Brunobkr/llama.cpp_AlgMor24_github

1#include "server-context.h"2#include "server-http.h"3#include "server-models.h"4#include "server-cors-proxy.h"5#include "server-stream.h"6#include "server-tools.h"7 8#include "arg.h"9#include "build-info.h"10#include "common.h"11#include "fit.h"12#include "llama.h"13#include "log.h"14 15#include <atomic>16#include <clocale>17#include <exception>18#include <signal.h>19#include <thread> // for std::thread::hardware_concurrency20 21#if defined(_WIN32)22#include <windows.h>23#endif24 25static std::function<void(int)> shutdown_handler;26static std::atomic_flag is_terminating = ATOMIC_FLAG_INIT;27 28static inline void signal_handler(int signal) {29    if (is_terminating.test_and_set()) {30        // in case it hangs, we can force terminate the server by hitting Ctrl+C twice31        // this is for better developer experience, we can remove when the server is stable enough32        fprintf(stderr, "Received second interrupt, terminating immediately.\n");33        exit(1);34    }35 36    shutdown_handler(signal);37}38 39// satisfies -Wmissing-declarations (used by llama command)40int llama_server(int argc, char ** argv);41 42// to be used via CLI (argc / argv are used by router mode only)43int llama_server(common_params & params, int argc, char ** argv);44void llama_server_terminate();45void llama_server_terminate() {46    if (shutdown_handler) {47        shutdown_handler(0);48    }49}50 51 52// wrapper function that handles exceptions and logs errors53// this is to make sure handler_t never throws exceptions; instead, it returns an error response54static server_http_context::handler_t ex_wrapper(server_http_context::handler_t func) {55    return [func = std::move(func)](const server_http_req & req) -> server_http_res_ptr {56        std::string message;57        error_type error;58        try {59            return func(req);60        } catch (const std::invalid_argument & e) {61            // treat invalid_argument as invalid request (400)62            error = ERROR_TYPE_INVALID_REQUEST;63            message = e.what();64        } catch (const std::exception & e) {65            // treat other exceptions as server error (500)66            error = ERROR_TYPE_SERVER;67            message = e.what();68        } catch (...) {69            error = ERROR_TYPE_SERVER;70            message = "unknown error";71        }72 73        auto res = std::make_unique<server_http_res>();74        res->status = 500;75        try {76            json error_data = format_error_response(message, error);77            res->status = json_value(error_data, "code", 500);78            res->data = safe_json_to_str({{ "error", error_data }});79            SRV_WRN("got exception: %s\n", res->data.c_str());80        } catch (const std::exception & e) {81            SRV_ERR("got another exception: %s | while handling exception: %s\n", e.what(), message.c_str());82            res->data = "Internal Server Error";83        }84        return res;85    };86}87 88int llama_server(int argc, char ** argv) {89    std::setlocale(LC_NUMERIC, "C");90 91#ifndef _WIN3292    // Ignore SIGPIPE so the server does not crash if a child (MCP server, tools runtime) exits while we are writing to its stdin93    signal(SIGPIPE, SIG_IGN);94#endif95 96    // own arguments required by this example97    common_params params;98 99    common_init();100 101    // start the stream session manager GC right after common init, before any HTTP route can102    // touch it. lifecycle is symmetric, stop_gc() runs in clean_up() before backend free103    server_stream_session_manager_start();104 105    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SERVER)) {106        return 1;107    }108 109    llama_backend_init();110    llama_numa_init(params.numa);111 112    return llama_server(params, argc, argv);113}114 115int llama_server(common_params & params, int argc, char ** argv) {116    bool is_run_by_cli = (argv == nullptr);117 118    common_models_handler models_handler;119 120    // note: router mode also accepts -hf remote-preset, so we need to check that first121    if (!is_run_by_cli && !params.model.hf_repo.empty()) {122        try {123            models_handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);124            if (common_models_handler_is_preset_repo(models_handler)) {125                // apply the preset and start the server in router mode126                common_models_handler_apply(models_handler, params);127            }128        } catch (const std::exception & e) {129            SRV_ERR("failed to fetch model metadata: %s\n", e.what());130            return 1;131        }132    }133 134    // router server never loads a model and must not touch the GPU135    const bool is_router_server = params.model.path.empty()136                               && params.model.hf_repo.empty();137 138    // skip device enumeration so the CUDA primary context stays uncreated139    common_params_print_info(params, !is_router_server);140 141    if (!is_router_server) {142        // validate batch size for embeddings143        // embeddings require all tokens to be processed in a single ubatch144        // see https://github.com/ggml-org/llama.cpp/issues/12836145        if (params.embedding && params.n_batch > params.n_ubatch) {146            SRV_WRN("embeddings enabled with n_batch (%d) > n_ubatch (%d)\n", params.n_batch, params.n_ubatch);147            SRV_WRN("setting n_batch = n_ubatch = %d to avoid assertion failure\n", params.n_ubatch);148            params.n_batch = params.n_ubatch;149        }150 151        if (params.n_parallel < 0) {152            SRV_TRC("%s", "n_parallel is set to auto, using n_parallel = 4 and kv_unified = true\n");153 154            params.n_parallel = 4;155            params.kv_unified = true;156        }157    }158 159    // for consistency between server router mode and single-model mode, we set the same model name as alias160    auto model_name = params.model.get_name();161    if (params.model_alias.empty() && !model_name.empty()) {162        params.model_alias.insert(model_name);163    }164 165    // note: this is guaranteed to out-live ctx_http and tools166    server_mcp mcp_mgr;167 168    // struct that contains llama context and inference169    server_context ctx_server;170 171    server_http_context ctx_http;172    if (!ctx_http.init(params)) {173        SRV_ERR("%s", "failed to initialize HTTP server\n");174        return 1;175    }176 177    //178    // Router179    //180 181    // register API routes182    server_child child; // only used in non-router mode183    server_routes routes(params, ctx_server);184    server_tools tools;185 186    std::optional<server_models_routes> models_routes{};187    if (is_router_server) {188        // setup server instances manager189        try {190            models_routes.emplace(params, argc, argv);191        } catch (const std::exception & e) {192            SRV_ERR("failed to initialize router models: %s\n", e.what());193            return 1;194        }195 196        // proxy handlers197        // note: routes.get_health stays the same198        routes.get_metrics                 = models_routes->proxy_get;199        routes.post_props                  = models_routes->proxy_post;200        routes.post_completions            = models_routes->proxy_post;201        routes.post_completions_oai        = models_routes->proxy_post;202        routes.post_chat_completions       = models_routes->proxy_post;203        routes.post_control                = models_routes->proxy_post;204        routes.post_responses_oai          = models_routes->proxy_post;205        routes.post_transcriptions_oai     = models_routes->proxy_post;206        routes.post_anthropic_messages     = models_routes->proxy_post;207        routes.post_anthropic_count_tokens = models_routes->proxy_post;208        routes.post_infill                 = models_routes->proxy_post;209        routes.post_embeddings             = models_routes->proxy_post;210        routes.post_embeddings_oai         = models_routes->proxy_post;211        routes.post_rerank                 = models_routes->proxy_post;212        routes.post_tokenize               = models_routes->proxy_post;213        routes.post_detokenize             = models_routes->proxy_post;214        routes.post_apply_template         = models_routes->proxy_post;215        routes.post_chat_completions_tok   = models_routes->proxy_post;216        routes.post_responses_tok_oai      = models_routes->proxy_post;217        routes.get_lora_adapters           = models_routes->proxy_get;218        routes.post_lora_adapters          = models_routes->proxy_post;219        routes.get_slots                   = models_routes->proxy_get;220        routes.post_slots                  = models_routes->proxy_post;221 222        // custom routes for router223        routes.get_props                   = models_routes->get_router_props;224        routes.get_models                  = models_routes->get_router_models;225 226        ctx_http.post("/models",               ex_wrapper(models_routes->post_router_models));227        ctx_http.post("/models/load",          ex_wrapper(models_routes->post_router_models_load));228        ctx_http.post("/models/unload",        ex_wrapper(models_routes->post_router_models_unload));229        ctx_http.get ("/models/sse",           ex_wrapper(models_routes->get_router_models_sse));230        ctx_http.del ("/models",               ex_wrapper(models_routes->del_router_models));231    }232 233    ctx_http.get ("/health",                   ex_wrapper(routes.get_health)); // public endpoint (no API key check)234    ctx_http.get ("/v1/health",                ex_wrapper(routes.get_health)); // public endpoint (no API key check)235    ctx_http.get ("/metrics",                  ex_wrapper(routes.get_metrics));236    ctx_http.get ("/props",                    ex_wrapper(routes.get_props));237    ctx_http.post("/props",                    ex_wrapper(routes.post_props));238    ctx_http.get ("/models",                   ex_wrapper(routes.get_models)); // public endpoint (no API key check)239    ctx_http.get ("/v1/models",                ex_wrapper(routes.get_models)); // public endpoint (no API key check)240    ctx_http.post("/completion",               ex_wrapper(routes.post_completions)); // legacy241    ctx_http.post("/completions",              ex_wrapper(routes.post_completions));242    ctx_http.post("/v1/completions",           ex_wrapper(routes.post_completions_oai));243    ctx_http.post("/chat/completions",         ex_wrapper(routes.post_chat_completions));244    ctx_http.post("/v1/chat/completions",      ex_wrapper(routes.post_chat_completions));245    ctx_http.post("/v1/chat/completions/control", ex_wrapper(routes.post_control));246    ctx_http.post("/v1/responses",             ex_wrapper(routes.post_responses_oai));247    ctx_http.post("/responses",                ex_wrapper(routes.post_responses_oai));248    ctx_http.post("/v1/audio/transcriptions",  ex_wrapper(routes.post_transcriptions_oai));249    ctx_http.post("/audio/transcriptions",     ex_wrapper(routes.post_transcriptions_oai));250    ctx_http.post("/v1/messages",              ex_wrapper(routes.post_anthropic_messages)); // anthropic messages API251    ctx_http.post("/infill",                   ex_wrapper(routes.post_infill));252    ctx_http.post("/embedding",                ex_wrapper(routes.post_embeddings)); // legacy253    ctx_http.post("/embeddings",               ex_wrapper(routes.post_embeddings));254    ctx_http.post("/v1/embeddings",            ex_wrapper(routes.post_embeddings_oai));255    ctx_http.post("/rerank",                   ex_wrapper(routes.post_rerank));256    ctx_http.post("/reranking",                ex_wrapper(routes.post_rerank));257    ctx_http.post("/v1/rerank",                ex_wrapper(routes.post_rerank));258    ctx_http.post("/v1/reranking",             ex_wrapper(routes.post_rerank));259    ctx_http.post("/tokenize",                 ex_wrapper(routes.post_tokenize));260    ctx_http.post("/detokenize",               ex_wrapper(routes.post_detokenize));261    ctx_http.post("/apply-template",           ex_wrapper(routes.post_apply_template));262    // token counting263    ctx_http.post("/chat/completions/input_tokens",    ex_wrapper(routes.post_chat_completions_tok));264    ctx_http.post("/v1/chat/completions/input_tokens", ex_wrapper(routes.post_chat_completions_tok));265    ctx_http.post("/responses/input_tokens",           ex_wrapper(routes.post_responses_tok_oai));266    ctx_http.post("/v1/responses/input_tokens",        ex_wrapper(routes.post_responses_tok_oai));267    ctx_http.post("/v1/messages/count_tokens",         ex_wrapper(routes.post_anthropic_count_tokens)); // anthropic token counting268    // LoRA adapters hotswap269    ctx_http.get ("/lora-adapters",            ex_wrapper(routes.get_lora_adapters));270    ctx_http.post("/lora-adapters",            ex_wrapper(routes.post_lora_adapters));271    // Save & load slots272    ctx_http.get ("/slots",                    ex_wrapper(routes.get_slots));273    ctx_http.post("/slots/:id_slot",           ex_wrapper(routes.post_slots));274 275    // resumable streaming: a child binds the local session factories, the router binds276    // proxies that resolve the owning child, see server-stream.h277    server_http_context::handler_t stream_get_h;278    server_http_context::handler_t streams_lookup_h;279    server_http_context::handler_t stream_delete_h;280    if (is_router_server) {281        stream_get_h     = models_routes->router_stream_get;282        streams_lookup_h = models_routes->router_streams_lookup;283        stream_delete_h  = models_routes->router_stream_delete;284    } else {285        stream_get_h     = server_stream_make_get_handler();286        streams_lookup_h = server_stream_make_lookup_handler();287        stream_delete_h  = server_stream_make_delete_handler();288    }289    ctx_http.get ("/v1/stream",                ex_wrapper(stream_get_h));290    ctx_http.post("/v1/streams/lookup",        ex_wrapper(streams_lookup_h));291    ctx_http.del ("/v1/stream",                ex_wrapper(stream_delete_h));292 293    // Google Cloud Platform (Vertex AI) compat294    ctx_http.register_gcp_compat();295 296    // return 403 for disabled features297    server_http_context::handler_t res_403 = [](const server_http_req &) {298        auto res = std::make_unique<server_http_res>();299        res->status = 403;300        res->data = safe_json_to_str({301            {"error", {302                {"message", "this feature is disabled"},303                {"type", "feature_disabled"},304            }}305        });306        return res;307    };308 309    if (params.cors_origins == "*" && params.api_keys.empty()) {310        SRV_WRN("%s", "-----------------\n");311        SRV_WRN("%s", "CORS is set to allow all origins ('*') and no API key is set\n");312        SRV_WRN("%s", "this can be a security risk (cross-origin attacks)\n");313        SRV_WRN("%s", "more info: https://github.com/ggml-org/llama.cpp/pull/25655\n");314        SRV_WRN("%s", "-----------------\n");315    }316 317    // CORS proxy (EXPERIMENTAL, only used by the Web UI for MCP)318    std::vector<std::string> warn_names;319    if (is_router_server) {320        warn_names.push_back("router mode");321    }322 323    if (params.ui_mcp_proxy) {324        ctx_http.get ("/cors-proxy",      ex_wrapper(proxy_handler_get));325        ctx_http.post("/cors-proxy",      ex_wrapper(proxy_handler_post));326        warn_names.push_back("MCP proxy (experimental)");327    } else {328        ctx_http.get ("/cors-proxy",      ex_wrapper(res_403));329        ctx_http.post("/cors-proxy",      ex_wrapper(res_403));330    }331 332    try {333        mcp_mgr.start(params);334    } catch (const std::exception & e) {335        SRV_ERR("MCP starting failed: %s\n", e.what());336        return 1;337    }338 339    if (!params.server_tools.empty() || !mcp_mgr.empty()) {340        try {341            tools.setup(params.server_tools, mcp_mgr, params.server_tools_runtime);342        } catch (const std::exception & e) {343            SRV_ERR("tools setup failed: %s\n", e.what());344            return 1;345        }346        ctx_http.get ("/tools",           ex_wrapper(tools.handle_get));347        ctx_http.post("/tools",           ex_wrapper(tools.handle_post));348        if (!params.server_tools.empty()) {349            warn_names.push_back("built-in tools (experimental)");350        }351        if (!params.server_tools_runtime.empty()) {352            warn_names.push_back("tools runtime (experimental)");353        }354        if (!mcp_mgr.empty()) {355            warn_names.push_back("MCP servers (experimental)");356        }357    } else {358        ctx_http.get ("/tools",           ex_wrapper(res_403));359        ctx_http.post("/tools",           ex_wrapper(res_403));360    }361 362    if (warn_names.size() > 0) {363        SRV_WRN("%s", "-----------------\n");364        SRV_WRN("%s", "the following feature(s) are enabled:\n");365        for (const auto & name : warn_names) {366            SRV_WRN("    %s\n", name.c_str());367        }368        SRV_WRN("%s", "do not expose the server to untrusted environments\n");369        SRV_WRN("%s", "-----------------\n");370    }371 372    //373    // Handle downloading model374    //375 376    if (child.is_child() && child.get_mode() == SERVER_CHILD_MODE_DOWNLOAD) {377        return child.run_download(params);378    } else if (!is_router_server && !is_run_by_cli) {379        // single-model mode (NOT spawned by router)380        // if this is invoked by CLI, model downloading should be already handled381        try {382            common_models_handler_apply(models_handler, params);383        } catch (const std::exception & e) {384            SRV_ERR("failed to download model: %s\n", e.what());385            return 1;386        }387    }388 389    //390    // Start the server391    //392 393    std::function<void()> clean_up;394 395    if (is_router_server) {396        SRV_INF("%s", "starting server in router mode. models will be automatically loaded on-demand\n");397 398        clean_up = [&models_routes, &mcp_mgr]() {399            SRV_INF("%s: cleaning up before exit...\n", __func__);400            // stop the session GC first, it finalizes live sessions and wakes pending readers401            server_stream_session_manager_stop();402            if (models_routes.has_value()) {403                models_routes->stopping.store(true); // maybe redundant, but just to be safe404                models_routes->models.unload_all();405            }406            mcp_mgr.shutdown();407            llama_backend_free();408        };409 410        if (!ctx_http.start()) {411            clean_up();412            SRV_ERR("%s", "exiting due to HTTP server error\n");413            return 1;414        }415        ctx_http.is_ready.store(true);416 417        shutdown_handler = [&](int) {418            if (models_routes.has_value()) {419                // important to disconnect any SSE clients420                models_routes->stopping.store(true);421            }422            mcp_mgr.shutdown();423            ctx_http.stop();424        };425 426    } else {427        // setup clean up function, to be called before exit428        clean_up = [&ctx_http, &ctx_server, &mcp_mgr]() {429            SRV_INF("%s: cleaning up before exit...\n", __func__);430            // stop the session GC first, it finalizes live sessions and wakes pending readers431            server_stream_session_manager_stop();432            ctx_http.stop();433            ctx_server.terminate();434            mcp_mgr.shutdown();435            llama_backend_free();436        };437 438        // start the HTTP server before loading the model to be able to serve /health requests439        if (!ctx_http.start()) {440            clean_up();441            SRV_ERR("%s", "exiting due to HTTP server error\n");442            return 1;443        }444 445        // setup communication child --> router if necessary446        if (child.is_child()) {447            ctx_server.set_state_callback([&](server_state state, json payload) {448                child.notify_to_router(server_state_to_str(state), payload);449            });450        }451 452        if (!ctx_server.load_model(params)) {453            clean_up();454            if (ctx_http.thread.joinable()) {455                ctx_http.thread.join();456            }457            SRV_ERR("%s", "exiting due to model loading error\n");458            return 1;459        }460 461        routes.update_meta(ctx_server);462        ctx_http.is_ready.store(true);463 464        SRV_INF("%s", "model loaded\n");465 466        shutdown_handler = [&](int) {467            mcp_mgr.shutdown();468            // this will unblock start_loop()469            ctx_server.terminate();470        };471    }472 473    // register signal handler if not running by CLI474    if (!is_run_by_cli) {475#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))476        struct sigaction sigint_action;477        sigint_action.sa_handler = signal_handler;478        sigemptyset (&sigint_action.sa_mask);479        sigint_action.sa_flags = 0;480        sigaction(SIGINT, &sigint_action, NULL);481        sigaction(SIGTERM, &sigint_action, NULL);482#elif defined (_WIN32)483        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {484            return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;485        };486        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);487#endif488    }489 490    SRV_INF("listening on %s\n", ctx_http.listening_address.c_str());491 492    // TODO: remove this in the future493    // check the string to also handle the .sock case494    if (string_ends_with(ctx_http.listening_address, ":8080")) {495        SRV_WRN("%s", "NOTICE: server default port will be changed to :9931 in a future release\n");496        SRV_WRN("%s", "        ref: https://github.com/ggml-org/llama.cpp/pull/26508\n");497    }498 499    if (is_router_server) {500        if (!params.models_preset_hf.empty()) {501            SRV_WRN(      "NOTE: using preset.ini from HF repo '%s'\n", params.models_preset_hf.c_str());502            SRV_WRN("%s", "      please only use presets that you can trust! Unknown presets may be unsafe\n");503        }504 505        if (ctx_http.thread.joinable()) {506            ctx_http.thread.join(); // keep the main thread alive507        }508 509        // when the HTTP server stops, clean up and exit510        clean_up();511    } else {512        // optionally, notify router server that this instance is ready513        std::thread monitor_thread;514        if (child.is_child()) {515            monitor_thread = child.setup(shutdown_handler);516            child.notify_to_router(server_state_to_str(SERVER_STATE_READY), routes.get_model_info());517        }518 519        // this call blocks the main thread until queue_tasks.terminate() is called520        ctx_server.start_loop();521 522        clean_up();523        if (ctx_http.thread.joinable()) {524            ctx_http.thread.join();525        }526        if (monitor_thread.joinable()) {527            monitor_thread.join();528        }529 530        auto * ll_ctx = ctx_server.get_llama_context();531        if (ll_ctx != nullptr) {532            common_memory_breakdown_print(ll_ctx);533        }534    }535 536    return 0;537}538