ggml-webgpu: add adapter toggles for F16 on Vulkan + NVIDIA

* server: refactor child --> router communication

* fix wakeup case

* add docs

* improve update_status()

* nits

common/arg.cpp

@@ -17,6 +17,7 @@
 #   define NOMINMAX
 #endif
 #include <windows.h>
+#include <shellapi.h>
 #endif
 
 #define JSON_ASSERT GGML_ASSERT
@@ -302,7 +303,6 @@ static handle_model_result common_params_handle_model(struct common_params_model
 
     if (!model.docker_repo.empty()) {
         model.path = common_docker_resolve_model(model.docker_repo);
-        model.name = model.docker_repo;
     } else if (!model.hf_repo.empty()) {
         // If -m was used with -hf, treat the model "path" as the hf_file to download
         if (model.hf_file.empty() && !model.path.empty()) {
@@ -322,7 +322,6 @@ static handle_model_result common_params_handle_model(struct common_params_model
             throw std::runtime_error("failed to download model from Hugging Face");
         }
 
-        model.name = model.hf_repo;
         model.path = download_result.model_path;
 
         if (!download_result.mmproj_path.empty()) {
@@ -893,7 +892,44 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
     return true;
 }
 
+#ifdef _WIN32
+struct utf8_argv {
+    std::vector<std::string> buf;
+    std::vector<char*> ptrs;
+};
+
+static utf8_argv make_utf8_argv() {
+    utf8_argv out;
+    int wargc = 0;
+    LPWSTR* wargv = CommandLineToArgvW(GetCommandLineW(), &wargc);
+    if (!wargv) return out;
+
+    out.buf.reserve(wargc);
+    for (int i = 0; i < wargc; ++i) {
+        int n = WideCharToMultiByte(CP_UTF8, WC_ERR_INVALID_CHARS, wargv[i], -1, nullptr, 0, nullptr, nullptr);
+        if (n <= 0) { out.buf.emplace_back(); continue; }
+        auto& s = out.buf.emplace_back();
+        s.resize(static_cast<size_t>(n - 1));
+        (void)WideCharToMultiByte(CP_UTF8, 0, wargv[i], -1, s.data(), n, nullptr, nullptr);
+    }
+    LocalFree(wargv);
+
+    out.ptrs.reserve(out.buf.size() + 1);
+    for (auto& s : out.buf) out.ptrs.push_back(s.data());
+    out.ptrs.push_back(nullptr);
+    return out;
+}
+#endif
+
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
+#ifdef _WIN32
+    auto utf8 = make_utf8_argv();
+    if (!utf8.ptrs.empty()) {
+        argc = static_cast<int>(utf8.buf.size());
+        argv = utf8.ptrs.data();
+    }
+#endif
+
     auto ctx_arg = common_params_parser_init(params, ex, print_usage);
     const common_params params_org = ctx_arg.params; // the example can modify the default params
 
@@ -2911,7 +2947,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_API_KEY"));
     add_opt(common_arg(
         {"--api-key-file"}, "FNAME",
-        "path to file containing API keys (default: none)",
+        "path to file containing API keys, one per line; lines starting with a hash are treated as comments (default: none)",
         [](common_params & params, const std::string & value) {
             std::ifstream key_file(value);
             if (!key_file) {
@@ -2919,7 +2955,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
             std::string key;
             while (std::getline(key_file, key)) {
-                if (!key.empty()) {
+                if (!key.empty() && key[0] != '#') {
                     params.api_keys.push_back(key);
                 }
             }

common/common.cpp

@@ -1074,6 +1074,18 @@ std::vector<common_file_info> fs_list(const std::string & path, bool include_dir
     return files;
 }
 
+std::ifstream fs_open_ifstream(const std::string & fname, std::ios_base::openmode mode) {
+#ifdef _WIN32
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, fname.c_str(), -1, NULL, 0);
+    if (!wlen) { return std::ifstream(); }
+    std::vector<wchar_t> wfname(wlen);
+    (void)MultiByteToWideChar(CP_UTF8, 0, fname.c_str(), -1, wfname.data(), wlen);
+    return std::ifstream(wfname.data(), mode);
+#else
+    return std::ifstream(fname, mode);
+#endif
+}
+
 //
 // TTY utils
 //

common/common.h

@@ -295,7 +295,16 @@ struct common_params_model {
     std::string hf_repo     = ""; // HF repo                                                // NOLINT
     std::string hf_file     = ""; // HF file                                                // NOLINT
     std::string docker_repo = ""; // Docker repo                                            // NOLINT
-    std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
+
+    std::string get_name() {
+        if (!hf_repo.empty()) {
+            return hf_repo;
+        }
+        if (!docker_repo.empty()) {
+            return docker_repo;
+        }
+        return path;
+    }
 };
 
 // draft-model-based speculative decoding parameters
@@ -842,6 +851,9 @@ struct common_file_info {
 };
 std::vector<common_file_info> fs_list(const std::string & path, bool include_directories);
 
+// fs open, also handle UTF8 on Windows
+std::ifstream fs_open_ifstream(const std::string & fname, std::ios_base::openmode mode);
+
 //
 // TTY utils
 //

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -3788,7 +3788,7 @@ static void ggml_webgpu_init_memset_pipeline(webgpu_global_context & ctx) {
     ctx->memset_pipeline = ggml_webgpu_create_pipeline(ctx->device, wgsl_memset, "memset", constants);
 }
 
-static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
+static void ggml_backend_webgpu_request_adapter(wgpu::Instance & instance, wgpu::Adapter & adapter) {
     wgpu::RequestAdapterOptions options = {};
 
 #ifndef __EMSCRIPTEN__
@@ -3800,17 +3800,20 @@ static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
     options.nextInChain                   = &adapterTogglesDesc;
 #endif
 
-    ctx->webgpu_global_ctx->instance.WaitAny(
-        ctx->webgpu_global_ctx->instance.RequestAdapter(
-            &options, wgpu::CallbackMode::AllowSpontaneous,
-            [&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) {
-                if (status != wgpu::RequestAdapterStatus::Success) {
-                    GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
-                    return;
-                }
-                ctx->webgpu_global_ctx->adapter = std::move(adapter);
-            }),
-        UINT64_MAX);
+    instance.WaitAny(instance.RequestAdapter(
+                         &options, wgpu::CallbackMode::AllowSpontaneous,
+                         [&adapter](wgpu::RequestAdapterStatus status, wgpu::Adapter _adapter, const char * message) {
+                             if (status != wgpu::RequestAdapterStatus::Success) {
+                                 GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
+                                 return;
+                             }
+                             adapter = std::move(_adapter);
+                         }),
+                     UINT64_MAX);
+}
+
+static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
+    ggml_backend_webgpu_request_adapter(ctx->webgpu_global_ctx->instance, ctx->webgpu_global_ctx->adapter);
     GGML_ASSERT(ctx->webgpu_global_ctx->adapter != nullptr);
 
     ctx->webgpu_global_ctx->adapter.GetLimits(&ctx->webgpu_global_ctx->capabilities.limits);
@@ -4543,20 +4546,7 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {
     // Probe for adapter support
     wgpu::Adapter adapter;
     if (ctx->webgpu_global_ctx->instance != nullptr) {
-        wgpu::RequestAdapterOptions options = {};
-
-        // probe for adapter support
-        ctx->webgpu_global_ctx->instance.WaitAny(
-            ctx->webgpu_global_ctx->instance.RequestAdapter(
-                &options, wgpu::CallbackMode::AllowSpontaneous,
-                [&adapter](wgpu::RequestAdapterStatus status, wgpu::Adapter _adapter, const char * message) {
-                    if (status != wgpu::RequestAdapterStatus::Success) {
-                        GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
-                        return;
-                    }
-                    adapter = std::move(_adapter);
-                }),
-            UINT64_MAX);
+        ggml_backend_webgpu_request_adapter(ctx->webgpu_global_ctx->instance, adapter);
     }
 
     // WebGPU backend requires f16 support and, on native, implicit device synchronization.

ggml/src/ggml.c

@@ -600,18 +600,15 @@ FILE * ggml_fopen(const char * fname, const char * mode) {
     // convert fname (UTF-8)
     wchar_t * wfname = ggml_mbstowcs(fname);
     if (wfname) {
-        // convert mode (ANSI)
-        wchar_t * wmode = GGML_MALLOC((strlen(mode) + 1) * sizeof(wchar_t));
-        wchar_t * wmode_p = wmode;
-        do {
-            *wmode_p++ = (wchar_t)*mode;
-        } while (*mode++);
-
-        // open file
-        file = _wfopen(wfname, wmode);
+        // convert mode (UTF-8)
+        wchar_t * wmode = ggml_mbstowcs(mode);
+        if (wmode) {
+            // open file
+            file = _wfopen(wfname, wmode);
+            GGML_FREE(wmode);
+        }
 
         GGML_FREE(wfname);
-        GGML_FREE(wmode);
     }
 
     return file;

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "refs/tags/v0.47.0"
+HTTPLIB_VERSION = "refs/tags/v0.48.0"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",

tools/cli/cli.cpp

@@ -202,7 +202,7 @@ struct cli_context {
 
     // TODO: support remote files in the future (http, https, etc)
     std::string load_input_file(const std::string & fname, bool is_media) {
-        std::ifstream file(fname, std::ios::binary);
+        std::ifstream file = fs_open_ifstream(fname, std::ios::binary);
         if (!file) {
             return "";
         }

tools/mtmd/clip-impl.h

@@ -13,6 +13,14 @@
 #include <sstream>
 #include <vector>
 #include <memory>
+#include <fstream>
+
+#ifdef _WIN32
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#endif
 
 // Internal header for clip.cpp
 
@@ -661,6 +669,22 @@ struct clip_image_f32_batch {
 // common utils
 //
 
+#ifdef _WIN32
+static std::ifstream open_ifstream_binary(const std::string & fname) {
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, fname.c_str(), -1, NULL, 0);
+    if (!wlen) {
+        throw std::runtime_error("failed to convert filename to UTF-16: " + fname);
+    }
+    std::vector<wchar_t> wfname(wlen);
+    (void)MultiByteToWideChar(CP_UTF8, 0, fname.c_str(), -1, wfname.data(), wlen);
+    return std::ifstream(wfname.data(), std::ios::binary);
+}
+#else
+static std::ifstream open_ifstream_binary(const std::string & fname) {
+    return std::ifstream(fname, std::ios::binary);
+}
+#endif
+
 static std::string string_format(const char * fmt, ...) {
     va_list ap;
     va_list ap2;

tools/mtmd/clip.cpp

@@ -1752,7 +1752,7 @@ struct clip_model_loader {
         std::map<std::string, size_t> tensor_offset;
         std::vector<ggml_tensor *> tensors_to_load;
 
-        auto fin = std::ifstream(fname, std::ios::binary);
+        auto fin = open_ifstream_binary(fname);
         if (!fin) {
             throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
         }

tools/mtmd/mtmd-cli.cpp

@@ -396,6 +396,9 @@ int main(int argc, char ** argv) {
 
     int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
 
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
     // Ctrl+C handling
     {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))

tools/mtmd/mtmd-helper.cpp

@@ -582,13 +582,29 @@ mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx,
 }
 
 mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname, bool placeholder) {
-    std::vector<unsigned char> buf;
+#ifdef _WIN32
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
+    if (!wlen) {
+        LOG_ERR("Unable to convert filename to UTF-16: %s\n", fname);
+        return {nullptr, nullptr};
+    }
+    std::vector<wchar_t> wfname(wlen);
+    wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wfname.data(), wlen);
+    if (!wlen) {
+        LOG_ERR("Unable to convert filename to UTF-16: %s\n", fname);
+        return {nullptr, nullptr};
+    }
+    FILE * f = _wfopen(wfname.data(), L"rb");
+#else
     FILE * f = fopen(fname, "rb");
+#endif
     if (!f) {
         LOG_ERR("Unable to open file %s: %s\n", fname, strerror(errno));
         return {nullptr, nullptr};
     }
 
+    std::vector<unsigned char> buf;
+
     fseek(f, 0, SEEK_END);
     long file_size = ftell(f);
     fseek(f, 0, SEEK_SET);

tools/server/README-dev.md

@@ -180,6 +180,17 @@ That requires `JSON.stringify` when formatted to message content:
 }
 ```
 
+### Router mode: how child <--> router communicates
+
+Upon spawning a new child process using `subprocess`, both child and router listen to the stdout/stderr (combined)
+
+For the direction from child to router:
+- Generic messages are logs, it will be forwarded to router's stdout
+- Special state update messages are prefixed by `cmd_child_to_router:state:`, followed by a JSON. See `server_models::handle_child_state` for more
+
+For the direction from router to child:
+- When server sends `cmd_router_to_child:exit`, the child should exit gracefully --> if after `DEFAULT_STOP_TIMEOUT` and the child is still running, force-kill it
+
 ### Model management API (router mode)
 
 Model management API was added via PR [#23976](https://github.com/ggml-org/llama.cpp/pull/23976)

tools/server/README.md

@@ -198,7 +198,7 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
 | `--rerank, --reranking` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
 | `--api-key KEY` | API key to use for authentication, multiple keys can be provided as a comma-separated list (default: none)<br/>(env: LLAMA_API_KEY) |
-| `--api-key-file FNAME` | path to file containing API keys (default: none)<br/>(env: LLAMA_ARG_API_KEY_FILE) |
+| `--api-key-file FNAME` | path to file containing API keys, one per line; lines starting with a hash are treated as comments (default: none)<br/>(env: LLAMA_ARG_API_KEY_FILE) |
 | `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key<br/>(env: LLAMA_ARG_SSL_KEY_FILE) |
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
 | `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_KWARGS) |

tools/server/server-common.cpp

@@ -12,6 +12,7 @@
 #include <random>
 #include <sstream>
 #include <fstream>
+#include <limits>
 
 json format_error_response(const std::string & message, const enum error_type type) {
     std::string type_str;
@@ -1238,7 +1239,7 @@ json format_response_rerank(
 // other utils
 //
 
-std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx) {
+std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx, size_t n_top) {
     std::vector<llama_token_data> cur;
 
     const auto * logits = llama_get_logits_ith(ctx, idx);
@@ -1257,21 +1258,34 @@ std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int i
         }
     }
 
-    // sort tokens by logits
-    std::sort(cur.begin(), cur.end(), [](const llama_token_data & a, const llama_token_data & b) {
-        return a.logit > b.logit;
-    });
+    // sort tokens by logits (partial: only the leading `n_top` need ordering)
+    if (n_top > cur.size()) {
+        n_top = cur.size();
+    }
+    if (n_top > 0) {
+        std::partial_sort(cur.begin(), cur.begin() + n_top, cur.end(),
+            [](const llama_token_data & a, const llama_token_data & b) {
+                return a.logit > b.logit;
+            });
+    }
 
     // apply softmax
-    float max_l = cur[0].logit;
+    float max_l = -std::numeric_limits<float>::infinity();
+    if (n_top > 0) {
+        max_l = cur[0].logit; // partial_sort guarantees the absolute maximum is at index 0
+    } else {
+        for (const auto & t : cur) {
+            max_l = std::max(max_l, t.logit);
+        }
+    }
     float cum_sum = 0.0f;
-    for (size_t i = 0; i < cur.size(); ++i) {
-        float p = expf(cur[i].logit - max_l);
-        cur[i].p = p;
+    for (auto & t : cur) {
+        float p = expf(t.logit - max_l);
+        t.p = p;
         cum_sum += p;
     }
-    for (size_t i = 0; i < cur.size(); ++i) {
-        cur[i].p /= cum_sum;
+    for (auto & t : cur) {
+        t.p /= cum_sum;
     }
 
     return cur;

tools/server/server-common.h

@@ -326,7 +326,7 @@ json format_response_rerank(
 // other utils
 //
 
-std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx);
+std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx, size_t n_top);
 
 std::string safe_json_to_str(const json & data);
 

tools/server/server-context.cpp

@@ -63,11 +63,6 @@ enum slot_state {
     SLOT_STATE_GENERATING,
 };
 
-enum server_state {
-    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
-    SERVER_STATE_READY,          // Server is ready and model is loaded
-};
-
 struct server_slot {
     int id;
 
@@ -773,6 +768,8 @@ public:
     // note: chat_params must not be refreshed upon existing sleeping state
     server_chat_params chat_params;
 
+    server_state_callback_t callback_state = [](server_state, json) -> void {};
+
     server_context_impl() {
         mtmd_helper_log_set(common_log_default_callback, nullptr);
     }
@@ -825,8 +822,7 @@ private:
 
     server_metrics metrics;
 
-    json json_ui_settings = json::object();    // Primary: new name
-    json json_webui_settings = json::object();    // Deprecated: use json_ui_settings instead (kept for compat)
+    json json_ui_settings = json::object();
 
     // Necessary similarity of prompt for slot selection
     float slot_prompt_similarity = 0.0f;
@@ -1245,8 +1241,8 @@ private:
         if (!params_base.model_alias.empty()) {
             // backward compat: use first alias as model name
             model_name = *params_base.model_alias.begin();
-        } else if (!params_base.model.name.empty()) {
-            model_name = params_base.model.name;
+        } else if (!params_base.model.get_name().empty()) {
+            model_name = params_base.model.get_name();
         } else {
             // fallback: derive model name from file name
             auto model_path = std::filesystem::path(params_base.model.path);
@@ -1308,7 +1304,6 @@ private:
                 try {
                     json json_settings = json::parse(cfg);
                     json_ui_settings = json_settings;
-                    json_webui_settings = json_settings; // deprecated: keep in sync
                 } catch (const std::exception & e) {
                     SRV_ERR("%s: failed to parse UI config: %s\n", __func__, e.what());
                     return false;
@@ -1826,8 +1821,7 @@ private:
                 });
             }
         } else {
-            // TODO: optimize this with min-p optimization
-            std::vector<llama_token_data> cur = get_token_probabilities(ctx_tgt, idx);
+            std::vector<llama_token_data> cur = get_token_probabilities(ctx_tgt, idx, n_probs_request);
             const size_t max_probs = cur.size();
             const size_t n_probs = std::min(max_probs, n_probs_request);
 
@@ -3687,7 +3681,6 @@ server_context_meta server_context::get_meta() const {
         /* has_inp_audio          */ impl->chat_params.allow_audio,
         /* has_inp_video          */ impl->chat_params.allow_video,
         /* json_ui_settings       */ impl->json_ui_settings,
-        /* json_webui_settings    */ impl->json_webui_settings,  // Deprecated
         /* slot_n_ctx             */ impl->get_slot_n_ctx(),
         /* pooling_type           */ llama_pooling_type(impl->ctx_tgt),
 
@@ -3738,8 +3731,11 @@ struct server_res_generator : server_http_res {
     }
 };
 
-void server_context::on_sleeping_changed(std::function<void(bool)> callback) {
-    impl->queue_tasks.on_sleeping_state(std::move(callback));
+void server_context::set_state_callback(server_state_callback_t callback) {
+    impl->callback_state = std::move(callback);
+    impl->queue_tasks.on_sleeping_state([this](bool sleeping) {
+        impl->callback_state(sleeping ? SERVER_STATE_SLEEPING : SERVER_STATE_READY, {});
+    });
 }
 
 // compute the number of tokens before the last user message in the prompt
@@ -4300,12 +4296,8 @@ void server_routes::init_routes() {
             { "endpoint_slots",              params.endpoint_slots },
             { "endpoint_props",              params.endpoint_props },
             { "endpoint_metrics",            params.endpoint_metrics },
-            // New keys
             { "ui",                          params.ui },
             { "ui_settings",                 meta->json_ui_settings },
-            // Deprecated: use ui/ui_settings instead (kept for backward compat)
-            { "webui",                       params.ui },
-            { "webui_settings",              meta->json_ui_settings },
             { "chat_template",               tmpl_default },
             { "chat_template_caps",          meta->chat_template_caps },
             { "bos_token",                   meta->bos_token_str },

tools/server/server-context.h

@@ -22,8 +22,7 @@ struct server_context_meta {
     bool has_inp_image;
     bool has_inp_audio;
     bool has_inp_video;
-    json json_ui_settings;            // Primary: new name
-    json json_webui_settings;            // Deprecated: use json_ui_settings instead (kept for backward compat)
+    json json_ui_settings;
     int slot_n_ctx;
     enum llama_pooling_type pooling_type;
 
@@ -53,6 +52,31 @@ struct server_context_meta {
     uint64_t model_size;
 };
 
+enum server_state {
+    // SERVER_STATE_DOWNLOADING,
+    SERVER_STATE_LOADING,
+    SERVER_STATE_READY,
+    SERVER_STATE_SLEEPING,
+};
+
+static std::string server_state_to_str(server_state state) {
+    switch (state) {
+        case SERVER_STATE_LOADING:     return "loading";
+        case SERVER_STATE_READY:       return "ready";
+        case SERVER_STATE_SLEEPING:    return "sleeping";
+        default: GGML_ASSERT(false && "invalid server_state");
+    }
+}
+
+static server_state server_state_from_str(const std::string & str) {
+    if (str == "loading")     return SERVER_STATE_LOADING;
+    if (str == "ready")       return SERVER_STATE_READY;
+    if (str == "sleeping")    return SERVER_STATE_SLEEPING;
+    GGML_ASSERT(false && "invalid server_state string");
+}
+
+using server_state_callback_t = std::function<void(server_state, json /* payload */)>;
+
 struct server_context {
     std::unique_ptr<server_context_impl> impl;
 
@@ -80,9 +104,8 @@ struct server_context {
     // not thread-safe, should only be used from the main thread
     server_context_meta get_meta() const;
 
-    // register a callback to be called when sleeping state changes
-    // must be set before load_model() is called
-    void on_sleeping_changed(std::function<void(bool)> callback);
+    // note: must be set before load_model() is called
+    void set_state_callback(server_state_callback_t callback);
 };
 
 

tools/server/server-models.cpp

@@ -1,5 +1,6 @@
 #include "server-common.h"
 #include "server-models.h"
+#include "server-context.h"
 
 #include "build-info.h"
 #include "preset.h"
@@ -44,9 +45,7 @@ extern char **environ;
 #define DEFAULT_STOP_TIMEOUT 10 // seconds
 
 #define CMD_ROUTER_TO_CHILD_EXIT  "cmd_router_to_child:exit"
-#define CMD_CHILD_TO_ROUTER_READY "cmd_child_to_router:ready" // also sent when waking up from sleep
-#define CMD_CHILD_TO_ROUTER_SLEEP "cmd_child_to_router:sleep"
-#define CMD_CHILD_TO_ROUTER_INFO  "cmd_child_to_router:info:" // followed by json string
+#define CMD_CHILD_TO_ROUTER_STATE "cmd_child_to_router:state:" // followed by json string
 
 // address for child process, this is needed because router may run on 0.0.0.0
 // ref: https://github.com/ggml-org/llama.cpp/issues/17862
@@ -904,12 +903,8 @@ void server_models::load(const std::string & name) {
                 while (fgets(buffer, vec_buf.size(), stdout_file) != nullptr) {
                     LOG("[%5d] %s", port, buffer);
                     std::string str(buffer);
-                    if (string_starts_with(buffer, CMD_CHILD_TO_ROUTER_READY)) {
-                        this->update_status(name, SERVER_MODEL_STATUS_LOADED, 0);
-                    } else if (string_starts_with(buffer, CMD_CHILD_TO_ROUTER_INFO)) {
-                        this->update_loaded_info(name, str);
-                    } else if (string_starts_with(buffer, CMD_CHILD_TO_ROUTER_SLEEP)) {
-                        this->update_status(name, SERVER_MODEL_STATUS_SLEEPING, 0);
+                    if (string_starts_with(buffer, CMD_CHILD_TO_ROUTER_STATE)) {
+                        this->handle_child_state(name, str);
                     }
                 }
             } else {
@@ -976,7 +971,10 @@ void server_models::load(const std::string & name) {
         subprocess_destroy(&child_proc->get());
 
         // update status and exit code
-        this->update_status(name, SERVER_MODEL_STATUS_UNLOADED, exit_code);
+        this->update_status(name, {
+            SERVER_MODEL_STATUS_UNLOADED,
+            exit_code
+        });
         SRV_INF("instance name=%s exited with status %d\n", name.c_str(), exit_code);
     });
 
@@ -1016,7 +1014,8 @@ struct server_models_download_res : public common_download_callback {
             common_download_model(model, opts);
             is_ok = true;
         } catch (const std::exception & e) {
-            SRV_ERR("download failed for model name=%s: %s\n", model.name.c_str(), e.what());
+            auto model_name = model.get_name();
+            SRV_ERR("download failed for model name=%s: %s\n", model_name.c_str(), e.what());
             is_ok = false;
         }
         return is_ok;
@@ -1036,7 +1035,7 @@ struct server_models_download_res : public common_download_callback {
 };
 
 void server_models::download(common_params_model && model, common_download_opts && opts) {
-    std::string name = model.name;
+    std::string name = model.get_name();
     GGML_ASSERT(name == model.hf_repo);
 
     std::unique_lock<std::mutex> lk(mutex);
@@ -1064,9 +1063,10 @@ void server_models::download(common_params_model && model, common_download_opts
     inst.th = std::thread([this, dl = std::move(dl)]() {
         dl->opts.callback = dl.get();
         bool ok = dl->run();
+        auto model_name = dl->model.get_name();
         SRV_INF("download finished for model name=%s with status=%s\n",
-                    dl->model.name.c_str(), ok ? "success" : "failure");
-        update_download_progress(dl->model.name, {}, true, ok);
+                    model_name.c_str(), ok ? "success" : "failure");
+        update_download_progress(model_name, {}, true, ok);
         // need_reload is set inside update_download_progress under the mutex;
         // the next load_models() call will clean up this instance
     });
@@ -1130,21 +1130,27 @@ void server_models::unload_all() {
     }
 }
 
-void server_models::update_status(const std::string & name, server_model_status status, int exit_code) {
+void server_models::update_status(const std::string & name, const update_status_args & args) {
     std::unique_lock<std::mutex> lk(mutex);
     auto it = mapping.find(name);
     if (it != mapping.end()) {
         auto & meta = it->second.meta;
-        meta.status    = status;
-        meta.exit_code = exit_code;
+        meta.status      = args.status;
+        meta.exit_code   = args.exit_code;
+        if (!args.loaded_info.is_null()) {
+            meta.loaded_info = args.loaded_info;
+        }
     }
     // broadcast status change to SSE
     {
         json data = {
-            {"status", server_model_status_to_string(status)},
+            {"status", server_model_status_to_string(args.status)},
         };
-        if (status == SERVER_MODEL_STATUS_UNLOADED) {
-            data["exit_code"] = exit_code;
+        if (args.status == SERVER_MODEL_STATUS_UNLOADED) {
+            data["exit_code"] = args.exit_code;
+        }
+        if (!args.loaded_info.is_null()) {
+            data["info"] = args.loaded_info;
         }
         // note: notify_sse doesn't acquire the lock, so no deadlock here
         notify_sse("status_change", name, data);
@@ -1152,29 +1158,6 @@ void server_models::update_status(const std::string & name, server_model_status
     cv.notify_all();
 }
 
-void server_models::update_loaded_info(const std::string & name, std::string & raw_info) {
-    if (!string_starts_with(raw_info, CMD_CHILD_TO_ROUTER_INFO)) {
-        SRV_WRN("invalid loaded info format from child for model name=%s: %s\n", name.c_str(), raw_info.c_str());
-        return;
-    }
-
-    json info;
-    try {
-        info = json::parse(raw_info.substr(strlen(CMD_CHILD_TO_ROUTER_INFO)));
-    } catch (const std::exception & e) {
-        SRV_WRN("failed to parse loaded info from child for model name=%s: %s\n", name.c_str(), e.what());
-        return;
-    }
-
-    std::unique_lock<std::mutex> lk(mutex);
-    auto it = mapping.find(name);
-    if (it != mapping.end()) {
-        auto & meta = it->second.meta;
-        meta.loaded_info = info;
-    }
-    cv.notify_all();
-}
-
 void server_models::update_download_progress(const std::string & name, const common_download_progress & progress, bool done, bool ok) {
     json curr;
     {
@@ -1323,21 +1306,54 @@ server_http_res_ptr server_models::proxy_request(const server_http_req & req, co
     return proxy;
 }
 
-bool server_models::is_child_server() {
+void server_models::handle_child_state(const std::string & name, const std::string & raw_input) {
+    server_state state;
+    json payload;
+
+    try {
+        json data = json::parse(raw_input.substr(strlen(CMD_CHILD_TO_ROUTER_STATE)));
+        state = server_state_from_str(json_value(data, "state", std::string()));
+        payload = json_value(data, "payload", json{});
+    } catch (const std::exception & e) {
+        SRV_ERR("failed to parse child state update for name=%s: %s\n", name.c_str(), e.what());
+        return;
+    }
+
+    switch (state) {
+        case SERVER_STATE_LOADING:
+            {
+                // do nothing for now
+                // TODO: report loading progress for first load and wakeup from sleep
+            } break;
+        case SERVER_STATE_READY:
+            {
+                update_status(name, {
+                    SERVER_MODEL_STATUS_LOADED,
+                    0,
+                    // note: payload can be empty if this is a wakeup from sleep
+                    payload.size() > 0 ? payload : nullptr
+                });
+            } break;
+        case SERVER_STATE_SLEEPING:
+            {
+                update_status(name, { SERVER_MODEL_STATUS_SLEEPING });
+            } break;
+        default:
+            // should never happen, but just in case
+            GGML_ASSERT(false && "unexpected state from child server");
+    }
+}
+
+//
+// server_child
+//
+
+bool server_child::is_child() {
     const char * router_port = std::getenv("LLAMA_SERVER_ROUTER_PORT");
     return router_port != nullptr;
 }
 
-std::thread server_models::setup_child_server(const std::function<void(int)> & shutdown_handler, const json & model_info) {
-    // send a notification to the router server that a model instance is ready
-    common_log_pause(common_log_main());
-    fflush(stdout);
-    fprintf(stdout, "%s\n", CMD_CHILD_TO_ROUTER_READY);
-    fflush(stdout);
-    fprintf(stdout, "%s%s\n", CMD_CHILD_TO_ROUTER_INFO, safe_json_to_str(model_info).c_str());
-    fflush(stdout);
-    common_log_resume(common_log_main());
-
+std::thread server_child::setup(const std::function<void(int)> & shutdown_handler) {
     // setup thread for monitoring stdin
     return std::thread([shutdown_handler]() {
         // wait for EOF on stdin
@@ -1363,10 +1379,14 @@ std::thread server_models::setup_child_server(const std::function<void(int)> & s
     });
 }
 
-void server_models::notify_router_sleeping_state(bool is_sleeping) {
+void server_child::notify_to_router(const std::string & state, const json & payload) {
+    json data = {
+        {"state", state},
+        {"payload", payload},
+    };
     common_log_pause(common_log_main());
     fflush(stdout);
-    fprintf(stdout, "%s\n", is_sleeping ? CMD_CHILD_TO_ROUTER_SLEEP : CMD_CHILD_TO_ROUTER_READY);
+    fprintf(stdout, "%s%s\n", CMD_CHILD_TO_ROUTER_STATE, safe_json_to_str(data).c_str());
     fflush(stdout);
     common_log_resume(common_log_main());
 }
@@ -1474,7 +1494,6 @@ void server_models_routes::init_routes() {
                 }},
                 // New key
                 {"ui_settings",          ui_settings},
-                {"webui_settings",       webui_settings},
                 {"build_info",           std::string(llama_build_info())},
                 {"cors_proxy_enabled",   params.ui_mcp_proxy},
             });
@@ -1645,7 +1664,6 @@ void server_models_routes::init_routes() {
         common_params_model model;
         common_download_opts opts;
 
-        model.name           = name;
         model.hf_repo        = name;
         opts.bearer_token    = params.hf_token;
         opts.download_mmproj = true;

tools/server/server-models.h

@@ -171,8 +171,12 @@ public:
     void download(common_params_model && model, common_download_opts && opts);
 
     // update the status of a model instance (thread-safe)
-    void update_status(const std::string & name, server_model_status status, int exit_code);
-    void update_loaded_info(const std::string & name, std::string & raw_info);
+    struct update_status_args {
+        server_model_status status;
+        int exit_code = 0; // only valid if status == UNLOADED
+        json loaded_info = nullptr;
+    };
+    void update_status(const std::string & name, const update_status_args & args);
     void update_download_progress(const std::string & name, const common_download_progress & progress, bool done, bool ok = true);
 
     // remove a cache model from disk and update the list (thread-safe)
@@ -193,21 +197,32 @@ public:
     // proxy an HTTP request to the model instance
     server_http_res_ptr proxy_request(const server_http_req & req, const std::string & method, const std::string & name, bool update_last_used);
 
+    // handle message sent from server_child::notify_to_router()
+    // raw input must starts with CMD_CHILD_TO_ROUTER_STATE, followed by a JSON string
+    // this function is not thread-safe, must be called from instance's monitoring thread
+    // payload per state:
+    //     state = loading     -> payload = {} (TODO: add progress info)
+    //     state = ready       -> payload = model_info (json), or {} if wakeup from sleeping
+    //     state = sleeping    -> payload = {}
+    void handle_child_state(const std::string & name, const std::string & raw_input);
+};
+
+struct server_child {
     // return true if the current process is a child server instance
-    static bool is_child_server();
+    bool is_child();
 
-    // notify the router server that a model instance is ready
+    // register the shutdown_handler to be called by the router
     // return the monitoring thread (to be joined by the caller)
-    static std::thread setup_child_server(const std::function<void(int)> & shutdown_handler, const json & model_info);
+    std::thread setup(const std::function<void(int)> & shutdown_handler);
 
-    // notify the router server that the sleeping state has changed
-    static void notify_router_sleeping_state(bool sleeping);
+    // notify router server for status changes (e.g. loading, downloading, sleeping, etc.)
+    // message will be handled by server_models::handle_child_state() on the router side
+    void notify_to_router(const std::string & state_name, const json & payload);
 };
 
 struct server_models_routes {
     common_params params;
     json ui_settings = json::object();     // Primary: new name
-    json webui_settings = json::object();  // Deprecated: use ui_settings (kept for compat)
     std::atomic<bool> stopping = false;    // for graceful disconnecting SSE clients during shutdown
     server_models models;
     server_models_routes(const common_params & params, int argc, char ** argv)
@@ -217,7 +232,6 @@ struct server_models_routes {
             try {
                 json json_settings = json::parse(cfg);
                 ui_settings = json_settings;
-                webui_settings = json_settings;  // Deprecated: keep in sync
             } catch (const std::exception & e) {
                 LOG_ERR("%s: failed to parse UI config: %s\n", __func__, e.what());
                 throw;

tools/server/server.cpp

@@ -90,8 +90,10 @@ int llama_server(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // router server never loads a model and must not touch the GPU
+    const bool is_router_server = params.model.path.empty()
+                               && params.model.hf_repo.empty();
+
     // skip device enumeration so the CUDA primary context stays uncreated
-    const bool is_router_server = params.model.path.empty();
     common_params_print_info(params, !is_router_server);
 
     if (!is_router_server) {
@@ -113,8 +115,9 @@ int llama_server(int argc, char ** argv) {
     }
 
     // for consistency between server router mode and single-model mode, we set the same model name as alias
-    if (params.model_alias.empty() && !params.model.name.empty()) {
-        params.model_alias.insert(params.model.name);
+    auto model_name = params.model.get_name();
+    if (params.model_alias.empty() && !model_name.empty()) {
+        params.model_alias.insert(model_name);
     }
 
     // struct that contains llama context and inference
@@ -255,6 +258,7 @@ int llama_server(int argc, char ** argv) {
     // Start the server
     //
 
+    server_child child; // only used in non-router mode
     std::function<void()> clean_up;
 
     if (is_router_server) {
@@ -300,15 +304,16 @@ int llama_server(int argc, char ** argv) {
             return 1;
         }
 
-        // load the model
-        SRV_INF("%s", "loading model\n");
-
-        if (server_models::is_child_server()) {
-            ctx_server.on_sleeping_changed([&](bool sleeping) {
-                server_models::notify_router_sleeping_state(sleeping);
+        // setup communication child --> router if necessary
+        if (child.is_child()) {
+            ctx_server.set_state_callback([&](server_state state, json payload) {
+                child.notify_to_router(server_state_to_str(state), payload);
             });
         }
 
+        // load the model
+        SRV_INF("%s", "loading model\n");
+
         if (!ctx_server.load_model(params)) {
             clean_up();
             if (ctx_http.thread.joinable()) {
@@ -365,9 +370,9 @@ int llama_server(int argc, char ** argv) {
 
         // optionally, notify router server that this instance is ready
         std::thread monitor_thread;
-        if (server_models::is_child_server()) {
-            json model_info = routes.get_model_info();
-            monitor_thread = server_models::setup_child_server(shutdown_handler, model_info);
+        if (child.is_child()) {
+            monitor_thread = child.setup(shutdown_handler);
+            child.notify_to_router(server_state_to_str(SERVER_STATE_READY), routes.get_model_info());
         }
 
         // this call blocks the main thread until queue_tasks.terminate() is called

tools/server/tests/unit/test_basic.py

@@ -79,9 +79,9 @@ def test_load_split_model():
     assert match_regex("(little|girl)+", res.body["content"])
 
 
-def test_no_webui():
+def test_no_ui():
     global server
-    # default: webui enabled
+    # default: UI enabled
     server.start()
     url = f"http://{server.server_host}:{server.server_port}"
     res = requests.get(url)
@@ -89,8 +89,8 @@ def test_no_webui():
     assert "<!doctype html>" in res.text
     server.stop()
 
-    # with --no-webui
-    server.no_webui = True
+    # with --no-ui, the UI should be disabled
+    server.no_ui = True
     server.start()
     res = requests.get(url)
     assert res.status_code == 404

tools/server/tests/unit/test_proxy.py

@@ -12,7 +12,7 @@ def create_server():
 
 def test_mcp_no_proxy():
     global server
-    server.webui_mcp_proxy = False
+    server.ui_mcp_proxy = False
     server.start()
 
     res = server.make_request("GET", "/cors-proxy")
@@ -21,7 +21,7 @@ def test_mcp_no_proxy():
 
 def test_mcp_proxy():
     global server
-    server.webui_mcp_proxy = True
+    server.ui_mcp_proxy = True
     server.start()
 
     url = f"http://{server.server_host}:{server.server_port}/cors-proxy?url=http://example.com"
@@ -32,7 +32,7 @@ def test_mcp_proxy():
 
 def test_mcp_proxy_custom_port():
     global server
-    server.webui_mcp_proxy = True
+    server.ui_mcp_proxy = True
     server.start()
 
     # try getting the server's models API via the proxy

tools/server/tests/utils.py

@@ -94,7 +94,7 @@ class ServerProcess:
     enable_ctx_shift: int | None = False
     spec_draft_n_min: int | None = None
     spec_draft_n_max: int | None = None
-    no_webui: bool | None = None
+    no_ui: bool | None = None
     jinja: bool | None = None
     reasoning_format: Literal['deepseek', 'none', 'nothink'] | None = None
     reasoning: Literal['on', 'off', 'auto'] | None = None
@@ -107,7 +107,7 @@ class ServerProcess:
     cache_ram: int | None = None
     no_cache_idle_slots: bool = False
     log_path: str | None = None
-    webui_mcp_proxy: bool = False
+    ui_mcp_proxy: bool = False
     backend_sampling: bool = False
     gcp_compat: bool = False
 
@@ -225,8 +225,8 @@ class ServerProcess:
             server_args.extend(["--spec-draft-n-max", self.spec_draft_n_max])
         if self.spec_draft_n_min:
             server_args.extend(["--spec-draft-n-min", self.spec_draft_n_min])
-        if self.no_webui:
-            server_args.append("--no-webui")
+        if self.no_ui:
+            server_args.append("--no-ui")
         if self.no_models_autoload:
             server_args.append("--no-models-autoload")
         if self.jinja:
@@ -251,8 +251,8 @@ class ServerProcess:
             server_args.extend(["--cache-ram", self.cache_ram])
         if self.no_cache_idle_slots:
             server_args.append("--no-cache-idle-slots")
-        if self.webui_mcp_proxy:
-            server_args.append("--webui-mcp-proxy")
+        if self.ui_mcp_proxy:
+            server_args.append("--ui-mcp-proxy")
         if self.backend_sampling:
             server_args.append("--backend_sampling")
         if self.gcp_compat:

vendor/cpp-httplib/httplib.cpp

@@ -5809,11 +5809,9 @@ std::string decode_query_component(const std::string &component,
 
   for (size_t i = 0; i < component.size(); i++) {
     if (component[i] == '%' && i + 2 < component.size()) {
-      std::string hex = component.substr(i + 1, 2);
-      char *end;
-      unsigned long value = std::strtoul(hex.c_str(), &end, 16);
-      if (end == hex.c_str() + 2) {
-        result += static_cast<char>(value);
+      auto val = 0;
+      if (detail::from_hex_to_i(component, i + 1, 2, val)) {
+        result += static_cast<char>(val);
         i += 2;
       } else {
         result += component[i];
@@ -12551,6 +12549,21 @@ bool parse_ipv4(const std::string &str, unsigned char *out) {
   return *p == '\0';
 }
 
+// Parse an IP literal (IPv4 or IPv6) into raw network-order bytes.
+// `out` must have room for at least 16 bytes. Returns the address length
+// (4 for IPv4, 16 for IPv6) on success, or 0 if the string is not an IP
+// literal. Used to match a host against iPAddress SANs the same way the
+// OpenSSL backend does via X509_check_ip.
+size_t parse_ip_address(const std::string &str, unsigned char *out) {
+  if (is_ipv4_address(str)) { return parse_ipv4(str, out) ? 4 : 0; }
+  struct in6_addr addr6 = {};
+  if (inet_pton(AF_INET6, str.c_str(), &addr6) == 1) {
+    memcpy(out, &addr6, 16);
+    return 16;
+  }
+  return 0;
+}
+
 #ifdef _WIN32
 // Enumerate Windows system certificates and call callback with DER data
 template <typename Callback>
@@ -12852,6 +12865,30 @@ int openssl_verify_callback(int preverify_ok, X509_STORE_CTX *ctx) {
   return callback(verify_ctx) ? 1 : 0;
 }
 
+// X509_STORE_get0_objects is deprecated since OpenSSL 4.0 because it is not
+// thread-safe; X509_STORE_get1_objects (OpenSSL 3.3+) returns a snapshot
+// that must be released with release_store_objects
+#if !defined(OPENSSL_IS_BORINGSSL) && !defined(LIBRESSL_VERSION_NUMBER) &&     \
+    OPENSSL_VERSION_NUMBER >= 0x30300000L
+#define CPPHTTPLIB_HAS_X509_STORE_GET1_OBJECTS
+#endif
+
+STACK_OF(X509_OBJECT) * get_store_objects(X509_STORE *store) {
+#ifdef CPPHTTPLIB_HAS_X509_STORE_GET1_OBJECTS
+  return X509_STORE_get1_objects(store);
+#else
+  return X509_STORE_get0_objects(store);
+#endif
+}
+
+void release_store_objects(STACK_OF(X509_OBJECT) * objs) {
+#ifdef CPPHTTPLIB_HAS_X509_STORE_GET1_OBJECTS
+  sk_X509_OBJECT_pop_free(objs, X509_OBJECT_free);
+#else
+  (void)objs; // get0 variant returns an internal pointer; nothing to free
+#endif
+}
+
 } // namespace impl
 
 ctx_t create_client_context() {
@@ -13373,11 +13410,19 @@ std::string get_cert_subject_cn(cert_t cert) {
   auto subject_name = X509_get_subject_name(x509);
   if (!subject_name) return "";
 
-  char buf[256];
-  auto len =
-      X509_NAME_get_text_by_NID(subject_name, NID_commonName, buf, sizeof(buf));
-  if (len < 0) return "";
-  return std::string(buf, static_cast<size_t>(len));
+  // X509_NAME_get_text_by_NID is deprecated since OpenSSL 4.0
+  auto idx = X509_NAME_get_index_by_NID(subject_name, NID_commonName, -1);
+  if (idx < 0) return "";
+
+  auto entry = X509_NAME_get_entry(subject_name, idx);
+  if (!entry) return "";
+
+  auto data = X509_NAME_ENTRY_get_data(entry);
+  if (!data) return "";
+
+  return std::string(
+      reinterpret_cast<const char *>(ASN1_STRING_get0_data(data)),
+      static_cast<size_t>(ASN1_STRING_length(data)));
 }
 
 std::string get_cert_issuer_name(cert_t cert) {
@@ -13582,8 +13627,9 @@ size_t get_ca_certs(ctx_t ctx, std::vector<cert_t> &certs) {
   auto store = SSL_CTX_get_cert_store(ssl_ctx);
   if (!store) { return 0; }
 
-  auto objs = X509_STORE_get0_objects(store);
+  auto objs = impl::get_store_objects(store);
   if (!objs) { return 0; }
+  auto se = detail::scope_exit([&] { impl::release_store_objects(objs); });
 
   auto count = sk_X509_OBJECT_num(objs);
   for (decltype(count) i = 0; i < count; i++) {
@@ -13609,8 +13655,9 @@ std::vector<std::string> get_ca_names(ctx_t ctx) {
   auto store = SSL_CTX_get_cert_store(ssl_ctx);
   if (!store) { return names; }
 
-  auto objs = X509_STORE_get0_objects(store);
+  auto objs = impl::get_store_objects(store);
   if (!objs) { return names; }
+  auto se = detail::scope_exit([&] { impl::release_store_objects(objs); });
 
   auto count = sk_X509_OBJECT_num(objs);
   for (decltype(count) i = 0; i < count; i++) {
@@ -13716,110 +13763,6 @@ std::string verify_error_string(long error_code) {
 
 } // namespace tls
 
-bool SSLClient::verify_host(X509 *server_cert) const {
-  /* Quote from RFC2818 section 3.1 "Server Identity"
-
-     If a subjectAltName extension of type dNSName is present, that MUST
-     be used as the identity. Otherwise, the (most specific) Common Name
-     field in the Subject field of the certificate MUST be used. Although
-     the use of the Common Name is existing practice, it is deprecated and
-     Certification Authorities are encouraged to use the dNSName instead.
-
-     Matching is performed using the matching rules specified by
-     [RFC2459].  If more than one identity of a given type is present in
-     the certificate (e.g., more than one dNSName name, a match in any one
-     of the set is considered acceptable.) Names may contain the wildcard
-     character * which is considered to match any single domain name
-     component or component fragment. E.g., *.a.com matches foo.a.com but
-     not bar.foo.a.com. f*.com matches foo.com but not bar.com.
-
-     In some cases, the URI is specified as an IP address rather than a
-     hostname. In this case, the iPAddress subjectAltName must be present
-     in the certificate and must exactly match the IP in the URI.
-
-  */
-  return verify_host_with_subject_alt_name(server_cert) ||
-         verify_host_with_common_name(server_cert);
-}
-
-bool
-SSLClient::verify_host_with_subject_alt_name(X509 *server_cert) const {
-  auto ret = false;
-
-  auto type = GEN_DNS;
-
-  struct in6_addr addr6 = {};
-  struct in_addr addr = {};
-  size_t addr_len = 0;
-
-#ifndef __MINGW32__
-  if (inet_pton(AF_INET6, host_.c_str(), &addr6)) {
-    type = GEN_IPADD;
-    addr_len = sizeof(struct in6_addr);
-  } else if (inet_pton(AF_INET, host_.c_str(), &addr)) {
-    type = GEN_IPADD;
-    addr_len = sizeof(struct in_addr);
-  }
-#endif
-
-  auto alt_names = static_cast<const struct stack_st_GENERAL_NAME *>(
-      X509_get_ext_d2i(server_cert, NID_subject_alt_name, nullptr, nullptr));
-
-  if (alt_names) {
-    auto dsn_matched = false;
-    auto ip_matched = false;
-
-    auto count = sk_GENERAL_NAME_num(alt_names);
-
-    for (decltype(count) i = 0; i < count && !dsn_matched; i++) {
-      auto val = sk_GENERAL_NAME_value(alt_names, i);
-      if (!val || val->type != type) { continue; }
-
-      auto name =
-          reinterpret_cast<const char *>(ASN1_STRING_get0_data(val->d.ia5));
-      if (name == nullptr) { continue; }
-
-      auto name_len = static_cast<size_t>(ASN1_STRING_length(val->d.ia5));
-
-      switch (type) {
-      case GEN_DNS:
-        dsn_matched =
-            detail::match_hostname(std::string(name, name_len), host_);
-        break;
-
-      case GEN_IPADD:
-        if (!memcmp(&addr6, name, addr_len) || !memcmp(&addr, name, addr_len)) {
-          ip_matched = true;
-        }
-        break;
-      }
-    }
-
-    if (dsn_matched || ip_matched) { ret = true; }
-  }
-
-  GENERAL_NAMES_free(const_cast<STACK_OF(GENERAL_NAME) *>(
-      reinterpret_cast<const STACK_OF(GENERAL_NAME) *>(alt_names)));
-  return ret;
-}
-
-bool SSLClient::verify_host_with_common_name(X509 *server_cert) const {
-  const auto subject_name = X509_get_subject_name(server_cert);
-
-  if (subject_name != nullptr) {
-    char name[BUFSIZ];
-    auto name_len = X509_NAME_get_text_by_NID(subject_name, NID_commonName,
-                                              name, sizeof(name));
-
-    if (name_len != -1) {
-      return detail::match_hostname(
-          std::string(name, static_cast<size_t>(name_len)), host_);
-    }
-  }
-
-  return false;
-}
-
 #endif // CPPHTTPLIB_OPENSSL_SUPPORT
 
 /*
@@ -14622,10 +14565,10 @@ bool verify_hostname(cert_t cert, const char *hostname) {
   auto mcert = static_cast<const mbedtls_x509_crt *>(cert);
   std::string host_str(hostname);
 
-  // Check if hostname is an IP address
-  bool is_ip = impl::is_ipv4_address(host_str);
-  unsigned char ip_bytes[4];
-  if (is_ip) { impl::parse_ipv4(host_str, ip_bytes); }
+  // Check if hostname is an IP address (IPv4 or IPv6)
+  unsigned char ip_bytes[16];
+  auto ip_len = impl::parse_ip_address(host_str, ip_bytes);
+  auto is_ip = ip_len > 0;
 
   // Check Subject Alternative Names (SAN)
   // In Mbed TLS 3.x, subject_alt_names contains raw values without ASN.1 tags
@@ -14637,9 +14580,9 @@ bool verify_hostname(cert_t cert, const char *hostname) {
     size_t len = san->buf.len;
 
     if (is_ip) {
-      // Check if this SAN is an IPv4 address (4 bytes)
-      if (len == 4 && memcmp(p, ip_bytes, 4) == 0) { return true; }
-      // Check if this SAN is an IPv6 address (16 bytes) - skip for now
+      // For an IP host, only a matching iPAddress SAN of the same family
+      // (4 bytes for IPv4, 16 bytes for IPv6) may authenticate it.
+      if (len == ip_len && memcmp(p, ip_bytes, ip_len) == 0) { return true; }
     } else {
       // Check if this SAN is a DNS name (printable ASCII string)
       bool is_dns = len > 0;
@@ -14654,21 +14597,25 @@ bool verify_hostname(cert_t cert, const char *hostname) {
     san = san->next;
   }
 
-  // Fallback: Check Common Name (CN) in subject
-  char cn[256];
-  int ret = mbedtls_x509_dn_gets(cn, sizeof(cn), &mcert->subject);
-  if (ret > 0) {
-    std::string cn_str(cn);
+  // Fallback: Check Common Name (CN) in subject. Skipped for IP-literal hosts:
+  // an IP identity is only valid via an iPAddress SAN, never the CN (RFC 9110;
+  // the OpenSSL backend's X509_check_ip behaves the same way).
+  if (!is_ip) {
+    char cn[256];
+    int ret = mbedtls_x509_dn_gets(cn, sizeof(cn), &mcert->subject);
+    if (ret > 0) {
+      std::string cn_str(cn);
 
-    // Look for "CN=" in the DN string
-    size_t cn_pos = cn_str.find("CN=");
-    if (cn_pos != std::string::npos) {
-      size_t start = cn_pos + 3;
-      size_t end = cn_str.find(',', start);
-      std::string cn_value =
-          cn_str.substr(start, end == std::string::npos ? end : end - start);
+      // Look for "CN=" in the DN string
+      size_t cn_pos = cn_str.find("CN=");
+      if (cn_pos != std::string::npos) {
+        size_t start = cn_pos + 3;
+        size_t end = cn_str.find(',', start);
+        std::string cn_value =
+            cn_str.substr(start, end == std::string::npos ? end : end - start);
 
-      if (detail::match_hostname(cn_value, host_str)) { return true; }
+        if (detail::match_hostname(cn_value, host_str)) { return true; }
+      }
     }
   }
 
@@ -15774,10 +15721,10 @@ bool verify_hostname(cert_t cert, const char *hostname) {
   auto x509 = static_cast<WOLFSSL_X509 *>(cert);
   std::string host_str(hostname);
 
-  // Check if hostname is an IP address
-  bool is_ip = impl::is_ipv4_address(host_str);
-  unsigned char ip_bytes[4];
-  if (is_ip) { impl::parse_ipv4(host_str, ip_bytes); }
+  // Check if hostname is an IP address (IPv4 or IPv6)
+  unsigned char ip_bytes[16];
+  auto ip_len = impl::parse_ip_address(host_str, ip_bytes);
+  auto is_ip = ip_len > 0;
 
   // Check Subject Alternative Names
   auto *san_names = static_cast<WOLF_STACK_OF(WOLFSSL_GENERAL_NAME) *>(
@@ -15804,10 +15751,12 @@ bool verify_hostname(cert_t cert, const char *hostname) {
           }
         }
       } else if (is_ip && names->type == WOLFSSL_GEN_IPADD) {
-        // IP address
+        // IP address: only an iPAddress SAN of the same family (4 bytes for
+        // IPv4, 16 bytes for IPv6) may authenticate the host.
         unsigned char *ip_data = wolfSSL_ASN1_STRING_data(names->d.iPAddress);
-        int ip_len = wolfSSL_ASN1_STRING_length(names->d.iPAddress);
-        if (ip_data && ip_len == 4 && memcmp(ip_data, ip_bytes, 4) == 0) {
+        auto san_ip_len = wolfSSL_ASN1_STRING_length(names->d.iPAddress);
+        if (ip_data && san_ip_len == static_cast<int>(ip_len) &&
+            memcmp(ip_data, ip_bytes, ip_len) == 0) {
           wolfSSL_sk_free(san_names);
           return true;
         }
@@ -15816,8 +15765,10 @@ bool verify_hostname(cert_t cert, const char *hostname) {
     wolfSSL_sk_free(san_names);
   }
 
-  // Fallback: Check Common Name (CN) in subject
-  WOLFSSL_X509_NAME *subject = wolfSSL_X509_get_subject_name(x509);
+  // Fallback: Check Common Name (CN) in subject. Skipped for IP-literal hosts:
+  // an IP identity is only valid via an iPAddress SAN, never the CN (RFC 9110;
+  // the OpenSSL backend's X509_check_ip behaves the same way).
+  auto subject = is_ip ? nullptr : wolfSSL_X509_get_subject_name(x509);
   if (subject) {
     char cn[256] = {};
     int cn_len = wolfSSL_X509_NAME_get_text_by_NID(subject, NID_commonName, cn,

vendor/cpp-httplib/httplib.h

@@ -8,8 +8,8 @@
 #ifndef CPPHTTPLIB_HTTPLIB_H
 #define CPPHTTPLIB_HTTPLIB_H
 
-#define CPPHTTPLIB_VERSION "0.47.0"
-#define CPPHTTPLIB_VERSION_NUM "0x002f00"
+#define CPPHTTPLIB_VERSION "0.48.0"
+#define CPPHTTPLIB_VERSION_NUM "0x003000"
 
 #ifdef _WIN32
 #if defined(_WIN32_WINNT) && _WIN32_WINNT < 0x0A00
@@ -686,18 +686,70 @@ inline from_chars_result<T> from_chars(const char *first, const char *last,
   return {p, std::errc{}};
 }
 
-// from_chars for double (simple wrapper for strtod)
+// from_chars for double (hand-written, locale-independent)
+//
+// The only double consumed by this library is the HTTP quality value, whose
+// grammar is (RFC 9110 12.4.2):
+//   qvalue = ( "0" [ "." 0*3DIGIT ] ) / ( "1" [ "." 0*3("0") ] )
+// i.e. a non-negative decimal with no sign, exponent, "inf"/"nan", or wide
+// magnitude. So this parser recognizes exactly  1*DIGIT [ "." *DIGIT ]  with
+// '.' always the decimal separator (std::strtod would instead read it from the
+// global C locale, mis-parsing q-values once an embedder calls
+// setlocale(LC_ALL, "") into a comma-decimal locale). The caller range-checks
+// the result to [0, 1], so inputs outside that range need not be distinguished
+// here. Allocation-free, single pass, and free of the overflow/rounding edge
+// cases that exponent and wide-range handling would introduce.
 inline from_chars_result<double> from_chars(const char *first, const char *last,
                                             double &value) {
-  std::string s(first, last);
-  char *endptr = nullptr;
-  errno = 0;
-  value = std::strtod(s.c_str(), &endptr);
-  if (endptr == s.c_str()) { return {first, std::errc::invalid_argument}; }
-  if (errno == ERANGE) {
-    return {first + (endptr - s.c_str()), std::errc::result_out_of_range};
+  value = 0.0;
+  const char *p = first;
+
+  // Each 1eN is exactly representable, so a single final division by the
+  // matching entry yields a correctly-rounded result.
+  static const double powers_of_ten[] = {
+      1e0,  1e1,  1e2,  1e3,  1e4,  1e5,  1e6,  1e7,  1e8, 1e9,
+      1e10, 1e11, 1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18};
+  const int max_frac_digits =
+      static_cast<int>(sizeof(powers_of_ten) / sizeof(powers_of_ten[0])) - 1;
+
+  // Accumulate digits into a 64-bit integer and remember how many were
+  // fractional. Two independent caps keep this bounded and safe:
+  //   * accumulation saturates before mantissa could overflow uint64_t, and
+  //   * frac_digits is capped at max_frac_digits so it is always a valid index
+  //     into powers_of_ten (without this an input like "0.000...0" would never
+  //     grow mantissa, so the saturation cap alone would not bound it).
+  // Both caps only drop digits far beyond the precision a q-value needs; any
+  // value they would change is well outside [0, 1] and rejected by the caller.
+  uint64_t mantissa = 0;
+  int frac_digits = 0;
+  bool seen_digit = false;
+
+  const uint64_t limit = ((std::numeric_limits<uint64_t>::max)() - 9) / 10;
+  auto accumulate = [&](char c) {
+    if (mantissa <= limit) {
+      mantissa = mantissa * 10 + static_cast<uint64_t>(c - '0');
+      return true;
+    }
+    return false;
+  };
+
+  for (; p != last && '0' <= *p && *p <= '9'; ++p) {
+    seen_digit = true;
+    accumulate(*p);
   }
-  return {first + (endptr - s.c_str()), std::errc{}};
+
+  if (p != last && *p == '.') {
+    ++p;
+    for (; p != last && '0' <= *p && *p <= '9'; ++p) {
+      seen_digit = true;
+      if (frac_digits < max_frac_digits && accumulate(*p)) { ++frac_digits; }
+    }
+  }
+
+  if (!seen_digit) { return {first, std::errc::invalid_argument}; }
+
+  value = static_cast<double>(mantissa) / powers_of_ten[frac_digits];
+  return {p, std::errc{}};
 }
 
 inline bool parse_port(const char *s, size_t len, int &port) {
@@ -2826,13 +2878,6 @@ private:
 #endif
 
   friend class ClientImpl;
-
-#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
-private:
-  bool verify_host(X509 *server_cert) const;
-  bool verify_host_with_subject_alt_name(X509 *server_cert) const;
-  bool verify_host_with_common_name(X509 *server_cert) const;
-#endif
 };
 #endif // CPPHTTPLIB_SSL_ENABLED