = inst.cpu_strict; |
| 1328 | poll = inst.poll; |
| 1329 | type_k = inst.type_k; |
| 1330 | type_v = inst.type_v; |
| 1331 | n_gpu_layers = inst.n_gpu_layers; |
| 1332 | n_cpu_moe = inst.n_cpu_moe; |
| 1333 | split_mode = inst.split_mode; |
| 1334 | main_gpu = inst.main_gpu; |
| 1335 | no_kv_offload = inst.no_kv_offload; |
| 1336 | flash_attn = inst.flash_attn; |
| 1337 | devices = inst.devices; |
| 1338 | tensor_split = inst.tensor_split; |
| 1339 | tensor_buft_overrides = inst.tensor_buft_overrides; |
| 1340 | use_mmap = inst.use_mmap; |
| 1341 | embeddings = inst.embeddings; |
| 1342 | no_op_offload = inst.no_op_offload; |
| 1343 | no_host = inst.no_host; |
| 1344 | n_prompt = inst.n_prompt; |
| 1345 | n_gen = inst.n_gen; |
| 1346 | n_depth = inst.n_depth; |
| 1347 | // RFC 3339 date-time format |
| 1348 | time_t t = time(NULL); |
| 1349 | std::strftime(s: buf, maxsize: sizeof(buf), format: "%FT%TZ", tp: gmtime(timer: &t)); |
| 1350 | test_time = buf; |
| 1351 | |
| 1352 | (void) ctx; |
| 1353 | } |
| 1354 | |
| 1355 | uint64_t avg_ns() const { return ::avg(v: samples_ns); } |
| 1356 | |
| 1357 | uint64_t stdev_ns() const { return ::stdev(v: samples_ns); } |
| 1358 | |
| 1359 | std::vector<double> get_ts() const { |
| 1360 | int n_tokens = n_prompt + n_gen; |
| 1361 | std::vector<double> ts; |
| 1362 | std::transform(first: samples_ns.begin(), last: samples_ns.end(), result: std::back_inserter(x&: ts), |
| 1363 | unary_op: [n_tokens](uint64_t t) { return 1e9 * n_tokens / t; }); |
| 1364 | return ts; |
| 1365 | } |
| 1366 | |
| 1367 | double avg_ts() const { return ::avg(v: get_ts()); } |
| 1368 | |
| 1369 | double stdev_ts() const { return ::stdev(v: get_ts()); } |
| 1370 | |
| 1371 | static std::string get_backend() { |
| 1372 | std::vector<std::string> backends; |
| 1373 | bool rpc_used = false; |
| 1374 | for (size_t i = 0; i < ggml_backend_reg_count(); i++) { |
| 1375 | auto * reg = ggml_backend_reg_get(index: i); |
| 1376 | std::string name = ggml_backend_reg_name(reg); |
| 1377 | if (string_starts_with(str: name, prefix: "RPC")) { |
| 1378 | if (ggml_backend_reg_dev_count(reg) > 0) { |
| 1379 | rpc_used = true; |
| 1380 | } |
| 1381 | } else { |
| 1382 | if (name != "CPU") { |
| 1383 | backends.push_back(x: ggml_backend_reg_name(reg)); |
| 1384 | } |
| 1385 | } |
| 1386 | } |
| 1387 | if (rpc_used) { |
| 1388 | backends.push_back(x: "RPC"); |
| 1389 | } |
| 1390 | return backends.empty() ? "CPU": join(values: backends, delim: ","); |
| 1391 | } |
| 1392 | |
| 1393 | static const std::vector<std::string> & get_fields() { |
| 1394 | static const std::vector<std::string> fields = { |
| 1395 | "build_commit", "build_number", "cpu_info", "gpu_info", "backends", |
| 1396 | "model_filename", "model_type", "model_size", "model_n_params", "n_batch", |
| 1397 | "n_ubatch", "n_threads", "cpu_mask", "cpu_strict", "poll", |
| 1398 | "type_k", "type_v", "n_gpu_layers", "n_cpu_moe", "split_mode", |
| 1399 | "main_gpu", "no_kv_offload", "flash_attn", "devices", "tensor_split", |
| 1400 | "tensor_buft_overrides", "use_mmap", "embeddings", "no_op_offload", |
| 1401 | "no_host", "n_prompt", "n_gen", "n_depth", "test_time", |
| 1402 | "avg_ns", "stddev_ns", "avg_ts", "stddev_ts" |
| 1403 | }; |
| 1404 | return fields; |
| 1405 | } |
| 1406 | |
| 1407 | enum field_type { STRING, BOOL, INT, FLOAT }; |
| 1408 | |
| 1409 | static field_type get_field_type(const std::string & field) { |
| 1410 | if (field == "build_number"|| field == "n_batch"|| field == "n_ubatch"|| field == "n_threads"|| |
| 1411 | field == "poll"|| field == "model_size"|| field == "model_n_params"|| field == "n_gpu_layers"|| |
| 1412 | field == "main_gpu"|| field == "n_prompt"|| field == "n_gen"|| field == "n_depth"|| field == "avg_ns"|| |
| 1413 | field == "stddev_ns"|| field == "no_op_offload"|| field == "n_cpu_moe") { |
| 1414 | return INT; |
| 1415 | } |
| 1416 | if (field == "f16_kv"|| field == "no_kv_offload"|| field == "cpu_strict"|| field == "flash_attn"|| |
| 1417 | field == "use_mmap"|| field == "embeddings"|| field == "no_host") { |
| 1418 | return BOOL; |
| 1419 | } |
| 1420 | if (field == "avg_ts"|| field == "stddev_ts") { |
| 1421 | return FLOAT; |
| 1422 | } |
| 1423 | return STRING; |
| 1424 | } |
| 1425 | |
| 1426 | std::vector<std::string> get_values() const { |
| 1427 | std::string tensor_split_str; |
| 1428 | std::string tensor_buft_overrides_str; |
| 1429 | int max_nonzero = 0; |
| 1430 | for (size_t i = 0; i < llama_max_devices(); i++) { |
| 1431 | if (tensor_split[i] > 0) { |
| 1432 | max_nonzero = i; |
| 1433 | } |
| 1434 | } |
| 1435 | for (int i = 0; i <= max_nonzero; i++) { |
| 1436 | char buf[32]; |
| 1437 | snprintf(s: buf, maxlen: sizeof(buf), format: "%.2f", tensor_split[i]); |
| 1438 | tensor_split_str += buf; |
| 1439 | if (i < max_nonzero) { |
| 1440 | tensor_split_str += "/"; |
| 1441 | } |
| 1442 | } |
| 1443 | if (tensor_buft_overrides.size() == 1) { |
| 1444 | // Last element of tensor_buft_overrides is always a null pattern |
| 1445 | // so if it is only one element long, it must be a null pattern. |
| 1446 | GGML_ASSERT(tensor_buft_overrides[0].pattern == nullptr); |
| 1447 | tensor_buft_overrides_str += "none"; |
| 1448 | } else { |
| 1449 | for (size_t i = 0; i < tensor_buft_overrides.size()-1; i++) { |
| 1450 | // Last element of tensor_buft_overrides is always a null pattern |
| 1451 | if (tensor_buft_overrides[i].pattern == nullptr) { |
| 1452 | tensor_buft_overrides_str += "none"; |
| 1453 | } else { |
| 1454 | tensor_buft_overrides_str += tensor_buft_overrides[i].pattern; |
| 1455 | tensor_buft_overrides_str += "="; |
| 1456 | tensor_buft_overrides_str += ggml_backend_buft_name(buft: tensor_buft_overrides[i].buft); |
| 1457 | } |
| 1458 | if (i + 2 < tensor_buft_overrides.size()) { |
| 1459 | tensor_buft_overrides_str += ";"; |
| 1460 | } |
| 1461 | } |
| 1462 | } |
| 1463 | std::vector<std::string> values = { build_commit, |
| 1464 | std::to_string(val: build_number), |
| 1465 | cpu_info, |
| 1466 | gpu_info, |
| 1467 | get_backend(), |
| 1468 | model_filename, |
| 1469 | model_type, |
| 1470 | std::to_string(val: model_size), |
| 1471 | std::to_string(val: model_n_params), |
| 1472 | std::to_string(val: n_batch), |
| 1473 | std::to_string(val: n_ubatch), |
| 1474 | std::to_string(val: n_threads), |
| 1475 | cpu_mask, |
| 1476 | std::to_string(val: cpu_strict), |
| 1477 | std::to_string(val: poll), |
| 1478 | ggml_type_name(type: type_k), |
| 1479 | ggml_type_name(type: type_v), |
| 1480 | std::to_string(val: n_gpu_layers), |
| 1481 | std::to_string(val: n_cpu_moe), |
| 1482 | split_mode_str(mode: split_mode), |
| 1483 | std::to_string(val: main_gpu), |
| 1484 | std::to_string(val: no_kv_offload), |
| 1485 | std::to_string(val: flash_attn), |
| 1486 | devices_to_string(devices), |
| 1487 | tensor_split_str, |
| 1488 | tensor_buft_overrides_str, |
| 1489 | std::to_string(val: use_mmap), |
| 1490 | std::to_string(val: embeddings), |
| 1491 | std::to_string(val: no_op_offload), |
| 1492 | std::to_string(val: no_host), |
| 1493 | std::to_string(val: n_prompt), |
| 1494 | std::to_string(val: n_gen), |
| 1495 | std::to_string(val: n_depth), |
| 1496 | test_time, |
| 1497 | std::to_string(val: avg_ns()), |
| 1498 | std::to_string(val: stdev_ns()), |
| 1499 | std::to_string(val: avg_ts()), |
| 1500 | std::to_string(val: stdev_ts()) }; |
| 1501 | return values; |
| 1502 | } |
| 1503 | |
| 1504 | std::map<std::string, std::string> get_map() const { |
| 1505 | std::map<std::string, std::string> map; |
| 1506 | auto fields = get_fields(); |
| 1507 | auto values = get_values(); |
| 1508 | std::transform(first1: fields.begin(), last1: fields.end(), first2: values.begin(), result: std::inserter(x&: map, i: map.end()), |
| 1509 | binary_op: std::make_pair<const std::string &, const std::string &>); |
| 1510 | return map; |
| 1511 | } |
| 1512 | }; |
| 1513 | |
| 1514 | const std::string test::build_commit = LLAMA_COMMIT; |
| 1515 | const int test::build_number = LLAMA_BUILD_NUMBER; |
| 1516 | |
| 1517 | struct printer { |
| 1518 | virtual ~printer() {} |
| 1519 | |
| 1520 | FILE * fout; |
| 1521 | |
| 1522 | virtual void print_header(const cmd_params & params) { (void) params; } |
| 1523 | |
| 1524 | virtual void print_test(const test & t) = 0; |
| 1525 | |
| 1526 | virtual void print_footer() {} |
| 1527 | }; |
| 1528 | |
| 1529 | struct csv_printer : public printer { |
| 1530 | static std::string escape_csv(const std::string & field) { |
| 1531 | std::string escaped = "\""; |
| 1532 | for (auto c : field) { |
| 1533 | if (c == '"') { |
| 1534 | escaped += "\""; |
| 1535 | } |
| 1536 | escaped += c; |
| 1537 | } |
| 1538 | escaped += "\""; |
| 1539 | return escaped; |
| 1540 | } |
| 1541 | |
| 1542 | void print_header(const cmd_params & params) override { |
| 1543 | std::vector<std::string> fields = test::get_fields(); |
| 1544 | fprintf(stream: fout, format: "%s\n", join(values: fields, delim: ",").c_str()); |
| 1545 | (void) params; |
| 1546 | } |
| 1547 | |
| 1548 | void print_test(const test & t) override { |
| 1549 | std::vector<std::string> values = t.get_values(); |
| 1550 | std::transform(first: values.begin(), last: values.end(), result: values.begin(), unary_op: escape_csv); |
| 1551 | fprintf(stream: fout, format: "%s\n", join(values, delim: ",").c_str()); |
| 1552 | } |
| 1553 | }; |
| 1554 | |
| 1555 | static std::string escape_json(const std::string & value) { |
| 1556 | std::string escaped; |
| 1557 | for (auto c : value) { |
| 1558 | if (c == '"') { |
| 1559 | escaped += "\\\""; |
| 1560 | } else if (c == '\\') { |
| 1561 | escaped += "\\\\"; |
| 1562 | } else if (c <= 0x1f) { |
| 1563 | char buf[8]; |
| 1564 | snprintf(s: buf, maxlen: sizeof(buf), format: "\\u%04x", c); |
| 1565 | escaped += buf; |
| 1566 | } else { |
| 1567 | escaped += c; |
| 1568 | } |
| 1569 | } |
| 1570 | return escaped; |
| 1571 | } |
| 1572 | |
| 1573 | static std::string format_json_value(const std::string & field, const std::string & value) { |
| 1574 | switch (test::get_field_type(field)) { |
| 1575 | case test::STRING: |
| 1576 | return "\""+ escape_json(value) + "\""; |
| 1577 | case test::BOOL: |
| 1578 | return value == "0"? "false": "true"; |
| 1579 | default: |
| 1580 | return value; |
| 1581 | } |
| 1582 | } |
| 1583 | |
| 1584 | struct json_printer : public printer { |
| 1585 | bool first = true; |
| 1586 | |
| 1587 | void print_header(const cmd_params & params) override { |
| 1588 | fprintf(stream: fout, format: "[\n"); |
| 1589 | (void) params; |
| 1590 | } |
| 1591 | |
| 1592 | void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) { |
| 1593 | assert(fields.size() == values.size()); |
| 1594 | for (size_t i = 0; i < fields.size(); i++) { |
| 1595 | fprintf(stream: fout, format: " \"%s\": %s,\n", fields.at(n: i).c_str(), |
| 1596 | format_json_value(field: fields.at(n: i), value: values.at(n: i)).c_str()); |
| 1597 | } |
| 1598 | } |
| 1599 | |
| 1600 | void print_test(const test & t) override { |
| 1601 | if (first) { |
| 1602 | first = false; |
| 1603 | } else { |
| 1604 | fprintf(stream: fout, format: ",\n"); |
| 1605 | } |
| 1606 | fprintf(stream: fout, format: " {\n"); |
| 1607 | print_fields(fields: test::get_fields(), values: t.get_values()); |
| 1608 | fprintf(stream: fout, format: " \"samples_ns\": [ %s ],\n", join(values: t.samples_ns, delim: ", ").c_str()); |
| 1609 | fprintf(stream: fout, format: " \"samples_ts\": [ %s ]\n", join(values: t.get_ts(), delim: ", ").c_str()); |
| 1610 | fprintf(stream: fout, format: " }"); |
| 1611 | fflush(stream: fout); |
| 1612 | } |
| 1613 | |
| 1614 | void print_footer() override { fprintf(stream: fout, format: "\n]\n"); } |
| 1615 | }; |
| 1616 | |
| 1617 | struct jsonl_printer : public printer { |
| 1618 | void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) { |
| 1619 | assert(fields.size() == values.size()); |
| 1620 | for (size_t i = 0; i < fields.size(); i++) { |
| 1621 | fprintf(stream: fout, format: "\"%s\": %s, ", fields.at(n: i).c_str(), format_json_value(field: fields.at(n: i), value: values.at(n: i)).c_str()); |
| 1622 | } |
| 1623 | } |
| 1624 | |
| 1625 | void print_test(const test & t) override { |
| 1626 | fprintf(stream: fout, format: "{"); |
| 1627 | print_fields(fields: test::get_fields(), values: t.get_values()); |
| 1628 | fprintf(stream: fout, format: "\"samples_ns\": [ %s ],", join(values: t.samples_ns, delim: ", ").c_str()); |
| 1629 | fprintf(stream: fout, format: "\"samples_ts\": [ %s ]", join(values: t.get_ts(), delim: ", ").c_str()); |
| 1630 | fprintf(stream: fout, format: "}\n"); |
| 1631 | fflush(stream: fout); |
| 1632 | } |
| 1633 | }; |
| 1634 | |
| 1635 | struct markdown_printer : public printer { |
| 1636 | std::vector<std::string> fields; |
| 1637 | |
| 1638 | static int get_field_width(const std::string & field) { |
| 1639 | if (field == "model") { |
| 1640 | return -30; |
| 1641 | } |
| 1642 | if (field == "t/s") { |
| 1643 | return 20; |
| 1644 | } |
| 1645 | if (field == "size"|| field == "params") { |
| 1646 | return 10; |
| 1647 | } |
| 1648 | if (field == "n_gpu_layers") { |
| 1649 | return 3; |
| 1650 | } |
| 1651 | if (field == "n_threads") { |
| 1652 | return 7; |
| 1653 | } |
| 1654 | if (field == "n_batch") { |
| 1655 | return 7; |
| 1656 | } |
| 1657 | if (field == "n_ubatch") { |
| 1658 | return 8; |
| 1659 | } |
| 1660 | if (field == "type_k"|| field == "type_v") { |
| 1661 | return 6; |
| 1662 | } |
| 1663 | if (field == "split_mode") { |
| 1664 | return 5; |
| 1665 | } |
| 1666 | if (field == "flash_attn") { |
| 1667 | return 2; |
| 1668 | } |
| 1669 | if (field == "devices") { |
| 1670 | return -12; |
| 1671 | } |
| 1672 | if (field == "use_mmap") { |
| 1673 | return 4; |
| 1674 | } |
| 1675 | if (field == "test") { |
| 1676 | return 15; |
| 1677 | } |
| 1678 | if (field == "no_op_offload") { |
| 1679 | return 4; |
| 1680 | } |
| 1681 | if (field == "no_host") { |
| 1682 | return 4; |
| 1683 | } |
| 1684 | |
| 1685 | int width = std::max(a: (int) field.length(), b: 10); |
| 1686 | |
| 1687 | if (test::get_field_type(field) == test::STRING) { |
| 1688 | return -width; |
| 1689 | } |
| 1690 | return width; |
| 1691 | } |
| 1692 | |
| 1693 | static std::string get_field_display_name(const std::string & field) { |
| 1694 | if (field == "n_gpu_layers") { |
| 1695 | return "ngl"; |
| 1696 | } |
| 1697 | if (field == "split_mode") { |
| 1698 | return "sm"; |
| 1699 | } |
| 1700 | if (field == "n_threads") { |
| 1701 | return "threads"; |
| 1702 | } |
| 1703 | if (field == "no_kv_offload") { |
| 1704 | return "nkvo"; |
| 1705 | } |
| 1706 | if (field == "flash_attn") { |
| 1707 | return "fa"; |
| 1708 | } |
| 1709 | if (field == "use_mmap") { |
| 1710 | return "mmap"; |
| 1711 | } |
| 1712 | if (field == "embeddings") { |
| 1713 | return "embd"; |
| 1714 | } |
| 1715 | if (field == "no_op_offload") { |
| 1716 | return "nopo"; |
| 1717 | } |
| 1718 | if (field == "no_host") { |
| 1719 | return "noh"; |
| 1720 | } |
| 1721 | if (field == "devices") { |
| 1722 | return "dev"; |
| 1723 | } |
| 1724 | if (field == "tensor_split") { |
| 1725 | return "ts"; |
| 1726 | } |
| 1727 | if (field == "tensor_buft_overrides") { |
| 1728 | return "ot"; |
| 1729 | } |
| 1730 | return field; |
| 1731 | } |
| 1732 | |
| 1733 | void print_header(const cmd_params & params) override { |
| 1734 | // select fields to print |
| 1735 | fields.emplace_back(args: "model"); |
| 1736 | fields.emplace_back(args: "size"); |
| 1737 | fields.emplace_back(args: "params"); |
| 1738 | fields.emplace_back(args: "backend"); |
| 1739 | bool is_cpu_backend = test::get_backend().find(s: "CPU") != std::string::npos || |
| 1740 | test::get_backend().find(s: "BLAS") != std::string::npos; |
| 1741 | if (!is_cpu_backend) { |
| 1742 | fields.emplace_back(args: "n_gpu_layers"); |
| 1743 | } |
| 1744 | if (params.n_cpu_moe.size() > 1) { |
| 1745 | fields.emplace_back(args: "n_cpu_moe"); |
| 1746 | } |
| 1747 | if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) { |
| 1748 | fields.emplace_back(args: "n_threads"); |
| 1749 | } |
| 1750 | if (params.cpu_mask.size() > 1 || params.cpu_mask != cmd_params_defaults.cpu_mask) { |
| 1751 | fields.emplace_back(args: "cpu_mask"); |
| 1752 | } |
| 1753 | if (params.cpu_strict.size() > 1 || params.cpu_strict != cmd_params_defaults.cpu_strict) { |
| 1754 | fields.emplace_back(args: "cpu_strict"); |
| 1755 | } |
| 1756 | if (params.poll.size() > 1 || params.poll != cmd_params_defaults.poll) { |
| 1757 | fields.emplace_back(args: "poll"); |
| 1758 | } |
| 1759 | if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) { |
| 1760 | fields.emplace_back(args: "n_batch"); |
| 1761 | } |
| 1762 | if (params.n_ubatch.size() > 1 || params.n_ubatch != cmd_params_defaults.n_ubatch) { |
| 1763 | fields.emplace_back(args: "n_ubatch"); |
| 1764 | } |
| 1765 | if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) { |
| 1766 | fields.emplace_back(args: "type_k"); |
| 1767 | } |
| 1768 | if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) { |
| 1769 | fields.emplace_back(args: "type_v"); |
| 1770 | } |
| 1771 | if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) { |
| 1772 | fields.emplace_back(args: "main_gpu"); |
| 1773 | } |
| 1774 | if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) { |
| 1775 | fields.emplace_back(args: "split_mode"); |
| 1776 | } |
| 1777 | if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) { |
| 1778 | fields.emplace_back(args: "no_kv_offload"); |
| 1779 | } |
| 1780 | if (params.flash_attn.size() > 1 || params.flash_attn != cmd_params_defaults.flash_attn) { |
| 1781 | fields.emplace_back(args: "flash_attn"); |
| 1782 | } |
| 1783 | if (params.devices.size() > 1 || params.devices != cmd_params_defaults.devices) { |
| 1784 | fields.emplace_back(args: "devices"); |
| 1785 | } |
| 1786 | if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) { |
| 1787 | fields.emplace_back(args: "tensor_split"); |
| 1788 | } |
| 1789 | if (params.tensor_buft_overrides.size() > 1 || !vec_vec_tensor_buft_override_equal(a: params.tensor_buft_overrides, b: cmd_params_defaults.tensor_buft_overrides)) { |
| 1790 | fields.emplace_back(args: "tensor_buft_overrides"); |
| 1791 | } |
| 1792 | if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) { |
| 1793 | fields.emplace_back(args: "use_mmap"); |
| 1794 | } |
| 1795 | if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) { |
| 1796 | fields.emplace_back(args: "embeddings"); |
| 1797 | } |
| 1798 | if (params.no_op_offload.size() > 1 || params.no_op_offload != cmd_params_defaults.no_op_offload) { |
| 1799 | fields.emplace_back(args: "no_op_offload"); |
| 1800 | } |
| 1801 | if (params.no_host.size() > 1 || params.no_host != cmd_params_defaults.no_host) { |
| 1802 | fields.emplace_back(args: "no_host"); |
| 1803 | } |
| 1804 | fields.emplace_back(args: "test"); |
| 1805 | fields.emplace_back(args: "t/s"); |
| 1806 | |
| 1807 | fprintf(stream: fout, format: "|"); |
| 1808 | for (const auto & field : fields) { |
| 1809 | fprintf(stream: fout, format: " %*s |", get_field_width(field), get_field_display_name(field).c_str()); |
| 1810 | } |
| 1811 | fprintf(stream: fout, format: "\n"); |
| 1812 | fprintf(stream: fout, format: "|"); |
| 1813 | for (const auto & field : fields) { |
| 1814 | int width = get_field_width(field); |
| 1815 | fprintf(stream: fout, format: " %s%s |", std::string(std::abs(x: width) - 1, '-').c_str(), width > 0 ? ":": "-"); |
| 1816 | } |
| 1817 | fprintf(stream: fout, format: "\n"); |
| 1818 | } |
| 1819 | |
| 1820 | void print_test(const test & t) override { |
| 1821 | std::map<std::string, std::string> vmap = t.get_map(); |
| 1822 | |
| 1823 | fprintf(stream: fout, format: "|"); |
| 1824 | for (const auto & field : fields) { |
| 1825 | std::string value; |
| 1826 | char buf[128]; |
| 1827 | if (field == "model") { |
| 1828 | value = t.model_type; |
| 1829 | } else if (field == "size") { |
| 1830 | if (t.model_size < 1024 * 1024 * 1024) { |
| 1831 | snprintf(s: buf, maxlen: sizeof(buf), format: "%.2f MiB", t.model_size / 1024.0 / 1024.0); |
| 1832 | } else { |
| 1833 | snprintf(s: buf, maxlen: sizeof(buf), format: "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0); |
| 1834 | } |
| 1835 | value = buf; |
| 1836 | } else if (field == "params") { |
| 1837 | if (t.model_n_params < 1000 * 1000 * 1000) { |
| 1838 | snprintf(s: buf, maxlen: sizeof(buf), format: "%.2f M", t.model_n_params / 1e6); |
| 1839 | } else { |
| 1840 | snprintf(s: buf, maxlen: sizeof(buf), format: "%.2f B", t.model_n_params / 1e9); |
| 1841 | } |
| 1842 | value = buf; |
| 1843 | } else if (field == "backend") { |
| 1844 | value = test::get_backend(); |
| 1845 | } else if (field == "test") { |
| 1846 | if (t.n_prompt > 0 && t.n_gen == 0) { |
| 1847 | snprintf(s: buf, maxlen: sizeof(buf), format: "pp%d", t.n_prompt); |
| 1848 | } else if (t.n_gen > 0 && t.n_prompt == 0) { |
| 1849 | snprintf(s: buf, maxlen: sizeof(buf), format: "tg%d", t.n_gen); |
| 1850 | } else { |
| 1851 | snprintf(s: buf, maxlen: sizeof(buf), format: "pp%d+tg%d", t.n_prompt, t.n_gen); |
| 1852 | } |
| 1853 | if (t.n_depth > 0) { |
| 1854 | int len = strlen(s: buf); |
| 1855 | snprintf(s: buf + len, maxlen: sizeof(buf) - len, format: " @ d%d", t.n_depth); |
| 1856 | } |
| 1857 | value = buf; |
| 1858 | } else if (field == "t/s") { |
| 1859 | snprintf(s: buf, maxlen: sizeof(buf), format: "%.2f ± %.2f", t.avg_ts(), t.stdev_ts()); |
| 1860 | value = buf; |
| 1861 | } else if (vmap.find(x: field) != vmap.end()) { |
| 1862 | value = vmap.at(k: field); |
| 1863 | } else { |
| 1864 | assert(false); |
| 1865 | exit(status: 1); |
| 1866 | } |
| 1867 | |
| 1868 | int width = get_field_width(field); |
| 1869 | if (field == "t/s") { |
| 1870 | // HACK: the utf-8 character is 2 bytes |
| 1871 | width += 1; |
| 1872 | } |
| 1873 | fprintf(stream: fout, format: " %*s |", width, value.c_str()); |
| 1874 | } |
| 1875 | fprintf(stream: fout, format: "\n"); |
| 1876 | } |
| 1877 | |
| 1878 | void print_footer() override { |
| 1879 | fprintf(stream: fout, format: "\nbuild: %s (%d)\n", test::build_commit.c_str(), test::build_number); |
| 1880 | } |
| 1881 | }; |
| 1882 | |
| 1883 | struct sql_printer : public printer { |
| 1884 | static std::string get_sql_field_type(const std::string & field) { |
| 1885 | switch (test::get_field_type(field)) { |
| 1886 | case test::STRING: |
| 1887 | return "TEXT"; |
| 1888 | case test::BOOL: |
| 1889 | case test::INT: |
| 1890 | return "INTEGER"; |
| 1891 | case test::FLOAT: |
| 1892 | return "REAL"; |
| 1893 | default: |
| 1894 | assert(false); |
| 1895 | exit(status: 1); |
| 1896 | } |
| 1897 | } |
| 1898 | |
| 1899 | void print_header(const cmd_params & params) override { |
| 1900 | std::vector<std::string> fields = test::get_fields(); |
| 1901 | fprintf(stream: fout, format: "CREATE TABLE IF NOT EXISTS llama_bench (\n"); |
| 1902 | for (size_t i = 0; i < fields.size(); i++) { |
| 1903 | fprintf(stream: fout, format: " %s %s%s\n", fields.at(n: i).c_str(), get_sql_field_type(field: fields.at(n: i)).c_str(), |
| 1904 | i < fields.size() - 1 ? ",": ""); |
| 1905 | } |
| 1906 | fprintf(stream: fout, format: ");\n"); |
| 1907 | fprintf(stream: fout, format: "\n"); |
| 1908 | (void) params; |
| 1909 | } |
| 1910 | |
| 1911 | void print_test(const test & t) override { |
| 1912 | fprintf(stream: fout, format: "INSERT INTO llama_bench (%s) ", join(values: test::get_fields(), delim: ", ").c_str()); |
| 1913 | fprintf(stream: fout, format: "VALUES ("); |
| 1914 | std::vector<std::string> values = t.get_values(); |
| 1915 | for (size_t i = 0; i < values.size(); i++) { |
| 1916 | fprintf(stream: fout, format: "'%s'%s", values.at(n: i).c_str(), i < values.size() - 1 ? ", ": ""); |
| 1917 | } |
| 1918 | fprintf(stream: fout, format: ");\n"); |
| 1919 | } |
| 1920 | }; |
| 1921 | |
| 1922 | struct ctx_state { |
| 1923 | int depth = 0; // in tokens |
| 1924 | |
| 1925 | std::vector<uint8_t> buf; // the llama_context state buffer |
| 1926 | }; |
| 1927 | |
| 1928 | static bool test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) { |
| 1929 | llama_set_n_threads(ctx, n_threads, n_threads_batch: n_threads); |
| 1930 | |
| 1931 | const llama_model * model = llama_get_model(ctx); |
| 1932 | const llama_vocab * vocab = llama_model_get_vocab(model); |
| 1933 | const int32_t n_vocab = llama_vocab_n_tokens(vocab); |
| 1934 | |
| 1935 | std::vector<llama_token> tokens(n_batch); |
| 1936 | |
| 1937 | int n_processed = 0; |
| 1938 | |
| 1939 | while (n_processed < n_prompt) { |
| 1940 | int n_tokens = std::min(a: n_prompt - n_processed, b: n_batch); |
| 1941 | tokens[0] = n_processed == 0 && llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab; |
| 1942 | for (int i = 1; i < n_tokens; i++) { |
| 1943 | tokens[i] = std::rand() % n_vocab; |
| 1944 | } |
| 1945 | int res = llama_decode(ctx, batch: llama_batch_get_one(tokens: tokens.data(), n_tokens)); |
| 1946 | if (res != 0) { |
| 1947 | fprintf(stderr, format: "%s: failed to decode prompt batch, res = %d\n", __func__, res); |
| 1948 | return false; |
| 1949 | } |
| 1950 | n_processed += n_tokens; |
| 1951 | } |
| 1952 | |
| 1953 | llama_synchronize(ctx); |
| 1954 | return true; |
| 1955 | } |
| 1956 | |
| 1957 | static bool test_gen(llama_context * ctx, int n_gen, int n_threads) { |
| 1958 | llama_set_n_threads(ctx, n_threads, n_threads_batch: n_threads); |
| 1959 | |
| 1960 | const llama_model * model = llama_get_model(ctx); |
| 1961 | const llama_vocab * vocab = llama_model_get_vocab(model); |
| 1962 | const int32_t n_vocab = llama_vocab_n_tokens(vocab); |
| 1963 | |
| 1964 | llama_token token = llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab; |
| 1965 | |
| 1966 | for (int i = 0; i < n_gen; i++) { |
| 1967 | int res = llama_decode(ctx, batch: llama_batch_get_one(tokens: &token, n_tokens: 1)); |
| 1968 | if (res != 0) { |
| 1969 | fprintf(stderr, format: "%s: failed to decode generation batch, res = %d\n", __func__, res); |
| 1970 | return false; |
| 1971 | } |
| 1972 | llama_synchronize(ctx); |
| 1973 | token = std::rand() % n_vocab; |
| 1974 | } |
| 1975 | return true; |
| 1976 | } |
| 1977 | |
| 1978 | static void llama_null_log_callback(enum ggml_log_level level, const char * text, void * user_data) { |
| 1979 | (void) level; |
| 1980 | (void) text; |
| 1981 | (void) user_data; |
| 1982 | } |
| 1983 | |
| 1984 | static std::unique_ptr<printer> create_printer(output_formats format) { |
| 1985 | switch (format) { |
| 1986 | case NONE: |
| 1987 | return nullptr; |
| 1988 | case CSV: |
| 1989 | return std::unique_ptr<printer>(new csv_printer()); |
| 1990 | case JSON: |
| 1991 | return std::unique_ptr<printer>(new json_printer()); |
| 1992 | case JSONL: |
| 1993 | return std::unique_ptr<printer>(new jsonl_printer()); |
| 1994 | case MARKDOWN: |
| 1995 | return std::unique_ptr<printer>(new markdown_printer()); |
| 1996 | case SQL: |
| 1997 | return std::unique_ptr<printer>(new sql_printer()); |
| 1998 | } |
| 1999 | GGML_ABORT("fatal error"); |
| 2000 | } |
| 2001 | |
| 2002 | int main(int argc, char ** argv) { |
| 2003 | // try to set locale for unicode characters in markdown |
| 2004 | setlocale(LC_CTYPE, locale: ".UTF-8"); |
| 2005 | |
| 2006 | #if !defined(NDEBUG) |
| 2007 | fprintf(stderr, "warning: asserts enabled, performance may be affected\n"); |
| 2008 | #endif |
| 2009 | |
| 2010 | #if (defined(_MSC_VER) && defined(_DEBUG)) || (!defined(_MSC_VER) && !defined(__OPTIMIZE__)) |
| 2011 | fprintf(stderr, "warning: debug build, performance may be affected\n"); |
| 2012 | #endif |
| 2013 | |
| 2014 | #if defined(__SANITIZE_ADDRESS__) || defined(__SANITIZE_THREAD__) |
| 2015 | fprintf(stderr, "warning: sanitizer enabled, performance may be affected\n"); |
| 2016 | #endif |
| 2017 | |
| 2018 | // initialize backends |
| 2019 | ggml_backend_load_all(); |
| 2020 | |
| 2021 | cmd_params params = parse_cmd_params(argc, argv); |
| 2022 | |
| 2023 | auto * cpu_dev = ggml_backend_dev_by_type(type: GGML_BACKEND_DEVICE_TYPE_CPU); |
| 2024 | if (!cpu_dev) { |
| 2025 | fprintf(stderr, format: "%s: error: CPU backend is not loaded\n", __func__); |
| 2026 | return 1; |
| 2027 | } |
| 2028 | auto * cpu_reg = ggml_backend_dev_backend_reg(device: cpu_dev); |
| 2029 | auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg: cpu_reg, name: "ggml_threadpool_new"); |
| 2030 | auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg: cpu_reg, name: "ggml_threadpool_free"); |
| 2031 | |
| 2032 | // initialize llama.cpp |
| 2033 | if (!params.verbose) { |
| 2034 | llama_log_set(log_callback: llama_null_log_callback, NULL); |
| 2035 | } |
| 2036 | llama_backend_init(); |
| 2037 | llama_numa_init(numa: params.numa); |
| 2038 | |
| 2039 | set_process_priority(params.prio); |
| 2040 | |
| 2041 | // initialize printer |
| 2042 | std::unique_ptr<printer> p = create_printer(format: params.output_format); |
| 2043 | std::unique_ptr<printer> p_err = create_printer(format: params.output_format_stderr); |
| 2044 | |
| 2045 | if (p) { |
| 2046 | p->fout = stdout; |
| 2047 | p->print_header(params); |
| 2048 | } |
| 2049 | |
| 2050 | if (p_err) { |
| 2051 | p_err->fout = stderr; |
| 2052 | p_err->print_header(params); |
| 2053 | } |
| 2054 | |
| 2055 | std::vector<cmd_params_instance> params_instances = get_cmd_params_instances(params); |
| 2056 | |
| 2057 | llama_model * lmodel = nullptr; |
| 2058 | const cmd_params_instance * prev_inst = nullptr; |
| 2059 | |
| 2060 | // store the llama_context state at the previous depth that we performed a test |
| 2061 | // ref: https://github.com/ggml-org/llama.cpp/pull/16944#issuecomment-3478151721 |
| 2062 | ctx_state cstate; |
| 2063 | |
| 2064 | int params_idx = 0; |
| 2065 | auto params_count = params_instances.size(); |
| 2066 | for (const auto & inst : params_instances) { |
| 2067 | params_idx++; |
| 2068 | if (params.progress) { |
| 2069 | fprintf(stderr, format: "llama-bench: benchmark %d/%zu: starting\n", params_idx, params_count); |
| 2070 | } |
| 2071 | // keep the same model between tests when possible |
| 2072 | if (!lmodel || !prev_inst || !inst.equal_mparams(other: *prev_inst)) { |
| 2073 | if (lmodel) { |
| 2074 | llama_model_free(model: lmodel); |
| 2075 | } |
| 2076 | |
| 2077 | lmodel = llama_model_load_from_file(path_model: inst.model.c_str(), params: inst.to_llama_mparams()); |
| 2078 | if (lmodel == NULL) { |
| 2079 | fprintf(stderr, format: "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str()); |
| 2080 | return 1; |
| 2081 | } |
| 2082 | prev_inst = &inst; |
| 2083 | } |
| 2084 | |
| 2085 | llama_context * ctx = llama_init_from_model(model: lmodel, params: inst.to_llama_cparams()); |
| 2086 | if (ctx == NULL) { |
| 2087 | fprintf(stderr, format: "%s: error: failed to create context with model '%s'\n", __func__, inst.model.c_str()); |
| 2088 | llama_model_free(model: lmodel); |
| 2089 | return 1; |
| 2090 | } |
| 2091 | |
| 2092 | test t(inst, lmodel, ctx); |
| 2093 | |
| 2094 | llama_memory_clear(mem: llama_get_memory(ctx), data: false); |
| 2095 | |
| 2096 | // cool off before the test |
| 2097 | if (params.delay) { |
| 2098 | std::this_thread::sleep_for(rtime: std::chrono::seconds(params.delay)); |
| 2099 | } |
| 2100 | |
| 2101 | struct ggml_threadpool_params tpp = ggml_threadpool_params_default(n_threads: t.n_threads); |
| 2102 | if (!parse_cpu_mask(mask: t.cpu_mask, boolmask&: tpp.cpumask)) { |
| 2103 | fprintf(stderr, format: "%s: failed to parse cpu-mask: %s\n", __func__, t.cpu_mask.c_str()); |
| 2104 | exit(status: 1); |
| 2105 | } |
| 2106 | tpp.strict_cpu = t.cpu_strict; |
| 2107 | tpp.poll = t.poll; |
| 2108 | tpp.prio = params.prio; |
| 2109 | |
| 2110 | struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp); |
| 2111 | if (!threadpool) { |
| 2112 | fprintf(stderr, format: "%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads); |
| 2113 | exit(status: 1); |
| 2114 | } |
| 2115 | |
| 2116 | llama_attach_threadpool(ctx, threadpool, NULL); |
| 2117 | |
| 2118 | // warmup run |
| 2119 | if (!params.no_warmup) { |
| 2120 | if (t.n_prompt > 0) { |
| 2121 | if (params.progress) { |
| 2122 | fprintf(stderr, format: "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count); |
| 2123 | } |
| 2124 | //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads); |
| 2125 | bool res = test_prompt(ctx, n_prompt: t.n_prompt, n_batch: t.n_batch, n_threads: t.n_threads); |
| 2126 | if (!res) { |
| 2127 | fprintf(stderr, format: "%s: error: failed to run prompt warmup\n", __func__); |
| 2128 | exit(status: 1); |
| 2129 | } |
| 2130 | } |
| 2131 | if (t.n_gen > 0) { |
| 2132 | if (params.progress) { |
| 2133 | fprintf(stderr, format: "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count); |
| 2134 | } |
| 2135 | bool res = test_gen(ctx, n_gen: 1, n_threads: t.n_threads); |
| 2136 | if (!res) { |
| 2137 | fprintf(stderr, format: "%s: error: failed to run gen warmup\n", __func__); |
| 2138 | exit(status: 1); |
| 2139 | } |
| 2140 | } |
| 2141 | } |
| 2142 | |
| 2143 | for (int i = 0; i < params.reps; i++) { |
| 2144 | llama_memory_clear(mem: llama_get_memory(ctx), data: false); |
| 2145 | |
| 2146 | if (t.n_depth > 0) { |
| 2147 | bool is_cached = t.n_depth == cstate.depth; |
| 2148 | |
| 2149 | if (is_cached) { |
| 2150 | // if previously we have computed at this depth, just restore the state |
| 2151 | const size_t ret = llama_state_seq_set_data(ctx, src: cstate.buf.data(), size: cstate.buf.size(), dest_seq_id: 0); |
| 2152 | if (ret == 0) { |
| 2153 | // if the old state is incompatible with the current context - reprocess from scratch |
| 2154 | is_cached = false; |
| 2155 | } |
| 2156 | } |
| 2157 | |
| 2158 | if (!is_cached) { |
| 2159 | if (params.progress) { |
| 2160 | fprintf(stderr, format: "llama-bench: benchmark %d/%zu: depth run %d/%d\n", params_idx, params_count, |
| 2161 | i + 1, params.reps); |
| 2162 | } |
| 2163 | bool res = test_prompt(ctx, n_prompt: t.n_depth, n_batch: t.n_batch, n_threads: t.n_threads); |
| 2164 | if (!res) { |
| 2165 | fprintf(stderr, format: "%s: error: failed to run depth\n", __func__); |
| 2166 | exit(status: 1); |
| 2167 | } |
| 2168 | |
| 2169 | // store the context state for reuse in later runs |
| 2170 | cstate.depth = t.n_depth; |
| 2171 | cstate.buf.resize(new_size: llama_state_seq_get_size(ctx, seq_id: 0)); |
| 2172 | llama_state_seq_get_data(ctx, dst: cstate.buf.data(), size: cstate.buf.size(), seq_id: 0); |
| 2173 | } else { |
| 2174 | if (params.progress) { |
| 2175 | fprintf(stderr, format: "llama-bench: benchmark %d/%zu: depth run %d/%d (cached)\n", params_idx, params_count, |
| 2176 | i + 1, params.reps); |
| 2177 | } |
| 2178 | } |
| 2179 | } |
| 2180 | |
| 2181 | uint64_t t_start = get_time_ns(); |
| 2182 | |
| 2183 | if (t.n_prompt > 0) { |
| 2184 | if (params.progress) { |
| 2185 | fprintf(stderr, format: "llama-bench: benchmark %d/%zu: prompt run %d/%d\n", params_idx, params_count, |
| 2186 | i + 1, params.reps); |
| 2187 | } |
| 2188 | bool res = test_prompt(ctx, n_prompt: t.n_prompt, n_batch: t.n_batch, n_threads: t.n_threads); |
| 2189 | if (!res) { |
| 2190 | fprintf(stderr, format: "%s: error: failed to run prompt\n", __func__); |
| 2191 | exit(status: 1); |
| 2192 | } |
| 2193 | } |
| 2194 | if (t.n_gen > 0) { |
| 2195 | if (params.progress) { |
| 2196 | fprintf(stderr, format: "llama-bench: benchmark %d/%zu: generation run %d/%d\n", params_idx, params_count, |
| 2197 | i + 1, params.reps); |
| 2198 | } |
| 2199 | bool res = test_gen(ctx, n_gen: t.n_gen, n_threads: t.n_threads); |
| 2200 | if (!res) { |
| 2201 | fprintf(stderr, format: "%s: error: failed to run gen\n", __func__); |
| 2202 | exit(status: 1); |
| 2203 | } |
| 2204 | } |
| 2205 | |
| 2206 | uint64_t t_ns = get_time_ns() - t_start; |
| 2207 | t.samples_ns.push_back(x: t_ns); |
| 2208 | } |
| 2209 | |
| 2210 | if (p) { |
| 2211 | p->print_test(t); |
| 2212 | fflush(stream: p->fout); |
| 2213 | } |
| 2214 | |
| 2215 | if (p_err) { |
| 2216 | p_err->print_test(t); |
| 2217 | fflush(stream: p_err->fout); |
| 2218 | } |
| 2219 | |
| 2220 | llama_perf_context_print(ctx); |
| 2221 | |
| 2222 | llama_free(ctx); |
| 2223 | |
| 2224 | ggml_threadpool_free_fn(threadpool); |
| 2225 | } |
| 2226 | |
| 2227 | llama_model_free(model: lmodel); |
| 2228 | |
| 2229 | if (p) { |
| 2230 | p->print_footer(); |
| 2231 | } |
| 2232 | |
| 2233 | if (p_err) { |
| 2234 | p_err->print_footer(); |
| 2235 | } |
| 2236 | |
| 2237 | llama_backend_free(); |
| 2238 | |
| 2239 | return 0; |
| 2240 | } |
| 2241 |
Generated on 2025-Nov-08 from project llama.cpp revision b6989
Powered by 
Code Browser 2.1
Generator usage only permitted with license.