Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.github/workflows/build.yml
#	.github/workflows/release.yml
#	.github/workflows/winget.yml
#	CODEOWNERS
#	README.md
#	ci/run.sh
#	docs/build.md
#	docs/ops.md
#	docs/ops/Vulkan.csv
#	ggml/CMakeLists.txt
#	ggml/src/ggml-cann/ggml-cann.cpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	scripts/sync_vendor.py
#	src/CMakeLists.txt
#	tests/test-json-schema-to-grammar.cpp
#	tests/test-quantize-stats.cpp
#	tools/server/CMakeLists.txt
#	tools/server/README.md

common/arg.cpp

@@ -214,13 +214,13 @@ struct handle_model_result {
 static handle_model_result common_params_handle_model(
         struct common_params_model & model,
         const std::string & bearer_token,
-        const std::string & model_path_default,
         bool offline) {
     handle_model_result result;
     // handle pre-fill default model path and url based on hf_repo and hf_file
     {
         if (!model.docker_repo.empty()) {  // Handle Docker URLs by resolving them to local paths
             model.path = common_docker_resolve_model(model.docker_repo);
+            model.name = model.docker_repo; // set name for consistency
         } else if (!model.hf_repo.empty()) {
             // short-hand to avoid specifying --hf-file -> default it to --model
             if (model.hf_file.empty()) {
@@ -229,7 +229,8 @@ static handle_model_result common_params_handle_model(
                     if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) {
                         exit(1); // built without CURL, error message already printed
                     }
-                    model.hf_repo = auto_detected.repo;
+                    model.name    = model.hf_repo;      // repo name with tag
+                    model.hf_repo = auto_detected.repo; // repo name without tag
                     model.hf_file = auto_detected.ggufFile;
                     if (!auto_detected.mmprojFile.empty()) {
                         result.found_mmproj   = true;
@@ -259,8 +260,6 @@ static handle_model_result common_params_handle_model(
                 model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
             }
 
-        } else if (model.path.empty()) {
-            model.path = model_path_default;
         }
     }
 
@@ -407,7 +406,7 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
 
     // handle model and download
     {
-        auto res = common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH, params.offline);
+        auto res = common_params_handle_model(params.model, params.hf_token, params.offline);
         if (params.no_mmproj) {
             params.mmproj = {};
         } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
@@ -417,12 +416,18 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         // only download mmproj if the current example is using it
         for (auto & ex : mmproj_examples) {
             if (ctx_arg.ex == ex) {
-                common_params_handle_model(params.mmproj,    params.hf_token, "", params.offline);
+                common_params_handle_model(params.mmproj,    params.hf_token, params.offline);
                 break;
             }
         }
-        common_params_handle_model(params.speculative.model, params.hf_token, "", params.offline);
-        common_params_handle_model(params.vocoder.model,     params.hf_token, "", params.offline);
+        common_params_handle_model(params.speculative.model, params.hf_token, params.offline);
+        common_params_handle_model(params.vocoder.model,     params.hf_token, params.offline);
+    }
+
+    // model is required (except for server)
+    // TODO @ngxson : maybe show a list of available models in CLI in this case
+    if (params.model.path.empty() && ctx_arg.ex != LLAMA_EXAMPLE_SERVER) {
+        throw std::invalid_argument("error: --model is required\n");
     }
 
     if (params.escape) {
@@ -1223,7 +1228,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -2092,11 +2097,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"-m", "--model"}, "FNAME",
         ex == LLAMA_EXAMPLE_EXPORT_LORA
-            ? std::string("model path from which to load base model")
-            : string_format(
-                "model path (default: `models/$filename` with filename from `--hf-file` "
-                "or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH
-            ),
+            ? "model path from which to load base model"
+            : "model path to load",
         [](common_params & params, const std::string & value) {
             params.model.path = value;
         }
@@ -2488,12 +2490,50 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "path to save slot kv cache (default: disabled)",
         [](common_params & params, const std::string & value) {
             params.slot_save_path = value;
+            if (!fs_is_directory(params.slot_save_path)) {
+                throw std::invalid_argument("not a directory: " + value);
+            }
             // if doesn't end with DIRECTORY_SEPARATOR, add it
             if (!params.slot_save_path.empty() && params.slot_save_path[params.slot_save_path.size() - 1] != DIRECTORY_SEPARATOR) {
                 params.slot_save_path += DIRECTORY_SEPARATOR;
             }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--media-path"}, "PATH",
+        "directory for loading local media files; files can be accessed via file:// URLs using relative paths (default: disabled)",
+        [](common_params & params, const std::string & value) {
+            params.media_path = value;
+            if (!fs_is_directory(params.media_path)) {
+                throw std::invalid_argument("not a directory: " + value);
+            }
+            // if doesn't end with DIRECTORY_SEPARATOR, add it
+            if (!params.media_path.empty() && params.media_path[params.media_path.size() - 1] != DIRECTORY_SEPARATOR) {
+                params.media_path += DIRECTORY_SEPARATOR;
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--models-dir"}, "PATH",
+        "directory containing models for the router server (default: disabled)",
+        [](common_params & params, const std::string & value) {
+            params.models_dir = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODELS_DIR"));
+    add_opt(common_arg(
+        {"--models-max"}, "N",
+        string_format("for router server, maximum number of models to load simultaneously (default: %d, 0 = unlimited)", params.models_max),
+        [](common_params & params, int value) {
+            params.models_max = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODELS_MAX"));
+    add_opt(common_arg(
+        {"--no-models-autoload"},
+        "disables automatic loading of models (default: enabled)",
+        [](common_params & params) {
+            params.models_autoload = false;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_MODELS_AUTOLOAD"));
     add_opt(common_arg(
         {"--jinja"},
         string_format("use jinja template for chat (default: %s)\n", params.use_jinja ? "enabled" : "disabled"),
@@ -2676,7 +2716,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_OFFLINE"));
     add_opt(common_arg(
         {"-lv", "--verbosity", "--log-verbosity"}, "N",
-        "Set the verbosity threshold. Messages with a higher verbosity will be ignored.",
+        string_format("Set the verbosity threshold. Messages with a higher verbosity will be ignored. Values:\n"
+            " - 0: generic output\n"
+            " - 1: error\n"
+            " - 2: warning\n"
+            " - 3: info\n"
+            " - 4: debug\n"
+            "(default: %d)\n", params.verbosity),
         [](common_params & params, int value) {
             params.verbosity = value;
             common_log_set_verbosity_thold(value);

common/chat.cpp

@@ -163,7 +163,7 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
     if (tool_choice == "required") {
         return COMMON_CHAT_TOOL_CHOICE_REQUIRED;
     }
-    throw std::runtime_error("Invalid tool_choice: " + tool_choice);
+    throw std::invalid_argument("Invalid tool_choice: " + tool_choice);
 }
 
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates) {
@@ -186,17 +186,17 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
     try {
 
         if (!messages.is_array()) {
-            throw std::runtime_error("Expected 'messages' to be an array, got " + messages.dump());
+            throw std::invalid_argument("Expected 'messages' to be an array, got " + messages.dump());
         }
 
         for (const auto & message : messages) {
             if (!message.is_object()) {
-                throw std::runtime_error("Expected 'message' to be an object, got " + message.dump());
+                throw std::invalid_argument("Expected 'message' to be an object, got " + message.dump());
             }
 
             common_chat_msg msg;
             if (!message.contains("role")) {
-                throw std::runtime_error("Missing 'role' in message: " + message.dump());
+                throw std::invalid_argument("Missing 'role' in message: " + message.dump());
             }
             msg.role = message.at("role");
 
@@ -209,11 +209,11 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
                 } else if (content.is_array()) {
                     for (const auto & part : content) {
                         if (!part.contains("type")) {
-                            throw std::runtime_error("Missing content part type: " + part.dump());
+                            throw std::invalid_argument("Missing content part type: " + part.dump());
                         }
                         const auto & type = part.at("type");
                         if (type != "text") {
-                            throw std::runtime_error("Unsupported content part type: " + type.dump());
+                            throw std::invalid_argument("Unsupported content part type: " + type.dump());
                         }
                         common_chat_msg_content_part msg_part;
                         msg_part.type = type;
@@ -221,25 +221,25 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
                         msg.content_parts.push_back(msg_part);
                     }
                 } else if (!content.is_null()) {
-                    throw std::runtime_error("Invalid 'content' type: expected string or array, got " + content.dump() + " (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
+                    throw std::invalid_argument("Invalid 'content' type: expected string or array, got " + content.dump() + " (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
                 }
             }
             if (has_tool_calls) {
                 for (const auto & tool_call : message.at("tool_calls")) {
                     common_chat_tool_call tc;
                     if (!tool_call.contains("type")) {
-                        throw std::runtime_error("Missing tool call type: " + tool_call.dump());
+                        throw std::invalid_argument("Missing tool call type: " + tool_call.dump());
                     }
                     const auto & type = tool_call.at("type");
                     if (type != "function") {
-                        throw std::runtime_error("Unsupported tool call type: " + tool_call.dump());
+                        throw std::invalid_argument("Unsupported tool call type: " + tool_call.dump());
                     }
                     if (!tool_call.contains("function")) {
-                        throw std::runtime_error("Missing tool call function: " + tool_call.dump());
+                        throw std::invalid_argument("Missing tool call function: " + tool_call.dump());
                     }
                     const auto & fc = tool_call.at("function");
                     if (!fc.contains("name")) {
-                        throw std::runtime_error("Missing tool call name: " + tool_call.dump());
+                        throw std::invalid_argument("Missing tool call name: " + tool_call.dump());
                     }
                     tc.name = fc.at("name");
                     tc.arguments = fc.at("arguments");
@@ -250,7 +250,7 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
                 }
             }
             if (!has_content && !has_tool_calls) {
-                throw std::runtime_error("Expected 'content' or 'tool_calls' (ref: https://github.com/ggml-org/llama.cpp/issues/8367 & https://github.com/ggml-org/llama.cpp/issues/12279)");
+                throw std::invalid_argument("Expected 'content' or 'tool_calls' (ref: https://github.com/ggml-org/llama.cpp/issues/8367 & https://github.com/ggml-org/llama.cpp/issues/12279)");
             }
             if (message.contains("reasoning_content")) {
                 msg.reasoning_content = message.at("reasoning_content");
@@ -353,18 +353,18 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     try {
         if (!tools.is_null()) {
             if (!tools.is_array()) {
-                throw std::runtime_error("Expected 'tools' to be an array, got " + tools.dump());
+                throw std::invalid_argument("Expected 'tools' to be an array, got " + tools.dump());
             }
             for (const auto & tool : tools) {
                 if (!tool.contains("type")) {
-                    throw std::runtime_error("Missing tool type: " + tool.dump());
+                    throw std::invalid_argument("Missing tool type: " + tool.dump());
                 }
                 const auto & type = tool.at("type");
                 if (!type.is_string() || type != "function") {
-                    throw std::runtime_error("Unsupported tool type: " + tool.dump());
+                    throw std::invalid_argument("Unsupported tool type: " + tool.dump());
                 }
                 if (!tool.contains("function")) {
-                    throw std::runtime_error("Missing tool function: " + tool.dump());
+                    throw std::invalid_argument("Missing tool function: " + tool.dump());
                 }
 
                 const auto & function = tool.at("function");

common/common.cpp

@@ -702,7 +702,7 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
 
 // Validate if a filename is safe to use
 // To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
-bool fs_validate_filename(const std::string & filename) {
+bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
     if (!filename.length()) {
         // Empty filename invalid
         return false;
@@ -762,10 +762,14 @@ bool fs_validate_filename(const std::string & filename) {
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
-            || c == '/' || c == '\\' || c == ':' || c == '*' // Illegal characters
+            || c == ':' || c == '*' // Illegal characters
             || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
             return false;
         }
+        if (!allow_subdirs && (c == '/' || c == '\\')) {
+            // Subdirectories not allowed, reject path separators
+            return false;
+        }
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -867,6 +871,11 @@ bool fs_create_directory_with_parents(const std::string & path) {
 #endif // _WIN32
 }
 
+bool fs_is_directory(const std::string & path) {
+    std::filesystem::path dir(path);
+    return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
+}
+
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -920,7 +929,7 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
-std::vector<common_file_info> fs_list_files(const std::string & path) {
+std::vector<common_file_info> fs_list(const std::string & path, bool include_directories) {
     std::vector<common_file_info> files;
     if (path.empty()) return files;
 
@@ -935,14 +944,22 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
             const auto & p = entry.path();
             if (std::filesystem::is_regular_file(p)) {
                 common_file_info info;
-                info.path = p.string();
-                info.name = p.filename().string();
+                info.path   = p.string();
+                info.name   = p.filename().string();
+                info.is_dir = false;
                 try {
                     info.size = static_cast<size_t>(std::filesystem::file_size(p));
                 } catch (const std::filesystem::filesystem_error &) {
                     info.size = 0;
                 }
                 files.push_back(std::move(info));
+            } else if (include_directories && std::filesystem::is_directory(p)) {
+                common_file_info info;
+                info.path   = p.string();
+                info.name   = p.filename().string();
+                info.size   = 0; // Directories have no size
+                info.is_dir = true;
+                files.push_back(std::move(info));
             }
         } catch (const std::filesystem::filesystem_error &) {
             // skip entries we cannot inspect

common/common.h

@@ -26,8 +26,6 @@
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
-#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
-
 struct common_time_meas {
     common_time_meas(int64_t & t_acc, bool disable = false);
     ~common_time_meas();
@@ -219,6 +217,7 @@ 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
 };
 
 struct common_params_speculative {
@@ -365,7 +364,7 @@ struct common_params {
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
-    int32_t verbosity                  = 0;
+    int32_t verbosity                  = 3;  // LOG_LEVEL_INFO
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
     bool    offline                    = false;
@@ -474,9 +473,15 @@ struct common_params {
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
+    // router server configs
+    std::string models_dir = ""; // directory containing models for the router server
+    int models_max = 4;          // maximum number of models to load simultaneously
+    bool models_autoload = true; // automatically load models when requested via the router server
+
     bool log_json = false;
 
     std::string slot_save_path;
+    std::string media_path; // path to directory for loading media files
 
     float slot_prompt_similarity = 0.1f;
 
@@ -627,8 +632,9 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 // Filesystem utils
 //
 
-bool fs_validate_filename(const std::string & filename);
+bool fs_validate_filename(const std::string & filename, bool allow_subdirs = false);
 bool fs_create_directory_with_parents(const std::string & path);
+bool fs_is_directory(const std::string & path);
 
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
@@ -637,8 +643,9 @@ struct common_file_info {
     std::string path;
     std::string name;
     size_t      size = 0; // in bytes
+    bool        is_dir = false;
 };
-std::vector<common_file_info> fs_list_files(const std::string & path);
+std::vector<common_file_info> fs_list(const std::string & path, bool include_directories);
 
 //
 // Model utils

common/download.cpp

@@ -430,7 +430,7 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
     curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
     curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
     curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
-    curl_easy_setopt(curl.get(), CURLOPT_VERBOSE, 1L);
+    curl_easy_setopt(curl.get(), CURLOPT_VERBOSE, 0L);
     typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
     auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
         auto data_vec = static_cast<std::vector<char> *>(data);
@@ -1054,7 +1054,7 @@ std::string common_docker_resolve_model(const std::string &) {
 std::vector<common_cached_model_info> common_list_cached_models() {
     std::vector<common_cached_model_info> models;
     const std::string cache_dir = fs_get_cache_directory();
-    const std::vector<common_file_info> files = fs_list_files(cache_dir);
+    const std::vector<common_file_info> files = fs_list(cache_dir, false);
     for (const auto & file : files) {
         if (string_starts_with(file.name, "manifest=") && string_ends_with(file.name, ".json")) {
             common_cached_model_info model_info;

common/download.h

@@ -14,8 +14,10 @@ struct common_cached_model_info {
     std::string model;
     std::string tag;
     size_t      size = 0; // GGUF size in bytes
+    // return string representation like "user/model:tag"
+    // if tag is "latest", it will be omitted
     std::string to_string() const {
-        return user + "/" + model + ":" + tag;
+        return user + "/" + model + (tag == "latest" ? "" : ":" + tag);
     }
 };
 

common/json-schema-to-grammar.cpp

@@ -974,7 +974,7 @@ public:
 
     void check_errors() {
         if (!_errors.empty()) {
-            throw std::runtime_error("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
+            throw std::invalid_argument("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
         }
         if (!_warnings.empty()) {
             fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", string_join(_warnings, "; ").c_str());

common/log.cpp

@@ -443,8 +443,22 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
 
+static int common_get_verbosity(enum ggml_log_level level) {
+    switch (level) {
+        case GGML_LOG_LEVEL_DEBUG: return LOG_LEVEL_DEBUG;
+        case GGML_LOG_LEVEL_INFO:  return LOG_LEVEL_INFO;
+        case GGML_LOG_LEVEL_WARN:  return LOG_LEVEL_WARN;
+        case GGML_LOG_LEVEL_ERROR: return LOG_LEVEL_ERROR;
+        case GGML_LOG_LEVEL_CONT:  return LOG_LEVEL_INFO; // same as INFO
+        case GGML_LOG_LEVEL_NONE:
+        default:
+            return LOG_LEVEL_OUTPUT;
+    }
+}
+
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
-    if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+    auto verbosity = common_get_verbosity(level);
+    if (verbosity <= common_log_verbosity_thold) {
         common_log_add(common_log_main(), level, "%s", text);
     }
 }

common/log.h

@@ -21,8 +21,14 @@
 #    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
-#define LOG_DEFAULT_DEBUG 1
-#define LOG_DEFAULT_LLAMA 0
+#define LOG_LEVEL_DEBUG  4
+#define LOG_LEVEL_INFO   3
+#define LOG_LEVEL_WARN   2
+#define LOG_LEVEL_ERROR  1
+#define LOG_LEVEL_OUTPUT 0 // output data from tools
+
+#define LOG_DEFAULT_DEBUG LOG_LEVEL_DEBUG
+#define LOG_DEFAULT_LLAMA LOG_LEVEL_INFO
 
 enum log_colors {
     LOG_COLORS_AUTO     = -1,
@@ -67,10 +73,11 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
 //   0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
 //   0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
 //
-// I - info    (stdout, V = 0)
-// W - warning (stderr, V = 0)
-// E - error   (stderr, V = 0)
 // D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
+// I - info    (stdout, V = LOG_DEFAULT_INFO)
+// W - warning (stderr, V = LOG_DEFAULT_WARN)
+// E - error   (stderr, V = LOG_DEFAULT_ERROR)
+// O - output  (stdout, V = LOG_DEFAULT_OUTPUT)
 //
 
 void common_log_set_file      (struct common_log * log, const char * file); // not thread-safe
@@ -95,14 +102,14 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps);   // w
         } \
     } while (0)
 
-#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, 0,         __VA_ARGS__)
-#define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity, __VA_ARGS__)
+#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, LOG_LEVEL_OUTPUT, __VA_ARGS__)
+#define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity,        __VA_ARGS__)
 
-#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  0,                 __VA_ARGS__)
-#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
+#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_LEVEL_DEBUG,  __VA_ARGS__)
+#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  LOG_LEVEL_INFO,   __VA_ARGS__)
+#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  LOG_LEVEL_WARN,   __VA_ARGS__)
+#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, LOG_LEVEL_ERROR,  __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  LOG_LEVEL_INFO,   __VA_ARGS__) // same as INFO
 
 #define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
 #define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)

convert_hf_to_gguf.py

@@ -1581,10 +1581,27 @@ class MmprojModel(ModelBase):
 
         # load preprocessor config
         self.preprocessor_config = {}
-        if not self.is_mistral_format:
-            with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
+
+        # prefer preprocessor_config.json if possible
+        preprocessor_config_path = self.dir_model / "preprocessor_config.json"
+        if preprocessor_config_path.is_file():
+            with open(preprocessor_config_path, "r", encoding="utf-8") as f:
                 self.preprocessor_config = json.load(f)
 
+        # prefer processor_config.json if possible
+        processor_config_path = self.dir_model / "processor_config.json"
+        if processor_config_path.is_file():
+            with open(processor_config_path, "r", encoding="utf-8") as f:
+                cfg = json.load(f)
+                # move image_processor to root level for compat
+                if "image_processor" in cfg:
+                    cfg = {
+                        **cfg,
+                        **cfg["image_processor"],
+                    }
+                # merge configs
+                self.preprocessor_config = {**self.preprocessor_config, **cfg}
+
     def get_vision_config(self) -> dict[str, Any] | None:
         config_name = "vision_config" if not self.is_mistral_format else "vision_encoder"
         return self.global_config.get(config_name)
@@ -2797,9 +2814,38 @@ class Llama4VisionModel(MmprojModel):
 
 @ModelBase.register("Mistral3ForConditionalGeneration")
 class Mistral3Model(LlamaModel):
-    model_arch = gguf.MODEL_ARCH.LLAMA
+    model_arch = gguf.MODEL_ARCH.MISTRAL3
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # for compatibility, we use LLAMA arch for older models
+        # TODO: remove this once everyone has migrated to newer version of llama.cpp
+        if self.hparams.get("model_type") != "ministral3":
+            self.model_arch = gguf.MODEL_ARCH.LLAMA
+            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
+            self.gguf_writer.add_architecture()
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        rope_params = self.hparams.get("rope_parameters")
+        if self.hparams.get("model_type") == "ministral3":
+            assert rope_params is not None, "ministral3 must have 'rope_parameters' config"
+            assert rope_params["rope_type"] == "yarn", "ministral3 rope_type must be 'yarn'"
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_params["factor"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_fast(rope_params["beta_fast"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_slow(rope_params["beta_slow"])
+            self.gguf_writer.add_rope_scaling_yarn_log_mul(rope_params["mscale_all_dim"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_params["original_max_position_embeddings"])
+            self.gguf_writer.add_rope_freq_base(rope_params["rope_theta"])
+            self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        # TODO: probably not worth supporting quantized weight, as official BF16 is also available
+        if name.endswith("weight_scale_inv"):
+            raise ValueError("This is a quantized weight, please use BF16 weight instead")
+
         name = name.replace("language_model.", "")
         if "multi_modal_projector" in name or "vision_tower" in name:
             return []
@@ -9809,12 +9855,22 @@ class ApertusModel(LlamaModel):
 
 
 class MistralModel(LlamaModel):
-    model_arch = gguf.MODEL_ARCH.LLAMA
+    model_arch = gguf.MODEL_ARCH.MISTRAL3
     model_name = "Mistral"
     hf_arch = ""
     is_mistral_format = True
     undo_permute = False
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # for compatibility, we use LLAMA arch for older models
+        # TODO: remove this once everyone migrates to newer version of llama.cpp
+        if "llama_4_scaling" not in self.hparams:
+            self.model_arch = gguf.MODEL_ARCH.LLAMA
+            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
+            self.gguf_writer.add_architecture()
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
     @staticmethod
     def get_community_chat_template(vocab: MistralVocab, templates_dir: Path, is_mistral_format: bool):
         assert TokenizerVersion is not None and Tekkenizer is not None and SentencePieceTokenizer is not None, _mistral_import_error_msg
@@ -9854,6 +9910,20 @@ class MistralModel(LlamaModel):
 
         return template
 
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if "yarn" in self.hparams:
+            yarn_params = self.hparams["yarn"]
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(yarn_params["factor"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_fast(yarn_params["beta"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_slow(yarn_params["alpha"])
+            self.gguf_writer.add_rope_scaling_yarn_log_mul(1.0) # mscale_all_dim
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(yarn_params["original_max_position_embeddings"])
+
+        if "llama_4_scaling" in self.hparams:
+            self.gguf_writer.add_attn_temperature_scale(self.hparams["llama_4_scaling"]["beta"])
+
 
 class PixtralModel(LlavaVisionModel):
     model_name = "Pixtral"

embd_res/klite.embd

@@ -4050,6 +4050,15 @@ Current version indicated by LITEVER below.
 		"system":"<|start_of_role|>system<|end_of_role|>",
 		"system_end":"<|end_of_text|>\\n",
 	},
+	{
+		"name":"Kimi ChatML",
+		"user":"<|im_user|>user<|im_middle|>",
+		"user_end":"<|im_end|>",
+		"assistant":"<|im_assistant|>assistant<|im_middle|>",
+		"assistant_end":"<|im_end|>",
+		"system":"<|im_system|>system<|im_middle|>",
+		"system_end":"<|im_end|>",
+	},
 	{
 		"name":"Llama 2 Chat",
 		"user":"[INST] ",
@@ -21925,7 +21934,7 @@ Current version indicated by LITEVER below.
 			let box2 = document.getElementById("chat_msg_body");
 			let box3 = document.getElementById("corpostylemain");
 			function isScrolledToBottom(element) {
-				return element.scrollHeight - element.scrollTop - element.clientHeight <= (autoscrollEnabled?220:60);
+				return element.scrollHeight - element.scrollTop - element.clientHeight <= (autoscrollEnabled?180:60);
 			}
 			suppressScrollEvent = true;
 			if(alwaysscroll || isScrolledToBottom(box1))

ggml/src/ggml-backend.cpp

@@ -1246,10 +1246,8 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                                 tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
                                 ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
                             }
-                            if (sched->n_copies > 1) {
-                                ggml_set_input(tensor_copy);
-                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
-                            }
+                            ggml_set_input(tensor_copy);
+                            ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             tensor_id_copy(src_id, src_backend_id, c) = tensor_copy;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }

ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp

@@ -8,6 +8,10 @@
 #include <sys/sysctl.h>
 #endif
 
+#if !defined(HWCAP2_SVE2)
+#define HWCAP2_SVE2 (1 << 1)
+#endif
+
 #if !defined(HWCAP2_I8MM)
 #define HWCAP2_I8MM (1 << 13)
 #endif

ggml/src/ggml-cpu/ggml-cpu.c

@@ -688,22 +688,14 @@ bool ggml_is_numa(void) {
 }
 
 #if defined(__ARM_ARCH)
-
-#if defined(__linux__) && defined(__aarch64__)
-#include <sys/auxv.h>
-#endif
-
-static void ggml_init_arm_arch_features(void) {
 #if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
-#if defined(__linux__)
-    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
-#else
-    // TODO: add support of SVE for non-linux systems
-#error "TODO: SVE is not supported on this platform. To use SVE, sve_cnt needs to be initialized here."
-#endif
-#endif
+#include <arm_sve.h>
+static void ggml_init_arm_arch_features(void) {
+    ggml_arm_arch_features.sve_cnt = svcntb();
 }
-
+#else
+static void ggml_init_arm_arch_features(void) {}
+#endif
 #endif // __ARM_ARCH
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {

ggml/src/ggml-cuda/common.cuh

@@ -994,6 +994,10 @@ struct ggml_cuda_concurrent_event {
     int                                          n_streams = 0;
     std::unordered_map<const ggml_tensor *, int> stream_mapping;
 
+    // Original order of nodes in this concurrent region (before interleaving)
+    // Used to restore grouping for fusion within streams
+    std::vector<const ggml_tensor *> original_order;
+
     const ggml_tensor * join_node;
 
     ggml_cuda_concurrent_event() = default;
@@ -1016,6 +1020,7 @@ struct ggml_cuda_concurrent_event {
     , fork_event(other.fork_event)
     , n_streams(other.n_streams)
     , stream_mapping(std::move(other.stream_mapping))
+    , original_order(std::move(other.original_order))
     , join_node(other.join_node) {
         other.fork_event = nullptr;
     }
@@ -1126,11 +1131,9 @@ struct ggml_cuda_concurrent_event {
 };
 
 struct ggml_cuda_stream_context {
-    std::vector<const ggml_tensor *>                                    original_nodes;
     std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> concurrent_events;
 
     void reset() {
-        original_nodes.clear();
         concurrent_events.clear();
     }
 };

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -3251,9 +3251,56 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                 }
             }
             if (should_launch_concurrent_events) {
-                //Restore the original graph to enable fusion within the streams
-                cgraph->nodes   = const_cast<ggml_tensor **>(stream_ctx.original_nodes.data());
-                cgraph->n_nodes = (int) stream_ctx.original_nodes.size();
+                // Restore original node order within each concurrent region to enable fusion within streams
+
+                std::unordered_map<const ggml_tensor *, int> node_to_idx;
+                node_to_idx.reserve(cgraph->n_nodes);
+                for (int i = 0; i < cgraph->n_nodes; ++i) {
+                    node_to_idx[cgraph->nodes[i]] = i;
+                }
+
+                for (auto & [fork_node, event] : stream_ctx.concurrent_events) {
+                    // Find positions of all nodes from this event in the current graph
+                    std::vector<int> positions;
+                    positions.reserve(event.original_order.size());
+
+                    bool all_found = true;
+                    for (const ggml_tensor * orig_node : event.original_order) {
+                        auto it = node_to_idx.find(orig_node);
+                        if (it != node_to_idx.end()) {
+                            positions.push_back(it->second);
+                        } else {
+                            all_found = false;
+                            break;
+                        }
+                    }
+
+                    if (!all_found || positions.size() != event.original_order.size()) {
+                        continue;
+                    }
+
+                    // Sort positions to get contiguous range
+                    std::vector<int> sorted_positions = positions;
+                    std::sort(sorted_positions.begin(), sorted_positions.end());
+
+                    bool is_contiguous = true;
+                    for (size_t i = 1; i < sorted_positions.size(); ++i) {
+                        if (sorted_positions[i] != sorted_positions[i-1] + 1) {
+                            is_contiguous = false;
+                            break;
+                        }
+                    }
+
+                    if (!is_contiguous) {
+                        continue;
+                    }
+
+                    // Restore original order at the sorted positions
+                    int start_pos = sorted_positions[0];
+                    for (size_t i = 0; i < event.original_order.size(); ++i) {
+                        cgraph->nodes[start_pos + i] = const_cast<ggml_tensor *>(event.original_order[i]);
+                    }
+                }
             }
 
             for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -3818,14 +3865,6 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
     // store {fork_idx, join_idx}
     std::vector<std::pair<int, int>> concurrent_node_ranges;
 
-    // save the original nodes
-    std::vector<const ggml_tensor *> original_nodes;
-    original_nodes.reserve(cgraph->n_nodes);
-    for (int i = 0; i < cgraph->n_nodes; ++i) {
-        original_nodes.push_back(cgraph->nodes[i]);
-    }
-    cuda_ctx->stream_context().original_nodes = std::move(original_nodes);
-
     for (const auto & [root_node, count] : fan_out) {
         if (count >= min_fan_out && count <= max_fan_out) {
             const int root_node_idx = node_indices[root_node];
@@ -3930,6 +3969,13 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
                     continue;
                 }
 
+                // Save the original order of nodes in this region before interleaving
+                // This is used later to restore grouping for fusion within streams
+                concurrent_event.original_order.reserve(total_branch_nodes);
+                for (int i = fork_node_idx + 1; i < join_node_idx; ++i) {
+                    concurrent_event.original_order.push_back(cgraph->nodes[i]);
+                }
+
                 std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> & concurrent_events = cuda_ctx->stream_context().concurrent_events;
                 GGML_ASSERT(concurrent_events.find(root_node) == concurrent_events.end());
                 concurrent_events.emplace(root_node, std::move(concurrent_event));

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1243,6 +1243,7 @@ struct vk_op_topk_push_constants {
     uint32_t orig_ncols;
     uint32_t ncols_input;
     uint32_t ncols_output;
+    uint32_t k;
     uint32_t nrows;
     uint32_t first_pass;
     uint32_t last_pass;
@@ -1689,6 +1690,14 @@ class vk_perf_logger {
             timings[name.str()].push_back(time);
             return;
         }
+        if (node->op == GGML_OP_TOP_K) {
+            std::stringstream name;
+            name << ggml_op_name(node->op) <<
+                " K=" << node->ne[0] <<
+                " (" << node->src[0]->ne[0] << "," << node->src[0]->ne[1] << "," << node->src[0]->ne[2] << "," << node->src[0]->ne[3] << ")";
+            timings[name.str()].push_back(time);
+            return;
+        }
         timings[ggml_op_name(node->op)].push_back(time);
     }
   private:
@@ -10375,17 +10384,8 @@ static void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     uint32_t nrows = ggml_nrows(src0);
     uint32_t k = dst->ne[0];
 
-    vk_op_topk_push_constants pc { ncols, ncols, k, nrows, 0, 0 };
+    vk_op_topk_push_constants pc { ncols, ncols, ncols, k, nrows, 0, 0 };
 
-    // Reserve space for ivec2 per element, double buffered
-    const size_t dbl_buf_size = size_t{ncols} * nrows * 2 * sizeof(int);
-    const size_t x_sz = dbl_buf_size * 2;
-    uint32_t dbl_buf_index = 0;
-
-    if (ctx->prealloc_size_x < x_sz) {
-        ctx->prealloc_size_x = x_sz;
-        ggml_vk_preallocate_buffers(ctx, subctx);
-    }
     if (ctx->prealloc_x_need_sync) {
         ggml_vk_sync_buffers(ctx, subctx);
     }
@@ -10400,8 +10400,9 @@ static void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     // largest elements. Repeat until we have the top K elements.
     // Need to do at least one iteration to write out the results.
     bool done_one_iter = false;
+    uint32_t dbl_buf_index = 0;
+    size_t dbl_buf_size;
     while (num_elements > k || !done_one_iter) {
-        done_one_iter = true;
 
         // Prefer going as small as num_topk_pipelines - 3 for perf reasons.
         // But if K is larger, then we need a larger workgroup
@@ -10441,6 +10442,21 @@ static void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, cons
         // Number of elements remaining after this pass
         uint32_t num_dst_elements = (num_elements / pipeline->wg_denoms[0]) * k + std::min(k, num_elements % pipeline->wg_denoms[0]);
 
+        pc2.ncols_output = num_dst_elements;
+
+        if (!done_one_iter) {
+            // Reserve space for ivec2 per element, double buffered
+            // K per workgroup per row
+            dbl_buf_size = num_dst_elements * nrows * 2 * sizeof(int);
+            dbl_buf_size = ROUNDUP_POW2(dbl_buf_size, ctx->device->properties.limits.minStorageBufferOffsetAlignment);
+            const size_t x_sz = dbl_buf_size * 2;
+
+            if (ctx->prealloc_size_x < x_sz) {
+                ctx->prealloc_size_x = x_sz;
+                ggml_vk_preallocate_buffers(ctx, subctx);
+            }
+        }
+
         vk_subbuffer src_buf;
         vk_subbuffer dst_buf;
 
@@ -10466,6 +10482,7 @@ static void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, cons
         if (num_elements > k) {
             ggml_vk_sync_buffers(ctx, subctx);
         }
+        done_one_iter = true;
     }
     ctx->prealloc_x_need_sync = true;
 }

ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp

@@ -19,6 +19,7 @@ layout (push_constant) uniform parameter {
     uint orig_ncols;
     uint ncols_input;
     uint ncols_output;
+    uint k;
     uint nrows;
     uint first_pass;
     uint last_pass;
@@ -36,7 +37,7 @@ void topk(bool needs_bounds_check, const uint row) {
             const uint row_offset = row * p.ncols_input;
             dst_row[col] = ivec2(gl_GlobalInvocationID.x, floatBitsToInt(data_a[row_offset + gl_GlobalInvocationID.x]));
         } else {
-            const uint row_offset = row * p.orig_ncols;
+            const uint row_offset = row * p.ncols_input;
             dst_row[col] = data_s[row_offset + gl_GlobalInvocationID.x];
         }
     } else {
@@ -44,7 +45,7 @@ void topk(bool needs_bounds_check, const uint row) {
     }
     barrier();
 
-    if (p.ncols_output == 1) {
+    if (p.k == 1) {
         // Fast path for single output - just do a max reduction
         [[unroll]] for (int s = BLOCK_SIZE / 2; s >= 1; s /= 2) {
             if (col < s) {
@@ -84,13 +85,17 @@ void topk(bool needs_bounds_check, const uint row) {
         }
     }
 
-    if (col < p.ncols_output && gl_GlobalInvocationID.x < p.orig_ncols) {
+    if (col < p.k) {
         if (p.last_pass != 0) {
-            const uint row_offset = row * p.ncols_output;
-            data_d[row_offset + col] = dst_row[col].x;
+            if (gl_GlobalInvocationID.x < p.ncols_input) {
+                const uint row_offset = row * p.k;
+                data_d[row_offset + col] = dst_row[col].x;
+            }
         } else {
-            const uint row_offset = row * p.orig_ncols + gl_WorkGroupID.x * p.ncols_output;
-            data_t[row_offset + col] = dst_row[col];
+            if (gl_WorkGroupID.x * p.k + col < p.ncols_output) {
+                const uint row_offset = row * p.ncols_output + gl_WorkGroupID.x * p.k;
+                data_t[row_offset + col] = dst_row[col];
+            }
         }
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp

@@ -25,6 +25,7 @@ layout (push_constant) uniform parameter {
     uint orig_ncols;
     uint ncols_input;
     uint ncols_output;
+    uint k;
     uint nrows;
     uint first_pass;
     uint last_pass;
@@ -60,7 +61,7 @@ void topk(const uint row) {
             const uint row_offset = row * p.ncols_input;
             dst_row[tid] = ivec2(gl_GlobalInvocationID.x, floatBitsToInt(data_a[row_offset + gl_GlobalInvocationID.x]));
         } else {
-            const uint row_offset = row * p.orig_ncols;
+            const uint row_offset = row * p.ncols_input;
             dst_row[tid] = data_s[row_offset + gl_GlobalInvocationID.x];
         }
     } else {
@@ -68,7 +69,7 @@ void topk(const uint row) {
     }
     barrier();
 
-    if (p.ncols_output == 1) {
+    if (p.k == 1) {
         // Fast path for single output - just do a max reduction
         [[unroll]] for (int s = BLOCK_SIZE / 2; s >= 1; s /= 2) {
             if (tid < s) {
@@ -98,7 +99,7 @@ void topk(const uint row) {
         uint range_max = 0xFF800000;
         // How many are above the current range, and how many we need to find.
         uint total = 0;
-        uint limit = min(p.ncols_output, p.ncols_input - gl_WorkGroupID.x * BLOCK_SIZE);
+        uint limit = min(p.k, p.ncols_input - gl_WorkGroupID.x * BLOCK_SIZE);
 
         while (mask != 0) {
             barrier();
@@ -139,7 +140,7 @@ void topk(const uint row) {
             range_max = range_min + ((min_idx + 1) << shift);
             range_min = range_min + (min_idx << shift);
 
-            if (total == p.ncols_output) {
+            if (total == p.k) {
                 break;
             }
             total -= counts[min_idx];
@@ -179,13 +180,17 @@ void topk(const uint row) {
         barrier();
     }
 
-    if (tid < p.ncols_output && gl_GlobalInvocationID.x < p.orig_ncols) {
+    if (tid < p.k) {
         if (p.last_pass != 0) {
-            const uint row_offset = row * p.ncols_output;
-            data_d[row_offset + tid] = dst_row[tid].x;
+            if (gl_GlobalInvocationID.x < p.ncols_input) {
+                const uint row_offset = row * p.k;
+                data_d[row_offset + tid] = dst_row[tid].x;
+            }
         } else {
-            const uint row_offset = row * p.orig_ncols + gl_WorkGroupID.x * p.ncols_output;
-            data_t[row_offset + tid] = dst_row[tid];
+            if (gl_WorkGroupID.x * p.k + tid < p.ncols_output) {
+                const uint row_offset = row * p.ncols_output + gl_WorkGroupID.x * p.k;
+                data_t[row_offset + tid] = dst_row[tid];
+            }
         }
     }
 }

ggml/src/gguf.cpp

@@ -1213,7 +1213,7 @@ void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const vo
 struct gguf_writer_base {
     size_t written_bytes {0u};
 
-    ~gguf_writer_base(void) {}
+    ~gguf_writer_base(void) = default;
 
     // we bet on devirtualization
     virtual void write(int8_t val) = 0;

gguf-py/gguf/constants.py

@@ -175,6 +175,7 @@ class Keys:
         VALUE_LENGTH_MLA             = "{arch}.attention.value_length_mla"
         SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
         SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
+        TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
 
     class Rope:
         DIMENSION_COUNT          = "{arch}.rope.dimension_count"
@@ -444,6 +445,7 @@ class MODEL_ARCH(IntEnum):
     MINIMAXM2        = auto()
     RND1             = auto()
     PANGU_EMBED      = auto()
+    MISTRAL3         = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -817,6 +819,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.COGVLM:           "cogvlm",
     MODEL_ARCH.RND1:             "rnd1",
     MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
+    MODEL_ARCH.MISTRAL3:         "mistral3",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -3071,6 +3074,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.MISTRAL3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
     # TODO
 }
 

gguf-py/gguf/gguf_writer.py

@@ -904,6 +904,9 @@ class GGUFWriter:
     def add_attn_temperature_length(self, value: int) -> None:
         self.add_uint32(Keys.Attention.TEMPERATURE_LENGTH.format(arch=self.arch), value)
 
+    def add_attn_temperature_scale(self, value: float) -> None:
+        self.add_float32(Keys.Attention.TEMPERATURE_SCALE.format(arch=self.arch), value)
+
     def add_pooling_type(self, value: PoolingType) -> None:
         self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
 

src/llama-arch.cpp

@@ -111,6 +111,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COGVLM,           "cogvlm"           },
     { LLM_ARCH_RND1,             "rnd1"             },
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
+    { LLM_ARCH_MISTRAL3,         "mistral3"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -204,6 +205,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
     { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
+    { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
 
@@ -853,7 +855,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
             { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
-            { LLM_TENSOR_SSM_A,              "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_A_NOSCAN,       "blk.%d.ssm_a" },
             { LLM_TENSOR_SSM_CONV1D,         "blk.%d.ssm_conv1d" },
             { LLM_TENSOR_SSM_DT,             "blk.%d.ssm_dt" },
             { LLM_TENSOR_SSM_BETA_ALPHA,     "blk.%d.ssm_ba" },
@@ -2512,6 +2514,32 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_MISTRAL3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2611,6 +2639,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_ACT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}},
     {LLM_TENSOR_SSM_CONV1D,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SSM_A,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}},
+    {LLM_TENSOR_SSM_A_NOSCAN,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // a version of SSM_A used for MUL instead of SSM_SCAN
     {LLM_TENSOR_SSM_DT_NORM,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_SSM_B_NORM,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_SSM_C_NORM,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},

src/llama-arch.h

@@ -115,6 +115,7 @@ enum llm_arch {
     LLM_ARCH_COGVLM,
     LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
+    LLM_ARCH_MISTRAL3,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -208,6 +209,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_OUTPUT_SCALE,
     LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
+    LLM_KV_ATTENTION_TEMPERATURE_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
@@ -377,6 +379,7 @@ enum llm_tensor {
     LLM_TENSOR_SSM_DT,
     LLM_TENSOR_SSM_DT_NORM,
     LLM_TENSOR_SSM_A,
+    LLM_TENSOR_SSM_A_NOSCAN,        // qwen3next special case with MUL instead of SSM_SCAN
     LLM_TENSOR_SSM_B_NORM,
     LLM_TENSOR_SSM_C_NORM,
     LLM_TENSOR_SSM_D,

src/llama-graph.cpp

@@ -71,6 +71,9 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && attn_scale) {
         const int64_t n_tokens = ubatch->n_tokens;
 
+        GGML_ASSERT(f_attn_temp_scale != 0.0f);
+        GGML_ASSERT(n_attn_temp_floor_scale != 0);
+
         std::vector<float> attn_scale_data(n_tokens, 0.0f);
         for (int i = 0; i < n_tokens; ++i) {
             const float pos = ubatch->pos[i];

src/llama-hparams.h

@@ -162,8 +162,8 @@ struct llama_hparams {
     // llama4 smallthinker
     uint32_t n_moe_layer_step        = 0;
     uint32_t n_no_rope_layer_step    = 4;
-    uint32_t n_attn_temp_floor_scale = 8192;
-    float    f_attn_temp_scale       = 0.1;
+    uint32_t n_attn_temp_floor_scale = 0;
+    float    f_attn_temp_scale       = 0.0f;
 
     // gemma3n altup
     uint32_t n_altup      = 4; // altup_num_inputs

src/llama-impl.h

@@ -37,7 +37,7 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 template <typename T>
 struct no_init {
     T value;
-    no_init() { /* do nothing */ }
+    no_init() = default;
 };
 
 struct time_meas {

src/llama-mmap.cpp

@@ -491,7 +491,7 @@ struct llama_mlock::impl {
         if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
             suggest = false;
         }
-        if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) {
+        if (suggest && ((uint64_t)lock_limit.rlim_max > (uint64_t)lock_limit.rlim_cur + size)) {
             suggest = false;
         }
 #endif

src/llama-model.cpp

@@ -123,6 +123,7 @@
 #include "models/t5-enc.cpp"
 #include "models/wavtokenizer-dec.cpp"
 #include "models/xverse.cpp"
+#include "models/mistral3.cpp"
 #include "models/graph-context-mamba.cpp"
 
 #if defined(GGML_USE_CLBLAST)
@@ -527,8 +528,8 @@ static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode s
 }
 
 struct llama_model::impl {
-    impl() {}
-    ~impl() {}
+    impl() = default;
+    ~impl() = default;
 
     uint64_t n_elements = 0;
 
@@ -565,7 +566,7 @@ llama_model::llama_model(const llama_model_params & params) : params(params), pi
     pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
 }
 
-llama_model::~llama_model() {}
+llama_model::~llama_model() = default;
 
 void llama_model::load_stats(llama_model_loader & ml) {
     pimpl->n_elements = ml.n_elements;
@@ -767,8 +768,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
                     hparams.n_no_rope_layer_step = hparams.n_layer; // always use rope
                 } else {
-                    hparams.swa_type      = LLAMA_SWA_TYPE_CHUNKED;
-                    hparams.n_swa         = 8192;
+                    hparams.swa_type                = LLAMA_SWA_TYPE_CHUNKED;
+                    hparams.n_swa                   = 8192;
+                    hparams.n_attn_temp_floor_scale = 8192;
+                    hparams.f_attn_temp_scale       = 0.1f;
                     hparams.set_swa_pattern(4);   // pattern: 3 chunked - 1 full
                 }
 
@@ -2351,6 +2354,42 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MISTRAL3:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale, false);
+
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST,   hparams.yarn_beta_fast, false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,   hparams.yarn_beta_slow, false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     hparams.rope_yarn_log_mul, false);
+
+                // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
+                if (hparams.f_attn_temp_scale != 0.0f) {
+                    hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;
+                    if (hparams.n_attn_temp_floor_scale == 0) {
+                        throw std::runtime_error("invalid n_ctx_orig_yarn for attention temperature scaling");
+                    }
+                }
+
+                // TODO: this seems to be correct with the case of mscale == mscale_all_dims == 1.0f
+                //       but may need further verification with other values
+                if (hparams.rope_yarn_log_mul != 0.0f) {
+                    float factor = 1.0f / hparams.rope_freq_scale_train;
+                    float mscale = 1.0f;
+                    float mscale_all_dims = hparams.rope_yarn_log_mul;
+                    static auto get_mscale = [](float scale, float mscale) {
+                        return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
+                    };
+                    hparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
+                }
+
+                switch (hparams.n_layer) {
+                    case 26: type = LLM_TYPE_3B; break;
+                    case 34: type = LLM_TYPE_8B; break;
+                    case 40: type = LLM_TYPE_14B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -2677,6 +2716,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_MINICPM:
             case LLM_ARCH_GRANITE:
             case LLM_ARCH_GRANITE_MOE:
+            case LLM_ARCH_MISTRAL3:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -6644,7 +6684,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.ssm_in         = create_tensor(tn(LLM_TENSOR_SSM_IN,         "weight", i), { n_embd, qkvz_dim }, 0);
                             layer.ssm_conv1d     = create_tensor(tn(LLM_TENSOR_SSM_CONV1D,     "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
                             layer.ssm_dt         = create_tensor(tn(LLM_TENSOR_SSM_DT,         "bias",   i), { hparams.ssm_dt_rank }, 0);
-                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A,                    i), { hparams.ssm_dt_rank }, 0);
+                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN,             i), { hparams.ssm_dt_rank }, 0);
                             layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0);
                             layer.ssm_norm       = create_tensor(tn(LLM_TENSOR_SSM_NORM,       "weight", i), { head_v_dim }, 0);
                             layer.ssm_out        = create_tensor(tn(LLM_TENSOR_SSM_OUT,        "weight", i), { value_dim, n_embd }, 0);
@@ -7683,6 +7723,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_qwen3next>(*this, params);
             } break;
+        case LLM_ARCH_MISTRAL3:
+            {
+                llm = std::make_unique<llm_build_mistral3>(*this, params);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -7851,6 +7895,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARCEE:
         case LLM_ARCH_ERNIE4_5:
         case LLM_ARCH_ERNIE4_5_MOE:
+        case LLM_ARCH_MISTRAL3:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

src/llama-vocab.cpp

@@ -3529,8 +3529,7 @@ void llama_vocab::impl::print_info() const {
 llama_vocab::llama_vocab() : pimpl(new impl(*this)) {
 }
 
-llama_vocab::~llama_vocab() {
-}
+llama_vocab::~llama_vocab() = default;
 
 void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) {
     pimpl->load(ml, kv);

src/models/mistral3.cpp

@@ -0,0 +1,160 @@
+#include "models.h"
+
+llm_build_mistral3::llm_build_mistral3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // (optional) temperature tuning
+    ggml_tensor * inp_attn_scale = nullptr;
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        inp_attn_scale = build_inp_attn_scale();
+    }
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            if (inp_attn_scale) {
+                // apply llama 4 temperature scaling
+                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                cb(Qcur, "Qcur_attn_temp_scaled", il);
+            }
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network (non-MoE)
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/models.h

@@ -322,6 +322,10 @@ struct llm_build_minimax_m2 : public llm_graph_context {
     llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_mistral3 : public llm_graph_context {
+    llm_build_mistral3(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_mpt : public llm_graph_context {
     llm_build_mpt(const llama_model & model, const llm_graph_params & params);
 };

tools/mtmd/clip.cpp

@@ -456,6 +456,7 @@ struct clip_ctx {
     int max_nodes = 8192;
     ggml_backend_sched_ptr sched;
     clip_flash_attn_type flash_attn_type = CLIP_FLASH_ATTN_TYPE_AUTO;
+    bool is_allocated = false;
 
     // for debugging
     bool debug_graph = false;
@@ -3366,12 +3367,30 @@ struct clip_model_loader {
     };
 
     static void warmup(clip_ctx & ctx_clip) {
+        // create a fake batch
+        const auto & hparams = ctx_clip.model.hparams;
+        clip_image_f32_batch batch;
+        clip_image_f32_ptr img(clip_image_f32_init());
+        if (ctx_clip.model.modality == CLIP_MODALITY_VISION) {
+            img->nx = hparams.warmup_image_size;
+            img->ny = hparams.warmup_image_size;
+            LOG_INF("%s: warmup with image size = %d x %d\n", __func__, img->nx, img->ny);
+        } else {
+            img->nx = hparams.warmup_audio_size;
+            img->ny = hparams.n_mel_bins;
+            LOG_INF("%s: warmup with audio size = %d\n", __func__, img->nx);
+        }
+        batch.entries.push_back(std::move(img));
+        warmup(ctx_clip, batch);
+    }
+
+    static void warmup(clip_ctx & ctx_clip, const clip_image_f32_batch & batch) {
         support_info_graph info;
 
         if (ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_AUTO) {
             // try to enable flash attention to see if it's supported
             ctx_clip.flash_attn_type = CLIP_FLASH_ATTN_TYPE_ENABLED;
-            info = alloc_compute_meta(ctx_clip);
+            info = alloc_compute_meta(ctx_clip, batch);
             if (!info.fattn && info.fattn_op) {
                 auto op = info.fattn_op;
                 LOG_WRN("%s: *****************************************************************\n", __func__);
@@ -3390,15 +3409,17 @@ struct clip_model_loader {
                 LOG_WRN("%s: please report this on github as an issue\n", __func__);
                 LOG_WRN("%s: *****************************************************************\n", __func__);
                 ctx_clip.flash_attn_type = CLIP_FLASH_ATTN_TYPE_DISABLED;
-                alloc_compute_meta(ctx_clip);
+                alloc_compute_meta(ctx_clip, batch);
             }
         } else {
-            info = alloc_compute_meta(ctx_clip);
+            info = alloc_compute_meta(ctx_clip, batch);
             if (!info.fattn && ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
                 LOG_WRN("%s: flash attention is not supported by the current backend; falling back to CPU (performance will be degraded)\n", __func__);
             }
         }
 
+        ctx_clip.is_allocated = true; // mark buffers as allocated
+
         LOG_INF("%s: flash attention is %s\n", __func__,
             (ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) ? "enabled" : "disabled");
 
@@ -3430,24 +3451,9 @@ struct clip_model_loader {
         }
     }
 
-    static support_info_graph alloc_compute_meta(clip_ctx & ctx_clip) {
-        const auto & hparams = ctx_clip.model.hparams;
+    static support_info_graph alloc_compute_meta(clip_ctx & ctx_clip, const clip_image_f32_batch & batch) {
         ctx_clip.buf_compute_meta.resize(ctx_clip.max_nodes * ggml_tensor_overhead() + ggml_graph_overhead());
 
-        // create a fake batch
-        clip_image_f32_batch batch;
-        clip_image_f32_ptr img(clip_image_f32_init());
-        if (ctx_clip.model.modality == CLIP_MODALITY_VISION) {
-            img->nx = hparams.warmup_image_size;
-            img->ny = hparams.warmup_image_size;
-            LOG_INF("%s: warmup with image size = %d x %d\n", __func__, img->nx, img->ny);
-        } else {
-            img->nx = hparams.warmup_audio_size;
-            img->ny = hparams.n_mel_bins;
-            LOG_INF("%s: warmup with audio size = %d\n", __func__, img->nx);
-        }
-        batch.entries.push_back(std::move(img));
-
         ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch);
         ggml_backend_sched_reserve(ctx_clip.sched.get(), gf);
 
@@ -3587,14 +3593,18 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
             ctx_vision = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_vision->model, CLIP_MODALITY_VISION);
             loader.load_tensors(*ctx_vision);
-            loader.warmup(*ctx_vision);
+            if (ctx_params.warmup) {
+                loader.warmup(*ctx_vision);
+            }
         }
 
         if (loader.has_audio) {
             ctx_audio = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_audio->model, CLIP_MODALITY_AUDIO);
             loader.load_tensors(*ctx_audio);
-            loader.warmup(*ctx_audio);
+            if (ctx_params.warmup) {
+                loader.warmup(*ctx_audio);
+            }
         }
 
     } catch (const std::exception & e) {
@@ -4807,6 +4817,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         return false; // only support batch size of 1
     }
 
+    // if buffers are not allocated, we need to do a warmup run to allocate them
+    if (!ctx->is_allocated) {
+        clip_model_loader::warmup(*ctx, *imgs_c_ptr);
+    }
+
     // build the inference graph
     ctx->debug_print_tensors.clear();
     ggml_backend_sched_reset(ctx->sched.get());
@@ -5192,7 +5207,23 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     try {
         loader.load_hparams(ctx_clip->model, CLIP_MODALITY_VISION);
         loader.load_tensors(*ctx_clip);
-        loader.alloc_compute_meta(*ctx_clip);
+
+        // create a fake batch
+        const auto & hparams = ctx_clip->model.hparams;
+        clip_image_f32_batch batch;
+        clip_image_f32_ptr img(clip_image_f32_init());
+        if (ctx_clip->model.modality == CLIP_MODALITY_VISION) {
+            img->nx = hparams.warmup_image_size;
+            img->ny = hparams.warmup_image_size;
+            LOG_INF("%s: warmup with image size = %d x %d\n", __func__, img->nx, img->ny);
+        } else {
+            img->nx = hparams.warmup_audio_size;
+            img->ny = hparams.n_mel_bins;
+            LOG_INF("%s: warmup with audio size = %d\n", __func__, img->nx);
+        }
+        batch.entries.push_back(std::move(img));
+
+        loader.alloc_compute_meta(*ctx_clip,batch);
     } catch (const std::exception & e) {
         printf("%s: failed to load model '%s': %s\n", __func__, fname_inp, e.what());
         if (ctx_clip) {

tools/mtmd/clip.h

@@ -34,6 +34,7 @@ struct clip_context_params {
     enum clip_flash_attn_type flash_attn_type;
     int image_min_tokens;
     int image_max_tokens;
+    bool warmup;
 };
 
 struct clip_init_result {

tools/mtmd/mtmd-cli.cpp

@@ -136,6 +136,7 @@ struct mtmd_cli_context {
         mparams.print_timings    = true;
         mparams.n_threads        = params.cpuparams.n_threads;
         mparams.flash_attn_type  = params.flash_attn_type;
+        mparams.warmup           = params.warmup;
         mparams.image_min_tokens = params.image_min_tokens;
         mparams.image_max_tokens = params.image_max_tokens;
         ctx_vision.reset(mtmd_init_from_file(clip_path, model, mparams));

tools/mtmd/mtmd.cpp

@@ -108,6 +108,7 @@ mtmd_context_params mtmd_context_params_default() {
         /* image_marker      */ MTMD_DEFAULT_IMAGE_MARKER,
         /* media_marker      */ mtmd_default_marker(),
         /* flash_attn_type   */ LLAMA_FLASH_ATTN_TYPE_AUTO,
+        /* warmup            */ true,
         /* image_min_tokens  */ -1,
         /* image_max_tokens  */ -1,
     };
@@ -177,6 +178,7 @@ struct mtmd_context {
             /* flash_attn_type   */ CLIP_FLASH_ATTN_TYPE_AUTO,
             /* image_min_tokens  */ ctx_params.image_min_tokens,
             /* image_max_tokens  */ ctx_params.image_max_tokens,
+            /* warmup            */ ctx_params.warmup,
         };
 
         auto res = clip_init(mmproj_fname, ctx_clip_params);

tools/mtmd/mtmd.h

@@ -82,6 +82,7 @@ struct mtmd_context_params {
     const char * image_marker; // deprecated, use media_marker instead
     const char * media_marker;
     enum llama_flash_attn_type flash_attn_type;
+    bool warmup; // whether to run a warmup encode pass after initialization
 
     // limit number of image tokens, only for vision models with dynamic resolution
     int image_min_tokens; // minimum number of tokens for image input (default: read from metadata)

tools/server/server-common.cpp

@@ -11,6 +11,7 @@
 
 #include <random>
 #include <sstream>
+#include <fstream>
 
 json format_error_response(const std::string & message, const enum error_type type) {
     std::string type_str;
@@ -774,6 +775,65 @@ json oaicompat_completion_params_parse(const json & body) {
     return llama_params;
 }
 
+// media_path always end with '/', see arg.cpp
+static void handle_media(
+        std::vector<raw_buffer> & out_files,
+        json & media_obj,
+        const std::string & media_path) {
+    std::string url = json_value(media_obj, "url", std::string());
+    if (string_starts_with(url, "http")) {
+        // download remote image
+        // TODO @ngxson : maybe make these params configurable
+        common_remote_params params;
+        params.headers.push_back("User-Agent: llama.cpp/" + build_info);
+        params.max_size = 1024 * 1024 * 10; // 10MB
+        params.timeout  = 10; // seconds
+        SRV_INF("downloading image from '%s'\n", url.c_str());
+        auto res = common_remote_get_content(url, params);
+        if (200 <= res.first && res.first < 300) {
+            SRV_INF("downloaded %ld bytes\n", res.second.size());
+            raw_buffer data;
+            data.insert(data.end(), res.second.begin(), res.second.end());
+            out_files.push_back(data);
+        } else {
+            throw std::runtime_error("Failed to download image");
+        }
+
+    } else if (string_starts_with(url, "file://")) {
+        if (media_path.empty()) {
+            throw std::invalid_argument("file:// URLs are not allowed unless --media-path is specified");
+        }
+        // load local image file
+        std::string file_path = url.substr(7); // remove "file://"
+        raw_buffer data;
+        if (!fs_validate_filename(file_path, true)) {
+            throw std::invalid_argument("file path is not allowed: " + file_path);
+        }
+        SRV_INF("loading image from local file '%s'\n", (media_path + file_path).c_str());
+        std::ifstream file(media_path + file_path, std::ios::binary);
+        if (!file) {
+            throw std::invalid_argument("file does not exist or cannot be opened: " + file_path);
+        }
+        data.assign((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+        out_files.push_back(data);
+
+    } else {
+        // try to decode base64 image
+        std::vector<std::string> parts = string_split<std::string>(url, /*separator*/ ',');
+        if (parts.size() != 2) {
+            throw std::runtime_error("Invalid url value");
+        } else if (!string_starts_with(parts[0], "data:image/")) {
+            throw std::runtime_error("Invalid url format: " + parts[0]);
+        } else if (!string_ends_with(parts[0], "base64")) {
+            throw std::runtime_error("url must be base64 encoded");
+        } else {
+            auto base64_data = parts[1];
+            auto decoded_data = base64_decode(base64_data);
+            out_files.push_back(decoded_data);
+        }
+    }
+}
+
 // used by /chat/completions endpoint
 json oaicompat_chat_params_parse(
     json & body, /* openai api json semantics */
@@ -819,26 +879,26 @@ json oaicompat_chat_params_parse(
             auto schema_wrapper = json_value(response_format, "json_schema", json::object());
             json_schema = json_value(schema_wrapper, "schema", json::object());
         } else if (!response_type.empty() && response_type != "text") {
-            throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
+            throw std::invalid_argument("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
         }
     }
 
     // get input files
     if (!body.contains("messages")) {
-        throw std::runtime_error("'messages' is required");
+        throw std::invalid_argument("'messages' is required");
     }
     json & messages = body.at("messages");
     if (!messages.is_array()) {
-        throw std::runtime_error("Expected 'messages' to be an array");
+        throw std::invalid_argument("Expected 'messages' to be an array");
     }
     for (auto & msg : messages) {
         std::string role = json_value(msg, "role", std::string());
         if (role != "assistant" && !msg.contains("content")) {
-            throw std::runtime_error("All non-assistant messages must contain 'content'");
+            throw std::invalid_argument("All non-assistant messages must contain 'content'");
         }
         if (role == "assistant") {
             if (!msg.contains("content") && !msg.contains("tool_calls")) {
-                throw std::runtime_error("Assistant message must contain either 'content' or 'tool_calls'!");
+                throw std::invalid_argument("Assistant message must contain either 'content' or 'tool_calls'!");
             }
             if (!msg.contains("content")) {
                 continue; // avoid errors with no content
@@ -850,7 +910,7 @@ json oaicompat_chat_params_parse(
         }
 
         if (!content.is_array()) {
-            throw std::runtime_error("Expected 'content' to be a string or an array");
+            throw std::invalid_argument("Expected 'content' to be a string or an array");
         }
 
         for (auto & p : content) {
@@ -860,41 +920,8 @@ json oaicompat_chat_params_parse(
                     throw std::runtime_error("image input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
                 }
 
-                json image_url  = json_value(p, "image_url", json::object());
-                std::string url = json_value(image_url, "url", std::string());
-                if (string_starts_with(url, "http")) {
-                    // download remote image
-                    // TODO @ngxson : maybe make these params configurable
-                    common_remote_params params;
-                    params.headers.push_back("User-Agent: llama.cpp/" + build_info);
-                    params.max_size = 1024 * 1024 * 10; // 10MB
-                    params.timeout  = 10; // seconds
-                    SRV_INF("downloading image from '%s'\n", url.c_str());
-                    auto res = common_remote_get_content(url, params);
-                    if (200 <= res.first && res.first < 300) {
-                        SRV_INF("downloaded %ld bytes\n", res.second.size());
-                        raw_buffer data;
-                        data.insert(data.end(), res.second.begin(), res.second.end());
-                        out_files.push_back(data);
-                    } else {
-                        throw std::runtime_error("Failed to download image");
-                    }
-
-                } else {
-                    // try to decode base64 image
-                    std::vector<std::string> parts = string_split<std::string>(url, /*separator*/ ',');
-                    if (parts.size() != 2) {
-                        throw std::runtime_error("Invalid image_url.url value");
-                    } else if (!string_starts_with(parts[0], "data:image/")) {
-                        throw std::runtime_error("Invalid image_url.url format: " + parts[0]);
-                    } else if (!string_ends_with(parts[0], "base64")) {
-                        throw std::runtime_error("image_url.url must be base64 encoded");
-                    } else {
-                        auto base64_data = parts[1];
-                        auto decoded_data = base64_decode(base64_data);
-                        out_files.push_back(decoded_data);
-                    }
-                }
+                json image_url = json_value(p, "image_url", json::object());
+                handle_media(out_files, image_url, opt.media_path);
 
                 // replace this chunk with a marker
                 p["type"] = "text";
@@ -911,18 +938,20 @@ json oaicompat_chat_params_parse(
                 std::string format = json_value(input_audio, "format", std::string());
                 // while we also support flac, we don't allow it here so we matches the OAI spec
                 if (format != "wav" && format != "mp3") {
-                    throw std::runtime_error("input_audio.format must be either 'wav' or 'mp3'");
+                    throw std::invalid_argument("input_audio.format must be either 'wav' or 'mp3'");
                 }
                 auto decoded_data = base64_decode(data); // expected to be base64 encoded
                 out_files.push_back(decoded_data);
 
+                // TODO: add audio_url support by reusing handle_media()
+
                 // replace this chunk with a marker
                 p["type"] = "text";
                 p["text"] = mtmd_default_marker();
                 p.erase("input_audio");
 
             } else if (type != "text") {
-                throw std::runtime_error("unsupported content[].type");
+                throw std::invalid_argument("unsupported content[].type");
             }
         }
     }
@@ -940,7 +969,7 @@ json oaicompat_chat_params_parse(
     inputs.enable_thinking       = opt.enable_thinking;
     if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
         if (body.contains("grammar")) {
-            throw std::runtime_error("Cannot use custom grammar constraints with tools.");
+            throw std::invalid_argument("Cannot use custom grammar constraints with tools.");
         }
         llama_params["parse_tool_calls"] = true;
     }
@@ -959,7 +988,7 @@ json oaicompat_chat_params_parse(
     } else if (enable_thinking_kwarg == "false") {
         inputs.enable_thinking = false;
     } else if (!enable_thinking_kwarg.empty() && enable_thinking_kwarg[0] == '"') {
-        throw std::runtime_error("invalid type for \"enable_thinking\" (expected boolean, got string)");
+        throw std::invalid_argument("invalid type for \"enable_thinking\" (expected boolean, got string)");
     }
 
     // if the assistant message appears at the end of list, we do not add end-of-turn token
@@ -972,14 +1001,14 @@ json oaicompat_chat_params_parse(
 
         /* sanity check, max one assistant message at the end of the list */
         if (!inputs.messages.empty() && inputs.messages.back().role == "assistant"){
-            throw std::runtime_error("Cannot have 2 or more assistant messages at the end of the list.");
+            throw std::invalid_argument("Cannot have 2 or more assistant messages at the end of the list.");
         }
 
         /* TODO: test this properly */
         inputs.reasoning_format = COMMON_REASONING_FORMAT_NONE;
 
         if ( inputs.enable_thinking ) {
-            throw std::runtime_error("Assistant response prefill is incompatible with enable_thinking.");
+            throw std::invalid_argument("Assistant response prefill is incompatible with enable_thinking.");
         }
 
         inputs.add_generation_prompt = true;
@@ -1020,18 +1049,18 @@ json oaicompat_chat_params_parse(
     // Handle "n" field
     int n_choices = json_value(body, "n", 1);
     if (n_choices != 1) {
-        throw std::runtime_error("Only one completion choice is allowed");
+        throw std::invalid_argument("Only one completion choice is allowed");
     }
 
     // Handle "logprobs" field
     // TODO: The response format of this option is not yet OAI-compatible, but seems like no one really using it; We may need to fix it in the future
     if (json_value(body, "logprobs", false)) {
         if (has_tools && stream) {
-            throw std::runtime_error("logprobs is not supported with tools + stream");
+            throw std::invalid_argument("logprobs is not supported with tools + stream");
         }
         llama_params["n_probs"] = json_value(body, "top_logprobs", 20);
     } else if (body.contains("top_logprobs") && !body.at("top_logprobs").is_null()) {
-        throw std::runtime_error("top_logprobs requires logprobs to be set to true");
+        throw std::invalid_argument("top_logprobs requires logprobs to be set to true");
     }
 
     // Copy remaining properties to llama_params
@@ -1263,7 +1292,11 @@ json convert_anthropic_to_oai(const json & body) {
     return oai_body;
 }
 
-json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64) {
+json format_embeddings_response_oaicompat(
+        const json & request,
+        const std::string & model_name,
+        const json & embeddings,
+        bool use_base64) {
     json data = json::array();
     int32_t n_tokens = 0;
     int i = 0;
@@ -1293,7 +1326,7 @@ json format_embeddings_response_oaicompat(const json & request, const json & emb
     }
 
     json res = json {
-        {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+        {"model", json_value(request, "model", model_name)},
         {"object", "list"},
         {"usage", json {
             {"prompt_tokens", n_tokens},
@@ -1307,6 +1340,7 @@ json format_embeddings_response_oaicompat(const json & request, const json & emb
 
 json format_response_rerank(
         const json & request,
+        const std::string & model_name,
         const json & ranks,
         bool is_tei_format,
         std::vector<std::string> & texts,
@@ -1338,7 +1372,7 @@ json format_response_rerank(
     if (is_tei_format) return results;
 
     json res = json{
-        {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+        {"model", json_value(request, "model", model_name)},
         {"object", "list"},
         {"usage", json{
             {"prompt_tokens", n_tokens},

tools/server/server-common.h

@@ -13,8 +13,6 @@
 #include <vector>
 #include <cinttypes>
 
-#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo"
-
 const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
 
 using json = nlohmann::ordered_json;
@@ -286,6 +284,7 @@ struct oaicompat_parser_options {
     bool allow_image;
     bool allow_audio;
     bool enable_thinking = true;
+    std::string media_path;
 };
 
 // used by /chat/completions endpoint
@@ -298,11 +297,16 @@ json oaicompat_chat_params_parse(
 json convert_anthropic_to_oai(const json & body);
 
 // TODO: move it to server-task.cpp
-json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64 = false);
+json format_embeddings_response_oaicompat(
+    const json & request,
+    const std::string & model_name,
+    const json & embeddings,
+    bool use_base64 = false);
 
 // TODO: move it to server-task.cpp
 json format_response_rerank(
         const json & request,
+        const std::string & model_name,
         const json & ranks,
         bool is_tei_format,
         std::vector<std::string> & texts,

tools/server/server-context.cpp

@@ -17,6 +17,7 @@
 #include <cinttypes>
 #include <memory>
 #include <unordered_set>
+#include <filesystem>
 
 // fix problem with std::min and std::max
 #if defined(_WIN32)
@@ -518,6 +519,8 @@ struct server_context_impl {
     // Necessary similarity of prompt for slot selection
     float slot_prompt_similarity = 0.0f;
 
+    std::string model_name; // name of the loaded model, to be used by API
+
     common_chat_templates_ptr chat_templates;
     oaicompat_parser_options  oai_parser_opt;
 
@@ -621,6 +624,7 @@ struct server_context_impl {
             mparams.print_timings    = false;
             mparams.n_threads        = params_base.cpuparams.n_threads;
             mparams.flash_attn_type  = params_base.flash_attn_type;
+            mparams.warmup           = params_base.warmup;
             mparams.image_min_tokens = params_base.image_min_tokens;
             mparams.image_max_tokens = params_base.image_max_tokens;
             mctx = mtmd_init_from_file(mmproj_path.c_str(), model, mparams);
@@ -757,6 +761,18 @@ struct server_context_impl {
         }
         SRV_WRN("%s", "for more info see https://github.com/ggml-org/llama.cpp/pull/16391\n");
 
+        if (!params_base.model_alias.empty()) {
+            // user explicitly specified model name
+            model_name = params_base.model_alias;
+        } else if (!params_base.model.name.empty()) {
+            // use model name in registry format (for models in cache)
+            model_name = params_base.model.name;
+        } else {
+            // fallback: derive model name from file name
+            auto model_path = std::filesystem::path(params_base.model.path);
+            model_name = model_path.filename().string();
+        }
+
         // thinking is enabled if:
         // 1. It's not explicitly disabled (reasoning_budget == 0)
         // 2. The chat template supports it
@@ -772,6 +788,7 @@ struct server_context_impl {
             /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
             /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
             /* enable_thinking       */ enable_thinking,
+            /* media_path            */ params_base.media_path,
         };
 
         // print sample chat example to make it clear which template is used
@@ -2610,7 +2627,7 @@ static std::unique_ptr<server_res_generator> handle_completions_impl(
             // OAI-compat
             task.params.res_type          = res_type;
             task.params.oaicompat_cmpl_id = completion_id;
-            // oaicompat_model is already populated by params_from_json_cmpl
+            task.params.oaicompat_model   = ctx_server.model_name;
 
             tasks.push_back(std::move(task));
         }
@@ -2938,7 +2955,7 @@ void server_routes::init_routes() {
         json data = {
             { "default_generation_settings", default_generation_settings_for_props },
             { "total_slots",                 ctx_server.params_base.n_parallel },
-            { "model_alias",                 ctx_server.params_base.model_alias },
+            { "model_alias",                 ctx_server.model_name },
             { "model_path",                  ctx_server.params_base.model.path },
             { "modalities",                  json {
                 {"vision", ctx_server.oai_parser_opt.allow_image},
@@ -3180,8 +3197,8 @@ void server_routes::init_routes() {
         json models = {
             {"models", {
                 {
-                    {"name", params.model_alias.empty() ? params.model.path : params.model_alias},
-                    {"model", params.model_alias.empty() ? params.model.path : params.model_alias},
+                    {"name", ctx_server.model_name},
+                    {"model", ctx_server.model_name},
                     {"modified_at", ""},
                     {"size", ""},
                     {"digest", ""}, // dummy value, llama.cpp does not support managing model file's hash
@@ -3203,7 +3220,7 @@ void server_routes::init_routes() {
             {"object", "list"},
             {"data", {
                 {
-                    {"id",       params.model_alias.empty() ? params.model.path : params.model_alias},
+                    {"id",       ctx_server.model_name},
                     {"object",   "model"},
                     {"created",  std::time(0)},
                     {"owned_by", "llamacpp"},
@@ -3350,6 +3367,7 @@ void server_routes::init_routes() {
         // write JSON response
         json root = format_response_rerank(
             body,
+            ctx_server.model_name,
             responses,
             is_tei_format,
             documents,
@@ -3612,7 +3630,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_embeddings_impl(cons
 
     // write JSON response
     json root = res_type == TASK_RESPONSE_TYPE_OAI_EMBD
-        ? format_embeddings_response_oaicompat(body, responses, use_base64)
+        ? format_embeddings_response_oaicompat(body, ctx_server.model_name, responses, use_base64)
         : json(responses);
     res->ok(root);
     return res;

tools/server/server-models.cpp

@@ -0,0 +1,975 @@
+#include "server-common.h"
+#include "server-models.h"
+
+#include "download.h"
+
+#include <cpp-httplib/httplib.h> // TODO: remove this once we use HTTP client from download.h
+#include <sheredom/subprocess.h>
+
+#include <functional>
+#include <thread>
+#include <mutex>
+#include <condition_variable>
+#include <cstring>
+#include <atomic>
+#include <chrono>
+#include <queue>
+
+#ifdef _WIN32
+#include <winsock2.h>
+#else
+#include <sys/socket.h>
+#include <netinet/in.h>
+#include <arpa/inet.h>
+#include <unistd.h>
+#endif
+
+#if defined(__APPLE__) && defined(__MACH__)
+// macOS: use _NSGetExecutablePath to get the executable path
+#include <mach-o/dyld.h>
+#include <limits.h>
+#endif
+
+#define CMD_EXIT "exit"
+
+static std::filesystem::path get_server_exec_path() {
+#if defined(_WIN32)
+    wchar_t buf[32768] = { 0 };  // Large buffer to handle long paths
+    DWORD len = GetModuleFileNameW(nullptr, buf, _countof(buf));
+    if (len == 0 || len >= _countof(buf)) {
+        throw std::runtime_error("GetModuleFileNameW failed or path too long");
+    }
+    return std::filesystem::path(buf);
+#elif defined(__APPLE__) && defined(__MACH__)
+    char small_path[PATH_MAX];
+    uint32_t size = sizeof(small_path);
+
+    if (_NSGetExecutablePath(small_path, &size) == 0) {
+        // resolve any symlinks to get absolute path
+        try {
+            return std::filesystem::canonical(std::filesystem::path(small_path));
+        } catch (...) {
+            return std::filesystem::path(small_path);
+        }
+    } else {
+        // buffer was too small, allocate required size and call again
+        std::vector<char> buf(size);
+        if (_NSGetExecutablePath(buf.data(), &size) == 0) {
+            try {
+                return std::filesystem::canonical(std::filesystem::path(buf.data()));
+            } catch (...) {
+                return std::filesystem::path(buf.data());
+            }
+        }
+        throw std::runtime_error("_NSGetExecutablePath failed after buffer resize");
+    }
+#else
+    char path[FILENAME_MAX];
+    ssize_t count = readlink("/proc/self/exe", path, FILENAME_MAX);
+    if (count <= 0) {
+        throw std::runtime_error("failed to resolve /proc/self/exe");
+    }
+    return std::filesystem::path(std::string(path, count));
+#endif
+}
+
+struct local_model {
+    std::string name;
+    std::string path;
+    std::string path_mmproj;
+};
+
+static std::vector<local_model> list_local_models(const std::string & dir) {
+    if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
+        throw std::runtime_error(string_format("error: '%s' does not exist or is not a directory\n", dir.c_str()));
+    }
+
+    std::vector<local_model> models;
+    auto scan_subdir = [&models](const std::string & subdir_path, const std::string & name) {
+        auto files = fs_list(subdir_path, false);
+        common_file_info model_file;
+        common_file_info first_shard_file;
+        common_file_info mmproj_file;
+        for (const auto & file : files) {
+            if (string_ends_with(file.name, ".gguf")) {
+                if (file.name.find("mmproj") != std::string::npos) {
+                    mmproj_file = file;
+                } else if (file.name.find("-00001-of-") != std::string::npos) {
+                    first_shard_file = file;
+                } else {
+                    model_file = file;
+                }
+            }
+        }
+        // single file model
+        local_model model{
+            /* name        */ name,
+            /* path        */ first_shard_file.path.empty() ? model_file.path : first_shard_file.path,
+            /* path_mmproj */ mmproj_file.path // can be empty
+        };
+        if (!model.path.empty()) {
+            models.push_back(model);
+        }
+    };
+
+    auto files = fs_list(dir, true);
+    for (const auto & file : files) {
+        if (file.is_dir) {
+            scan_subdir(file.path, file.name);
+        } else if (string_ends_with(file.name, ".gguf")) {
+            // single file model
+            std::string name = file.name;
+            string_replace_all(name, ".gguf", "");
+            local_model model{
+                /* name        */ name,
+                /* path        */ file.path,
+                /* path_mmproj */ ""
+            };
+            models.push_back(model);
+        }
+    }
+    return models;
+}
+
+//
+// server_models
+//
+
+server_models::server_models(
+        const common_params & params,
+        int argc,
+        char ** argv,
+        char ** envp) : base_params(params) {
+    for (int i = 0; i < argc; i++) {
+        base_args.push_back(std::string(argv[i]));
+    }
+    for (char ** env = envp; *env != nullptr; env++) {
+        base_env.push_back(std::string(*env));
+    }
+    GGML_ASSERT(!base_args.empty());
+    // set binary path
+    try {
+        base_args[0] = get_server_exec_path().string();
+    } catch (const std::exception & e) {
+        LOG_WRN("failed to get server executable path: %s\n", e.what());
+        LOG_WRN("using original argv[0] as fallback: %s\n", base_args[0].c_str());
+    }
+    // TODO: allow refreshing cached model list
+    // add cached models
+    auto cached_models = common_list_cached_models();
+    for (const auto & model : cached_models) {
+        server_model_meta meta{
+            /* name        */ model.to_string(),
+            /* path        */ model.manifest_path,
+            /* path_mmproj */ "", // auto-detected when loading
+            /* in_cache    */ true,
+            /* port        */ 0,
+            /* status      */ SERVER_MODEL_STATUS_UNLOADED,
+            /* last_used   */ 0,
+            /* args        */ std::vector<std::string>(),
+            /* exit_code   */ 0
+        };
+        mapping[meta.name] = instance_t{
+            /* subproc */ std::make_shared<subprocess_s>(),
+            /* th      */ std::thread(),
+            /* meta    */ meta
+        };
+    }
+    // add local models specificed via --models-dir
+    if (!params.models_dir.empty()) {
+        auto local_models = list_local_models(params.models_dir);
+        for (const auto & model : local_models) {
+            if (mapping.find(model.name) != mapping.end()) {
+                // already exists in cached models, skip
+                continue;
+            }
+            server_model_meta meta{
+                /* name        */ model.name,
+                /* path        */ model.path,
+                /* path_mmproj */ model.path_mmproj,
+                /* in_cache    */ false,
+                /* port        */ 0,
+                /* status      */ SERVER_MODEL_STATUS_UNLOADED,
+                /* last_used   */ 0,
+                /* args        */ std::vector<std::string>(),
+                /* exit_code   */ 0
+            };
+            mapping[meta.name] = instance_t{
+                /* subproc */ std::make_shared<subprocess_s>(),
+                /* th      */ std::thread(),
+                /* meta    */ meta
+            };
+        }
+    }
+}
+
+void server_models::update_meta(const std::string & name, const server_model_meta & meta) {
+    std::lock_guard<std::mutex> lk(mutex);
+    auto it = mapping.find(name);
+    if (it != mapping.end()) {
+        it->second.meta = meta;
+    }
+    cv.notify_all(); // notify wait_until_loaded
+}
+
+bool server_models::has_model(const std::string & name) {
+    std::lock_guard<std::mutex> lk(mutex);
+    return mapping.find(name) != mapping.end();
+}
+
+std::optional<server_model_meta> server_models::get_meta(const std::string & name) {
+    std::lock_guard<std::mutex> lk(mutex);
+    auto it = mapping.find(name);
+    if (it != mapping.end()) {
+        return it->second.meta;
+    }
+    return std::nullopt;
+}
+
+static int get_free_port() {
+#ifdef _WIN32
+    WSADATA wsaData;
+    if (WSAStartup(MAKEWORD(2, 2), &wsaData) != 0) {
+        return -1;
+    }
+    typedef SOCKET native_socket_t;
+#define INVALID_SOCKET_VAL INVALID_SOCKET
+#define CLOSE_SOCKET(s) closesocket(s)
+#else
+    typedef int native_socket_t;
+#define INVALID_SOCKET_VAL -1
+#define CLOSE_SOCKET(s) close(s)
+#endif
+
+    native_socket_t sock = socket(AF_INET, SOCK_STREAM, 0);
+    if (sock == INVALID_SOCKET_VAL) {
+#ifdef _WIN32
+        WSACleanup();
+#endif
+        return -1;
+    }
+
+    struct sockaddr_in serv_addr;
+    std::memset(&serv_addr, 0, sizeof(serv_addr));
+    serv_addr.sin_family = AF_INET;
+    serv_addr.sin_addr.s_addr = htonl(INADDR_ANY);
+    serv_addr.sin_port = htons(0);
+
+    if (bind(sock, (struct sockaddr*)&serv_addr, sizeof(serv_addr)) != 0) {
+        CLOSE_SOCKET(sock);
+#ifdef _WIN32
+        WSACleanup();
+#endif
+        return -1;
+    }
+
+#ifdef _WIN32
+    int namelen = sizeof(serv_addr);
+#else
+    socklen_t namelen = sizeof(serv_addr);
+#endif
+    if (getsockname(sock, (struct sockaddr*)&serv_addr, &namelen) != 0) {
+        CLOSE_SOCKET(sock);
+#ifdef _WIN32
+        WSACleanup();
+#endif
+        return -1;
+    }
+
+    int port = ntohs(serv_addr.sin_port);
+
+    CLOSE_SOCKET(sock);
+#ifdef _WIN32
+    WSACleanup();
+#endif
+
+    return port;
+}
+
+// helper to convert vector<string> to char **
+// pointers are only valid as long as the original vector is valid
+static std::vector<char *> to_char_ptr_array(const std::vector<std::string> & vec) {
+    std::vector<char *> result;
+    result.reserve(vec.size() + 1);
+    for (const auto & s : vec) {
+        result.push_back(const_cast<char*>(s.c_str()));
+    }
+    result.push_back(nullptr);
+    return result;
+}
+
+std::vector<server_model_meta> server_models::get_all_meta() {
+    std::lock_guard<std::mutex> lk(mutex);
+    std::vector<server_model_meta> result;
+    result.reserve(mapping.size());
+    for (const auto & [name, inst] : mapping) {
+        result.push_back(inst.meta);
+    }
+    return result;
+}
+
+void server_models::unload_lru() {
+    if (base_params.models_max <= 0) {
+        return; // no limit
+    }
+    // remove one of the servers if we passed the models_max (least recently used - LRU)
+    std::string lru_model_name = "";
+    int64_t lru_last_used = ggml_time_ms();
+    size_t count_active = 0;
+    {
+        std::lock_guard<std::mutex> lk(mutex);
+        for (const auto & m : mapping) {
+            if (m.second.meta.is_active()) {
+                count_active++;
+                if (m.second.meta.last_used < lru_last_used) {
+                    lru_model_name = m.first;
+                    lru_last_used = m.second.meta.last_used;
+                }
+            }
+        }
+    }
+    if (!lru_model_name.empty() && count_active >= (size_t)base_params.models_max) {
+        SRV_INF("models_max limit reached, removing LRU name=%s\n", lru_model_name.c_str());
+        unload(lru_model_name);
+    }
+}
+
+static void add_or_replace_arg(std::vector<std::string> & args, const std::string & key, const std::string & value) {
+    for (size_t i = 0; i < args.size(); i++) {
+        if (args[i] == key && i + 1 < args.size()) {
+            args[i + 1] = value;
+            return;
+        }
+    }
+    // not found, append
+    args.push_back(key);
+    args.push_back(value);
+}
+
+void server_models::load(const std::string & name, bool auto_load) {
+    if (!has_model(name)) {
+        throw std::runtime_error("model name=" + name + " is not found");
+    }
+    unload_lru();
+
+    std::lock_guard<std::mutex> lk(mutex);
+
+    auto meta = mapping[name].meta;
+    if (meta.status != SERVER_MODEL_STATUS_UNLOADED) {
+        SRV_INF("model %s is not ready\n", name.c_str());
+        return;
+    }
+
+    // prepare new instance info
+    instance_t inst;
+    inst.meta           = meta;
+    inst.meta.port      = get_free_port();
+    inst.meta.status    = SERVER_MODEL_STATUS_LOADING;
+    inst.meta.last_used = ggml_time_ms();
+
+    if (inst.meta.port <= 0) {
+        throw std::runtime_error("failed to get a port number");
+    }
+
+    inst.subproc = std::make_shared<subprocess_s>();
+    {
+        SRV_INF("spawning server instance with name=%s on port %d\n", inst.meta.name.c_str(), inst.meta.port);
+
+        std::vector<std::string> child_args;
+        if (auto_load && !meta.args.empty()) {
+            child_args = meta.args; // copy previous args
+        } else {
+            child_args = base_args; // copy
+            if (inst.meta.in_cache) {
+                add_or_replace_arg(child_args, "-hf", inst.meta.name);
+            } else {
+                add_or_replace_arg(child_args, "-m", inst.meta.path);
+                if (!inst.meta.path_mmproj.empty()) {
+                    add_or_replace_arg(child_args, "--mmproj", inst.meta.path_mmproj);
+                }
+            }
+        }
+
+        // set model args
+        add_or_replace_arg(child_args, "--port", std::to_string(inst.meta.port));
+        add_or_replace_arg(child_args, "--alias", inst.meta.name);
+
+        std::vector<std::string> child_env = base_env; // copy
+        child_env.push_back("LLAMA_SERVER_ROUTER_PORT=" + std::to_string(base_params.port));
+
+        SRV_INF("%s", "spawning server instance with args:\n");
+        for (const auto & arg : child_args) {
+            SRV_INF("  %s\n", arg.c_str());
+        }
+        inst.meta.args = child_args; // save for debugging
+
+        std::vector<char *> argv = to_char_ptr_array(child_args);
+        std::vector<char *> envp = to_char_ptr_array(child_env);
+
+        int options = subprocess_option_no_window | subprocess_option_combined_stdout_stderr;
+        int result = subprocess_create_ex(argv.data(), options, envp.data(), inst.subproc.get());
+        if (result != 0) {
+            throw std::runtime_error("failed to spawn server instance");
+        }
+
+        inst.stdin_file = subprocess_stdin(inst.subproc.get());
+    }
+
+    // start a thread to manage the child process
+    // captured variables are guaranteed to be destroyed only after the thread is joined
+    inst.th = std::thread([this, name, child_proc = inst.subproc, port = inst.meta.port]() {
+        // read stdout/stderr and forward to main server log
+        FILE * p_stdout_stderr = subprocess_stdout(child_proc.get());
+        if (p_stdout_stderr) {
+            char buffer[4096];
+            while (fgets(buffer, sizeof(buffer), p_stdout_stderr) != nullptr) {
+                LOG("[%5d] %s", port, buffer);
+            }
+        } else {
+            SRV_ERR("failed to get stdout/stderr of child process for name=%s\n", name.c_str());
+        }
+        // we reach here when the child process exits
+        int exit_code = 0;
+        subprocess_join(child_proc.get(), &exit_code);
+        subprocess_destroy(child_proc.get());
+        // update PID and status
+        {
+            std::lock_guard<std::mutex> lk(mutex);
+            auto it = mapping.find(name);
+            if (it != mapping.end()) {
+                auto & meta = it->second.meta;
+                meta.exit_code = exit_code;
+                meta.status    = SERVER_MODEL_STATUS_UNLOADED;
+            }
+            cv.notify_all();
+        }
+        SRV_INF("instance name=%s exited with status %d\n", name.c_str(), exit_code);
+    });
+
+    // clean up old process/thread if exists
+    {
+        auto & old_instance = mapping[name];
+        // old process should have exited already, but just in case, we clean it up here
+        if (subprocess_alive(old_instance.subproc.get())) {
+            SRV_WRN("old process for model name=%s is still alive, this is unexpected\n", name.c_str());
+            subprocess_terminate(old_instance.subproc.get()); // force kill
+        }
+        if (old_instance.th.joinable()) {
+            old_instance.th.join();
+        }
+    }
+
+    mapping[name] = std::move(inst);
+    cv.notify_all();
+}
+
+static void interrupt_subprocess(FILE * stdin_file) {
+    // because subprocess.h does not provide a way to send SIGINT,
+    // we will send a command to the child process to exit gracefully
+    if (stdin_file) {
+        fprintf(stdin_file, "%s\n", CMD_EXIT);
+        fflush(stdin_file);
+    }
+}
+
+void server_models::unload(const std::string & name) {
+    std::lock_guard<std::mutex> lk(mutex);
+    auto it = mapping.find(name);
+    if (it != mapping.end()) {
+        if (it->second.meta.is_active()) {
+            SRV_INF("unloading model instance name=%s\n", name.c_str());
+            interrupt_subprocess(it->second.stdin_file);
+            // status change will be handled by the managing thread
+        } else {
+            SRV_WRN("model instance name=%s is not loaded\n", name.c_str());
+        }
+    }
+}
+
+void server_models::unload_all() {
+    std::vector<std::thread> to_join;
+    {
+        std::lock_guard<std::mutex> lk(mutex);
+        for (auto & [name, inst] : mapping) {
+            if (inst.meta.is_active()) {
+                SRV_INF("unloading model instance name=%s\n", name.c_str());
+                interrupt_subprocess(inst.stdin_file);
+                // status change will be handled by the managing thread
+            }
+            // moving the thread to join list to avoid deadlock
+            to_join.push_back(std::move(inst.th));
+        }
+    }
+    for (auto & th : to_join) {
+        if (th.joinable()) {
+            th.join();
+        }
+    }
+}
+
+void server_models::update_status(const std::string & name, server_model_status status) {
+    // for now, we only allow updating to LOADED status
+    if (status != SERVER_MODEL_STATUS_LOADED) {
+        throw std::runtime_error("invalid status value");
+    }
+    auto meta = get_meta(name);
+    if (meta.has_value()) {
+        meta->status = status;
+        update_meta(name, meta.value());
+    }
+}
+
+void server_models::wait_until_loaded(const std::string & name) {
+    std::unique_lock<std::mutex> lk(mutex);
+    cv.wait(lk, [this, &name]() {
+        auto it = mapping.find(name);
+        if (it != mapping.end()) {
+            return it->second.meta.status != SERVER_MODEL_STATUS_LOADING;
+        }
+        return false;
+    });
+}
+
+bool server_models::ensure_model_loaded(const std::string & name) {
+    auto meta = get_meta(name);
+    if (!meta.has_value()) {
+        throw std::runtime_error("model name=" + name + " is not found");
+    }
+    if (meta->status == SERVER_MODEL_STATUS_LOADED) {
+        return false; // already loaded
+    }
+    if (meta->status == SERVER_MODEL_STATUS_UNLOADED) {
+        SRV_INF("model name=%s is not loaded, loading...\n", name.c_str());
+        load(name, true);
+    }
+
+    SRV_INF("waiting until model name=%s is fully loaded...\n", name.c_str());
+    wait_until_loaded(name);
+
+    // check final status
+    meta = get_meta(name);
+    if (!meta.has_value() || meta->is_failed()) {
+        throw std::runtime_error("model name=" + name + " failed to load");
+    }
+
+    return true;
+}
+
+server_http_res_ptr server_models::proxy_request(const server_http_req & req, const std::string & method, const std::string & name, bool update_last_used) {
+    auto meta = get_meta(name);
+    if (!meta.has_value()) {
+        throw std::runtime_error("model name=" + name + " is not found");
+    }
+    if (meta->status != SERVER_MODEL_STATUS_LOADED) {
+        throw std::invalid_argument("model name=" + name + " is not loaded");
+    }
+    if (update_last_used) {
+        std::unique_lock<std::mutex> lk(mutex);
+        mapping[name].meta.last_used = ggml_time_ms();
+    }
+    SRV_INF("proxying request to model %s on port %d\n", name.c_str(), meta->port);
+    auto proxy = std::make_unique<server_http_proxy>(
+            method,
+            base_params.hostname,
+            meta->port,
+            req.path,
+            req.headers,
+            req.body,
+            req.should_stop);
+    return proxy;
+}
+
+std::thread server_models::setup_child_server(const common_params & base_params, int router_port, const std::string & name, std::function<void(int)> & shutdown_handler) {
+    // send a notification to the router server that a model instance is ready
+    // TODO @ngxson : use HTTP client from libcommon
+    httplib::Client cli(base_params.hostname, router_port);
+    cli.set_connection_timeout(0, 200000); // 200 milliseconds
+
+    httplib::Request req;
+    req.method = "POST";
+    req.path   = "/models/status";
+    req.set_header("Content-Type", "application/json");
+    if (!base_params.api_keys.empty()) {
+        req.set_header("Authorization", "Bearer " + base_params.api_keys[0]);
+    }
+
+    json body;
+    body["model"] = name;
+    body["value"] = server_model_status_to_string(SERVER_MODEL_STATUS_LOADED);
+    req.body = body.dump();
+
+    SRV_INF("notifying router server (port=%d) that model %s is ready\n", router_port, name.c_str());
+    auto result = cli.send(std::move(req));
+    if (result.error() != httplib::Error::Success) {
+        auto err_str = httplib::to_string(result.error());
+        SRV_ERR("failed to notify router server: %s\n", err_str.c_str());
+        exit(1); // force exit
+    }
+
+    // setup thread for monitoring stdin
+    return std::thread([shutdown_handler]() {
+        // wait for EOF on stdin
+        SRV_INF("%s", "child server monitoring thread started, waiting for EOF on stdin...\n");
+        bool eof = false;
+        while (true) {
+            std::string line;
+            if (!std::getline(std::cin, line)) {
+                // EOF detected, that means the router server is unexpectedly exit or killed
+                eof = true;
+                break;
+            }
+            if (line.find(CMD_EXIT) != std::string::npos) {
+                SRV_INF("%s", "exit command received, exiting...\n");
+                shutdown_handler(0);
+                break;
+            }
+        }
+        if (eof) {
+            SRV_INF("%s", "EOF on stdin detected, forcing shutdown...\n");
+            exit(1);
+        }
+    });
+}
+
+
+
+//
+// server_models_routes
+//
+
+static void res_ok(std::unique_ptr<server_http_res> & res, const json & response_data) {
+    res->status = 200;
+    res->data = safe_json_to_str(response_data);
+}
+
+static void res_err(std::unique_ptr<server_http_res> & res, const json & error_data) {
+    res->status = json_value(error_data, "code", 500);
+    res->data = safe_json_to_str({{ "error", error_data }});
+}
+
+static bool router_validate_model(const std::string & name, server_models & models, bool models_autoload, std::unique_ptr<server_http_res> & res) {
+    if (name.empty()) {
+        res_err(res, format_error_response("model name is missing from the request", ERROR_TYPE_INVALID_REQUEST));
+        return false;
+    }
+    auto meta = models.get_meta(name);
+    if (!meta.has_value()) {
+        res_err(res, format_error_response("model not found", ERROR_TYPE_INVALID_REQUEST));
+        return false;
+    }
+    if (models_autoload) {
+        models.ensure_model_loaded(name);
+    } else {
+        if (meta->status != SERVER_MODEL_STATUS_LOADED) {
+            res_err(res, format_error_response("model is not loaded", ERROR_TYPE_INVALID_REQUEST));
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool is_autoload(const common_params & params, const server_http_req & req) {
+    std::string autoload = req.get_param("autoload");
+    if (autoload.empty()) {
+        return params.models_autoload;
+    } else {
+        return autoload == "true" || autoload == "1";
+    }
+}
+
+void server_models_routes::init_routes() {
+    this->get_router_props = [this](const server_http_req & req) {
+        std::string name = req.get_param("model");
+        if (name.empty()) {
+            // main instance
+            auto res = std::make_unique<server_http_res>();
+            res_ok(res, {
+                // TODO: add support for this on web UI
+                {"role",          "router"},
+                {"max_instances", 4}, // dummy value for testing
+                // this is a dummy response to make sure webui doesn't break
+                {"model_alias", "llama-server"},
+                {"model_path",  "none"},
+                {"default_generation_settings", {
+                    {"params", json{}},
+                    {"n_ctx",  0},
+                }},
+            });
+            return res;
+        }
+        return proxy_get(req);
+    };
+
+    this->proxy_get = [this](const server_http_req & req) {
+        std::string method = "GET";
+        std::string name = req.get_param("model");
+        bool autoload = is_autoload(params, req);
+        auto error_res = std::make_unique<server_http_res>();
+        if (!router_validate_model(name, models, autoload, error_res)) {
+            return error_res;
+        }
+        return models.proxy_request(req, method, name, false);
+    };
+
+    this->proxy_post = [this](const server_http_req & req) {
+        std::string method = "POST";
+        json body = json::parse(req.body);
+        std::string name = json_value(body, "model", std::string());
+        bool autoload = is_autoload(params, req);
+        auto error_res = std::make_unique<server_http_res>();
+        if (!router_validate_model(name, models, autoload, error_res)) {
+            return error_res;
+        }
+        return models.proxy_request(req, method, name, true); // update last usage for POST request only
+    };
+
+    this->get_router_models = [this](const server_http_req &) {
+        auto res = std::make_unique<server_http_res>();
+        json models_json = json::array();
+        auto all_models = models.get_all_meta();
+        std::time_t t = std::time(0);
+        for (const auto & meta : all_models) {
+            json status {
+                {"value", server_model_status_to_string(meta.status)},
+                {"args",  meta.args},
+            };
+            if (meta.is_failed()) {
+                status["exit_code"] = meta.exit_code;
+                status["failed"]    = true;
+            }
+            models_json.push_back(json {
+                {"id",       meta.name},
+                {"object",   "model"},    // for OAI-compat
+                {"owned_by", "llamacpp"}, // for OAI-compat
+                {"created",  t},          // for OAI-compat
+                {"in_cache", meta.in_cache},
+                {"path",     meta.path},
+                {"status",   status},
+                // TODO: add other fields, may require reading GGUF metadata
+            });
+        }
+        res_ok(res, {
+            {"data", models_json},
+            {"object", "list"},
+        });
+        return res;
+    };
+
+    this->post_router_models_load = [this](const server_http_req & req) {
+        auto res = std::make_unique<server_http_res>();
+        json body = json::parse(req.body);
+        std::string name = json_value(body, "model", std::string());
+        auto model = models.get_meta(name);
+        if (!model.has_value()) {
+            res_err(res, format_error_response("model is not found", ERROR_TYPE_NOT_FOUND));
+            return res;
+        }
+        if (model->status == SERVER_MODEL_STATUS_LOADED) {
+            res_err(res, format_error_response("model is already loaded", ERROR_TYPE_INVALID_REQUEST));
+            return res;
+        }
+        models.load(name, false);
+        res_ok(res, {{"success", true}});
+        return res;
+    };
+
+    // used by child process to notify the router about status change
+    // TODO @ngxson : maybe implement authentication for this endpoint in the future
+    this->post_router_models_status = [this](const server_http_req & req) {
+        auto res = std::make_unique<server_http_res>();
+        json body = json::parse(req.body);
+        std::string model = json_value(body, "model", std::string());
+        std::string value = json_value(body, "value", std::string());
+        models.update_status(model, server_model_status_from_string(value));
+        res_ok(res, {{"success", true}});
+        return res;
+    };
+
+    this->get_router_models = [this](const server_http_req &) {
+        auto res = std::make_unique<server_http_res>();
+        json models_json = json::array();
+        auto all_models = models.get_all_meta();
+        std::time_t t = std::time(0);
+        for (const auto & meta : all_models) {
+            json status {
+                {"value", server_model_status_to_string(meta.status)},
+                {"args",  meta.args},
+            };
+            if (meta.is_failed()) {
+                status["exit_code"] = meta.exit_code;
+                status["failed"]    = true;
+            }
+            models_json.push_back(json {
+                {"id",       meta.name},
+                {"object",   "model"},    // for OAI-compat
+                {"owned_by", "llamacpp"}, // for OAI-compat
+                {"created",  t},          // for OAI-compat
+                {"in_cache", meta.in_cache},
+                {"path",     meta.path},
+                {"status",   status},
+                // TODO: add other fields, may require reading GGUF metadata
+            });
+        }
+        res_ok(res, {
+            {"data", models_json},
+            {"object", "list"},
+        });
+        return res;
+    };
+
+    this->post_router_models_unload = [this](const server_http_req & req) {
+        auto res = std::make_unique<server_http_res>();
+        json body = json::parse(req.body);
+        std::string name = json_value(body, "model", std::string());
+        auto model = models.get_meta(name);
+        if (!model.has_value()) {
+            res_err(res, format_error_response("model is not found", ERROR_TYPE_INVALID_REQUEST));
+            return res;
+        }
+        if (model->status != SERVER_MODEL_STATUS_LOADED) {
+            res_err(res, format_error_response("model is not loaded", ERROR_TYPE_INVALID_REQUEST));
+            return res;
+        }
+        models.unload(name);
+        res_ok(res, {{"success", true}});
+        return res;
+    };
+}
+
+
+
+//
+// server_http_proxy
+//
+
+// simple implementation of a pipe
+// used for streaming data between threads
+template<typename T>
+struct pipe_t {
+    std::mutex mutex;
+    std::condition_variable cv;
+    std::queue<T> queue;
+    std::atomic<bool> writer_closed{false};
+    std::atomic<bool> reader_closed{false};
+    void close_write() {
+        writer_closed.store(true, std::memory_order_relaxed);
+        cv.notify_all();
+    }
+    void close_read() {
+        reader_closed.store(true, std::memory_order_relaxed);
+        cv.notify_all();
+    }
+    bool read(T & output, const std::function<bool()> & should_stop) {
+        std::unique_lock<std::mutex> lk(mutex);
+        constexpr auto poll_interval = std::chrono::milliseconds(500);
+        while (true) {
+            if (!queue.empty()) {
+                output = std::move(queue.front());
+                queue.pop();
+                return true;
+            }
+            if (writer_closed.load()) {
+                return false; // clean EOF
+            }
+            if (should_stop()) {
+                close_read(); // signal broken pipe to writer
+                return false; // cancelled / reader no longer alive
+            }
+            cv.wait_for(lk, poll_interval);
+        }
+    }
+    bool write(T && data) {
+        std::lock_guard<std::mutex> lk(mutex);
+        if (reader_closed.load()) {
+            return false; // broken pipe
+        }
+        queue.push(std::move(data));
+        cv.notify_one();
+        return true;
+    }
+};
+
+server_http_proxy::server_http_proxy(
+        const std::string & method,
+        const std::string & host,
+        int port,
+        const std::string & path,
+        const std::map<std::string, std::string> & headers,
+        const std::string & body,
+        const std::function<bool()> should_stop) {
+    // shared between reader and writer threads
+    auto cli  = std::make_shared<httplib::Client>(host, port);
+    auto pipe = std::make_shared<pipe_t<msg_t>>();
+
+    // setup Client
+    cli->set_connection_timeout(0, 200000); // 200 milliseconds
+    this->status = 500; // to be overwritten upon response
+    this->cleanup = [pipe]() {
+        pipe->close_read();
+        pipe->close_write();
+    };
+
+    // wire up the receive end of the pipe
+    this->next = [pipe, should_stop](std::string & out) -> bool {
+        msg_t msg;
+        bool has_next = pipe->read(msg, should_stop);
+        if (!msg.data.empty()) {
+            out = std::move(msg.data);
+        }
+        return has_next; // false if EOF or pipe broken
+    };
+
+    // wire up the HTTP client
+    // note: do NOT capture `this` pointer, as it may be destroyed before the thread ends
+    httplib::ResponseHandler response_handler = [pipe, cli](const httplib::Response & response) {
+        msg_t msg;
+        msg.status = response.status;
+        for (const auto & [key, value] : response.headers) {
+            msg.headers[key] = value;
+        }
+        return pipe->write(std::move(msg)); // send headers first
+    };
+    httplib::ContentReceiverWithProgress content_receiver = [pipe](const char * data, size_t data_length, size_t, size_t) {
+        // send data chunks
+        // returns false if pipe is closed / broken (signal to stop receiving)
+        return pipe->write({{}, 0, std::string(data, data_length)});
+    };
+
+    // prepare the request to destination server
+    httplib::Request req;
+    {
+        req.method = method;
+        req.path = path;
+        for (const auto & [key, value] : headers) {
+            req.set_header(key, value);
+        }
+        req.body = body;
+        req.response_handler = response_handler;
+        req.content_receiver = content_receiver;
+    }
+
+    // start the proxy thread
+    SRV_DBG("start proxy thread %s %s\n", req.method.c_str(), req.path.c_str());
+    this->thread = std::thread([cli, pipe, req]() {
+        auto result = cli->send(std::move(req));
+        if (result.error() != httplib::Error::Success) {
+            auto err_str = httplib::to_string(result.error());
+            SRV_ERR("http client error: %s\n", err_str.c_str());
+            pipe->write({{}, 500, ""}); // header
+            pipe->write({{}, 0, "proxy error: " + err_str}); // body
+        }
+        pipe->close_write(); // signal EOF to reader
+        SRV_DBG("%s", "client request thread ended\n");
+    });
+    this->thread.detach();
+
+    // wait for the first chunk (headers)
+    msg_t header;
+    if (pipe->read(header, should_stop)) {
+        SRV_DBG("%s", "received response headers\n");
+        this->status  = header.status;
+        this->headers = header.headers;
+    } else {
+        SRV_DBG("%s", "no response headers received (request cancelled?)\n");
+    }
+}

tools/server/server-models.h

@@ -0,0 +1,174 @@
+#pragma once
+
+#include "common.h"
+#include "server-http.h"
+
+#include <mutex>
+#include <condition_variable>
+#include <functional>
+#include <memory>
+
+/**
+ * state diagram:
+ *
+ * UNLOADED ──► LOADING ──► LOADED
+ *  ▲            │            │
+ *  └───failed───┘            │
+ *  ▲                         │
+ *  └────────unloaded─────────┘
+ */
+enum server_model_status {
+    // TODO: also add downloading state when the logic is added
+    SERVER_MODEL_STATUS_UNLOADED,
+    SERVER_MODEL_STATUS_LOADING,
+    SERVER_MODEL_STATUS_LOADED
+};
+
+static server_model_status server_model_status_from_string(const std::string & status_str) {
+    if (status_str == "unloaded") {
+        return SERVER_MODEL_STATUS_UNLOADED;
+    }
+    if (status_str == "loading") {
+        return SERVER_MODEL_STATUS_LOADING;
+    }
+    if (status_str == "loaded") {
+        return SERVER_MODEL_STATUS_LOADED;
+    }
+    throw std::runtime_error("invalid server model status");
+}
+
+static std::string server_model_status_to_string(server_model_status status) {
+    switch (status) {
+        case SERVER_MODEL_STATUS_UNLOADED: return "unloaded";
+        case SERVER_MODEL_STATUS_LOADING:  return "loading";
+        case SERVER_MODEL_STATUS_LOADED:   return "loaded";
+        default:                           return "unknown";
+    }
+}
+
+struct server_model_meta {
+    std::string name;
+    std::string path;
+    std::string path_mmproj; // only available if in_cache=false
+    bool in_cache = false; // if true, use -hf; use -m otherwise
+    int port = 0;
+    server_model_status status = SERVER_MODEL_STATUS_UNLOADED;
+    int64_t last_used = 0; // for LRU unloading
+    std::vector<std::string> args; // additional args passed to the model instance (used for debugging)
+    int exit_code = 0; // exit code of the model instance process (only valid if status == FAILED)
+
+    bool is_active() const {
+        return status == SERVER_MODEL_STATUS_LOADED || status == SERVER_MODEL_STATUS_LOADING;
+    }
+
+    bool is_failed() const {
+        return status == SERVER_MODEL_STATUS_UNLOADED && exit_code != 0;
+    }
+};
+
+struct subprocess_s;
+
+struct server_models {
+private:
+    struct instance_t {
+        std::shared_ptr<subprocess_s> subproc; // shared between main thread and monitoring thread
+        std::thread th;
+        server_model_meta meta;
+        FILE * stdin_file = nullptr;
+    };
+
+    std::mutex mutex;
+    std::condition_variable cv;
+    std::map<std::string, instance_t> mapping;
+
+    common_params base_params;
+    std::vector<std::string> base_args;
+    std::vector<std::string> base_env;
+
+    void update_meta(const std::string & name, const server_model_meta & meta);
+
+    // unload least recently used models if the limit is reached
+    void unload_lru();
+
+public:
+    server_models(const common_params & params, int argc, char ** argv, char ** envp);
+
+    // check if a model instance exists
+    bool has_model(const std::string & name);
+
+    // return a copy of model metadata
+    std::optional<server_model_meta> get_meta(const std::string & name);
+
+    // return a copy of all model metadata
+    std::vector<server_model_meta> get_all_meta();
+
+    // if auto_load is true, load the model with previous args if any
+    void load(const std::string & name, bool auto_load);
+    void unload(const std::string & name);
+    void unload_all();
+
+    // update the status of a model instance
+    void update_status(const std::string & name, server_model_status status);
+
+    // wait until the model instance is fully loaded
+    // return when the model is loaded or failed to load
+    void wait_until_loaded(const std::string & name);
+
+    // load the model if not loaded, otherwise do nothing
+    // return false if model is already loaded; return true otherwise (meta may need to be refreshed)
+    bool ensure_model_loaded(const std::string & name);
+
+    // 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);
+
+    // notify the router server that a model instance is ready
+    // return the monitoring thread (to be joined by the caller)
+    static std::thread setup_child_server(const common_params & base_params, int router_port, const std::string & name, std::function<void(int)> & shutdown_handler);
+};
+
+struct server_models_routes {
+    common_params params;
+    server_models models;
+    server_models_routes(const common_params & params, int argc, char ** argv, char ** envp)
+            : params(params), models(params, argc, argv, envp) {
+        init_routes();
+    }
+
+    void init_routes();
+    // handlers using lambda function, so that they can capture `this` without `std::bind`
+    server_http_context::handler_t get_router_props;
+    server_http_context::handler_t proxy_get;
+    server_http_context::handler_t proxy_post;
+    server_http_context::handler_t get_router_models;
+    server_http_context::handler_t post_router_models_load;
+    server_http_context::handler_t post_router_models_status;
+    server_http_context::handler_t post_router_models_unload;
+};
+
+/**
+ * A simple HTTP proxy that forwards requests to another server
+ * and relays the responses back.
+ */
+struct server_http_proxy : server_http_res {
+    std::function<void()> cleanup = nullptr;
+public:
+    server_http_proxy(const std::string & method,
+                      const std::string & host,
+                      int port,
+                      const std::string & path,
+                      const std::map<std::string, std::string> & headers,
+                      const std::string & body,
+                      const std::function<bool()> should_stop);
+    ~server_http_proxy() {
+        if (cleanup) {
+            cleanup();
+        }
+    }
+private:
+    std::thread thread;
+    struct msg_t {
+        std::map<std::string, std::string> headers;
+        int status = 0;
+        std::string data;
+    };
+};

tools/server/server-task.cpp

@@ -450,9 +450,6 @@ task_params server_task::params_from_json_cmpl(
         }
     }
 
-    std::string model_name = params_base.model_alias.empty() ? DEFAULT_OAICOMPAT_MODEL : params_base.model_alias;
-    params.oaicompat_model = json_value(data, "model", model_name);
-
     return params;
 }
 

tools/server/server.cpp

@@ -1,5 +1,6 @@
 #include "server-context.h"
 #include "server-http.h"
+#include "server-models.h"
 
 #include "arg.h"
 #include "common.h"
@@ -33,30 +34,38 @@ static inline void signal_handler(int signal) {
 static server_http_context::handler_t ex_wrapper(server_http_context::handler_t func) {
     return [func = std::move(func)](const server_http_req & req) -> server_http_res_ptr {
         std::string message;
+        error_type error;
         try {
             return func(req);
+        } catch (const std::invalid_argument & e) {
+            // treat invalid_argument as invalid request (400)
+            error = ERROR_TYPE_INVALID_REQUEST;
+            message = e.what();
         } catch (const std::exception & e) {
+            // treat other exceptions as server error (500)
+            error = ERROR_TYPE_SERVER;
             message = e.what();
         } catch (...) {
+            error = ERROR_TYPE_SERVER;
             message = "unknown error";
         }
 
         auto res = std::make_unique<server_http_res>();
         res->status = 500;
         try {
-            json error_data = format_error_response(message, ERROR_TYPE_SERVER);
+            json error_data = format_error_response(message, error);
             res->status = json_value(error_data, "code", 500);
             res->data = safe_json_to_str({{ "error", error_data }});
-            LOG_WRN("got exception: %s\n", res->data.c_str());
+            SRV_WRN("got exception: %s\n", res->data.c_str());
         } catch (const std::exception & e) {
-            LOG_ERR("got another exception: %s | while hanlding exception: %s\n", e.what(), message.c_str());
+            SRV_ERR("got another exception: %s | while handling exception: %s\n", e.what(), message.c_str());
             res->data = "Internal Server Error";
         }
         return res;
     };
 }
 
-int main(int argc, char ** argv) {
+int main(int argc, char ** argv, char ** envp) {
     // own arguments required by this example
     common_params params;
 
@@ -75,6 +84,11 @@ int main(int argc, char ** argv) {
         params.kv_unified = true;
     }
 
+    // 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 = params.model.name;
+    }
+
     common_init();
 
     // struct that contains llama context and inference
@@ -101,6 +115,42 @@ int main(int argc, char ** argv) {
     // register API routes
     server_routes routes(params, ctx_server, [&ctx_http]() { return ctx_http.is_ready.load(); });
 
+    bool is_router_server = params.model.path.empty();
+    std::optional<server_models_routes> models_routes{};
+    if (is_router_server) {
+        // setup server instances manager
+        models_routes.emplace(params, argc, argv, envp);
+
+        // proxy handlers
+        // note: routes.get_health stays the same
+        routes.get_metrics                 = models_routes->proxy_get;
+        routes.post_props                  = models_routes->proxy_post;
+        routes.get_api_show                = models_routes->proxy_get;
+        routes.post_completions            = models_routes->proxy_post;
+        routes.post_completions_oai        = models_routes->proxy_post;
+        routes.post_chat_completions       = models_routes->proxy_post;
+        routes.post_anthropic_messages     = models_routes->proxy_post;
+        routes.post_anthropic_count_tokens = models_routes->proxy_post;
+        routes.post_infill                 = models_routes->proxy_post;
+        routes.post_embeddings             = models_routes->proxy_post;
+        routes.post_embeddings_oai         = models_routes->proxy_post;
+        routes.post_rerank                 = models_routes->proxy_post;
+        routes.post_tokenize               = models_routes->proxy_post;
+        routes.post_detokenize             = models_routes->proxy_post;
+        routes.post_apply_template         = models_routes->proxy_post;
+        routes.get_lora_adapters           = models_routes->proxy_get;
+        routes.post_lora_adapters          = models_routes->proxy_post;
+        routes.get_slots                   = models_routes->proxy_get;
+        routes.post_slots                  = models_routes->proxy_post;
+
+        // custom routes for router
+        routes.get_props  = models_routes->get_router_props;
+        routes.get_models = models_routes->get_router_models;
+        ctx_http.post("/models/load",   ex_wrapper(models_routes->post_router_models_load));
+        ctx_http.post("/models/unload", ex_wrapper(models_routes->post_router_models_unload));
+        ctx_http.post("/models/status", ex_wrapper(models_routes->post_router_models_status));
+    }
+
     ctx_http.get ("/health",              ex_wrapper(routes.get_health)); // public endpoint (no API key check)
     ctx_http.get ("/v1/health",           ex_wrapper(routes.get_health)); // public endpoint (no API key check)
     ctx_http.get ("/metrics",             ex_wrapper(routes.get_metrics));
@@ -140,43 +190,69 @@ int main(int argc, char ** argv) {
     // Start the server
     //
 
-    // setup clean up function, to be called before exit
-    auto clean_up = [&ctx_http, &ctx_server]() {
-        SRV_INF("%s: cleaning up before exit...\n", __func__);
-        ctx_http.stop();
-        ctx_server.terminate();
-        llama_backend_free();
-    };
+    std::function<void()> clean_up;
 
-    // start the HTTP server before loading the model to be able to serve /health requests
-    if (!ctx_http.start()) {
-        clean_up();
-        LOG_ERR("%s: exiting due to HTTP server error\n", __func__);
-        return 1;
-    }
+    if (is_router_server) {
+        LOG_INF("%s: starting router server, no model will be loaded in this process\n", __func__);
 
-    // load the model
-    LOG_INF("%s: loading model\n", __func__);
+        clean_up = [&models_routes]() {
+            SRV_INF("%s: cleaning up before exit...\n", __func__);
+            if (models_routes.has_value()) {
+                models_routes->models.unload_all();
+            }
+            llama_backend_free();
+        };
 
-    if (!ctx_server.load_model(params)) {
-        clean_up();
-        if (ctx_http.thread.joinable()) {
-            ctx_http.thread.join();
+        if (!ctx_http.start()) {
+            clean_up();
+            LOG_ERR("%s: exiting due to HTTP server error\n", __func__);
+            return 1;
         }
-        LOG_ERR("%s: exiting due to model loading error\n", __func__);
-        return 1;
+        ctx_http.is_ready.store(true);
+
+        shutdown_handler = [&](int) {
+            ctx_http.stop();
+        };
+
+    } else {
+        // setup clean up function, to be called before exit
+        clean_up = [&ctx_http, &ctx_server]() {
+            SRV_INF("%s: cleaning up before exit...\n", __func__);
+            ctx_http.stop();
+            ctx_server.terminate();
+            llama_backend_free();
+        };
+
+        // start the HTTP server before loading the model to be able to serve /health requests
+        if (!ctx_http.start()) {
+            clean_up();
+            LOG_ERR("%s: exiting due to HTTP server error\n", __func__);
+            return 1;
+        }
+
+        // load the model
+        LOG_INF("%s: loading model\n", __func__);
+
+        if (!ctx_server.load_model(params)) {
+            clean_up();
+            if (ctx_http.thread.joinable()) {
+                ctx_http.thread.join();
+            }
+            LOG_ERR("%s: exiting due to model loading error\n", __func__);
+            return 1;
+        }
+
+        ctx_server.init();
+        ctx_http.is_ready.store(true);
+
+        LOG_INF("%s: model loaded\n", __func__);
+
+        shutdown_handler = [&](int) {
+            // this will unblock start_loop()
+            ctx_server.terminate();
+        };
     }
 
-    ctx_server.init();
-    ctx_http.is_ready.store(true);
-
-    LOG_INF("%s: model loaded\n", __func__);
-
-    shutdown_handler = [&](int) {
-        // this will unblock start_loop()
-        ctx_server.terminate();
-    };
-
     // TODO: refactor in common/console
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
     struct sigaction sigint_action;
@@ -192,16 +268,39 @@ int main(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-    LOG_INF("%s: server is listening on %s\n", __func__, ctx_http.listening_address.c_str());
-    LOG_INF("%s: starting the main loop...\n", __func__);
-    // this call blocks the main thread until ctx_server.terminate() is called
-    ctx_server.start_loop();
+    if (is_router_server) {
+        LOG_INF("%s: router server is listening on %s\n", __func__, ctx_http.listening_address.c_str());
+        LOG_INF("%s: NOTE: router mode is experimental\n", __func__);
+        LOG_INF("%s:       it is not recommended to use this mode in untrusted environments\n", __func__);
+        if (ctx_http.thread.joinable()) {
+            ctx_http.thread.join(); // keep the main thread alive
+        }
 
-    clean_up();
-    if (ctx_http.thread.joinable()) {
-        ctx_http.thread.join();
+        // when the HTTP server stops, clean up and exit
+        clean_up();
+    } else {
+        LOG_INF("%s: server is listening on %s\n", __func__, ctx_http.listening_address.c_str());
+        LOG_INF("%s: starting the main loop...\n", __func__);
+
+        // optionally, notify router server that this instance is ready
+        const char * router_port = std::getenv("LLAMA_SERVER_ROUTER_PORT");
+        std::thread monitor_thread;
+        if (router_port != nullptr) {
+            monitor_thread = server_models::setup_child_server(params, std::atoi(router_port), params.model_alias, shutdown_handler);
+        }
+
+        // this call blocks the main thread until queue_tasks.terminate() is called
+        ctx_server.start_loop();
+
+        clean_up();
+        if (ctx_http.thread.joinable()) {
+            ctx_http.thread.join();
+        }
+        if (monitor_thread.joinable()) {
+            monitor_thread.join();
+        }
+        llama_memory_breakdown_print(ctx_server.get_llama_context());
     }
-    llama_memory_breakdown_print(ctx_server.get_llama_context());
 
     return 0;
 }

tools/server/tests/unit/test_chat_completion.py

@@ -41,7 +41,8 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte
     assert res.status_code == 200
     assert "cmpl" in res.body["id"] # make sure the completion id has the expected format
     assert res.body["system_fingerprint"].startswith("b")
-    assert res.body["model"] == model if model is not None else server.model_alias
+    # we no longer reflect back the model name, see https://github.com/ggml-org/llama.cpp/pull/17668
+    # assert res.body["model"] == model if model is not None else server.model_alias
     assert res.body["usage"]["prompt_tokens"] == n_prompt
     assert res.body["usage"]["completion_tokens"] == n_predicted
     choice = res.body["choices"][0]
@@ -59,7 +60,7 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte
 )
 def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason):
     global server
-    server.model_alias = None # try using DEFAULT_OAICOMPAT_MODEL
+    server.model_alias = "llama-test-model"
     server.start()
     res = server.make_stream_request("POST", "/chat/completions", data={
         "max_tokens": max_tokens,
@@ -81,7 +82,7 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
             else:
                 assert "role" not in choice["delta"]
             assert data["system_fingerprint"].startswith("b")
-            assert "gpt-3.5" in data["model"] # DEFAULT_OAICOMPAT_MODEL, maybe changed in the future
+            assert data["model"] == "llama-test-model"
             if last_cmpl_id is None:
                 last_cmpl_id = data["id"]
             assert last_cmpl_id == data["id"] # make sure the completion id is the same for all events in the stream
@@ -198,7 +199,7 @@ def test_completion_with_response_format(response_format: dict, n_predicted: int
         choice = res.body["choices"][0]
         assert match_regex(re_content, choice["message"]["content"])
     else:
-        assert res.status_code != 200
+        assert res.status_code == 400
         assert "error" in res.body
 
 

tools/server/tests/unit/test_router.py

@@ -0,0 +1,50 @@
+import pytest
+from utils import *
+
+server: ServerProcess
+
+@pytest.fixture(autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.router()
+
+
+@pytest.mark.parametrize(
+    "model,success",
+    [
+        ("ggml-org/tinygemma3-GGUF:Q8_0", True),
+        ("non-existent/model", False),
+    ]
+)
+def test_router_chat_completion_stream(model: str, success: bool):
+    # TODO: make sure the model is in cache (ie. ServerProcess.load_all()) before starting the router server
+    global server
+    server.start()
+    content = ""
+    ex: ServerError | None = None
+    try:
+        res = server.make_stream_request("POST", "/chat/completions", data={
+            "model": model,
+            "max_tokens": 16,
+            "messages": [
+                {"role": "user", "content": "hello"},
+            ],
+            "stream": True,
+        })
+        for data in res:
+            if data["choices"]:
+                choice = data["choices"][0]
+                if choice["finish_reason"] in ["stop", "length"]:
+                    assert "content" not in choice["delta"]
+                else:
+                    assert choice["finish_reason"] is None
+                    content += choice["delta"]["content"] or ''
+    except ServerError as e:
+        ex = e
+
+    if success:
+        assert ex is None
+        assert len(content) > 0
+    else:
+        assert ex is not None
+        assert content == ""

tools/server/tests/unit/test_security.py

@@ -94,3 +94,34 @@ def test_cors_options(origin: str, cors_header: str, cors_header_value: str):
     assert res.status_code == 200
     assert cors_header in res.headers
     assert res.headers[cors_header] == cors_header_value
+
+
+@pytest.mark.parametrize(
+    "media_path, image_url, success",
+    [
+        (None,             "file://mtmd/test-1.jpeg",    False), # disabled media path, should fail
+        ("../../../tools", "file://mtmd/test-1.jpeg",    True),
+        ("../../../tools", "file:////mtmd//test-1.jpeg", True),  # should be the same file as above
+        ("../../../tools", "file://mtmd/notfound.jpeg",  False), # non-existent file
+        ("../../../tools", "file://../mtmd/test-1.jpeg", False), # no directory traversal
+    ]
+)
+def test_local_media_file(media_path, image_url, success,):
+    server = ServerPreset.tinygemma3()
+    server.media_path = media_path
+    server.start()
+    res = server.make_request("POST", "/chat/completions", data={
+        "max_tokens": 1,
+        "messages": [
+            {"role": "user", "content": [
+                {"type": "text", "text": "test"},
+                {"type": "image_url", "image_url": {
+                    "url": image_url,
+                }},
+            ]},
+        ],
+    })
+    if success:
+        assert res.status_code == 200
+    else:
+        assert res.status_code == 400

tools/server/tests/utils.py

@@ -46,7 +46,7 @@ class ServerProcess:
     debug: bool = False
     server_port: int = 8080
     server_host: str = "127.0.0.1"
-    model_hf_repo: str = "ggml-org/models"
+    model_hf_repo: str | None = "ggml-org/models"
     model_hf_file: str | None = "tinyllamas/stories260K.gguf"
     model_alias: str = "tinyllama-2"
     temperature: float = 0.8
@@ -95,6 +95,7 @@ class ServerProcess:
     chat_template_file: str | None = None
     server_path: str | None = None
     mmproj_url: str | None = None
+    media_path: str | None = None
 
     # session variables
     process: subprocess.Popen | None = None
@@ -217,6 +218,8 @@ class ServerProcess:
             server_args.extend(["--chat-template-file", self.chat_template_file])
         if self.mmproj_url:
             server_args.extend(["--mmproj-url", self.mmproj_url])
+        if self.media_path:
+            server_args.extend(["--media-path", self.media_path])
 
         args = [str(arg) for arg in [server_path, *server_args]]
         print(f"tests: starting server with: {' '.join(args)}")
@@ -521,9 +524,8 @@ class ServerPreset:
         server = ServerProcess()
         server.offline = True # will be downloaded by load_all()
         # mmproj is already provided by HF registry API
-        server.model_hf_repo = "ggml-org/tinygemma3-GGUF"
-        server.model_hf_file = "tinygemma3-Q8_0.gguf"
-        server.mmproj_url = "https://huggingface.co/ggml-org/tinygemma3-GGUF/resolve/main/mmproj-tinygemma3.gguf"
+        server.model_hf_file = None
+        server.model_hf_repo = "ggml-org/tinygemma3-GGUF:Q8_0"
         server.model_alias = "tinygemma3"
         server.n_ctx = 1024
         server.n_batch = 32
@@ -532,6 +534,21 @@ class ServerPreset:
         server.seed = 42
         return server
 
+    @staticmethod
+    def router() -> ServerProcess:
+        server = ServerProcess()
+        # router server has no models
+        server.model_file = None
+        server.model_alias = None
+        server.model_hf_repo = None
+        server.model_hf_file = None
+        server.n_ctx = 1024
+        server.n_batch = 16
+        server.n_slots = 1
+        server.n_predict = 16
+        server.seed = 42
+        return server
+
 
 def parallel_function_calls(function_list: List[Tuple[Callable[..., Any], Tuple[Any, ...]]]) -> List[Any]:
     """

tools/server/webui/.storybook/main.ts

@@ -1,7 +1,7 @@
 import type { StorybookConfig } from '@storybook/sveltekit';
 
 const config: StorybookConfig = {
-	stories: ['../src/**/*.mdx', '../src/**/*.stories.@(js|ts|svelte)'],
+	stories: ['../tests/stories/**/*.mdx', '../tests/stories/**/*.stories.@(js|ts|svelte)'],
 	addons: [
 		'@storybook/addon-svelte-csf',
 		'@chromatic-com/storybook',

tools/server/webui/README.md

@@ -2,65 +2,685 @@
 
 A modern, feature-rich web interface for llama.cpp built with SvelteKit. This UI provides an intuitive chat interface with advanced file handling, conversation management, and comprehensive model interaction capabilities.
 
+The WebUI supports two server operation modes:
+
+- **MODEL mode** - Single model operation (standard llama-server)
+- **ROUTER mode** - Multi-model operation with dynamic model loading/unloading
+
+---
+
+## Table of Contents
+
+- [Features](#features)
+- [Getting Started](#getting-started)
+- [Tech Stack](#tech-stack)
+- [Build Pipeline](#build-pipeline)
+- [Architecture](#architecture)
+- [Data Flows](#data-flows)
+- [Architectural Patterns](#architectural-patterns)
+- [Testing](#testing)
+
+---
+
 ## Features
 
-- **Modern Chat Interface** - Clean, responsive design with dark/light mode
-- **File Attachments** - Support for images, text files, PDFs, and audio with rich previews and drag-and-drop support
-- **Conversation Management** - Create, edit, branch, and search conversations
-- **Advanced Markdown** - Code highlighting, math formulas (KaTeX), and content blocks
-- **Reasoning Content** - Support for models with thinking blocks
-- **Keyboard Shortcuts** - Keyboard navigation (Shift+Ctrl/Cmd+O for new chat, Shift+Ctrl/Cmdt+E for edit conversation, Shift+Ctrl/Cmdt+D for delete conversation, Ctrl/Cmd+K for search, Ctrl/Cmd+V for paste, Ctrl/Cmd+B for opening/collapsing sidebar)
-- **Request Tracking** - Monitor processing with slots endpoint integration
-- **UI Testing** - Storybook component library with automated tests
+### Chat Interface
 
-## Development
+- **Streaming responses** with real-time updates
+- **Reasoning content** - Support for models with thinking/reasoning blocks
+- **Dark/light theme** with system preference detection
+- **Responsive design** for desktop and mobile
 
-Install dependencies:
+### File Attachments
+
+- **Images** - JPEG, PNG, GIF, WebP, SVG (with PNG conversion)
+- **Documents** - PDF (text extraction or image conversion for vision models)
+- **Audio** - MP3, WAV for audio-capable models
+- **Text files** - Source code, markdown, and other text formats
+- **Drag-and-drop** and paste support with rich previews
+
+### Conversation Management
+
+- **Branching** - Branch messages conversations at any point by editing messages or regenerating responses, navigate between branches
+- **Regeneration** - Regenerate responses with optional model switching (ROUTER mode)
+- **Import/Export** - JSON format for backup and sharing
+- **Search** - Find conversations by title or content
+
+### Advanced Rendering
+
+- **Syntax highlighting** - Code blocks with language detection
+- **Math formulas** - KaTeX rendering for LaTeX expressions
+- **Markdown** - Full GFM support with tables, lists, and more
+
+### Multi-Model Support (ROUTER mode)
+
+- **Model selector** with Loaded/Available groups
+- **Automatic loading** - Models load on selection
+- **Modality validation** - Prevents sending images to non-vision models
+- **LRU unloading** - Server auto-manages model cache
+
+### Keyboard Shortcuts
+
+| Shortcut           | Action               |
+| ------------------ | -------------------- |
+| `Shift+Ctrl/Cmd+O` | New chat             |
+| `Shift+Ctrl/Cmd+E` | Edit conversation    |
+| `Shift+Ctrl/Cmd+D` | Delete conversation  |
+| `Ctrl/Cmd+K`       | Search conversations |
+| `Ctrl/Cmd+B`       | Toggle sidebar       |
+
+### Developer Experience
+
+- **Request tracking** - Monitor token generation with `/slots` endpoint
+- **Storybook** - Component library with visual testing
+- **Hot reload** - Instant updates during development
+
+---
+
+## Getting Started
+
+### Prerequisites
+
+- **Node.js** 18+ (20+ recommended)
+- **npm** 9+
+- **llama-server** running locally (for API access)
+
+### 1. Install Dependencies
 
 ```bash
+cd tools/server/webui
 npm install
 ```
 
-Start the development server + Storybook:
+### 2. Start llama-server
+
+In a separate terminal, start the backend server:
+
+```bash
+# Single model (MODEL mode)
+./llama-server -m model.gguf
+
+# Multi-model (ROUTER mode)
+./llama-server --model-store /path/to/models
+```
+
+### 3. Start Development Servers
 
 ```bash
 npm run dev
 ```
 
-This will start both the SvelteKit dev server and Storybook on port 6006.
+This starts:
 
-## Building
+- **Vite dev server** at `http://localhost:5173` - The main WebUI
+- **Storybook** at `http://localhost:6006` - Component documentation
 
-Create a production build:
+The Vite dev server proxies API requests to `http://localhost:8080` (default llama-server port):
+
+```typescript
+// vite.config.ts proxy configuration
+proxy: {
+  '/v1': 'http://localhost:8080',
+  '/props': 'http://localhost:8080',
+  '/slots': 'http://localhost:8080',
+  '/models': 'http://localhost:8080'
+}
+```
+
+### Development Workflow
+
+1. Open `http://localhost:5173` in your browser
+2. Make changes to `.svelte`, `.ts`, or `.css` files
+3. Changes hot-reload instantly
+4. Use Storybook at `http://localhost:6006` for isolated component development
+
+---
+
+## Tech Stack
+
+| Layer             | Technology                      | Purpose                                                  |
+| ----------------- | ------------------------------- | -------------------------------------------------------- |
+| **Framework**     | SvelteKit + Svelte 5            | Reactive UI with runes (`$state`, `$derived`, `$effect`) |
+| **UI Components** | shadcn-svelte + bits-ui         | Accessible, customizable component library               |
+| **Styling**       | TailwindCSS 4                   | Utility-first CSS with design tokens                     |
+| **Database**      | IndexedDB (Dexie)               | Client-side storage for conversations and messages       |
+| **Build**         | Vite                            | Fast bundling with static adapter                        |
+| **Testing**       | Playwright + Vitest + Storybook | E2E, unit, and visual testing                            |
+| **Markdown**      | remark + rehype                 | Markdown processing with KaTeX and syntax highlighting   |
+
+### Key Dependencies
+
+```json
+{
+	"svelte": "^5.0.0",
+	"bits-ui": "^2.8.11",
+	"dexie": "^4.0.11",
+	"pdfjs-dist": "^5.4.54",
+	"highlight.js": "^11.11.1",
+	"rehype-katex": "^7.0.1"
+}
+```
+
+---
+
+## Build Pipeline
+
+### Development Build
+
+```bash
+npm run dev
+```
+
+Runs Vite in development mode with:
+
+- Hot Module Replacement (HMR)
+- Source maps
+- Proxy to llama-server
+
+### Production Build
 
 ```bash
 npm run build
 ```
 
-The build outputs static files to `../public` directory for deployment with llama.cpp server.
+The build process:
 
-## Testing
+1. **Vite Build** - Bundles all TypeScript, Svelte, and CSS
+2. **Static Adapter** - Outputs to `../public` (llama-server's static file directory)
+3. **Post-Build Script** - Cleans up intermediate files
+4. **Custom Plugin** - Creates `index.html.gz` with:
+   - Inlined favicon as base64
+   - GZIP compression (level 9)
+   - Deterministic output (zeroed timestamps)
 
-Run the test suite:
-
-```bash
-# E2E tests
-npm run test:e2e
-
-# Unit tests
-npm run test:unit
-
-# UI tests
-npm run test:ui
-
-# All tests
-npm run test
+```text
+tools/server/webui/        →  build  →  tools/server/public/
+├── src/                                 ├── index.html.gz  (served by llama-server)
+├── static/                              └── (favicon inlined)
+└── ...
 ```
 
+### SvelteKit Configuration
+
+```javascript
+// svelte.config.js
+adapter: adapter({
+  pages: '../public',      // Output directory
+  assets: '../public',     // Static assets
+  fallback: 'index.html',  // SPA fallback
+  strict: true
+}),
+output: {
+  bundleStrategy: 'inline' // Single-file bundle
+}
+```
+
+### Integration with llama-server
+
+The WebUI is embedded directly into the llama-server binary:
+
+1. `npm run build` outputs `index.html.gz` to `tools/server/public/`
+2. llama-server compiles this into the binary at build time
+3. When accessing `/`, llama-server serves the gzipped HTML
+4. All assets are inlined (CSS, JS, fonts, favicon)
+
+This results in a **single portable binary** with the full WebUI included.
+
+---
+
 ## Architecture
 
-- **Framework**: SvelteKit with Svelte 5 runes
-- **Components**: ShadCN UI + bits-ui design system
-- **Database**: IndexedDB with Dexie for local storage
-- **Build**: Static adapter for deployment with llama.cpp server
-- **Testing**: Playwright (E2E) + Vitest (unit) + Storybook (components)
+The WebUI follows a layered architecture with unidirectional data flow:
+
+```text
+Routes → Components → Hooks → Stores → Services → Storage/API
+```
+
+### High-Level Architecture
+
+See: [`docs/architecture/high-level-architecture-simplified.md`](docs/architecture/high-level-architecture-simplified.md)
+
+```mermaid
+flowchart TB
+    subgraph Routes["📍 Routes"]
+        R1["/ (Welcome)"]
+        R2["/chat/[id]"]
+        RL["+layout.svelte"]
+    end
+
+    subgraph Components["🧩 Components"]
+        C_Sidebar["ChatSidebar"]
+        C_Screen["ChatScreen"]
+        C_Form["ChatForm"]
+        C_Messages["ChatMessages"]
+        C_ModelsSelector["ModelsSelector"]
+        C_Settings["ChatSettings"]
+    end
+
+    subgraph Stores["🗄️ Stores"]
+        S1["chatStore"]
+        S2["conversationsStore"]
+        S3["modelsStore"]
+        S4["serverStore"]
+        S5["settingsStore"]
+    end
+
+    subgraph Services["⚙️ Services"]
+        SV1["ChatService"]
+        SV2["ModelsService"]
+        SV3["PropsService"]
+        SV4["DatabaseService"]
+    end
+
+    subgraph Storage["💾 Storage"]
+        ST1["IndexedDB"]
+        ST2["LocalStorage"]
+    end
+
+    subgraph APIs["🌐 llama-server"]
+        API1["/v1/chat/completions"]
+        API2["/props"]
+        API3["/models/*"]
+    end
+
+    R1 & R2 --> C_Screen
+    RL --> C_Sidebar
+    C_Screen --> C_Form & C_Messages & C_Settings
+    C_Screen --> S1 & S2
+    C_ModelsSelector --> S3 & S4
+    S1 --> SV1 & SV4
+    S3 --> SV2 & SV3
+    SV4 --> ST1
+    SV1 --> API1
+    SV2 --> API3
+    SV3 --> API2
+```
+
+### Layer Breakdown
+
+#### Routes (`src/routes/`)
+
+- **`/`** - Welcome screen, creates new conversation
+- **`/chat/[id]`** - Active chat interface
+- **`+layout.svelte`** - Sidebar, navigation, global initialization
+
+#### Components (`src/lib/components/`)
+
+Components are organized in `app/` (application-specific) and `ui/` (shadcn-svelte primitives).
+
+**Chat Components** (`app/chat/`):
+
+| Component          | Responsibility                                                              |
+| ------------------ | --------------------------------------------------------------------------- |
+| `ChatScreen/`      | Main chat container, coordinates message list, input form, and attachments  |
+| `ChatForm/`        | Message input textarea with file upload, paste handling, keyboard shortcuts |
+| `ChatMessages/`    | Message list with branch navigation, regenerate/continue/edit actions       |
+| `ChatAttachments/` | File attachment previews, drag-and-drop, PDF/image/audio handling           |
+| `ChatSettings/`    | Parameter sliders (temperature, top-p, etc.) with server default sync       |
+| `ChatSidebar/`     | Conversation list, search, import/export, navigation                        |
+
+**Dialog Components** (`app/dialogs/`):
+
+| Component                       | Responsibility                                           |
+| ------------------------------- | -------------------------------------------------------- |
+| `DialogChatSettings`            | Full-screen settings configuration                       |
+| `DialogModelInformation`        | Model details (context size, modalities, parallel slots) |
+| `DialogChatAttachmentPreview`   | Full preview for images, PDFs (text or page view), code  |
+| `DialogConfirmation`            | Generic confirmation for destructive actions             |
+| `DialogConversationTitleUpdate` | Edit conversation title                                  |
+
+**Server/Model Components** (`app/server/`, `app/models/`):
+
+| Component           | Responsibility                                            |
+| ------------------- | --------------------------------------------------------- |
+| `ServerErrorSplash` | Error display when server is unreachable                  |
+| `ModelsSelector`    | Model dropdown with Loaded/Available groups (ROUTER mode) |
+
+**Shared UI Components** (`app/misc/`):
+
+| Component                        | Responsibility                                                   |
+| -------------------------------- | ---------------------------------------------------------------- |
+| `MarkdownContent`                | Markdown rendering with KaTeX, syntax highlighting, copy buttons |
+| `SyntaxHighlightedCode`          | Code blocks with language detection and highlighting             |
+| `ActionButton`, `ActionDropdown` | Reusable action buttons and menus                                |
+| `BadgeModality`, `BadgeInfo`     | Status and capability badges                                     |
+
+#### Hooks (`src/lib/hooks/`)
+
+- **`useModelChangeValidation`** - Validates model switch against conversation modalities
+- **`useProcessingState`** - Tracks streaming progress and token generation
+
+#### Stores (`src/lib/stores/`)
+
+| Store                | Responsibility                                            |
+| -------------------- | --------------------------------------------------------- |
+| `chatStore`          | Message sending, streaming, abort control, error handling |
+| `conversationsStore` | CRUD for conversations, message branching, navigation     |
+| `modelsStore`        | Model list, selection, loading/unloading (ROUTER)         |
+| `serverStore`        | Server properties, role detection, modalities             |
+| `settingsStore`      | User preferences, parameter sync with server defaults     |
+
+#### Services (`src/lib/services/`)
+
+| Service                | Responsibility                                  |
+| ---------------------- | ----------------------------------------------- |
+| `ChatService`          | API calls to`/v1/chat/completions`, SSE parsing |
+| `ModelsService`        | `/models`, `/models/load`, `/models/unload`     |
+| `PropsService`         | `/props`, `/props?model=`                       |
+| `DatabaseService`      | IndexedDB operations via Dexie                  |
+| `ParameterSyncService` | Syncs settings with server defaults             |
+
+---
+
+## Data Flows
+
+### MODEL Mode (Single Model)
+
+See: [`docs/flows/data-flow-simplified-model-mode.md`](docs/flows/data-flow-simplified-model-mode.md)
+
+```mermaid
+sequenceDiagram
+    participant User
+    participant UI
+    participant Stores
+    participant DB as IndexedDB
+    participant API as llama-server
+
+    Note over User,API: Initialization
+    UI->>Stores: initialize()
+    Stores->>DB: load conversations
+    Stores->>API: GET /props
+    API-->>Stores: server config
+    Stores->>API: GET /v1/models
+    API-->>Stores: single model (auto-selected)
+
+    Note over User,API: Chat Flow
+    User->>UI: send message
+    Stores->>DB: save user message
+    Stores->>API: POST /v1/chat/completions (stream)
+    loop streaming
+        API-->>Stores: SSE chunks
+        Stores-->>UI: reactive update
+    end
+    Stores->>DB: save assistant message
+```
+
+### ROUTER Mode (Multi-Model)
+
+See: [`docs/flows/data-flow-simplified-router-mode.md`](docs/flows/data-flow-simplified-router-mode.md)
+
+```mermaid
+sequenceDiagram
+    participant User
+    participant UI
+    participant Stores
+    participant API as llama-server
+
+    Note over User,API: Initialization
+    Stores->>API: GET /props
+    API-->>Stores: {role: "router"}
+    Stores->>API: GET /models
+    API-->>Stores: models[] with status
+
+    Note over User,API: Model Selection
+    User->>UI: select model
+    alt model not loaded
+        Stores->>API: POST /models/load
+        loop poll status
+            Stores->>API: GET /models
+        end
+        Stores->>API: GET /props?model=X
+    end
+    Stores->>Stores: validate modalities
+
+    Note over User,API: Chat Flow
+    Stores->>API: POST /v1/chat/completions {model: X}
+    loop streaming
+        API-->>Stores: SSE chunks + model info
+    end
+```
+
+### Detailed Flow Diagrams
+
+| Flow          | Description                                | File                                                        |
+| ------------- | ------------------------------------------ | ----------------------------------------------------------- |
+| Chat          | Message lifecycle, streaming, regeneration | [`chat-flow.md`](docs/flows/chat-flow.md)                   |
+| Models        | Loading, unloading, modality caching       | [`models-flow.md`](docs/flows/models-flow.md)               |
+| Server        | Props fetching, role detection             | [`server-flow.md`](docs/flows/server-flow.md)               |
+| Conversations | CRUD, branching, import/export             | [`conversations-flow.md`](docs/flows/conversations-flow.md) |
+| Database      | IndexedDB schema, operations               | [`database-flow.md`](docs/flows/database-flow.md)           |
+| Settings      | Parameter sync, user overrides             | [`settings-flow.md`](docs/flows/settings-flow.md)           |
+
+---
+
+## Architectural Patterns
+
+### 1. Reactive State with Svelte 5 Runes
+
+All stores use Svelte 5's fine-grained reactivity:
+
+```typescript
+// Store with reactive state
+class ChatStore {
+	#isLoading = $state(false);
+	#currentResponse = $state('');
+
+	// Derived values auto-update
+	get isStreaming() {
+		return $derived(this.#isLoading && this.#currentResponse.length > 0);
+	}
+}
+
+// Exported reactive accessors
+export const isLoading = () => chatStore.isLoading;
+export const currentResponse = () => chatStore.currentResponse;
+```
+
+### 2. Unidirectional Data Flow
+
+Data flows in one direction, making state predictable:
+
+```mermaid
+flowchart LR
+    subgraph UI["UI Layer"]
+        A[User Action] --> B[Component]
+    end
+
+    subgraph State["State Layer"]
+        B --> C[Store Method]
+        C --> D[State Update]
+    end
+
+    subgraph IO["I/O Layer"]
+        C --> E[Service]
+        E --> F[API / IndexedDB]
+        F -.->|Response| D
+    end
+
+    D -->|Reactive| B
+```
+
+Components dispatch actions to stores, stores coordinate with services for I/O, and state updates reactively propagate back to the UI.
+
+### 3. Per-Conversation State
+
+Enables concurrent streaming across multiple conversations:
+
+```typescript
+class ChatStore {
+	chatLoadingStates = new Map<string, boolean>();
+	chatStreamingStates = new Map<string, { response: string; messageId: string }>();
+	abortControllers = new Map<string, AbortController>();
+}
+```
+
+### 4. Message Branching with Tree Structure
+
+Conversations are stored as a tree, not a linear list:
+
+```typescript
+interface DatabaseMessage {
+	id: string;
+	parent: string | null; // Points to parent message
+	children: string[]; // List of child message IDs
+	// ...
+}
+
+interface DatabaseConversation {
+	currentNode: string; // Currently viewed branch tip
+	// ...
+}
+```
+
+Navigation between branches updates `currentNode` without losing history.
+
+### 5. Layered Service Architecture
+
+Stores handle state; services handle I/O:
+
+```text
+┌─────────────────┐
+│     Stores      │  Business logic, state management
+├─────────────────┤
+│    Services     │  API calls, database operations
+├─────────────────┤
+│   Storage/API   │  IndexedDB, LocalStorage, HTTP
+└─────────────────┘
+```
+
+### 6. Server Role Abstraction
+
+Single codebase handles both MODEL and ROUTER modes:
+
+```typescript
+// serverStore.ts
+get isRouterMode() {
+  return this.role === ServerRole.ROUTER;
+}
+
+// Components conditionally render based on mode
+{#if isRouterMode()}
+  <ModelsSelector />
+{/if}
+```
+
+### 7. Modality Validation
+
+Prevents sending attachments to incompatible models:
+
+```typescript
+// useModelChangeValidation hook
+const validate = (modelId: string) => {
+	const modelModalities = modelsStore.getModelModalities(modelId);
+	const conversationModalities = conversationsStore.usedModalities;
+
+	// Check if model supports all used modalities
+	if (conversationModalities.hasImages && !modelModalities.vision) {
+		return { valid: false, reason: 'Model does not support images' };
+	}
+	// ...
+};
+```
+
+### 8. Persistent Storage Strategy
+
+Data is persisted across sessions using two storage mechanisms:
+
+```mermaid
+flowchart TB
+    subgraph Browser["Browser Storage"]
+        subgraph IDB["IndexedDB (Dexie)"]
+            C[Conversations]
+            M[Messages]
+        end
+        subgraph LS["LocalStorage"]
+            S[Settings Config]
+            O[User Overrides]
+            T[Theme Preference]
+        end
+    end
+
+    subgraph Stores["Svelte Stores"]
+        CS[conversationsStore] --> C
+        CS --> M
+        SS[settingsStore] --> S
+        SS --> O
+        SS --> T
+    end
+```
+
+- **IndexedDB**: Conversations and messages (large, structured data)
+- **LocalStorage**: Settings, user parameter overrides, theme (small key-value data)
+- **Memory only**: Server props, model list (fetched fresh on each session)
+
+---
+
+## Testing
+
+### Test Types
+
+| Type          | Tool               | Location                         | Command             |
+| ------------- | ------------------ | -------------------------------- | ------------------- |
+| **E2E**       | Playwright         | `tests/e2e/`                     | `npm run test:e2e`  |
+| **Unit**      | Vitest             | `tests/client/`, `tests/server/` | `npm run test:unit` |
+| **UI/Visual** | Storybook + Vitest | `tests/stories/`                 | `npm run test:ui`   |
+
+### Running Tests
+
+```bash
+# All tests
+npm run test
+
+# Individual test suites
+npm run test:e2e      # End-to-end (requires llama-server)
+npm run test:client   # Client-side unit tests
+npm run test:server   # Server-side unit tests
+npm run test:ui       # Storybook visual tests
+```
+
+### Storybook Development
+
+```bash
+npm run storybook     # Start Storybook dev server on :6006
+npm run build-storybook  # Build static Storybook
+```
+
+### Linting and Formatting
+
+```bash
+npm run lint          # Check code style
+npm run format        # Auto-format with Prettier
+npm run check         # TypeScript type checking
+```
+
+---
+
+## Project Structure
+
+```text
+tools/server/webui/
+├── src/
+│   ├── lib/
+│   │   ├── components/   # UI components (app/, ui/)
+│   │   ├── hooks/        # Svelte hooks
+│   │   ├── stores/       # State management
+│   │   ├── services/     # API and database services
+│   │   ├── types/        # TypeScript interfaces
+│   │   └── utils/        # Utility functions
+│   ├── routes/           # SvelteKit routes
+│   └── styles/           # Global styles
+├── static/               # Static assets
+├── tests/                # Test files
+├── docs/                 # Architecture diagrams
+│   ├── architecture/     # High-level architecture
+│   └── flows/            # Feature-specific flows
+└── .storybook/           # Storybook configuration
+```
+
+---
+
+## Related Documentation
+
+- [llama.cpp Server README](../README.md) - Full server documentation
+- [Multimodal Documentation](../../../docs/multimodal.md) - Image and audio support
+- [Function Calling](../../../docs/function-calling.md) - Tool use capabilities

tools/server/webui/docs/architecture/high-level-architecture-simplified.md

@@ -0,0 +1,102 @@
+```mermaid
+flowchart TB
+    subgraph Routes["📍 Routes"]
+        R1["/ (Welcome)"]
+        R2["/chat/[id]"]
+        RL["+layout.svelte"]
+    end
+
+    subgraph Components["🧩 Components"]
+        C_Sidebar["ChatSidebar"]
+        C_Screen["ChatScreen"]
+        C_Form["ChatForm"]
+        C_Messages["ChatMessages"]
+        C_ModelsSelector["ModelsSelector"]
+        C_Settings["ChatSettings"]
+    end
+
+    subgraph Hooks["🪝 Hooks"]
+        H1["useModelChangeValidation"]
+        H2["useProcessingState"]
+    end
+
+    subgraph Stores["🗄️ Stores"]
+        S1["chatStore<br/><i>Chat interactions & streaming</i>"]
+        S2["conversationsStore<br/><i>Conversation data & messages</i>"]
+        S3["modelsStore<br/><i>Model selection & loading</i>"]
+        S4["serverStore<br/><i>Server props & role detection</i>"]
+        S5["settingsStore<br/><i>User configuration</i>"]
+    end
+
+    subgraph Services["⚙️ Services"]
+        SV1["ChatService"]
+        SV2["ModelsService"]
+        SV3["PropsService"]
+        SV4["DatabaseService"]
+        SV5["ParameterSyncService"]
+    end
+
+    subgraph Storage["💾 Storage"]
+        ST1["IndexedDB<br/><i>conversations, messages</i>"]
+        ST2["LocalStorage<br/><i>config, userOverrides</i>"]
+    end
+
+    subgraph APIs["🌐 llama-server API"]
+        API1["/v1/chat/completions"]
+        API2["/props"]
+        API3["/models/*"]
+        API4["/v1/models"]
+    end
+
+    %% Routes → Components
+    R1 & R2 --> C_Screen
+    RL --> C_Sidebar
+
+    %% Component hierarchy
+    C_Screen --> C_Form & C_Messages & C_Settings
+    C_Form & C_Messages --> C_ModelsSelector
+
+    %% Components → Hooks → Stores
+    C_Form & C_Messages --> H1 & H2
+    H1 --> S3 & S4
+    H2 --> S1 & S5
+
+    %% Components → Stores
+    C_Screen --> S1 & S2
+    C_Sidebar --> S2
+    C_ModelsSelector --> S3 & S4
+    C_Settings --> S5
+
+    %% Stores → Services
+    S1 --> SV1 & SV4
+    S2 --> SV4
+    S3 --> SV2 & SV3
+    S4 --> SV3
+    S5 --> SV5
+
+    %% Services → Storage
+    SV4 --> ST1
+    SV5 --> ST2
+
+    %% Services → APIs
+    SV1 --> API1
+    SV2 --> API3 & API4
+    SV3 --> API2
+
+    %% Styling
+    classDef routeStyle fill:#e1f5fe,stroke:#01579b,stroke-width:2px
+    classDef componentStyle fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
+    classDef hookStyle fill:#fff8e1,stroke:#ff8f00,stroke-width:2px
+    classDef storeStyle fill:#fff3e0,stroke:#e65100,stroke-width:2px
+    classDef serviceStyle fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px
+    classDef storageStyle fill:#fce4ec,stroke:#c2185b,stroke-width:2px
+    classDef apiStyle fill:#e3f2fd,stroke:#1565c0,stroke-width:2px
+
+    class R1,R2,RL routeStyle
+    class C_Sidebar,C_Screen,C_Form,C_Messages,C_ModelsSelector,C_Settings componentStyle
+    class H1,H2 hookStyle
+    class S1,S2,S3,S4,S5 storeStyle
+    class SV1,SV2,SV3,SV4,SV5 serviceStyle
+    class ST1,ST2 storageStyle
+    class API1,API2,API3,API4 apiStyle
+```

tools/server/webui/docs/architecture/high-level-architecture.md

@@ -0,0 +1,269 @@
+```mermaid
+flowchart TB
+subgraph Routes["📍 Routes"]
+R1["/ (+page.svelte)"]
+R2["/chat/[id]"]
+RL["+layout.svelte"]
+end
+
+    subgraph Components["🧩 Components"]
+        direction TB
+        subgraph LayoutComponents["Layout"]
+            C_Sidebar["ChatSidebar"]
+            C_Screen["ChatScreen"]
+        end
+        subgraph ChatUIComponents["Chat UI"]
+            C_Form["ChatForm"]
+            C_Messages["ChatMessages"]
+            C_Message["ChatMessage"]
+            C_Attach["ChatAttachments"]
+            C_ModelsSelector["ModelsSelector"]
+            C_Settings["ChatSettings"]
+        end
+    end
+
+    subgraph Hooks["🪝 Hooks"]
+        H1["useModelChangeValidation"]
+        H2["useProcessingState"]
+        H3["isMobile"]
+    end
+
+    subgraph Stores["🗄️ Stores"]
+        direction TB
+        subgraph S1["chatStore"]
+            S1State["<b>State:</b><br/>isLoading, currentResponse<br/>errorDialogState<br/>activeProcessingState<br/>chatLoadingStates<br/>chatStreamingStates<br/>abortControllers<br/>processingStates<br/>activeConversationId<br/>isStreamingActive"]
+            S1LoadState["<b>Loading State:</b><br/>setChatLoading()<br/>isChatLoading()<br/>syncLoadingStateForChat()<br/>clearUIState()<br/>isChatLoadingPublic()<br/>getAllLoadingChats()<br/>getAllStreamingChats()"]
+            S1ProcState["<b>Processing State:</b><br/>setActiveProcessingConversation()<br/>getProcessingState()<br/>clearProcessingState()<br/>getActiveProcessingState()<br/>updateProcessingStateFromTimings()<br/>getCurrentProcessingStateSync()<br/>restoreProcessingStateFromMessages()"]
+            S1Stream["<b>Streaming:</b><br/>streamChatCompletion()<br/>startStreaming()<br/>stopStreaming()<br/>stopGeneration()<br/>isStreaming()"]
+            S1Error["<b>Error Handling:</b><br/>showErrorDialog()<br/>dismissErrorDialog()<br/>isAbortError()"]
+            S1Msg["<b>Message Operations:</b><br/>addMessage()<br/>sendMessage()<br/>updateMessage()<br/>deleteMessage()<br/>getDeletionInfo()"]
+            S1Regen["<b>Regeneration:</b><br/>regenerateMessage()<br/>regenerateMessageWithBranching()<br/>continueAssistantMessage()"]
+            S1Edit["<b>Editing:</b><br/>editAssistantMessage()<br/>editUserMessagePreserveResponses()<br/>editMessageWithBranching()"]
+            S1Utils["<b>Utilities:</b><br/>getApiOptions()<br/>parseTimingData()<br/>getOrCreateAbortController()<br/>getConversationModel()"]
+        end
+        subgraph S2["conversationsStore"]
+            S2State["<b>State:</b><br/>conversations<br/>activeConversation<br/>activeMessages<br/>usedModalities<br/>isInitialized<br/>titleUpdateConfirmationCallback"]
+            S2Modal["<b>Modalities:</b><br/>getModalitiesUpToMessage()<br/>calculateModalitiesFromMessages()"]
+            S2Lifecycle["<b>Lifecycle:</b><br/>initialize()<br/>loadConversations()<br/>clearActiveConversation()"]
+            S2ConvCRUD["<b>Conversation CRUD:</b><br/>createConversation()<br/>loadConversation()<br/>deleteConversation()<br/>updateConversationName()<br/>updateConversationTitleWithConfirmation()"]
+            S2MsgMgmt["<b>Message Management:</b><br/>refreshActiveMessages()<br/>addMessageToActive()<br/>updateMessageAtIndex()<br/>findMessageIndex()<br/>sliceActiveMessages()<br/>removeMessageAtIndex()<br/>getConversationMessages()"]
+            S2Nav["<b>Navigation:</b><br/>navigateToSibling()<br/>updateCurrentNode()<br/>updateConversationTimestamp()"]
+            S2Export["<b>Import/Export:</b><br/>downloadConversation()<br/>exportAllConversations()<br/>importConversations()<br/>triggerDownload()"]
+            S2Utils["<b>Utilities:</b><br/>setTitleUpdateConfirmationCallback()"]
+        end
+        subgraph S3["modelsStore"]
+            S3State["<b>State:</b><br/>models, routerModels<br/>selectedModelId<br/>selectedModelName<br/>loading, updating, error<br/>modelLoadingStates<br/>modelPropsCache<br/>modelPropsFetching<br/>propsCacheVersion"]
+            S3Getters["<b>Computed Getters:</b><br/>selectedModel<br/>loadedModelIds<br/>loadingModelIds<br/>singleModelName"]
+            S3Modal["<b>Modalities:</b><br/>getModelModalities()<br/>modelSupportsVision()<br/>modelSupportsAudio()<br/>getModelModalitiesArray()<br/>getModelProps()<br/>updateModelModalities()"]
+            S3Status["<b>Status Queries:</b><br/>isModelLoaded()<br/>isModelOperationInProgress()<br/>getModelStatus()<br/>isModelPropsFetching()"]
+            S3Fetch["<b>Data Fetching:</b><br/>fetch()<br/>fetchRouterModels()<br/>fetchModelProps()<br/>fetchModalitiesForLoadedModels()"]
+            S3Select["<b>Model Selection:</b><br/>selectModelById()<br/>selectModelByName()<br/>clearSelection()<br/>findModelByName()<br/>findModelById()<br/>hasModel()"]
+            S3LoadUnload["<b>Loading/Unloading Models:</b><br/>loadModel()<br/>unloadModel()<br/>ensureModelLoaded()<br/>waitForModelStatus()<br/>pollForModelStatus()"]
+            S3Utils["<b>Utilities:</b><br/>toDisplayName()<br/>clear()"]
+        end
+        subgraph S4["serverStore"]
+            S4State["<b>State:</b><br/>props<br/>loading, error<br/>role<br/>fetchPromise"]
+            S4Getters["<b>Getters:</b><br/>defaultParams<br/>contextSize<br/>isRouterMode<br/>isModelMode"]
+            S4Data["<b>Data Handling:</b><br/>fetch()<br/>getErrorMessage()<br/>clear()"]
+            S4Utils["<b>Utilities:</b><br/>detectRole()"]
+        end
+        subgraph S5["settingsStore"]
+            S5State["<b>State:</b><br/>config<br/>theme<br/>isInitialized<br/>userOverrides"]
+            S5Lifecycle["<b>Lifecycle:</b><br/>initialize()<br/>loadConfig()<br/>saveConfig()<br/>loadTheme()<br/>saveTheme()"]
+            S5Update["<b>Config Updates:</b><br/>updateConfig()<br/>updateMultipleConfig()<br/>updateTheme()"]
+            S5Reset["<b>Reset:</b><br/>resetConfig()<br/>resetTheme()<br/>resetAll()<br/>resetParameterToServerDefault()"]
+            S5Sync["<b>Server Sync:</b><br/>syncWithServerDefaults()<br/>forceSyncWithServerDefaults()"]
+            S5Utils["<b>Utilities:</b><br/>getConfig()<br/>getAllConfig()<br/>getParameterInfo()<br/>getParameterDiff()<br/>getServerDefaults()<br/>clearAllUserOverrides()"]
+        end
+
+        subgraph ReactiveExports["⚡ Reactive Exports"]
+            direction LR
+            subgraph ChatExports["chatStore"]
+                RE1["isLoading()"]
+                RE2["currentResponse()"]
+                RE3["errorDialog()"]
+                RE4["activeProcessingState()"]
+                RE5["isChatStreaming()"]
+                RE6["isChatLoading()"]
+                RE7["getChatStreaming()"]
+                RE8["getAllLoadingChats()"]
+                RE9["getAllStreamingChats()"]
+            end
+            subgraph ConvExports["conversationsStore"]
+                RE10["conversations()"]
+                RE11["activeConversation()"]
+                RE12["activeMessages()"]
+                RE13["isConversationsInitialized()"]
+                RE14["usedModalities()"]
+            end
+            subgraph ModelsExports["modelsStore"]
+                RE15["modelOptions()"]
+                RE16["routerModels()"]
+                RE17["modelsLoading()"]
+                RE18["modelsUpdating()"]
+                RE19["modelsError()"]
+                RE20["selectedModelId()"]
+                RE21["selectedModelName()"]
+                RE22["selectedModelOption()"]
+                RE23["loadedModelIds()"]
+                RE24["loadingModelIds()"]
+                RE25["propsCacheVersion()"]
+                RE26["singleModelName()"]
+            end
+            subgraph ServerExports["serverStore"]
+                RE27["serverProps()"]
+                RE28["serverLoading()"]
+                RE29["serverError()"]
+                RE30["serverRole()"]
+                RE31["defaultParams()"]
+                RE32["contextSize()"]
+                RE33["isRouterMode()"]
+                RE34["isModelMode()"]
+            end
+            subgraph SettingsExports["settingsStore"]
+                RE35["config()"]
+                RE36["theme()"]
+                RE37["isInitialized()"]
+            end
+        end
+    end
+
+    subgraph Services["⚙️ Services"]
+        direction TB
+        subgraph SV1["ChatService"]
+            SV1Msg["<b>Messaging:</b><br/>sendMessage()"]
+            SV1Stream["<b>Streaming:</b><br/>handleStreamResponse()<br/>parseSSEChunk()"]
+            SV1Convert["<b>Conversion:</b><br/>convertMessageToChatData()<br/>convertExtraToApiFormat()"]
+            SV1Utils["<b>Utilities:</b><br/>extractReasoningContent()<br/>getServerProps()<br/>getModels()"]
+        end
+        subgraph SV2["ModelsService"]
+            SV2List["<b>Listing:</b><br/>list()<br/>listRouter()"]
+            SV2LoadUnload["<b>Load/Unload:</b><br/>load()<br/>unload()"]
+            SV2Status["<b>Status:</b><br/>isModelLoaded()<br/>isModelLoading()"]
+        end
+        subgraph SV3["PropsService"]
+            SV3Fetch["<b>Fetching:</b><br/>fetch()<br/>fetchForModel()"]
+        end
+        subgraph SV4["DatabaseService"]
+            SV4Conv["<b>Conversations:</b><br/>createConversation()<br/>getConversation()<br/>getAllConversations()<br/>updateConversation()<br/>deleteConversation()"]
+            SV4Msg["<b>Messages:</b><br/>createMessageBranch()<br/>createRootMessage()<br/>getConversationMessages()<br/>updateMessage()<br/>deleteMessage()<br/>deleteMessageCascading()"]
+            SV4Node["<b>Navigation:</b><br/>updateCurrentNode()"]
+            SV4Import["<b>Import:</b><br/>importConversations()"]
+        end
+        subgraph SV5["ParameterSyncService"]
+            SV5Extract["<b>Extraction:</b><br/>extractServerDefaults()"]
+            SV5Merge["<b>Merging:</b><br/>mergeWithServerDefaults()"]
+            SV5Info["<b>Info:</b><br/>getParameterInfo()<br/>canSyncParameter()<br/>getSyncableParameterKeys()<br/>validateServerParameter()"]
+            SV5Diff["<b>Diff:</b><br/>createParameterDiff()"]
+        end
+    end
+
+    subgraph Storage["💾 Storage"]
+        ST1["IndexedDB"]
+        ST2["conversations"]
+        ST3["messages"]
+        ST5["LocalStorage"]
+        ST6["config"]
+        ST7["userOverrides"]
+    end
+
+    subgraph APIs["🌐 llama-server API"]
+        API1["/v1/chat/completions"]
+        API2["/props<br/>/props?model="]
+        API3["/models<br/>/models/load<br/>/models/unload"]
+        API4["/v1/models"]
+    end
+
+    %% Routes render Components
+    R1 --> C_Screen
+    R2 --> C_Screen
+    RL --> C_Sidebar
+
+    %% Component hierarchy
+    C_Screen --> C_Form & C_Messages & C_Settings
+    C_Messages --> C_Message
+    C_Message --> C_ModelsSelector
+    C_Form --> C_ModelsSelector
+    C_Form --> C_Attach
+    C_Message --> C_Attach
+
+    %% Components use Hooks
+    C_Form --> H1
+    C_Message --> H1 & H2
+    C_Screen --> H2
+
+    %% Hooks use Stores
+    H1 --> S3 & S4
+    H2 --> S1 & S5
+
+    %% Components use Stores
+    C_Screen --> S1 & S2
+    C_Messages --> S2
+    C_Message --> S1 & S2 & S3
+    C_Form --> S1 & S3
+    C_Sidebar --> S2
+    C_ModelsSelector --> S3 & S4
+    C_Settings --> S5
+
+    %% Stores export Reactive State
+    S1 -. exports .-> ChatExports
+    S2 -. exports .-> ConvExports
+    S3 -. exports .-> ModelsExports
+    S4 -. exports .-> ServerExports
+    S5 -. exports .-> SettingsExports
+
+    %% Stores use Services
+    S1 --> SV1 & SV4
+    S2 --> SV4
+    S3 --> SV2 & SV3
+    S4 --> SV3
+    S5 --> SV5
+
+    %% Services to Storage
+    SV4 --> ST1
+    ST1 --> ST2 & ST3
+    SV5 --> ST5
+    ST5 --> ST6 & ST7
+
+    %% Services to APIs
+    SV1 --> API1
+    SV2 --> API3 & API4
+    SV3 --> API2
+
+    %% Styling
+    classDef routeStyle fill:#e1f5fe,stroke:#01579b,stroke-width:2px
+    classDef componentStyle fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
+    classDef componentGroupStyle fill:#e1bee7,stroke:#7b1fa2,stroke-width:1px
+    classDef storeStyle fill:#fff3e0,stroke:#e65100,stroke-width:2px
+    classDef stateStyle fill:#ffe0b2,stroke:#e65100,stroke-width:1px
+    classDef methodStyle fill:#ffecb3,stroke:#e65100,stroke-width:1px
+    classDef reactiveStyle fill:#fffde7,stroke:#f9a825,stroke-width:1px
+    classDef serviceStyle fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px
+    classDef serviceMStyle fill:#c8e6c9,stroke:#2e7d32,stroke-width:1px
+    classDef storageStyle fill:#fce4ec,stroke:#c2185b,stroke-width:2px
+    classDef apiStyle fill:#e3f2fd,stroke:#1565c0,stroke-width:2px
+
+    class R1,R2,RL routeStyle
+    class C_Sidebar,C_Screen,C_Form,C_Messages,C_Message componentStyle
+    class C_ModelsSelector,C_Settings componentStyle
+    class C_Attach componentStyle
+    class H1,H2,H3 methodStyle
+    class LayoutComponents,ChatUIComponents componentGroupStyle
+    class Hooks storeStyle
+    class S1,S2,S3,S4,S5 storeStyle
+    class S1State,S2State,S3State,S4State,S5State stateStyle
+    class S1Msg,S1Regen,S1Edit,S1Stream,S1LoadState,S1ProcState,S1Error,S1Utils methodStyle
+    class S2Lifecycle,S2ConvCRUD,S2MsgMgmt,S2Nav,S2Modal,S2Export,S2Utils methodStyle
+    class S3Getters,S3Modal,S3Status,S3Fetch,S3Select,S3LoadUnload,S3Utils methodStyle
+    class S4Getters,S4Data,S4Utils methodStyle
+    class S5Lifecycle,S5Update,S5Reset,S5Sync,S5Utils methodStyle
+    class ChatExports,ConvExports,ModelsExports,ServerExports,SettingsExports reactiveStyle
+    class SV1,SV2,SV3,SV4,SV5 serviceStyle
+    class SV1Msg,SV1Stream,SV1Convert,SV1Utils serviceMStyle
+    class SV2List,SV2LoadUnload,SV2Status serviceMStyle
+    class SV3Fetch serviceMStyle
+    class SV4Conv,SV4Msg,SV4Node,SV4Import serviceMStyle
+    class SV5Extract,SV5Merge,SV5Info,SV5Diff serviceMStyle
+    class ST1,ST2,ST3,ST5,ST6,ST7 storageStyle
+    class API1,API2,API3,API4 apiStyle
+```

tools/server/webui/docs/flows/chat-flow.md

@@ -0,0 +1,174 @@
+```mermaid
+sequenceDiagram
+    participant UI as 🧩 ChatForm / ChatMessage
+    participant chatStore as 🗄️ chatStore
+    participant convStore as 🗄️ conversationsStore
+    participant settingsStore as 🗄️ settingsStore
+    participant ChatSvc as ⚙️ ChatService
+    participant DbSvc as ⚙️ DatabaseService
+    participant API as 🌐 /v1/chat/completions
+
+    Note over chatStore: State:<br/>isLoading, currentResponse<br/>errorDialogState, activeProcessingState<br/>chatLoadingStates (Map)<br/>chatStreamingStates (Map)<br/>abortControllers (Map)<br/>processingStates (Map)
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 💬 SEND MESSAGE
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>chatStore: sendMessage(content, extras)
+    activate chatStore
+
+    chatStore->>chatStore: setChatLoading(convId, true)
+    chatStore->>chatStore: clearChatStreaming(convId)
+
+    alt no active conversation
+        chatStore->>convStore: createConversation()
+        Note over convStore: → see conversations-flow.mmd
+    end
+
+    chatStore->>chatStore: addMessage("user", content, extras)
+    chatStore->>DbSvc: createMessageBranch(userMsg, parentId)
+    chatStore->>convStore: addMessageToActive(userMsg)
+    chatStore->>convStore: updateCurrentNode(userMsg.id)
+
+    chatStore->>chatStore: createAssistantMessage(userMsg.id)
+    chatStore->>DbSvc: createMessageBranch(assistantMsg, userMsg.id)
+    chatStore->>convStore: addMessageToActive(assistantMsg)
+
+    chatStore->>chatStore: streamChatCompletion(messages, assistantMsg)
+    deactivate chatStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 🌊 STREAMING
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    activate chatStore
+    chatStore->>chatStore: startStreaming()
+    Note right of chatStore: isStreamingActive = true
+
+    chatStore->>chatStore: setActiveProcessingConversation(convId)
+    chatStore->>chatStore: getOrCreateAbortController(convId)
+    Note right of chatStore: abortControllers.set(convId, new AbortController())
+
+    chatStore->>chatStore: getApiOptions()
+    Note right of chatStore: Merge from settingsStore.config:<br/>temperature, max_tokens, top_p, etc.
+
+    chatStore->>ChatSvc: sendMessage(messages, options, signal)
+    activate ChatSvc
+
+    ChatSvc->>ChatSvc: convertMessageToChatData(messages)
+    Note right of ChatSvc: DatabaseMessage[] → ApiChatMessageData[]<br/>Process attachments (images, PDFs, audio)
+
+    ChatSvc->>API: POST /v1/chat/completions
+    Note right of API: {messages, model?, stream: true, ...params}
+
+    loop SSE chunks
+        API-->>ChatSvc: data: {"choices":[{"delta":{...}}]}
+        ChatSvc->>ChatSvc: parseSSEChunk(line)
+
+        alt content chunk
+            ChatSvc-->>chatStore: onChunk(content)
+            chatStore->>chatStore: setChatStreaming(convId, response, msgId)
+            Note right of chatStore: currentResponse = $state(accumulated)
+            chatStore->>convStore: updateMessageAtIndex(idx, {content})
+        end
+
+        alt reasoning chunk
+            ChatSvc-->>chatStore: onReasoningChunk(reasoning)
+            chatStore->>convStore: updateMessageAtIndex(idx, {thinking})
+        end
+
+        alt tool_calls chunk
+            ChatSvc-->>chatStore: onToolCallChunk(toolCalls)
+            chatStore->>convStore: updateMessageAtIndex(idx, {toolCalls})
+        end
+
+        alt model info
+            ChatSvc-->>chatStore: onModel(modelName)
+            chatStore->>chatStore: recordModel(modelName)
+            chatStore->>DbSvc: updateMessage(msgId, {model})
+        end
+
+        alt timings (during stream)
+            ChatSvc-->>chatStore: onTimings(timings, promptProgress)
+            chatStore->>chatStore: updateProcessingStateFromTimings()
+        end
+
+        chatStore-->>UI: reactive $state update
+    end
+
+    API-->>ChatSvc: data: [DONE]
+    ChatSvc-->>chatStore: onComplete(content, reasoning, timings, toolCalls)
+    deactivate ChatSvc
+
+    chatStore->>chatStore: stopStreaming()
+    chatStore->>DbSvc: updateMessage(msgId, {content, timings, model})
+    chatStore->>convStore: updateCurrentNode(msgId)
+    chatStore->>chatStore: setChatLoading(convId, false)
+    chatStore->>chatStore: clearChatStreaming(convId)
+    chatStore->>chatStore: clearProcessingState(convId)
+    deactivate chatStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: ⏹️ STOP GENERATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>chatStore: stopGeneration()
+    activate chatStore
+    chatStore->>chatStore: savePartialResponseIfNeeded(convId)
+    Note right of chatStore: Save currentResponse to DB if non-empty
+    chatStore->>chatStore: abortControllers.get(convId).abort()
+    Note right of chatStore: fetch throws AbortError → caught by isAbortError()
+    chatStore->>chatStore: stopStreaming()
+    chatStore->>chatStore: setChatLoading(convId, false)
+    chatStore->>chatStore: clearChatStreaming(convId)
+    chatStore->>chatStore: clearProcessingState(convId)
+    deactivate chatStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 🔁 REGENERATE
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>chatStore: regenerateMessageWithBranching(msgId, model?)
+    activate chatStore
+    chatStore->>convStore: findMessageIndex(msgId)
+    chatStore->>chatStore: Get parent of target message
+    chatStore->>chatStore: createAssistantMessage(parentId)
+    chatStore->>DbSvc: createMessageBranch(newAssistantMsg, parentId)
+    chatStore->>convStore: refreshActiveMessages()
+    Note right of chatStore: Same streaming flow
+    chatStore->>chatStore: streamChatCompletion(...)
+    deactivate chatStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: ➡️ CONTINUE
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>chatStore: continueAssistantMessage(msgId)
+    activate chatStore
+    chatStore->>chatStore: Get existing content from message
+    chatStore->>chatStore: streamChatCompletion(..., existingContent)
+    Note right of chatStore: Appends to existing message content
+    deactivate chatStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: ✏️ EDIT USER MESSAGE
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>chatStore: editUserMessagePreserveResponses(msgId, newContent)
+    activate chatStore
+    chatStore->>chatStore: Get parent of target message
+    chatStore->>DbSvc: createMessageBranch(editedMsg, parentId)
+    chatStore->>convStore: refreshActiveMessages()
+    Note right of chatStore: Creates new branch, original preserved
+    deactivate chatStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: ❌ ERROR HANDLING
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over chatStore: On stream error (non-abort):
+    chatStore->>chatStore: showErrorDialog(type, message)
+    Note right of chatStore: errorDialogState = {type: 'timeout'|'server', message}
+    chatStore->>convStore: removeMessageAtIndex(failedMsgIdx)
+    chatStore->>DbSvc: deleteMessage(failedMsgId)
+```

tools/server/webui/docs/flows/conversations-flow.md

@@ -0,0 +1,155 @@
+```mermaid
+sequenceDiagram
+    participant UI as 🧩 ChatSidebar / ChatScreen
+    participant convStore as 🗄️ conversationsStore
+    participant chatStore as 🗄️ chatStore
+    participant DbSvc as ⚙️ DatabaseService
+    participant IDB as 💾 IndexedDB
+
+    Note over convStore: State:<br/>conversations: DatabaseConversation[]<br/>activeConversation: DatabaseConversation | null<br/>activeMessages: DatabaseMessage[]<br/>isInitialized: boolean<br/>usedModalities: $derived({vision, audio})
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: 🚀 INITIALIZATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over convStore: Auto-initialized in constructor (browser only)
+    convStore->>convStore: initialize()
+    activate convStore
+    convStore->>convStore: loadConversations()
+    convStore->>DbSvc: getAllConversations()
+    DbSvc->>IDB: SELECT * FROM conversations ORDER BY lastModified DESC
+    IDB-->>DbSvc: Conversation[]
+    DbSvc-->>convStore: conversations
+    convStore->>convStore: conversations = $state(data)
+    convStore->>convStore: isInitialized = true
+    deactivate convStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: ➕ CREATE CONVERSATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>convStore: createConversation(name?)
+    activate convStore
+    convStore->>DbSvc: createConversation(name || "New Chat")
+    DbSvc->>IDB: INSERT INTO conversations
+    IDB-->>DbSvc: conversation {id, name, lastModified, currNode: ""}
+    DbSvc-->>convStore: conversation
+    convStore->>convStore: conversations.unshift(conversation)
+    convStore->>convStore: activeConversation = $state(conversation)
+    convStore->>convStore: activeMessages = $state([])
+    deactivate convStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: 📂 LOAD CONVERSATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>convStore: loadConversation(convId)
+    activate convStore
+    convStore->>DbSvc: getConversation(convId)
+    DbSvc->>IDB: SELECT * FROM conversations WHERE id = ?
+    IDB-->>DbSvc: conversation
+    convStore->>convStore: activeConversation = $state(conversation)
+
+    convStore->>convStore: refreshActiveMessages()
+    convStore->>DbSvc: getConversationMessages(convId)
+    DbSvc->>IDB: SELECT * FROM messages WHERE convId = ?
+    IDB-->>DbSvc: allMessages[]
+    convStore->>convStore: filterByLeafNodeId(allMessages, currNode)
+    Note right of convStore: Filter to show only current branch path
+    convStore->>convStore: activeMessages = $state(filtered)
+
+    convStore->>chatStore: syncLoadingStateForChat(convId)
+    Note right of chatStore: Sync isLoading/currentResponse if streaming
+    deactivate convStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: 🌳 MESSAGE BRANCHING MODEL
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over IDB: Message Tree Structure:<br/>- Each message has parent (null for root)<br/>- Each message has children[] array<br/>- Conversation.currNode points to active leaf<br/>- filterByLeafNodeId() traverses from root to currNode
+
+    rect rgb(240, 240, 255)
+        Note over convStore: Example Branch Structure:
+        Note over convStore: root → user1 → assistant1 → user2 → assistant2a (currNode)<br/>                                    ↘ assistant2b (alt branch)
+    end
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: ↔️ BRANCH NAVIGATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>convStore: navigateToSibling(msgId, direction)
+    activate convStore
+    convStore->>convStore: Find message in activeMessages
+    convStore->>convStore: Get parent message
+    convStore->>convStore: Find sibling in parent.children[]
+    convStore->>convStore: findLeafNode(siblingId, allMessages)
+    Note right of convStore: Navigate to leaf of sibling branch
+    convStore->>convStore: updateCurrentNode(leafId)
+    convStore->>DbSvc: updateCurrentNode(convId, leafId)
+    DbSvc->>IDB: UPDATE conversations SET currNode = ?
+    convStore->>convStore: refreshActiveMessages()
+    deactivate convStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: 📝 UPDATE CONVERSATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>convStore: updateConversationName(convId, newName)
+    activate convStore
+    convStore->>DbSvc: updateConversation(convId, {name: newName})
+    DbSvc->>IDB: UPDATE conversations SET name = ?
+    convStore->>convStore: Update in conversations array
+    deactivate convStore
+
+    Note over convStore: Auto-title update (after first response):
+    convStore->>convStore: updateConversationTitleWithConfirmation()
+    convStore->>convStore: titleUpdateConfirmationCallback?()
+    Note right of convStore: Shows dialog if title would change
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: 🗑️ DELETE CONVERSATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>convStore: deleteConversation(convId)
+    activate convStore
+    convStore->>DbSvc: deleteConversation(convId)
+    DbSvc->>IDB: DELETE FROM conversations WHERE id = ?
+    DbSvc->>IDB: DELETE FROM messages WHERE convId = ?
+    convStore->>convStore: conversations.filter(c => c.id !== convId)
+    alt deleted active conversation
+        convStore->>convStore: clearActiveConversation()
+    end
+    deactivate convStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: 📊 MODALITY TRACKING
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over convStore: usedModalities = $derived.by(() => {<br/>  calculateModalitiesFromMessages(activeMessages)<br/>})
+
+    Note over convStore: Scans activeMessages for attachments:<br/>- IMAGE → vision: true<br/>- PDF (processedAsImages) → vision: true<br/>- AUDIO → audio: true
+
+    UI->>convStore: getModalitiesUpToMessage(msgId)
+    Note right of convStore: Used for regeneration validation<br/>Only checks messages BEFORE target
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,IDB: 📤 EXPORT / 📥 IMPORT
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>convStore: exportAllConversations()
+    activate convStore
+    convStore->>DbSvc: getAllConversations()
+    loop each conversation
+        convStore->>DbSvc: getConversationMessages(convId)
+    end
+    convStore->>convStore: triggerDownload(JSON blob)
+    deactivate convStore
+
+    UI->>convStore: importConversations(file)
+    activate convStore
+    convStore->>convStore: Parse JSON file
+    convStore->>DbSvc: importConversations(parsed)
+    DbSvc->>IDB: Bulk INSERT conversations + messages
+    convStore->>convStore: loadConversations()
+    deactivate convStore
+```

tools/server/webui/docs/flows/data-flow-simplified-model-mode.md

@@ -0,0 +1,45 @@
+```mermaid
+%% MODEL Mode Data Flow (single model)
+%% Detailed flows: ./flows/server-flow.mmd, ./flows/models-flow.mmd, ./flows/chat-flow.mmd
+
+sequenceDiagram
+    participant User as 👤 User
+    participant UI as 🧩 UI
+    participant Stores as 🗄️ Stores
+    participant DB as 💾 IndexedDB
+    participant API as 🌐 llama-server
+
+    Note over User,API: 🚀 Initialization (see: server-flow.mmd, models-flow.mmd)
+
+    UI->>Stores: initialize()
+    Stores->>DB: load conversations
+    Stores->>API: GET /props
+    API-->>Stores: server config + modalities
+    Stores->>API: GET /v1/models
+    API-->>Stores: single model (auto-selected)
+
+    Note over User,API: 💬 Chat Flow (see: chat-flow.mmd)
+
+    User->>UI: send message
+    UI->>Stores: sendMessage()
+    Stores->>DB: save user message
+    Stores->>API: POST /v1/chat/completions (stream)
+    loop streaming
+        API-->>Stores: SSE chunks
+        Stores-->>UI: reactive update
+    end
+    API-->>Stores: done + timings
+    Stores->>DB: save assistant message
+
+    Note over User,API: 🔁 Regenerate
+
+    User->>UI: regenerate
+    Stores->>DB: create message branch
+    Note right of Stores: same streaming flow
+
+    Note over User,API: ⏹️ Stop
+
+    User->>UI: stop
+    Stores->>Stores: abort stream
+    Stores->>DB: save partial response
+```

tools/server/webui/docs/flows/data-flow-simplified-router-mode.md

@@ -0,0 +1,77 @@
+```mermaid
+%% ROUTER Mode Data Flow (multi-model)
+%% Detailed flows: ./flows/server-flow.mmd, ./flows/models-flow.mmd, ./flows/chat-flow.mmd
+
+sequenceDiagram
+    participant User as 👤 User
+    participant UI as 🧩 UI
+    participant Stores as 🗄️ Stores
+    participant DB as 💾 IndexedDB
+    participant API as 🌐 llama-server
+
+    Note over User,API: 🚀 Initialization (see: server-flow.mmd, models-flow.mmd)
+
+    UI->>Stores: initialize()
+    Stores->>DB: load conversations
+    Stores->>API: GET /props
+    API-->>Stores: {role: "router"}
+    Stores->>API: GET /models
+    API-->>Stores: models[] with status (loaded/available)
+    loop each loaded model
+        Stores->>API: GET /props?model=X
+        API-->>Stores: modalities (vision/audio)
+    end
+
+    Note over User,API: 🔄 Model Selection (see: models-flow.mmd)
+
+    User->>UI: select model
+    alt model not loaded
+        Stores->>API: POST /models/load
+        loop poll status
+            Stores->>API: GET /models
+            API-->>Stores: check if loaded
+        end
+        Stores->>API: GET /props?model=X
+        API-->>Stores: cache modalities
+    end
+    Stores->>Stores: validate modalities vs conversation
+    alt valid
+        Stores->>Stores: select model
+    else invalid
+        Stores->>API: POST /models/unload
+        UI->>User: show error toast
+    end
+
+    Note over User,API: 💬 Chat Flow (see: chat-flow.mmd)
+
+    User->>UI: send message
+    UI->>Stores: sendMessage()
+    Stores->>DB: save user message
+    Stores->>API: POST /v1/chat/completions {model: X}
+    Note right of API: router forwards to model
+    loop streaming
+        API-->>Stores: SSE chunks + model info
+        Stores-->>UI: reactive update
+    end
+    API-->>Stores: done + timings
+    Stores->>DB: save assistant message + model used
+
+    Note over User,API: 🔁 Regenerate (optional: different model)
+
+    User->>UI: regenerate
+    Stores->>Stores: validate modalities up to this message
+    Stores->>DB: create message branch
+    Note right of Stores: same streaming flow
+
+    Note over User,API: ⏹️ Stop
+
+    User->>UI: stop
+    Stores->>Stores: abort stream
+    Stores->>DB: save partial response
+
+    Note over User,API: 🗑️ LRU Unloading
+
+    Note right of API: Server auto-unloads LRU models<br/>when cache full
+    User->>UI: select unloaded model
+    Note right of Stores: triggers load flow again
+```

tools/server/webui/docs/flows/database-flow.md

@@ -0,0 +1,155 @@
+```mermaid
+sequenceDiagram
+    participant Store as 🗄️ Stores
+    participant DbSvc as ⚙️ DatabaseService
+    participant Dexie as 📦 Dexie ORM
+    participant IDB as 💾 IndexedDB
+
+    Note over DbSvc: Stateless service - all methods static<br/>Database: "LlamacppWebui"
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over Store,IDB: 📊 SCHEMA
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    rect rgb(240, 248, 255)
+        Note over IDB: conversations table:<br/>id (PK), lastModified, currNode, name
+    end
+
+    rect rgb(255, 248, 240)
+        Note over IDB: messages table:<br/>id (PK), convId (FK), type, role, timestamp,<br/>parent, children[], content, thinking,<br/>toolCalls, extra[], model, timings
+    end
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over Store,IDB: 💬 CONVERSATIONS CRUD
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Store->>DbSvc: createConversation(name)
+    activate DbSvc
+    DbSvc->>DbSvc: Generate UUID
+    DbSvc->>Dexie: db.conversations.add({id, name, lastModified, currNode: ""})
+    Dexie->>IDB: INSERT
+    IDB-->>Dexie: success
+    DbSvc-->>Store: DatabaseConversation
+    deactivate DbSvc
+
+    Store->>DbSvc: getConversation(convId)
+    DbSvc->>Dexie: db.conversations.get(convId)
+    Dexie->>IDB: SELECT WHERE id = ?
+    IDB-->>DbSvc: DatabaseConversation
+
+    Store->>DbSvc: getAllConversations()
+    DbSvc->>Dexie: db.conversations.orderBy('lastModified').reverse().toArray()
+    Dexie->>IDB: SELECT ORDER BY lastModified DESC
+    IDB-->>DbSvc: DatabaseConversation[]
+
+    Store->>DbSvc: updateConversation(convId, updates)
+    DbSvc->>Dexie: db.conversations.update(convId, {...updates, lastModified})
+    Dexie->>IDB: UPDATE
+
+    Store->>DbSvc: deleteConversation(convId)
+    activate DbSvc
+    DbSvc->>Dexie: db.conversations.delete(convId)
+    Dexie->>IDB: DELETE FROM conversations
+    DbSvc->>Dexie: db.messages.where('convId').equals(convId).delete()
+    Dexie->>IDB: DELETE FROM messages WHERE convId = ?
+    deactivate DbSvc
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over Store,IDB: 📝 MESSAGES CRUD
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Store->>DbSvc: createRootMessage(convId)
+    activate DbSvc
+    DbSvc->>DbSvc: Create root message {type: "root", parent: null}
+    DbSvc->>Dexie: db.messages.add(rootMsg)
+    Dexie->>IDB: INSERT
+    DbSvc-->>Store: rootMessageId
+    deactivate DbSvc
+
+    Store->>DbSvc: createMessageBranch(message, parentId)
+    activate DbSvc
+    DbSvc->>DbSvc: Generate UUID for new message
+    DbSvc->>Dexie: db.messages.add({...message, id, parent: parentId})
+    Dexie->>IDB: INSERT message
+
+    alt parentId exists
+        DbSvc->>Dexie: db.messages.get(parentId)
+        Dexie->>IDB: SELECT parent
+        DbSvc->>DbSvc: parent.children.push(newId)
+        DbSvc->>Dexie: db.messages.update(parentId, {children})
+        Dexie->>IDB: UPDATE parent.children
+    end
+
+    DbSvc->>Dexie: db.conversations.update(convId, {currNode: newId})
+    Dexie->>IDB: UPDATE conversation.currNode
+    DbSvc-->>Store: DatabaseMessage
+    deactivate DbSvc
+
+    Store->>DbSvc: getConversationMessages(convId)
+    DbSvc->>Dexie: db.messages.where('convId').equals(convId).toArray()
+    Dexie->>IDB: SELECT WHERE convId = ?
+    IDB-->>DbSvc: DatabaseMessage[]
+
+    Store->>DbSvc: updateMessage(msgId, updates)
+    DbSvc->>Dexie: db.messages.update(msgId, updates)
+    Dexie->>IDB: UPDATE
+
+    Store->>DbSvc: deleteMessage(msgId)
+    DbSvc->>Dexie: db.messages.delete(msgId)
+    Dexie->>IDB: DELETE
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over Store,IDB: 🌳 BRANCHING OPERATIONS
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Store->>DbSvc: updateCurrentNode(convId, nodeId)
+    DbSvc->>Dexie: db.conversations.update(convId, {currNode: nodeId, lastModified})
+    Dexie->>IDB: UPDATE
+
+    Store->>DbSvc: deleteMessageCascading(msgId)
+    activate DbSvc
+    DbSvc->>DbSvc: findDescendantMessages(msgId, allMessages)
+    Note right of DbSvc: Recursively find all children
+    loop each descendant
+        DbSvc->>Dexie: db.messages.delete(descendantId)
+        Dexie->>IDB: DELETE
+    end
+    DbSvc->>Dexie: db.messages.delete(msgId)
+    Dexie->>IDB: DELETE target message
+    deactivate DbSvc
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over Store,IDB: 📥 IMPORT
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Store->>DbSvc: importConversations(data)
+    activate DbSvc
+    loop each conversation in data
+        DbSvc->>DbSvc: Generate new UUIDs (avoid conflicts)
+        DbSvc->>Dexie: db.conversations.add(conversation)
+        Dexie->>IDB: INSERT conversation
+        loop each message
+            DbSvc->>Dexie: db.messages.add(message)
+            Dexie->>IDB: INSERT message
+        end
+    end
+    deactivate DbSvc
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over Store,IDB: 🔗 MESSAGE TREE UTILITIES
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over DbSvc: Used by stores (imported from utils):
+
+    rect rgb(240, 255, 240)
+        Note over DbSvc: filterByLeafNodeId(messages, leafId)<br/>→ Returns path from root to leaf<br/>→ Used to display current branch
+    end
+
+    rect rgb(240, 255, 240)
+        Note over DbSvc: findLeafNode(startId, messages)<br/>→ Traverse to deepest child<br/>→ Used for branch navigation
+    end
+
+    rect rgb(240, 255, 240)
+        Note over DbSvc: findDescendantMessages(msgId, messages)<br/>→ Find all children recursively<br/>→ Used for cascading deletes
+    end
+```

tools/server/webui/docs/flows/models-flow.md

@@ -0,0 +1,181 @@
+```mermaid
+sequenceDiagram
+    participant UI as 🧩 ModelsSelector
+    participant Hooks as 🪝 useModelChangeValidation
+    participant modelsStore as 🗄️ modelsStore
+    participant serverStore as 🗄️ serverStore
+    participant convStore as 🗄️ conversationsStore
+    participant ModelsSvc as ⚙️ ModelsService
+    participant PropsSvc as ⚙️ PropsService
+    participant API as 🌐 llama-server
+
+    Note over modelsStore: State:<br/>models: ModelOption[]<br/>routerModels: ApiModelDataEntry[]<br/>selectedModelId, selectedModelName<br/>loading, updating, error<br/>modelLoadingStates (Map)<br/>modelPropsCache (Map)<br/>propsCacheVersion
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 🚀 INITIALIZATION (MODEL mode)
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>modelsStore: fetch()
+    activate modelsStore
+    modelsStore->>modelsStore: loading = true
+
+    alt serverStore.props not loaded
+        modelsStore->>serverStore: fetch()
+        Note over serverStore: → see server-flow.mmd
+    end
+
+    modelsStore->>ModelsSvc: list()
+    ModelsSvc->>API: GET /v1/models
+    API-->>ModelsSvc: ApiModelListResponse {data: [model]}
+
+    modelsStore->>modelsStore: models = $state(mapped)
+    Note right of modelsStore: Map to ModelOption[]:<br/>{id, name, model, description, capabilities}
+
+    Note over modelsStore: MODEL mode: Get modalities from serverStore.props
+    modelsStore->>modelsStore: modelPropsCache.set(model.id, serverStore.props)
+    modelsStore->>modelsStore: models[0].modalities = props.modalities
+
+    modelsStore->>modelsStore: Auto-select single model
+    Note right of modelsStore: selectedModelId = models[0].id
+    modelsStore->>modelsStore: loading = false
+    deactivate modelsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 🚀 INITIALIZATION (ROUTER mode)
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>modelsStore: fetch()
+    activate modelsStore
+    modelsStore->>ModelsSvc: list()
+    ModelsSvc->>API: GET /v1/models
+    API-->>ModelsSvc: ApiModelListResponse
+    modelsStore->>modelsStore: models = $state(mapped)
+    deactivate modelsStore
+
+    Note over UI: After models loaded, layout triggers:
+    UI->>modelsStore: fetchRouterModels()
+    activate modelsStore
+    modelsStore->>ModelsSvc: listRouter()
+    ModelsSvc->>API: GET /models
+    API-->>ModelsSvc: ApiRouterModelsListResponse
+    Note right of API: {data: [{id, status, path, in_cache}]}
+    modelsStore->>modelsStore: routerModels = $state(data)
+
+    modelsStore->>modelsStore: fetchModalitiesForLoadedModels()
+    loop each model where status === "loaded"
+        modelsStore->>PropsSvc: fetchForModel(modelId)
+        PropsSvc->>API: GET /props?model={modelId}
+        API-->>PropsSvc: ApiLlamaCppServerProps
+        modelsStore->>modelsStore: modelPropsCache.set(modelId, props)
+    end
+    modelsStore->>modelsStore: propsCacheVersion++
+    deactivate modelsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 🔄 MODEL SELECTION (ROUTER mode)
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>Hooks: useModelChangeValidation({getRequiredModalities, onSuccess?, onValidationFailure?})
+    Note over Hooks: Hook configured per-component:<br/>ChatForm: getRequiredModalities = usedModalities<br/>ChatMessage: getRequiredModalities = getModalitiesUpToMessage(msgId)
+
+    UI->>Hooks: handleModelChange(modelId, modelName)
+    activate Hooks
+    Hooks->>Hooks: previousSelectedModelId = modelsStore.selectedModelId
+    Hooks->>modelsStore: isModelLoaded(modelName)?
+
+    alt model NOT loaded
+        Hooks->>modelsStore: loadModel(modelName)
+        Note over modelsStore: → see LOAD MODEL section below
+    end
+
+    Note over Hooks: Always fetch props (from cache or API)
+    Hooks->>modelsStore: fetchModelProps(modelName)
+    modelsStore-->>Hooks: props
+
+    Hooks->>convStore: getRequiredModalities()
+    convStore-->>Hooks: {vision, audio}
+
+    Hooks->>Hooks: Validate: model.modalities ⊇ required?
+
+    alt validation PASSED
+        Hooks->>modelsStore: selectModelById(modelId)
+        Hooks-->>UI: return true
+    else validation FAILED
+        Hooks->>UI: toast.error("Model doesn't support required modalities")
+        alt model was just loaded
+            Hooks->>modelsStore: unloadModel(modelName)
+        end
+        alt onValidationFailure provided
+            Hooks->>modelsStore: selectModelById(previousSelectedModelId)
+        end
+        Hooks-->>UI: return false
+    end
+    deactivate Hooks
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: ⬆️ LOAD MODEL (ROUTER mode)
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    modelsStore->>modelsStore: loadModel(modelId)
+    activate modelsStore
+
+    alt already loaded
+        modelsStore-->>modelsStore: return (no-op)
+    end
+
+    modelsStore->>modelsStore: modelLoadingStates.set(modelId, true)
+    modelsStore->>ModelsSvc: load(modelId)
+    ModelsSvc->>API: POST /models/load {model: modelId}
+    API-->>ModelsSvc: {status: "loading"}
+
+    modelsStore->>modelsStore: pollForModelStatus(modelId, LOADED)
+    loop poll every 500ms (max 60 attempts)
+        modelsStore->>modelsStore: fetchRouterModels()
+        modelsStore->>ModelsSvc: listRouter()
+        ModelsSvc->>API: GET /models
+        API-->>ModelsSvc: models[]
+        modelsStore->>modelsStore: getModelStatus(modelId)
+        alt status === LOADED
+            Note right of modelsStore: break loop
+        else status === LOADING
+            Note right of modelsStore: wait 500ms, continue
+        end
+    end
+
+    modelsStore->>modelsStore: updateModelModalities(modelId)
+    modelsStore->>PropsSvc: fetchForModel(modelId)
+    PropsSvc->>API: GET /props?model={modelId}
+    API-->>PropsSvc: props with modalities
+    modelsStore->>modelsStore: modelPropsCache.set(modelId, props)
+    modelsStore->>modelsStore: propsCacheVersion++
+
+    modelsStore->>modelsStore: modelLoadingStates.set(modelId, false)
+    deactivate modelsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: ⬇️ UNLOAD MODEL (ROUTER mode)
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    modelsStore->>modelsStore: unloadModel(modelId)
+    activate modelsStore
+    modelsStore->>modelsStore: modelLoadingStates.set(modelId, true)
+    modelsStore->>ModelsSvc: unload(modelId)
+    ModelsSvc->>API: POST /models/unload {model: modelId}
+
+    modelsStore->>modelsStore: pollForModelStatus(modelId, UNLOADED)
+    loop poll until unloaded
+        modelsStore->>ModelsSvc: listRouter()
+        ModelsSvc->>API: GET /models
+    end
+
+    modelsStore->>modelsStore: modelLoadingStates.set(modelId, false)
+    deactivate modelsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 📊 COMPUTED GETTERS
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over modelsStore: Getters:<br/>- selectedModel: ModelOption | null<br/>- loadedModelIds: string[] (from routerModels)<br/>- loadingModelIds: string[] (from modelLoadingStates)<br/>- singleModelName: string | null (MODEL mode only)
+
+    Note over modelsStore: Modality helpers:<br/>- getModelModalities(modelId): {vision, audio}<br/>- modelSupportsVision(modelId): boolean<br/>- modelSupportsAudio(modelId): boolean
+```

tools/server/webui/docs/flows/server-flow.md

@@ -0,0 +1,76 @@
+```mermaid
+sequenceDiagram
+    participant UI as 🧩 +layout.svelte
+    participant serverStore as 🗄️ serverStore
+    participant PropsSvc as ⚙️ PropsService
+    participant API as 🌐 llama-server
+
+    Note over serverStore: State:<br/>props: ApiLlamaCppServerProps | null<br/>loading, error<br/>role: ServerRole | null (MODEL | ROUTER)<br/>fetchPromise (deduplication)
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 🚀 INITIALIZATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>serverStore: fetch()
+    activate serverStore
+
+    alt fetchPromise exists (already fetching)
+        serverStore-->>UI: return fetchPromise
+        Note right of serverStore: Deduplicate concurrent calls
+    end
+
+    serverStore->>serverStore: loading = true
+    serverStore->>serverStore: fetchPromise = new Promise()
+
+    serverStore->>PropsSvc: fetch()
+    PropsSvc->>API: GET /props
+    API-->>PropsSvc: ApiLlamaCppServerProps
+    Note right of API: {role, model_path, model_alias,<br/>modalities, default_generation_settings, ...}
+
+    PropsSvc-->>serverStore: props
+    serverStore->>serverStore: props = $state(data)
+
+    serverStore->>serverStore: detectRole(props)
+    Note right of serverStore: role = props.role === "router"<br/>  ? ServerRole.ROUTER<br/>  : ServerRole.MODEL
+
+    serverStore->>serverStore: loading = false
+    serverStore->>serverStore: fetchPromise = null
+    deactivate serverStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 📊 COMPUTED GETTERS
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over serverStore: Getters from props:
+
+    rect rgb(240, 255, 240)
+        Note over serverStore: defaultParams<br/>→ props.default_generation_settings.params<br/>(temperature, top_p, top_k, etc.)
+    end
+
+    rect rgb(240, 255, 240)
+        Note over serverStore: contextSize<br/>→ props.default_generation_settings.n_ctx
+    end
+
+    rect rgb(255, 240, 240)
+        Note over serverStore: isRouterMode<br/>→ role === ServerRole.ROUTER
+    end
+
+    rect rgb(255, 240, 240)
+        Note over serverStore: isModelMode<br/>→ role === ServerRole.MODEL
+    end
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: 🔗 RELATIONSHIPS
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over serverStore: Used by:
+    Note right of serverStore: - modelsStore: role detection, MODEL mode modalities<br/>- settingsStore: syncWithServerDefaults (defaultParams)<br/>- chatStore: contextSize for processing state<br/>- UI components: isRouterMode for conditional rendering
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,API: ❌ ERROR HANDLING
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over serverStore: getErrorMessage(): string | null<br/>Returns formatted error for UI display
+
+    Note over serverStore: clear(): void<br/>Resets all state (props, error, loading, role)
+```

tools/server/webui/docs/flows/settings-flow.md

@@ -0,0 +1,144 @@
+```mermaid
+sequenceDiagram
+    participant UI as 🧩 ChatSettings
+    participant settingsStore as 🗄️ settingsStore
+    participant serverStore as 🗄️ serverStore
+    participant ParamSvc as ⚙️ ParameterSyncService
+    participant LS as 💾 LocalStorage
+
+    Note over settingsStore: State:<br/>config: SettingsConfigType<br/>theme: string ("auto" | "light" | "dark")<br/>isInitialized: boolean<br/>userOverrides: Set&lt;string&gt;
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,LS: 🚀 INITIALIZATION
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over settingsStore: Auto-initialized in constructor (browser only)
+    settingsStore->>settingsStore: initialize()
+    activate settingsStore
+
+    settingsStore->>settingsStore: loadConfig()
+    settingsStore->>LS: get("llama-config")
+    LS-->>settingsStore: StoredConfig | null
+
+    alt config exists
+        settingsStore->>settingsStore: Merge with SETTING_CONFIG_DEFAULT
+        Note right of settingsStore: Fill missing keys with defaults
+    else no config
+        settingsStore->>settingsStore: config = SETTING_CONFIG_DEFAULT
+    end
+
+    settingsStore->>LS: get("llama-userOverrides")
+    LS-->>settingsStore: string[] | null
+    settingsStore->>settingsStore: userOverrides = new Set(data)
+
+    settingsStore->>settingsStore: loadTheme()
+    settingsStore->>LS: get("llama-theme")
+    LS-->>settingsStore: theme | "auto"
+
+    settingsStore->>settingsStore: isInitialized = true
+    deactivate settingsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,LS: 🔄 SYNC WITH SERVER DEFAULTS
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over UI: Triggered from +layout.svelte when serverStore.props loaded
+    UI->>settingsStore: syncWithServerDefaults()
+    activate settingsStore
+
+    settingsStore->>serverStore: defaultParams
+    serverStore-->>settingsStore: {temperature, top_p, top_k, ...}
+
+    settingsStore->>ParamSvc: extractServerDefaults(defaultParams)
+    ParamSvc-->>settingsStore: Record<string, value>
+
+    settingsStore->>ParamSvc: mergeWithServerDefaults(config, serverDefaults)
+    Note right of ParamSvc: For each syncable parameter:<br/>- If NOT in userOverrides → use server default<br/>- If in userOverrides → keep user value
+    ParamSvc-->>settingsStore: mergedConfig
+
+    settingsStore->>settingsStore: config = mergedConfig
+    settingsStore->>settingsStore: saveConfig()
+    deactivate settingsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,LS: ⚙️ UPDATE CONFIG
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>settingsStore: updateConfig(key, value)
+    activate settingsStore
+    settingsStore->>settingsStore: config[key] = value
+    settingsStore->>settingsStore: userOverrides.add(key)
+    Note right of settingsStore: Mark as user-modified (won't be overwritten by server)
+    settingsStore->>settingsStore: saveConfig()
+    settingsStore->>LS: set("llama-config", config)
+    settingsStore->>LS: set("llama-userOverrides", [...userOverrides])
+    deactivate settingsStore
+
+    UI->>settingsStore: updateMultipleConfig({key1: val1, key2: val2})
+    activate settingsStore
+    Note right of settingsStore: Batch update, single save
+    settingsStore->>settingsStore: For each key: config[key] = value
+    settingsStore->>settingsStore: For each key: userOverrides.add(key)
+    settingsStore->>settingsStore: saveConfig()
+    deactivate settingsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,LS: 🔄 RESET
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>settingsStore: resetConfig()
+    activate settingsStore
+    settingsStore->>settingsStore: config = SETTING_CONFIG_DEFAULT
+    settingsStore->>settingsStore: userOverrides.clear()
+    settingsStore->>settingsStore: syncWithServerDefaults()
+    Note right of settingsStore: Apply server defaults for syncable params
+    settingsStore->>settingsStore: saveConfig()
+    deactivate settingsStore
+
+    UI->>settingsStore: resetParameterToServerDefault(key)
+    activate settingsStore
+    settingsStore->>settingsStore: userOverrides.delete(key)
+    settingsStore->>serverStore: defaultParams[key]
+    settingsStore->>settingsStore: config[key] = serverDefault
+    settingsStore->>settingsStore: saveConfig()
+    deactivate settingsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,LS: 🎨 THEME
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>settingsStore: updateTheme(newTheme)
+    activate settingsStore
+    settingsStore->>settingsStore: theme = newTheme
+    settingsStore->>settingsStore: saveTheme()
+    settingsStore->>LS: set("llama-theme", theme)
+    deactivate settingsStore
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,LS: 📊 PARAMETER INFO
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    UI->>settingsStore: getParameterInfo(key)
+    settingsStore->>ParamSvc: getParameterInfo(key, config, serverDefaults, userOverrides)
+    ParamSvc-->>settingsStore: ParameterInfo
+    Note right of ParamSvc: {<br/>  currentValue,<br/>  serverDefault,<br/>  isUserOverride: boolean,<br/>  canSync: boolean,<br/>  isDifferentFromServer: boolean<br/>}
+
+    UI->>settingsStore: getParameterDiff()
+    settingsStore->>ParamSvc: createParameterDiff(config, serverDefaults, userOverrides)
+    ParamSvc-->>settingsStore: ParameterDiff[]
+    Note right of ParamSvc: Array of parameters where user != server
+
+    %% ═══════════════════════════════════════════════════════════════════════════
+    Note over UI,LS: 📋 CONFIG CATEGORIES
+    %% ═══════════════════════════════════════════════════════════════════════════
+
+    Note over settingsStore: Syncable with server (from /props):
+    rect rgb(240, 255, 240)
+        Note over settingsStore: temperature, top_p, top_k, min_p<br/>repeat_penalty, presence_penalty, frequency_penalty<br/>dynatemp_range, dynatemp_exponent<br/>typ_p, xtc_probability, xtc_threshold<br/>dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n
+    end
+
+    Note over settingsStore: UI-only (not synced):
+    rect rgb(255, 240, 240)
+        Note over settingsStore: systemMessage, custom (JSON)<br/>showStatistics, enableContinueGeneration<br/>autoMicOnEmpty, disableAutoScroll<br/>apiKey, pdfAsImage, disableReasoningFormat
+    end
+```

tools/server/webui/package-lock.json

@@ -64,7 +64,7 @@
 				"svelte": "^5.0.0",
 				"svelte-check": "^4.0.0",
 				"tailwind-merge": "^3.3.1",
-				"tailwind-variants": "^1.0.0",
+				"tailwind-variants": "^3.2.2",
 				"tailwindcss": "^4.0.0",
 				"tw-animate-css": "^1.3.5",
 				"typescript": "^5.0.0",
@@ -8324,31 +8324,23 @@
 			}
 		},
 		"node_modules/tailwind-variants": {
-			"version": "1.0.0",
-			"resolved": "https://registry.npmjs.org/tailwind-variants/-/tailwind-variants-1.0.0.tgz",
-			"integrity": "sha512-2WSbv4ulEEyuBKomOunut65D8UZwxrHoRfYnxGcQNnHqlSCp2+B7Yz2W+yrNDrxRodOXtGD/1oCcKGNBnUqMqA==",
+			"version": "3.2.2",
+			"resolved": "https://registry.npmjs.org/tailwind-variants/-/tailwind-variants-3.2.2.tgz",
+			"integrity": "sha512-Mi4kHeMTLvKlM98XPnK+7HoBPmf4gygdFmqQPaDivc3DpYS6aIY6KiG/PgThrGvii5YZJqRsPz0aPyhoFzmZgg==",
 			"dev": true,
 			"license": "MIT",
-			"dependencies": {
-				"tailwind-merge": "3.0.2"
-			},
 			"engines": {
 				"node": ">=16.x",
 				"pnpm": ">=7.x"
 			},
 			"peerDependencies": {
+				"tailwind-merge": ">=3.0.0",
 				"tailwindcss": "*"
-			}
-		},
-		"node_modules/tailwind-variants/node_modules/tailwind-merge": {
-			"version": "3.0.2",
-			"resolved": "https://registry.npmjs.org/tailwind-merge/-/tailwind-merge-3.0.2.tgz",
-			"integrity": "sha512-l7z+OYZ7mu3DTqrL88RiKrKIqO3NcpEO8V/Od04bNpvk0kiIFndGEoqfuzvj4yuhRkHKjRkII2z+KS2HfPcSxw==",
-			"dev": true,
-			"license": "MIT",
-			"funding": {
-				"type": "github",
-				"url": "https://github.com/sponsors/dcastil"
+			},
+			"peerDependenciesMeta": {
+				"tailwind-merge": {
+					"optional": true
+				}
 			}
 		},
 		"node_modules/tailwindcss": {

tools/server/webui/package.json

@@ -66,7 +66,7 @@
 		"svelte": "^5.0.0",
 		"svelte-check": "^4.0.0",
 		"tailwind-merge": "^3.3.1",
-		"tailwind-variants": "^1.0.0",
+		"tailwind-variants": "^3.2.2",
 		"tailwindcss": "^4.0.0",
 		"tw-animate-css": "^1.3.5",
 		"typescript": "^5.0.0",

tools/server/webui/playwright.config.ts

@@ -7,5 +7,5 @@ export default defineConfig({
 		timeout: 120000,
 		reuseExistingServer: false
 	},
-	testDir: 'e2e'
+	testDir: 'tests/e2e'
 });

tools/server/webui/scripts/dev.sh

@@ -49,7 +49,9 @@ trap cleanup SIGINT SIGTERM
 echo "🚀 Starting development servers..."
 echo "📝 Note: Make sure to start llama-server separately if needed"
 cd tools/server/webui
-storybook dev -p 6006 --ci & vite dev --host 0.0.0.0 &
+# Use --insecure-http-parser to handle malformed HTTP responses from llama-server
+# (some responses have both Content-Length and Transfer-Encoding headers)
+storybook dev -p 6006 --ci & NODE_OPTIONS="--insecure-http-parser" vite dev --host 0.0.0.0 &
 
 # Wait for all background processes
 wait

tools/server/webui/src/app.css

@@ -29,7 +29,7 @@
 	--chart-3: oklch(0.398 0.07 227.392);
 	--chart-4: oklch(0.828 0.189 84.429);
 	--chart-5: oklch(0.769 0.188 70.08);
-	--sidebar: oklch(0.985 0 0);
+	--sidebar: oklch(0.987 0 0);
 	--sidebar-foreground: oklch(0.145 0 0);
 	--sidebar-primary: oklch(0.205 0 0);
 	--sidebar-primary-foreground: oklch(0.985 0 0);
@@ -66,7 +66,7 @@
 	--chart-3: oklch(0.769 0.188 70.08);
 	--chart-4: oklch(0.627 0.265 303.9);
 	--chart-5: oklch(0.645 0.246 16.439);
-	--sidebar: oklch(0.205 0 0);
+	--sidebar: oklch(0.19 0 0);
 	--sidebar-foreground: oklch(0.985 0 0);
 	--sidebar-primary: oklch(0.488 0.243 264.376);
 	--sidebar-primary-foreground: oklch(0.985 0 0);

tools/server/webui/src/app.d.ts

@@ -4,27 +4,38 @@
 // Import chat types from dedicated module
 
 import type {
+	// API types
 	ApiChatCompletionRequest,
 	ApiChatCompletionResponse,
 	ApiChatCompletionStreamChunk,
+	ApiChatCompletionToolCall,
+	ApiChatCompletionToolCallDelta,
 	ApiChatMessageData,
 	ApiChatMessageContentPart,
 	ApiContextSizeError,
 	ApiErrorResponse,
 	ApiLlamaCppServerProps,
-	ApiProcessingState
-} from '$lib/types/api';
-
-import type {
+	ApiModelDataEntry,
+	ApiModelListResponse,
+	ApiProcessingState,
+	ApiRouterModelMeta,
+	ApiRouterModelsLoadRequest,
+	ApiRouterModelsLoadResponse,
+	ApiRouterModelsStatusRequest,
+	ApiRouterModelsStatusResponse,
+	ApiRouterModelsListResponse,
+	ApiRouterModelsUnloadRequest,
+	ApiRouterModelsUnloadResponse,
+	// Chat types
+	ChatAttachmentDisplayItem,
+	ChatAttachmentPreviewItem,
 	ChatMessageType,
 	ChatRole,
 	ChatUploadedFile,
 	ChatMessageSiblingInfo,
 	ChatMessagePromptProgress,
-	ChatMessageTimings
-} from '$lib/types/chat';
-
-import type {
+	ChatMessageTimings,
+	// Database types
 	DatabaseConversation,
 	DatabaseMessage,
 	DatabaseMessageExtra,
@@ -32,14 +43,20 @@ import type {
 	DatabaseMessageExtraImageFile,
 	DatabaseMessageExtraTextFile,
 	DatabaseMessageExtraPdfFile,
-	DatabaseMessageExtraLegacyContext
-} from '$lib/types/database';
-
-import type {
+	DatabaseMessageExtraLegacyContext,
+	ExportedConversation,
+	ExportedConversations,
+	// Model types
+	ModelModalities,
+	ModelOption,
+	// Settings types
+	SettingsChatServiceOptions,
 	SettingsConfigValue,
 	SettingsFieldConfig,
 	SettingsConfigType
-} from '$lib/types/settings';
+} from '$lib/types';
+
+import { ServerRole, ServerModelStatus, ModelModality } from '$lib/enums';
 
 declare global {
 	// namespace App {
@@ -51,22 +68,38 @@ declare global {
 	// }
 
 	export {
+		// API types
 		ApiChatCompletionRequest,
 		ApiChatCompletionResponse,
 		ApiChatCompletionStreamChunk,
+		ApiChatCompletionToolCall,
+		ApiChatCompletionToolCallDelta,
 		ApiChatMessageData,
 		ApiChatMessageContentPart,
 		ApiContextSizeError,
 		ApiErrorResponse,
 		ApiLlamaCppServerProps,
+		ApiModelDataEntry,
+		ApiModelListResponse,
 		ApiProcessingState,
-		ChatMessageData,
+		ApiRouterModelMeta,
+		ApiRouterModelsLoadRequest,
+		ApiRouterModelsLoadResponse,
+		ApiRouterModelsStatusRequest,
+		ApiRouterModelsStatusResponse,
+		ApiRouterModelsListResponse,
+		ApiRouterModelsUnloadRequest,
+		ApiRouterModelsUnloadResponse,
+		// Chat types
+		ChatAttachmentDisplayItem,
+		ChatAttachmentPreviewItem,
 		ChatMessagePromptProgress,
 		ChatMessageSiblingInfo,
 		ChatMessageTimings,
 		ChatMessageType,
 		ChatRole,
 		ChatUploadedFile,
+		// Database types
 		DatabaseConversation,
 		DatabaseMessage,
 		DatabaseMessageExtra,
@@ -75,9 +108,19 @@ declare global {
 		DatabaseMessageExtraTextFile,
 		DatabaseMessageExtraPdfFile,
 		DatabaseMessageExtraLegacyContext,
+		ExportedConversation,
+		ExportedConversations,
+		// Enum types
+		ModelModality,
+		ServerRole,
+		ServerModelStatus,
+		// Model types
+		ModelModalities,
+		ModelOption,
+		// Settings types
+		SettingsChatServiceOptions,
 		SettingsConfigValue,
 		SettingsFieldConfig,
-		SettingsConfigType,
-		SettingsChatServiceOptions
+		SettingsConfigType
 	};
 }

tools/server/webui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentPreview.svelte

@@ -1,9 +1,17 @@
 <script lang="ts">
-	import { FileText, Image, Music, FileIcon, Eye } from '@lucide/svelte';
-	import { FileTypeCategory, MimeTypeApplication } from '$lib/enums/files';
-	import { convertPDFToImage } from '$lib/utils/pdf-processing';
 	import { Button } from '$lib/components/ui/button';
-	import { getFileTypeCategory } from '$lib/utils/file-type';
+	import * as Alert from '$lib/components/ui/alert';
+	import { SyntaxHighlightedCode } from '$lib/components/app';
+	import { FileText, Image, Music, FileIcon, Eye, Info } from '@lucide/svelte';
+	import {
+		isTextFile,
+		isImageFile,
+		isPdfFile,
+		isAudioFile,
+		getLanguageFromFilename
+	} from '$lib/utils';
+	import { convertPDFToImage } from '$lib/utils/browser-only';
+	import { modelsStore } from '$lib/stores/models.svelte';
 
 	interface Props {
 		// Either an uploaded file or a stored attachment
@@ -12,53 +20,36 @@
 		// For uploaded files
 		preview?: string;
 		name?: string;
-		type?: string;
 		textContent?: string;
+		// For checking vision modality
+		activeModelId?: string;
 	}
 
-	let { uploadedFile, attachment, preview, name, type, textContent }: Props = $props();
+	let { uploadedFile, attachment, preview, name, textContent, activeModelId }: Props = $props();
+
+	let hasVisionModality = $derived(
+		activeModelId ? modelsStore.modelSupportsVision(activeModelId) : false
+	);
 
 	let displayName = $derived(uploadedFile?.name || attachment?.name || name || 'Unknown File');
 
-	let displayPreview = $derived(
-		uploadedFile?.preview || (attachment?.type === 'imageFile' ? attachment.base64Url : preview)
-	);
+	// Determine file type from uploaded file or attachment
+	let isAudio = $derived(isAudioFile(attachment, uploadedFile));
+	let isImage = $derived(isImageFile(attachment, uploadedFile));
+	let isPdf = $derived(isPdfFile(attachment, uploadedFile));
+	let isText = $derived(isTextFile(attachment, uploadedFile));
 
-	let displayType = $derived(
-		uploadedFile?.type ||
-			(attachment?.type === 'imageFile'
-				? 'image'
-				: attachment?.type === 'textFile'
-					? 'text'
-					: attachment?.type === 'audioFile'
-						? attachment.mimeType || 'audio'
-						: attachment?.type === 'pdfFile'
-							? MimeTypeApplication.PDF
-							: type || 'unknown')
+	let displayPreview = $derived(
+		uploadedFile?.preview ||
+			(isImage && attachment && 'base64Url' in attachment ? attachment.base64Url : preview)
 	);
 
 	let displayTextContent = $derived(
 		uploadedFile?.textContent ||
-			(attachment?.type === 'textFile'
-				? attachment.content
-				: attachment?.type === 'pdfFile'
-					? attachment.content
-					: textContent)
+			(attachment && 'content' in attachment ? attachment.content : textContent)
 	);
 
-	let isAudio = $derived(
-		getFileTypeCategory(displayType) === FileTypeCategory.AUDIO || displayType === 'audio'
-	);
-
-	let isImage = $derived(
-		getFileTypeCategory(displayType) === FileTypeCategory.IMAGE || displayType === 'image'
-	);
-
-	let isPdf = $derived(displayType === MimeTypeApplication.PDF);
-
-	let isText = $derived(
-		getFileTypeCategory(displayType) === FileTypeCategory.TEXT || displayType === 'text'
-	);
+	let language = $derived(getLanguageFromFilename(displayName));
 
 	let IconComponent = $derived(() => {
 		if (isImage) return Image;
@@ -87,15 +78,20 @@
 
 			if (uploadedFile?.file) {
 				file = uploadedFile.file;
-			} else if (attachment?.type === 'pdfFile') {
+			} else if (isPdf && attachment) {
 				// Check if we have pre-processed images
-				if (attachment.images && Array.isArray(attachment.images)) {
+				if (
+					'images' in attachment &&
+					attachment.images &&
+					Array.isArray(attachment.images) &&
+					attachment.images.length > 0
+				) {
 					pdfImages = attachment.images;
 					return;
 				}
 
 				// Convert base64 back to File for processing
-				if (attachment.base64Data) {
+				if ('base64Data' in attachment && attachment.base64Data) {
 					const base64Data = attachment.base64Data;
 					const byteCharacters = atob(base64Data);
 					const byteNumbers = new Array(byteCharacters.length);
@@ -103,7 +99,7 @@
 						byteNumbers[i] = byteCharacters.charCodeAt(i);
 					}
 					const byteArray = new Uint8Array(byteNumbers);
-					file = new File([byteArray], displayName, { type: MimeTypeApplication.PDF });
+					file = new File([byteArray], displayName, { type: 'application/pdf' });
 				}
 			}
 
@@ -181,6 +177,24 @@
 				/>
 			</div>
 		{:else if isPdf && pdfViewMode === 'pages'}
+			{#if !hasVisionModality && activeModelId}
+				<Alert.Root class="mb-4">
+					<Info class="h-4 w-4" />
+					<Alert.Title>Preview only</Alert.Title>
+					<Alert.Description>
+						<span class="inline-flex">
+							The selected model does not support vision. Only the extracted
+							<!-- svelte-ignore a11y_click_events_have_key_events -->
+							<!-- svelte-ignore a11y_no_static_element_interactions -->
+							<span class="mx-1 cursor-pointer underline" onclick={() => (pdfViewMode = 'text')}>
+								text
+							</span>
+							will be sent to the model.
+						</span>
+					</Alert.Description>
+				</Alert.Root>
+			{/if}
+
 			{#if pdfImagesLoading}
 				<div class="flex items-center justify-center p-8">
 					<div class="text-center">
@@ -227,28 +241,24 @@
 				</div>
 			{/if}
 		{:else if (isText || (isPdf && pdfViewMode === 'text')) && displayTextContent}
-			<div
-				class="max-h-[60vh] overflow-auto rounded-lg bg-muted p-4 font-mono text-sm break-words whitespace-pre-wrap"
-			>
-				{displayTextContent}
-			</div>
+			<SyntaxHighlightedCode code={displayTextContent} {language} maxWidth="69rem" />
 		{:else if isAudio}
 			<div class="flex items-center justify-center p-8">
 				<div class="w-full max-w-md text-center">
 					<Music class="mx-auto mb-4 h-16 w-16 text-muted-foreground" />
 
-					{#if attachment?.type === 'audioFile'}
+					{#if uploadedFile?.preview}
+						<audio controls class="mb-4 w-full" src={uploadedFile.preview}>
+							Your browser does not support the audio element.
+						</audio>
+					{:else if isAudio && attachment && 'mimeType' in attachment && 'base64Data' in attachment}
 						<audio
 							controls
 							class="mb-4 w-full"
-							src="data:{attachment.mimeType};base64,{attachment.base64Data}"
+							src={`data:${attachment.mimeType};base64,${attachment.base64Data}`}
 						>
 							Your browser does not support the audio element.
 						</audio>
-					{:else if uploadedFile?.preview}
-						<audio controls class="mb-4 w-full" src={uploadedFile.preview}>
-							Your browser does not support the audio element.
-						</audio>
 					{:else}
 						<p class="mb-4 text-muted-foreground">Audio preview not available</p>
 					{/if}

tools/server/webui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentThumbnailFile.svelte

@@ -1,7 +1,7 @@
 <script lang="ts">
 	import { RemoveButton } from '$lib/components/app';
-	import { formatFileSize, getFileTypeLabel, getPreviewText } from '$lib/utils/file-preview';
-	import { FileTypeCategory, MimeTypeText } from '$lib/enums/files';
+	import { getFileTypeLabel, getPreviewText, formatFileSize, isTextFile } from '$lib/utils';
+	import { AttachmentType } from '$lib/enums';
 
 	interface Props {
 		class?: string;
@@ -12,7 +12,9 @@
 		readonly?: boolean;
 		size?: number;
 		textContent?: string;
-		type: string;
+		// Either uploaded file or stored attachment
+		uploadedFile?: ChatUploadedFile;
+		attachment?: DatabaseMessageExtra;
 	}
 
 	let {
@@ -24,11 +26,41 @@
 		readonly = false,
 		size,
 		textContent,
-		type
+		uploadedFile,
+		attachment
 	}: Props = $props();
+
+	let isText = $derived(isTextFile(attachment, uploadedFile));
+
+	let fileTypeLabel = $derived.by(() => {
+		if (uploadedFile?.type) {
+			return getFileTypeLabel(uploadedFile.type);
+		}
+
+		if (attachment) {
+			if ('mimeType' in attachment && attachment.mimeType) {
+				return getFileTypeLabel(attachment.mimeType);
+			}
+
+			if (attachment.type) {
+				return getFileTypeLabel(attachment.type);
+			}
+		}
+
+		return getFileTypeLabel(name);
+	});
+
+	let pdfProcessingMode = $derived.by(() => {
+		if (attachment?.type === AttachmentType.PDF) {
+			const pdfAttachment = attachment as DatabaseMessageExtraPdfFile;
+
+			return pdfAttachment.processedAsImages ? 'Sent as Image' : 'Sent as Text';
+		}
+		return null;
+	});
 </script>
 
-{#if type === MimeTypeText.PLAIN || type === FileTypeCategory.TEXT}
+{#if isText}
 	{#if readonly}
 		<!-- Readonly mode (ChatMessage) -->
 		<button
@@ -45,7 +77,7 @@
 						<span class="text-xs text-muted-foreground">{formatFileSize(size)}</span>
 					{/if}
 
-					{#if textContent && type === 'text'}
+					{#if textContent}
 						<div class="relative mt-2 w-full">
 							<div
 								class="overflow-hidden font-mono text-xs leading-relaxed break-words whitespace-pre-wrap text-muted-foreground"
@@ -105,17 +137,21 @@
 		<div
 			class="flex h-8 w-8 items-center justify-center rounded bg-primary/10 text-xs font-medium text-primary"
 		>
-			{getFileTypeLabel(type)}
+			{fileTypeLabel}
 		</div>
 
-		<div class="flex flex-col gap-1">
+		<div class="flex flex-col gap-0.5">
 			<span
-				class="max-w-24 truncate text-sm font-medium text-foreground group-hover:pr-6 md:max-w-32"
+				class="max-w-24 truncate text-sm font-medium text-foreground {readonly
+					? ''
+					: 'group-hover:pr-6'} md:max-w-32"
 			>
 				{name}
 			</span>
 
-			{#if size}
+			{#if pdfProcessingMode}
+				<span class="text-left text-xs text-muted-foreground">{pdfProcessingMode}</span>
+			{:else if size}
 				<span class="text-left text-xs text-muted-foreground">{formatFileSize(size)}</span>
 			{/if}
 		</div>

tools/server/webui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentThumbnailImage.svelte

@@ -30,7 +30,9 @@
 	}: Props = $props();
 </script>
 
-<div class="group relative overflow-hidden rounded-lg border border-border bg-muted {className}">
+<div
+	class="group relative overflow-hidden rounded-lg bg-muted shadow-lg dark:border dark:border-muted {className}"
+>
 	{#if onClick}
 		<button
 			type="button"

tools/server/webui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsList.svelte

@@ -2,10 +2,8 @@
 	import { ChatAttachmentThumbnailImage, ChatAttachmentThumbnailFile } from '$lib/components/app';
 	import { Button } from '$lib/components/ui/button';
 	import { ChevronLeft, ChevronRight } from '@lucide/svelte';
-	import { FileTypeCategory } from '$lib/enums/files';
-	import { getFileTypeCategory } from '$lib/utils/file-type';
 	import { DialogChatAttachmentPreview, DialogChatAttachmentsViewAll } from '$lib/components/app';
-	import type { ChatAttachmentDisplayItem, ChatAttachmentPreviewItem } from '$lib/types/chat';
+	import { getAttachmentDisplayItems } from '$lib/utils';
 
 	interface Props {
 		class?: string;
@@ -22,6 +20,8 @@
 		imageWidth?: string;
 		// Limit display to single row with "+ X more" button
 		limitToSingleRow?: boolean;
+		// For vision modality check
+		activeModelId?: string;
 	}
 
 	let {
@@ -35,10 +35,11 @@
 		imageClass = '',
 		imageHeight = 'h-24',
 		imageWidth = 'w-auto',
-		limitToSingleRow = false
+		limitToSingleRow = false,
+		activeModelId
 	}: Props = $props();
 
-	let displayItems = $derived(getDisplayItems());
+	let displayItems = $derived(getAttachmentDisplayItems({ uploadedFiles, attachments }));
 
 	let canScrollLeft = $state(false);
 	let canScrollRight = $state(false);
@@ -49,81 +50,6 @@
 	let showViewAll = $derived(limitToSingleRow && displayItems.length > 0 && isScrollable);
 	let viewAllDialogOpen = $state(false);
 
-	function getDisplayItems(): ChatAttachmentDisplayItem[] {
-		const items: ChatAttachmentDisplayItem[] = [];
-
-		// Add uploaded files (ChatForm)
-		for (const file of uploadedFiles) {
-			items.push({
-				id: file.id,
-				name: file.name,
-				size: file.size,
-				preview: file.preview,
-				type: file.type,
-				isImage: getFileTypeCategory(file.type) === FileTypeCategory.IMAGE,
-				uploadedFile: file,
-				textContent: file.textContent
-			});
-		}
-
-		// Add stored attachments (ChatMessage)
-		for (const [index, attachment] of attachments.entries()) {
-			if (attachment.type === 'imageFile') {
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					preview: attachment.base64Url,
-					type: 'image',
-					isImage: true,
-					attachment,
-					attachmentIndex: index
-				});
-			} else if (attachment.type === 'textFile') {
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					type: 'text',
-					isImage: false,
-					attachment,
-					attachmentIndex: index,
-					textContent: attachment.content
-				});
-			} else if (attachment.type === 'context') {
-				// Legacy format from old webui - treat as text file
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					type: 'text',
-					isImage: false,
-					attachment,
-					attachmentIndex: index,
-					textContent: attachment.content
-				});
-			} else if (attachment.type === 'audioFile') {
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					type: attachment.mimeType || 'audio',
-					isImage: false,
-					attachment,
-					attachmentIndex: index
-				});
-			} else if (attachment.type === 'pdfFile') {
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					type: 'application/pdf',
-					isImage: false,
-					attachment,
-					attachmentIndex: index,
-					textContent: attachment.content
-				});
-			}
-		}
-
-		return items.reverse();
-	}
-
 	function openPreview(item: ChatAttachmentDisplayItem, event?: MouseEvent) {
 		event?.stopPropagation();
 		event?.preventDefault();
@@ -133,7 +59,6 @@
 			attachment: item.attachment,
 			preview: item.preview,
 			name: item.name,
-			type: item.type,
 			size: item.size,
 			textContent: item.textContent
 		};
@@ -181,26 +106,88 @@
 
 {#if displayItems.length > 0}
 	<div class={className} {style}>
-		<div class="relative">
-			<button
-				class="absolute top-1/2 left-4 z-10 flex h-6 w-6 -translate-y-1/2 items-center justify-center rounded-full bg-foreground/15 shadow-md backdrop-blur-xs transition-opacity hover:bg-foreground/35 {canScrollLeft
-					? 'opacity-100'
-					: 'pointer-events-none opacity-0'}"
-				onclick={scrollLeft}
-				aria-label="Scroll left"
-			>
-				<ChevronLeft class="h-4 w-4" />
-			</button>
+		{#if limitToSingleRow}
+			<div class="relative">
+				<button
+					class="absolute top-1/2 left-4 z-10 flex h-6 w-6 -translate-y-1/2 items-center justify-center rounded-full bg-foreground/15 shadow-md backdrop-blur-xs transition-opacity hover:bg-foreground/35 {canScrollLeft
+						? 'opacity-100'
+						: 'pointer-events-none opacity-0'}"
+					onclick={scrollLeft}
+					aria-label="Scroll left"
+				>
+					<ChevronLeft class="h-4 w-4" />
+				</button>
 
-			<div
-				class="scrollbar-hide flex items-start gap-3 overflow-x-auto"
-				bind:this={scrollContainer}
-				onscroll={updateScrollButtons}
-			>
+				<div
+					class="scrollbar-hide flex items-start gap-3 overflow-x-auto"
+					bind:this={scrollContainer}
+					onscroll={updateScrollButtons}
+				>
+					{#each displayItems as item (item.id)}
+						{#if item.isImage && item.preview}
+							<ChatAttachmentThumbnailImage
+								class="flex-shrink-0 cursor-pointer {limitToSingleRow
+									? 'first:ml-4 last:mr-4'
+									: ''}"
+								id={item.id}
+								name={item.name}
+								preview={item.preview}
+								{readonly}
+								onRemove={onFileRemove}
+								height={imageHeight}
+								width={imageWidth}
+								{imageClass}
+								onClick={(event) => openPreview(item, event)}
+							/>
+						{:else}
+							<ChatAttachmentThumbnailFile
+								class="flex-shrink-0 cursor-pointer {limitToSingleRow
+									? 'first:ml-4 last:mr-4'
+									: ''}"
+								id={item.id}
+								name={item.name}
+								size={item.size}
+								{readonly}
+								onRemove={onFileRemove}
+								textContent={item.textContent}
+								attachment={item.attachment}
+								uploadedFile={item.uploadedFile}
+								onClick={(event) => openPreview(item, event)}
+							/>
+						{/if}
+					{/each}
+				</div>
+
+				<button
+					class="absolute top-1/2 right-4 z-10 flex h-6 w-6 -translate-y-1/2 items-center justify-center rounded-full bg-foreground/15 shadow-md backdrop-blur-xs transition-opacity hover:bg-foreground/35 {canScrollRight
+						? 'opacity-100'
+						: 'pointer-events-none opacity-0'}"
+					onclick={scrollRight}
+					aria-label="Scroll right"
+				>
+					<ChevronRight class="h-4 w-4" />
+				</button>
+			</div>
+
+			{#if showViewAll}
+				<div class="mt-2 -mr-2 flex justify-end px-4">
+					<Button
+						type="button"
+						variant="ghost"
+						size="sm"
+						class="h-6 text-xs text-muted-foreground hover:text-foreground"
+						onclick={() => (viewAllDialogOpen = true)}
+					>
+						View all ({displayItems.length})
+					</Button>
+				</div>
+			{/if}
+		{:else}
+			<div class="flex flex-wrap items-start justify-end gap-3">
 				{#each displayItems as item (item.id)}
 					{#if item.isImage && item.preview}
 						<ChatAttachmentThumbnailImage
-							class="flex-shrink-0 cursor-pointer {limitToSingleRow ? 'first:ml-4 last:mr-4' : ''}"
+							class="cursor-pointer"
 							id={item.id}
 							name={item.name}
 							preview={item.preview}
@@ -213,43 +200,20 @@
 						/>
 					{:else}
 						<ChatAttachmentThumbnailFile
-							class="flex-shrink-0 cursor-pointer {limitToSingleRow ? 'first:ml-4 last:mr-4' : ''}"
+							class="cursor-pointer"
 							id={item.id}
 							name={item.name}
-							type={item.type}
 							size={item.size}
 							{readonly}
 							onRemove={onFileRemove}
 							textContent={item.textContent}
-							onClick={(event) => openPreview(item, event)}
+							attachment={item.attachment}
+							uploadedFile={item.uploadedFile}
+							onClick={(event?: MouseEvent) => openPreview(item, event)}
 						/>
 					{/if}
 				{/each}
 			</div>
-
-			<button
-				class="absolute top-1/2 right-4 z-10 flex h-6 w-6 -translate-y-1/2 items-center justify-center rounded-full bg-foreground/15 shadow-md backdrop-blur-xs transition-opacity hover:bg-foreground/35 {canScrollRight
-					? 'opacity-100'
-					: 'pointer-events-none opacity-0'}"
-				onclick={scrollRight}
-				aria-label="Scroll right"
-			>
-				<ChevronRight class="h-4 w-4" />
-			</button>
-		</div>
-
-		{#if showViewAll}
-			<div class="mt-2 -mr-2 flex justify-end px-4">
-				<Button
-					type="button"
-					variant="ghost"
-					size="sm"
-					class="h-6 text-xs text-muted-foreground hover:text-foreground"
-					onclick={() => (viewAllDialogOpen = true)}
-				>
-					View all
-				</Button>
-			</div>
 		{/if}
 	</div>
 {/if}
@@ -261,9 +225,9 @@
 		attachment={previewItem.attachment}
 		preview={previewItem.preview}
 		name={previewItem.name}
-		type={previewItem.type}
 		size={previewItem.size}
 		textContent={previewItem.textContent}
+		{activeModelId}
 	/>
 {/if}
 
@@ -275,4 +239,5 @@
 	{onFileRemove}
 	imageHeight="h-64"
 	{imageClass}
+	{activeModelId}
 />

tools/server/webui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsViewAll.svelte

@@ -4,9 +4,7 @@
 		ChatAttachmentThumbnailFile,
 		DialogChatAttachmentPreview
 	} from '$lib/components/app';
-	import { FileTypeCategory } from '$lib/enums/files';
-	import { getFileTypeCategory } from '$lib/utils/file-type';
-	import type { ChatAttachmentDisplayItem, ChatAttachmentPreviewItem } from '$lib/types/chat';
+	import { getAttachmentDisplayItems } from '$lib/utils';
 
 	interface Props {
 		uploadedFiles?: ChatUploadedFile[];
@@ -16,6 +14,7 @@
 		imageHeight?: string;
 		imageWidth?: string;
 		imageClass?: string;
+		activeModelId?: string;
 	}
 
 	let {
@@ -25,89 +24,17 @@
 		onFileRemove,
 		imageHeight = 'h-24',
 		imageWidth = 'w-auto',
-		imageClass = ''
+		imageClass = '',
+		activeModelId
 	}: Props = $props();
 
 	let previewDialogOpen = $state(false);
 	let previewItem = $state<ChatAttachmentPreviewItem | null>(null);
 
-	let displayItems = $derived(getDisplayItems());
+	let displayItems = $derived(getAttachmentDisplayItems({ uploadedFiles, attachments }));
 	let imageItems = $derived(displayItems.filter((item) => item.isImage));
 	let fileItems = $derived(displayItems.filter((item) => !item.isImage));
 
-	function getDisplayItems(): ChatAttachmentDisplayItem[] {
-		const items: ChatAttachmentDisplayItem[] = [];
-
-		for (const file of uploadedFiles) {
-			items.push({
-				id: file.id,
-				name: file.name,
-				size: file.size,
-				preview: file.preview,
-				type: file.type,
-				isImage: getFileTypeCategory(file.type) === FileTypeCategory.IMAGE,
-				uploadedFile: file,
-				textContent: file.textContent
-			});
-		}
-
-		for (const [index, attachment] of attachments.entries()) {
-			if (attachment.type === 'imageFile') {
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					preview: attachment.base64Url,
-					type: 'image',
-					isImage: true,
-					attachment,
-					attachmentIndex: index
-				});
-			} else if (attachment.type === 'textFile') {
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					type: 'text',
-					isImage: false,
-					attachment,
-					attachmentIndex: index,
-					textContent: attachment.content
-				});
-			} else if (attachment.type === 'context') {
-				// Legacy format from old webui - treat as text file
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					type: 'text',
-					isImage: false,
-					attachment,
-					attachmentIndex: index,
-					textContent: attachment.content
-				});
-			} else if (attachment.type === 'audioFile') {
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					type: attachment.mimeType || 'audio',
-					isImage: false,
-					attachment,
-					attachmentIndex: index
-				});
-			} else if (attachment.type === 'pdfFile') {
-				items.push({
-					id: `attachment-${index}`,
-					name: attachment.name,
-					type: 'application/pdf',
-					isImage: false,
-					attachment,
-					attachmentIndex: index,
-					textContent: attachment.content
-				});
-			}
-		}
-
-		return items.reverse();
-	}
-
 	function openPreview(item: (typeof displayItems)[0], event?: Event) {
 		if (event) {
 			event.preventDefault();
@@ -119,7 +46,6 @@
 			attachment: item.attachment,
 			preview: item.preview,
 			name: item.name,
-			type: item.type,
 			size: item.size,
 			textContent: item.textContent
 		};
@@ -138,12 +64,13 @@
 							class="cursor-pointer"
 							id={item.id}
 							name={item.name}
-							type={item.type}
 							size={item.size}
 							{readonly}
 							onRemove={onFileRemove}
 							textContent={item.textContent}
-							onClick={(event) => openPreview(item, event)}
+							attachment={item.attachment}
+							uploadedFile={item.uploadedFile}
+							onClick={(event?: MouseEvent) => openPreview(item, event)}
 						/>
 					{/each}
 				</div>
@@ -183,8 +110,8 @@
 		attachment={previewItem.attachment}
 		preview={previewItem.preview}
 		name={previewItem.name}
-		type={previewItem.type}
 		size={previewItem.size}
 		textContent={previewItem.textContent}
+		{activeModelId}
 	/>
 {/if}

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatForm.svelte

@@ -9,15 +9,13 @@
 	} from '$lib/components/app';
 	import { INPUT_CLASSES } from '$lib/constants/input-classes';
 	import { config } from '$lib/stores/settings.svelte';
-	import { FileTypeCategory, MimeTypeApplication } from '$lib/enums/files';
-	import {
-		AudioRecorder,
-		convertToWav,
-		createAudioFile,
-		isAudioRecordingSupported
-	} from '$lib/utils/audio-recording';
-	import { onMount } from 'svelte';
+	import { modelsStore, modelOptions, selectedModelId } from '$lib/stores/models.svelte';
+	import { isRouterMode } from '$lib/stores/server.svelte';
+	import { chatStore } from '$lib/stores/chat.svelte';
+	import { activeMessages } from '$lib/stores/conversations.svelte';
 	import {
+		FileTypeCategory,
+		MimeTypeApplication,
 		FileExtensionAudio,
 		FileExtensionImage,
 		FileExtensionPdf,
@@ -25,8 +23,15 @@
 		MimeTypeAudio,
 		MimeTypeImage,
 		MimeTypeText
-	} from '$lib/enums/files';
-	import { isIMEComposing } from '$lib/utils/is-ime-composing';
+	} from '$lib/enums';
+	import { isIMEComposing } from '$lib/utils';
+	import {
+		AudioRecorder,
+		convertToWav,
+		createAudioFile,
+		isAudioRecordingSupported
+	} from '$lib/utils/browser-only';
+	import { onMount } from 'svelte';
 
 	interface Props {
 		class?: string;
@@ -53,6 +58,7 @@
 	}: Props = $props();
 
 	let audioRecorder: AudioRecorder | undefined;
+	let chatFormActionsRef: ChatFormActions | undefined = $state(undefined);
 	let currentConfig = $derived(config());
 	let fileAcceptString = $state<string | undefined>(undefined);
 	let fileInputRef: ChatFormFileInputInvisible | undefined = $state(undefined);
@@ -63,18 +69,97 @@
 	let recordingSupported = $state(false);
 	let textareaRef: ChatFormTextarea | undefined = $state(undefined);
 
+	// Check if model is selected (in ROUTER mode)
+	let conversationModel = $derived(
+		chatStore.getConversationModel(activeMessages() as DatabaseMessage[])
+	);
+	let isRouter = $derived(isRouterMode());
+	let hasModelSelected = $derived(!isRouter || !!conversationModel || !!selectedModelId());
+
+	// Get active model ID for capability detection
+	let activeModelId = $derived.by(() => {
+		const options = modelOptions();
+
+		if (!isRouter) {
+			return options.length > 0 ? options[0].model : null;
+		}
+
+		// First try user-selected model
+		const selectedId = selectedModelId();
+		if (selectedId) {
+			const model = options.find((m) => m.id === selectedId);
+			if (model) return model.model;
+		}
+
+		// Fallback to conversation model
+		if (conversationModel) {
+			const model = options.find((m) => m.model === conversationModel);
+			if (model) return model.model;
+		}
+
+		return null;
+	});
+
+	// State for model props reactivity
+	let modelPropsVersion = $state(0);
+
+	// Fetch model props when active model changes (works for both MODEL and ROUTER mode)
+	$effect(() => {
+		if (activeModelId) {
+			const cached = modelsStore.getModelProps(activeModelId);
+			if (!cached) {
+				modelsStore.fetchModelProps(activeModelId).then(() => {
+					modelPropsVersion++;
+				});
+			}
+		}
+	});
+
+	// Derive modalities from active model (works for both MODEL and ROUTER mode)
+	let hasAudioModality = $derived.by(() => {
+		if (activeModelId) {
+			void modelPropsVersion; // Trigger reactivity on props fetch
+			return modelsStore.modelSupportsAudio(activeModelId);
+		}
+
+		return false;
+	});
+
+	let hasVisionModality = $derived.by(() => {
+		if (activeModelId) {
+			void modelPropsVersion; // Trigger reactivity on props fetch
+			return modelsStore.modelSupportsVision(activeModelId);
+		}
+
+		return false;
+	});
+
+	function checkModelSelected(): boolean {
+		if (!hasModelSelected) {
+			// Open the model selector
+			chatFormActionsRef?.openModelSelector();
+			return false;
+		}
+
+		return true;
+	}
+
 	function getAcceptStringForFileType(fileType: FileTypeCategory): string {
 		switch (fileType) {
 			case FileTypeCategory.IMAGE:
 				return [...Object.values(FileExtensionImage), ...Object.values(MimeTypeImage)].join(',');
+
 			case FileTypeCategory.AUDIO:
 				return [...Object.values(FileExtensionAudio), ...Object.values(MimeTypeAudio)].join(',');
+
 			case FileTypeCategory.PDF:
 				return [...Object.values(FileExtensionPdf), ...Object.values(MimeTypeApplication)].join(
 					','
 				);
+
 			case FileTypeCategory.TEXT:
 				return [...Object.values(FileExtensionText), MimeTypeText.PLAIN].join(',');
+
 			default:
 				return '';
 		}
@@ -103,6 +188,9 @@
 
 			if ((!message.trim() && uploadedFiles.length === 0) || disabled || isLoading) return;
 
+			// Check if model is selected first
+			if (!checkModelSelected()) return;
+
 			const messageToSend = message.trim();
 			const filesToSend = [...uploadedFiles];
 
@@ -131,6 +219,7 @@
 		if (files.length > 0) {
 			event.preventDefault();
 			onFileUpload?.(files);
+
 			return;
 		}
 
@@ -154,6 +243,7 @@
 	async function handleMicClick() {
 		if (!audioRecorder || !recordingSupported) {
 			console.warn('Audio recording not supported');
+
 			return;
 		}
 
@@ -187,6 +277,9 @@
 		event.preventDefault();
 		if ((!message.trim() && uploadedFiles.length === 0) || disabled || isLoading) return;
 
+		// Check if model is selected first
+		if (!checkModelSelected()) return;
+
 		const messageToSend = message.trim();
 		const filesToSend = [...uploadedFiles];
 
@@ -225,12 +318,16 @@
 <ChatFormFileInputInvisible
 	bind:this={fileInputRef}
 	bind:accept={fileAcceptString}
+	{hasAudioModality}
+	{hasVisionModality}
 	onFileSelect={handleFileSelect}
 />
 
 <form
 	onsubmit={handleSubmit}
-	class="{INPUT_CLASSES} border-radius-bottom-none mx-auto max-w-[48rem] overflow-hidden rounded-3xl backdrop-blur-md {className}"
+	class="{INPUT_CLASSES} border-radius-bottom-none mx-auto max-w-[48rem] overflow-hidden rounded-3xl backdrop-blur-md {disabled
+		? 'cursor-not-allowed opacity-60'
+		: ''} {className}"
 >
 	<ChatAttachmentsList
 		bind:uploadedFiles
@@ -238,6 +335,7 @@
 		limitToSingleRow
 		class="py-5"
 		style="scroll-padding: 1rem;"
+		activeModelId={activeModelId ?? undefined}
 	/>
 
 	<div
@@ -252,10 +350,13 @@
 		/>
 
 		<ChatFormActions
+			bind:this={chatFormActionsRef}
 			canSend={message.trim().length > 0 || uploadedFiles.length > 0}
+			hasText={message.trim().length > 0}
 			{disabled}
 			{isLoading}
 			{isRecording}
+			{uploadedFiles}
 			onFileUpload={handleFileUpload}
 			onMicClick={handleMicClick}
 			onStop={handleStop}

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionFileAttachments.svelte

@@ -1,22 +1,29 @@
 <script lang="ts">
-	import { Paperclip, Image, FileText, File, Volume2 } from '@lucide/svelte';
+	import { Paperclip } from '@lucide/svelte';
 	import { Button } from '$lib/components/ui/button';
 	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
 	import * as Tooltip from '$lib/components/ui/tooltip';
-	import { TOOLTIP_DELAY_DURATION } from '$lib/constants/tooltip-config';
-	import { FileTypeCategory } from '$lib/enums/files';
-	import { supportsAudio, supportsVision } from '$lib/stores/server.svelte';
+	import { FILE_TYPE_ICONS } from '$lib/constants/icons';
+	import { FileTypeCategory } from '$lib/enums';
 
 	interface Props {
 		class?: string;
 		disabled?: boolean;
+		hasAudioModality?: boolean;
+		hasVisionModality?: boolean;
 		onFileUpload?: (fileType?: FileTypeCategory) => void;
 	}
 
-	let { class: className = '', disabled = false, onFileUpload }: Props = $props();
+	let {
+		class: className = '',
+		disabled = false,
+		hasAudioModality = false,
+		hasVisionModality = false,
+		onFileUpload
+	}: Props = $props();
 
 	const fileUploadTooltipText = $derived.by(() => {
-		return !supportsVision()
+		return !hasVisionModality
 			? 'Text files and PDFs supported. Images, audio, and video require vision models.'
 			: 'Attach files';
 	});
@@ -29,7 +36,7 @@
 <div class="flex items-center gap-1 {className}">
 	<DropdownMenu.Root>
 		<DropdownMenu.Trigger name="Attach files">
-			<Tooltip.Root delayDuration={TOOLTIP_DELAY_DURATION}>
+			<Tooltip.Root>
 				<Tooltip.Trigger>
 					<Button
 						class="file-upload-button h-8 w-8 rounded-full bg-transparent p-0 text-muted-foreground hover:bg-foreground/10 hover:text-foreground"
@@ -49,40 +56,40 @@
 		</DropdownMenu.Trigger>
 
 		<DropdownMenu.Content align="start" class="w-48">
-			<Tooltip.Root delayDuration={TOOLTIP_DELAY_DURATION}>
+			<Tooltip.Root>
 				<Tooltip.Trigger class="w-full">
 					<DropdownMenu.Item
 						class="images-button flex cursor-pointer items-center gap-2"
-						disabled={!supportsVision()}
+						disabled={!hasVisionModality}
 						onclick={() => handleFileUpload(FileTypeCategory.IMAGE)}
 					>
-						<Image class="h-4 w-4" />
+						<FILE_TYPE_ICONS.image class="h-4 w-4" />
 
 						<span>Images</span>
 					</DropdownMenu.Item>
 				</Tooltip.Trigger>
 
-				{#if !supportsVision()}
+				{#if !hasVisionModality}
 					<Tooltip.Content>
 						<p>Images require vision models to be processed</p>
 					</Tooltip.Content>
 				{/if}
 			</Tooltip.Root>
 
-			<Tooltip.Root delayDuration={TOOLTIP_DELAY_DURATION}>
+			<Tooltip.Root>
 				<Tooltip.Trigger class="w-full">
 					<DropdownMenu.Item
 						class="audio-button flex cursor-pointer items-center gap-2"
-						disabled={!supportsAudio()}
+						disabled={!hasAudioModality}
 						onclick={() => handleFileUpload(FileTypeCategory.AUDIO)}
 					>
-						<Volume2 class="h-4 w-4" />
+						<FILE_TYPE_ICONS.audio class="h-4 w-4" />
 
 						<span>Audio Files</span>
 					</DropdownMenu.Item>
 				</Tooltip.Trigger>
 
-				{#if !supportsAudio()}
+				{#if !hasAudioModality}
 					<Tooltip.Content>
 						<p>Audio files require audio models to be processed</p>
 					</Tooltip.Content>
@@ -93,24 +100,24 @@
 				class="flex cursor-pointer items-center gap-2"
 				onclick={() => handleFileUpload(FileTypeCategory.TEXT)}
 			>
-				<FileText class="h-4 w-4" />
+				<FILE_TYPE_ICONS.text class="h-4 w-4" />
 
 				<span>Text Files</span>
 			</DropdownMenu.Item>
 
-			<Tooltip.Root delayDuration={TOOLTIP_DELAY_DURATION}>
+			<Tooltip.Root>
 				<Tooltip.Trigger class="w-full">
 					<DropdownMenu.Item
 						class="flex cursor-pointer items-center gap-2"
 						onclick={() => handleFileUpload(FileTypeCategory.PDF)}
 					>
-						<File class="h-4 w-4" />
+						<FILE_TYPE_ICONS.pdf class="h-4 w-4" />
 
 						<span>PDF Files</span>
 					</DropdownMenu.Item>
 				</Tooltip.Trigger>
 
-				{#if !supportsVision()}
+				{#if !hasVisionModality}
 					<Tooltip.Content>
 						<p>PDFs will be converted to text. Image-based PDFs may not work properly.</p>
 					</Tooltip.Content>

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionRecord.svelte

@@ -1,12 +1,12 @@
 <script lang="ts">
-	import { Mic } from '@lucide/svelte';
+	import { Mic, Square } from '@lucide/svelte';
 	import { Button } from '$lib/components/ui/button';
 	import * as Tooltip from '$lib/components/ui/tooltip';
-	import { supportsAudio } from '$lib/stores/server.svelte';
 
 	interface Props {
 		class?: string;
 		disabled?: boolean;
+		hasAudioModality?: boolean;
 		isLoading?: boolean;
 		isRecording?: boolean;
 		onMicClick?: () => void;
@@ -15,6 +15,7 @@
 	let {
 		class: className = '',
 		disabled = false,
+		hasAudioModality = false,
 		isLoading = false,
 		isRecording = false,
 		onMicClick
@@ -22,25 +23,27 @@
 </script>
 
 <div class="flex items-center gap-1 {className}">
-	<Tooltip.Root delayDuration={100}>
+	<Tooltip.Root>
 		<Tooltip.Trigger>
 			<Button
 				class="h-8 w-8 rounded-full p-0 {isRecording
 					? 'animate-pulse bg-red-500 text-white hover:bg-red-600'
-					: 'bg-transparent text-muted-foreground hover:bg-foreground/10 hover:text-foreground'} {!supportsAudio()
-					? 'cursor-not-allowed opacity-50'
 					: ''}"
-				disabled={disabled || isLoading || !supportsAudio()}
+				disabled={disabled || isLoading || !hasAudioModality}
 				onclick={onMicClick}
 				type="button"
 			>
 				<span class="sr-only">{isRecording ? 'Stop recording' : 'Start recording'}</span>
 
-				<Mic class="h-4 w-4" />
+				{#if isRecording}
+					<Square class="h-4 w-4 animate-pulse fill-white" />
+				{:else}
+					<Mic class="h-4 w-4" />
+				{/if}
 			</Button>
 		</Tooltip.Trigger>
 
-		{#if !supportsAudio()}
+		{#if !hasAudioModality}
 			<Tooltip.Content>
 				<p>Current model does not support audio</p>
 			</Tooltip.Content>

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionSubmit.svelte

@@ -0,0 +1,55 @@
+<script lang="ts">
+	import { ArrowUp } from '@lucide/svelte';
+	import { Button } from '$lib/components/ui/button';
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { cn } from '$lib/components/ui/utils';
+
+	interface Props {
+		canSend?: boolean;
+		disabled?: boolean;
+		isLoading?: boolean;
+		showErrorState?: boolean;
+		tooltipLabel?: string;
+	}
+
+	let {
+		canSend = false,
+		disabled = false,
+		isLoading = false,
+		showErrorState = false,
+		tooltipLabel
+	}: Props = $props();
+
+	let isDisabled = $derived(!canSend || disabled || isLoading);
+</script>
+
+{#snippet submitButton(props = {})}
+	<Button
+		type="submit"
+		disabled={isDisabled}
+		class={cn(
+			'h-8 w-8 rounded-full p-0',
+			showErrorState
+				? 'bg-red-400/10 text-red-400 hover:bg-red-400/20 hover:text-red-400 disabled:opacity-100'
+				: ''
+		)}
+		{...props}
+	>
+		<span class="sr-only">Send</span>
+		<ArrowUp class="h-12 w-12" />
+	</Button>
+{/snippet}
+
+{#if tooltipLabel}
+	<Tooltip.Root>
+		<Tooltip.Trigger>
+			{@render submitButton()}
+		</Tooltip.Trigger>
+
+		<Tooltip.Content>
+			<p>{tooltipLabel}</p>
+		</Tooltip.Content>
+	</Tooltip.Root>
+{:else}
+	{@render submitButton()}
+{/if}

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte

@@ -1,13 +1,20 @@
 <script lang="ts">
-	import { Square, ArrowUp } from '@lucide/svelte';
+	import { Square } from '@lucide/svelte';
 	import { Button } from '$lib/components/ui/button';
 	import {
 		ChatFormActionFileAttachments,
 		ChatFormActionRecord,
-		ChatFormModelSelector
+		ChatFormActionSubmit,
+		ModelsSelector
 	} from '$lib/components/app';
+	import { FileTypeCategory } from '$lib/enums';
+	import { getFileTypeCategory } from '$lib/utils';
 	import { config } from '$lib/stores/settings.svelte';
-	import type { FileTypeCategory } from '$lib/enums/files';
+	import { modelsStore, modelOptions, selectedModelId } from '$lib/stores/models.svelte';
+	import { isRouterMode } from '$lib/stores/server.svelte';
+	import { chatStore } from '$lib/stores/chat.svelte';
+	import { activeMessages, usedModalities } from '$lib/stores/conversations.svelte';
+	import { useModelChangeValidation } from '$lib/hooks/use-model-change-validation.svelte';
 
 	interface Props {
 		canSend?: boolean;
@@ -15,6 +22,8 @@
 		disabled?: boolean;
 		isLoading?: boolean;
 		isRecording?: boolean;
+		hasText?: boolean;
+		uploadedFiles?: ChatUploadedFile[];
 		onFileUpload?: (fileType?: FileTypeCategory) => void;
 		onMicClick?: () => void;
 		onStop?: () => void;
@@ -26,20 +35,150 @@
 		disabled = false,
 		isLoading = false,
 		isRecording = false,
+		hasText = false,
+		uploadedFiles = [],
 		onFileUpload,
 		onMicClick,
 		onStop
 	}: Props = $props();
 
 	let currentConfig = $derived(config());
+	let isRouter = $derived(isRouterMode());
+
+	let conversationModel = $derived(
+		chatStore.getConversationModel(activeMessages() as DatabaseMessage[])
+	);
+
+	let previousConversationModel: string | null = null;
+
+	$effect(() => {
+		if (conversationModel && conversationModel !== previousConversationModel) {
+			previousConversationModel = conversationModel;
+			modelsStore.selectModelByName(conversationModel);
+		}
+	});
+
+	let activeModelId = $derived.by(() => {
+		const options = modelOptions();
+
+		if (!isRouter) {
+			return options.length > 0 ? options[0].model : null;
+		}
+
+		const selectedId = selectedModelId();
+		if (selectedId) {
+			const model = options.find((m) => m.id === selectedId);
+			if (model) return model.model;
+		}
+
+		if (conversationModel) {
+			const model = options.find((m) => m.model === conversationModel);
+			if (model) return model.model;
+		}
+
+		return null;
+	});
+
+	let modelPropsVersion = $state(0); // Used to trigger reactivity after fetch
+
+	$effect(() => {
+		if (activeModelId) {
+			const cached = modelsStore.getModelProps(activeModelId);
+
+			if (!cached) {
+				modelsStore.fetchModelProps(activeModelId).then(() => {
+					modelPropsVersion++;
+				});
+			}
+		}
+	});
+
+	let hasAudioModality = $derived.by(() => {
+		if (activeModelId) {
+			void modelPropsVersion;
+
+			return modelsStore.modelSupportsAudio(activeModelId);
+		}
+
+		return false;
+	});
+
+	let hasVisionModality = $derived.by(() => {
+		if (activeModelId) {
+			void modelPropsVersion;
+
+			return modelsStore.modelSupportsVision(activeModelId);
+		}
+
+		return false;
+	});
+
+	let hasAudioAttachments = $derived(
+		uploadedFiles.some((file) => getFileTypeCategory(file.type) === FileTypeCategory.AUDIO)
+	);
+	let shouldShowRecordButton = $derived(
+		hasAudioModality && !hasText && !hasAudioAttachments && currentConfig.autoMicOnEmpty
+	);
+
+	let hasModelSelected = $derived(!isRouter || !!conversationModel || !!selectedModelId());
+
+	let isSelectedModelInCache = $derived.by(() => {
+		if (!isRouter) return true;
+
+		if (conversationModel) {
+			return modelOptions().some((option) => option.model === conversationModel);
+		}
+
+		const currentModelId = selectedModelId();
+		if (!currentModelId) return false;
+
+		return modelOptions().some((option) => option.id === currentModelId);
+	});
+
+	let submitTooltip = $derived.by(() => {
+		if (!hasModelSelected) {
+			return 'Please select a model first';
+		}
+
+		if (!isSelectedModelInCache) {
+			return 'Selected model is not available, please select another';
+		}
+
+		return '';
+	});
+
+	let selectorModelRef: ModelsSelector | undefined = $state(undefined);
+
+	export function openModelSelector() {
+		selectorModelRef?.open();
+	}
+
+	const { handleModelChange } = useModelChangeValidation({
+		getRequiredModalities: () => usedModalities(),
+		onValidationFailure: async (previousModelId) => {
+			if (previousModelId) {
+				await modelsStore.selectModelById(previousModelId);
+			}
+		}
+	});
 </script>
 
-<div class="flex w-full items-center gap-2 {className}">
-	<ChatFormActionFileAttachments class="mr-auto" {disabled} {onFileUpload} />
+<div class="flex w-full items-center gap-3 {className}" style="container-type: inline-size">
+	<ChatFormActionFileAttachments
+		class="mr-auto"
+		{disabled}
+		{hasAudioModality}
+		{hasVisionModality}
+		{onFileUpload}
+	/>
 
-	{#if currentConfig.modelSelectorEnabled}
-		<ChatFormModelSelector class="shrink-0" />
-	{/if}
+	<ModelsSelector
+		bind:this={selectorModelRef}
+		currentModel={conversationModel}
+		forceForegroundText={true}
+		useGlobalSelection={true}
+		onModelChange={handleModelChange}
+	/>
 
 	{#if isLoading}
 		<Button
@@ -50,16 +189,15 @@
 			<span class="sr-only">Stop</span>
 			<Square class="h-8 w-8 fill-destructive stroke-destructive" />
 		</Button>
+	{:else if shouldShowRecordButton}
+		<ChatFormActionRecord {disabled} {hasAudioModality} {isLoading} {isRecording} {onMicClick} />
 	{:else}
-		<ChatFormActionRecord {disabled} {isLoading} {isRecording} {onMicClick} />
-
-		<Button
-			type="submit"
-			disabled={!canSend || disabled || isLoading}
-			class="h-8 w-8 rounded-full p-0"
-		>
-			<span class="sr-only">Send</span>
-			<ArrowUp class="h-12 w-12" />
-		</Button>
+		<ChatFormActionSubmit
+			canSend={canSend && hasModelSelected && isSelectedModelInCache}
+			{disabled}
+			{isLoading}
+			tooltipLabel={submitTooltip}
+			showErrorState={hasModelSelected && !isSelectedModelInCache}
+		/>
 	{/if}
 </div>

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormFileInputInvisible.svelte

@@ -1,9 +1,11 @@
 <script lang="ts">
-	import { generateModalityAwareAcceptString } from '$lib/utils/modality-file-validation';
+	import { generateModalityAwareAcceptString } from '$lib/utils';
 
 	interface Props {
 		accept?: string;
 		class?: string;
+		hasAudioModality?: boolean;
+		hasVisionModality?: boolean;
 		multiple?: boolean;
 		onFileSelect?: (files: File[]) => void;
 	}
@@ -11,6 +13,8 @@
 	let {
 		accept = $bindable(),
 		class: className = '',
+		hasAudioModality = false,
+		hasVisionModality = false,
 		multiple = true,
 		onFileSelect
 	}: Props = $props();
@@ -18,7 +22,13 @@
 	let fileInputElement: HTMLInputElement | undefined;
 
 	// Use modality-aware accept string by default, but allow override
-	let finalAccept = $derived(accept ?? generateModalityAwareAcceptString());
+	let finalAccept = $derived(
+		accept ??
+			generateModalityAwareAcceptString({
+				hasVision: hasVisionModality,
+				hasAudio: hasAudioModality
+			})
+	);
 
 	export function click() {
 		fileInputElement?.click();

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormModelSelector.svelte

@@ -1,352 +0,0 @@
-<script lang="ts">
-	import { onMount, tick } from 'svelte';
-	import { ChevronDown, Loader2 } from '@lucide/svelte';
-	import { cn } from '$lib/components/ui/utils';
-	import { portalToBody } from '$lib/utils/portal-to-body';
-	import {
-		fetchModels,
-		modelOptions,
-		modelsError,
-		modelsLoading,
-		modelsUpdating,
-		selectModel,
-		selectedModelId
-	} from '$lib/stores/models.svelte';
-	import type { ModelOption } from '$lib/types/models';
-
-	interface Props {
-		class?: string;
-	}
-
-	let { class: className = '' }: Props = $props();
-
-	let options = $derived(modelOptions());
-	let loading = $derived(modelsLoading());
-	let updating = $derived(modelsUpdating());
-	let error = $derived(modelsError());
-	let activeId = $derived(selectedModelId());
-
-	let isMounted = $state(false);
-	let isOpen = $state(false);
-	let container: HTMLDivElement | null = null;
-	let triggerButton = $state<HTMLButtonElement | null>(null);
-	let menuRef = $state<HTMLDivElement | null>(null);
-	let menuPosition = $state<{
-		top: number;
-		left: number;
-		width: number;
-		placement: 'top' | 'bottom';
-		maxHeight: number;
-	} | null>(null);
-	let lockedWidth: number | null = null;
-
-	onMount(async () => {
-		try {
-			await fetchModels();
-		} catch (error) {
-			console.error('Unable to load models:', error);
-		} finally {
-			isMounted = true;
-		}
-	});
-
-	function handlePointerDown(event: PointerEvent) {
-		if (!container) return;
-
-		const target = event.target as Node | null;
-
-		if (target && !container.contains(target) && !(menuRef && menuRef.contains(target))) {
-			closeMenu();
-		}
-	}
-
-	function handleKeydown(event: KeyboardEvent) {
-		if (event.key === 'Escape') {
-			closeMenu();
-		}
-	}
-
-	function handleResize() {
-		if (isOpen) {
-			updateMenuPosition();
-		}
-	}
-
-	async function handleSelect(value: string | undefined) {
-		if (!value) return;
-
-		const option = options.find((item) => item.id === value);
-		if (!option) {
-			console.error('Model is no longer available');
-			return;
-		}
-
-		try {
-			await selectModel(option.id);
-		} catch (error) {
-			console.error('Failed to switch model:', error);
-		}
-	}
-
-	const VIEWPORT_GUTTER = 8;
-	const MENU_OFFSET = 6;
-	const MENU_MAX_WIDTH = 320;
-
-	async function openMenu() {
-		if (loading || updating) return;
-
-		isOpen = true;
-		await tick();
-		updateMenuPosition();
-		requestAnimationFrame(() => updateMenuPosition());
-	}
-
-	function toggleOpen() {
-		if (loading || updating) return;
-
-		if (isOpen) {
-			closeMenu();
-		} else {
-			void openMenu();
-		}
-	}
-
-	function closeMenu() {
-		if (!isOpen) return;
-
-		isOpen = false;
-		menuPosition = null;
-		lockedWidth = null;
-	}
-
-	async function handleOptionSelect(optionId: string) {
-		try {
-			await handleSelect(optionId);
-		} finally {
-			closeMenu();
-		}
-	}
-
-	$effect(() => {
-		if (loading || updating) {
-			closeMenu();
-		}
-	});
-
-	$effect(() => {
-		const optionCount = options.length;
-
-		if (!isOpen || optionCount <= 0) return;
-
-		queueMicrotask(() => updateMenuPosition());
-	});
-
-	function updateMenuPosition() {
-		if (!isOpen || !triggerButton || !menuRef) return;
-
-		const triggerRect = triggerButton.getBoundingClientRect();
-		const viewportWidth = window.innerWidth;
-		const viewportHeight = window.innerHeight;
-
-		if (viewportWidth === 0 || viewportHeight === 0) return;
-
-		const scrollWidth = menuRef.scrollWidth;
-		const scrollHeight = menuRef.scrollHeight;
-
-		const availableWidth = Math.max(0, viewportWidth - VIEWPORT_GUTTER * 2);
-		const constrainedMaxWidth = Math.min(MENU_MAX_WIDTH, availableWidth || MENU_MAX_WIDTH);
-		const safeMaxWidth =
-			constrainedMaxWidth > 0 ? constrainedMaxWidth : Math.min(MENU_MAX_WIDTH, viewportWidth);
-		const desiredMinWidth = Math.min(160, safeMaxWidth || 160);
-
-		let width = lockedWidth;
-		if (width === null) {
-			const naturalWidth = Math.min(scrollWidth, safeMaxWidth);
-			const baseWidth = Math.max(triggerRect.width, naturalWidth, desiredMinWidth);
-			width = Math.min(baseWidth, safeMaxWidth || baseWidth);
-			lockedWidth = width;
-		} else {
-			width = Math.min(Math.max(width, desiredMinWidth), safeMaxWidth || width);
-		}
-
-		if (width > 0) {
-			menuRef.style.width = `${width}px`;
-		}
-
-		const availableBelow = Math.max(
-			0,
-			viewportHeight - VIEWPORT_GUTTER - triggerRect.bottom - MENU_OFFSET
-		);
-		const availableAbove = Math.max(0, triggerRect.top - VIEWPORT_GUTTER - MENU_OFFSET);
-		const viewportAllowance = Math.max(0, viewportHeight - VIEWPORT_GUTTER * 2);
-		const fallbackAllowance = Math.max(1, viewportAllowance > 0 ? viewportAllowance : scrollHeight);
-
-		function computePlacement(placement: 'top' | 'bottom') {
-			const available = placement === 'bottom' ? availableBelow : availableAbove;
-			const allowedHeight =
-				available > 0 ? Math.min(available, fallbackAllowance) : fallbackAllowance;
-			const maxHeight = Math.min(scrollHeight, allowedHeight);
-			const height = Math.max(0, maxHeight);
-
-			let top: number;
-			if (placement === 'bottom') {
-				const rawTop = triggerRect.bottom + MENU_OFFSET;
-				const minTop = VIEWPORT_GUTTER;
-				const maxTop = viewportHeight - VIEWPORT_GUTTER - height;
-				if (maxTop < minTop) {
-					top = minTop;
-				} else {
-					top = Math.min(Math.max(rawTop, minTop), maxTop);
-				}
-			} else {
-				const rawTop = triggerRect.top - MENU_OFFSET - height;
-				const minTop = VIEWPORT_GUTTER;
-				const maxTop = viewportHeight - VIEWPORT_GUTTER - height;
-				if (maxTop < minTop) {
-					top = minTop;
-				} else {
-					top = Math.max(Math.min(rawTop, maxTop), minTop);
-				}
-			}
-
-			return { placement, top, height, maxHeight };
-		}
-
-		const belowMetrics = computePlacement('bottom');
-		const aboveMetrics = computePlacement('top');
-
-		let metrics = belowMetrics;
-		if (scrollHeight > belowMetrics.maxHeight && aboveMetrics.maxHeight > belowMetrics.maxHeight) {
-			metrics = aboveMetrics;
-		}
-
-		menuRef.style.maxHeight = metrics.maxHeight > 0 ? `${Math.round(metrics.maxHeight)}px` : '';
-
-		let left = triggerRect.right - width;
-		const maxLeft = viewportWidth - VIEWPORT_GUTTER - width;
-		if (maxLeft < VIEWPORT_GUTTER) {
-			left = VIEWPORT_GUTTER;
-		} else {
-			if (left > maxLeft) {
-				left = maxLeft;
-			}
-			if (left < VIEWPORT_GUTTER) {
-				left = VIEWPORT_GUTTER;
-			}
-		}
-
-		menuPosition = {
-			top: Math.round(metrics.top),
-			left: Math.round(left),
-			width: Math.round(width),
-			placement: metrics.placement,
-			maxHeight: Math.round(metrics.maxHeight)
-		};
-	}
-
-	function getDisplayOption(): ModelOption | undefined {
-		if (activeId) {
-			return options.find((option) => option.id === activeId);
-		}
-
-		return options[0];
-	}
-</script>
-
-<svelte:window onresize={handleResize} />
-
-<svelte:document onpointerdown={handlePointerDown} onkeydown={handleKeydown} />
-
-<div
-	class={cn('relative z-10 flex max-w-[200px] min-w-[120px] flex-col items-end gap-1', className)}
-	bind:this={container}
->
-	{#if loading && options.length === 0 && !isMounted}
-		<div class="flex items-center gap-2 text-xs text-muted-foreground">
-			<Loader2 class="h-4 w-4 animate-spin" />
-			Loading models…
-		</div>
-	{:else if options.length === 0}
-		<p class="text-xs text-muted-foreground">No models available.</p>
-	{:else}
-		{@const selectedOption = getDisplayOption()}
-
-		<div class="relative w-full">
-			<button
-				type="button"
-				class={cn(
-					'flex w-full items-center justify-end gap-2 rounded-md px-2 py-1 text-sm text-muted-foreground transition hover:text-foreground focus:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 disabled:cursor-not-allowed disabled:opacity-60',
-					isOpen ? 'text-foreground' : ''
-				)}
-				aria-haspopup="listbox"
-				aria-expanded={isOpen}
-				onclick={toggleOpen}
-				bind:this={triggerButton}
-				disabled={loading || updating}
-			>
-				<span class="max-w-[160px] truncate text-right font-medium">
-					{selectedOption?.name || 'Select model'}
-				</span>
-
-				{#if updating}
-					<Loader2 class="h-3.5 w-3.5 animate-spin text-muted-foreground" />
-				{:else}
-					<ChevronDown
-						class={cn(
-							'h-4 w-4 text-muted-foreground transition-transform',
-							isOpen ? 'rotate-180 text-foreground' : ''
-						)}
-					/>
-				{/if}
-			</button>
-
-			{#if isOpen}
-				<div
-					bind:this={menuRef}
-					use:portalToBody
-					class={cn(
-						'fixed z-[1000] overflow-hidden rounded-md border bg-popover shadow-lg transition-opacity',
-						menuPosition ? 'opacity-100' : 'pointer-events-none opacity-0'
-					)}
-					role="listbox"
-					style:top={menuPosition ? `${menuPosition.top}px` : undefined}
-					style:left={menuPosition ? `${menuPosition.left}px` : undefined}
-					style:width={menuPosition ? `${menuPosition.width}px` : undefined}
-					data-placement={menuPosition?.placement ?? 'bottom'}
-				>
-					<div
-						class="overflow-y-auto py-1"
-						style:max-height={menuPosition && menuPosition.maxHeight > 0
-							? `${menuPosition.maxHeight}px`
-							: undefined}
-					>
-						{#each options as option (option.id)}
-							<button
-								type="button"
-								class={cn(
-									'flex w-full flex-col items-start gap-0.5 px-3 py-2 text-left text-sm transition hover:bg-muted focus:bg-muted focus:outline-none',
-									option.id === selectedOption?.id ? 'bg-accent text-accent-foreground' : ''
-								)}
-								role="option"
-								aria-selected={option.id === selectedOption?.id}
-								onclick={() => handleOptionSelect(option.id)}
-							>
-								<span class="block w-full truncate font-medium" title={option.name}>
-									{option.name}
-								</span>
-
-								{#if option.description}
-									<span class="text-xs text-muted-foreground">{option.description}</span>
-								{/if}
-							</button>
-						{/each}
-					</div>
-				</div>
-			{/if}
-		</div>
-	{/if}
-
-	{#if error}
-		<p class="text-xs text-destructive">{error}</p>
-	{/if}
-</div>

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormTextarea.svelte

@@ -1,5 +1,5 @@
 <script lang="ts">
-	import autoResizeTextarea from '$lib/utils/autoresize-textarea';
+	import { autoResizeTextarea } from '$lib/utils';
 	import { onMount } from 'svelte';
 
 	interface Props {

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage.svelte

@@ -1,8 +1,6 @@
 <script lang="ts">
-	import { getDeletionInfo } from '$lib/stores/chat.svelte';
-	import { copyToClipboard } from '$lib/utils/copy';
-	import { isIMEComposing } from '$lib/utils/is-ime-composing';
-	import type { ApiChatCompletionToolCall } from '$lib/types/api';
+	import { chatStore } from '$lib/stores/chat.svelte';
+	import { copyToClipboard, isIMEComposing } from '$lib/utils';
 	import ChatMessageAssistant from './ChatMessageAssistant.svelte';
 	import ChatMessageUser from './ChatMessageUser.svelte';
 
@@ -20,7 +18,7 @@
 		) => void;
 		onEditUserMessagePreserveResponses?: (message: DatabaseMessage, newContent: string) => void;
 		onNavigateToSibling?: (siblingId: string) => void;
-		onRegenerateWithBranching?: (message: DatabaseMessage) => void;
+		onRegenerateWithBranching?: (message: DatabaseMessage, modelOverride?: string) => void;
 		siblingInfo?: ChatMessageSiblingInfo | null;
 	}
 
@@ -98,7 +96,7 @@
 	}
 
 	async function handleDelete() {
-		deletionInfo = await getDeletionInfo(message.id);
+		deletionInfo = await chatStore.getDeletionInfo(message.id);
 		showDeleteDialog = true;
 	}
 
@@ -133,8 +131,8 @@
 		}
 	}
 
-	function handleRegenerate() {
-		onRegenerateWithBranching?.(message);
+	function handleRegenerate(modelOverride?: string) {
+		onRegenerateWithBranching?.(message, modelOverride);
 	}
 
 	function handleContinue() {

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageActions.svelte

@@ -71,7 +71,7 @@
 			{/if}
 
 			{#if role === 'assistant' && onRegenerate}
-				<ActionButton icon={RefreshCw} tooltip="Regenerate" onclick={onRegenerate} />
+				<ActionButton icon={RefreshCw} tooltip="Regenerate" onclick={() => onRegenerate()} />
 			{/if}
 
 			{#if role === 'assistant' && onContinue}

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte

@@ -1,29 +1,26 @@
 <script lang="ts">
-	import { ChatMessageThinkingBlock, MarkdownContent } from '$lib/components/app';
-	import { useProcessingState } from '$lib/hooks/use-processing-state.svelte';
-	import { isLoading } from '$lib/stores/chat.svelte';
-	import autoResizeTextarea from '$lib/utils/autoresize-textarea';
-	import { fade } from 'svelte/transition';
 	import {
-		Check,
-		Copy,
-		Package,
-		X,
-		Gauge,
-		Clock,
-		WholeWord,
-		ChartNoAxesColumn,
-		Wrench
-	} from '@lucide/svelte';
+		ModelBadge,
+		ChatMessageActions,
+		ChatMessageStatistics,
+		ChatMessageThinkingBlock,
+		CopyToClipboardIcon,
+		MarkdownContent,
+		ModelsSelector
+	} from '$lib/components/app';
+	import { useProcessingState } from '$lib/hooks/use-processing-state.svelte';
+	import { useModelChangeValidation } from '$lib/hooks/use-model-change-validation.svelte';
+	import { isLoading } from '$lib/stores/chat.svelte';
+	import { autoResizeTextarea, copyToClipboard } from '$lib/utils';
+	import { fade } from 'svelte/transition';
+	import { Check, X, Wrench } from '@lucide/svelte';
 	import { Button } from '$lib/components/ui/button';
 	import { Checkbox } from '$lib/components/ui/checkbox';
 	import { INPUT_CLASSES } from '$lib/constants/input-classes';
-	import ChatMessageActions from './ChatMessageActions.svelte';
 	import Label from '$lib/components/ui/label/label.svelte';
 	import { config } from '$lib/stores/settings.svelte';
-	import { modelName as serverModelName } from '$lib/stores/server.svelte';
-	import { copyToClipboard } from '$lib/utils/copy';
-	import type { ApiChatCompletionToolCall } from '$lib/types/api';
+	import { conversationsStore } from '$lib/stores/conversations.svelte';
+	import { isRouterMode } from '$lib/stores/server.svelte';
 
 	interface Props {
 		class?: string;
@@ -46,7 +43,7 @@
 		onEditKeydown?: (event: KeyboardEvent) => void;
 		onEditedContentChange?: (content: string) => void;
 		onNavigateToSibling?: (siblingId: string) => void;
-		onRegenerate: () => void;
+		onRegenerate: (modelOverride?: string) => void;
 		onSaveEdit?: () => void;
 		onShowDeleteDialogChange: (show: boolean) => void;
 		onShouldBranchAfterEditChange?: (value: boolean) => void;
@@ -93,15 +90,18 @@
 
 	const processingState = useProcessingState();
 	let currentConfig = $derived(config());
-	let serverModel = $derived(serverModelName());
+	let isRouter = $derived(isRouterMode());
 	let displayedModel = $derived((): string | null => {
-		if (!currentConfig.showModelInfo) return null;
-
 		if (message.model) {
 			return message.model;
 		}
 
-		return serverModel;
+		return null;
+	});
+
+	const { handleModelChange } = useModelChangeValidation({
+		getRequiredModalities: () => conversationsStore.getModalitiesUpToMessage(message.id),
+		onSuccess: (modelName) => onRegenerate(modelName)
 	});
 
 	function handleCopyModel() {
@@ -244,21 +244,24 @@
 
 	<div class="info my-6 grid gap-4">
 		{#if displayedModel()}
-			<span class="inline-flex items-center gap-2 text-xs text-muted-foreground">
-				<span class="inline-flex items-center gap-1">
-					<Package class="h-3.5 w-3.5" />
+			<span class="inline-flex flex-wrap items-center gap-2 text-xs text-muted-foreground">
+				{#if isRouter}
+					<ModelsSelector
+						currentModel={displayedModel()}
+						onModelChange={handleModelChange}
+						disabled={isLoading()}
+						upToMessageId={message.id}
+					/>
+				{:else}
+					<ModelBadge model={displayedModel() || undefined} onclick={handleCopyModel} />
+				{/if}
 
-					<span>Model used:</span>
-				</span>
-
-				<button
-					class="inline-flex cursor-pointer items-center gap-1 rounded-sm bg-muted-foreground/15 px-1.5 py-0.75"
-					onclick={handleCopyModel}
-				>
-					{displayedModel()}
-
-					<Copy class="ml-1 h-3 w-3 " />
-				</button>
+				{#if currentConfig.showMessageStats && message.timings && message.timings.predicted_n && message.timings.predicted_ms}
+					<ChatMessageStatistics
+						predictedTokens={message.timings.predicted_n}
+						predictedMs={message.timings.predicted_ms}
+					/>
+				{/if}
 			</span>
 		{/if}
 
@@ -282,8 +285,10 @@
 								onclick={() => handleCopyToolCall(badge.copyValue)}
 							>
 								{badge.label}
-
-								<Copy class="ml-1 h-3 w-3" />
+								<CopyToClipboardIcon
+									text={badge.copyValue}
+									ariaLabel={`Copy tool call ${badge.label}`}
+								/>
 							</button>
 						{/each}
 					{:else if fallbackToolCalls}
@@ -295,45 +300,12 @@
 							onclick={() => handleCopyToolCall(fallbackToolCalls)}
 						>
 							{fallbackToolCalls}
-
-							<Copy class="ml-1 h-3 w-3" />
+							<CopyToClipboardIcon text={fallbackToolCalls} ariaLabel="Copy tool call payload" />
 						</button>
 					{/if}
 				</span>
 			{/if}
 		{/if}
-
-		{#if currentConfig.showMessageStats && message.timings && message.timings.predicted_n && message.timings.predicted_ms}
-			{@const tokensPerSecond = (message.timings.predicted_n / message.timings.predicted_ms) * 1000}
-			<span class="inline-flex items-center gap-2 text-xs text-muted-foreground">
-				<span class="inline-flex items-center gap-1">
-					<ChartNoAxesColumn class="h-3.5 w-3.5" />
-
-					<span>Statistics:</span>
-				</span>
-
-				<div class="inline-flex flex-wrap items-center gap-2 text-xs text-muted-foreground">
-					<span
-						class="inline-flex items-center gap-1 rounded-sm bg-muted-foreground/15 px-1.5 py-0.75"
-					>
-						<Gauge class="h-3 w-3" />
-						{tokensPerSecond.toFixed(2)} tokens/s
-					</span>
-					<span
-						class="inline-flex items-center gap-1 rounded-sm bg-muted-foreground/15 px-1.5 py-0.75"
-					>
-						<WholeWord class="h-3 w-3" />
-						{message.timings.predicted_n} tokens
-					</span>
-					<span
-						class="inline-flex items-center gap-1 rounded-sm bg-muted-foreground/15 px-1.5 py-0.75"
-					>
-						<Clock class="h-3 w-3" />
-						{(message.timings.predicted_ms / 1000).toFixed(2)}s
-					</span>
-				</div>
-			</span>
-		{/if}
 	</div>
 
 	{#if message.timestamp && !isEditing}

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageStatistics.svelte

@@ -0,0 +1,20 @@
+<script lang="ts">
+	import { Clock, Gauge, WholeWord } from '@lucide/svelte';
+	import { BadgeChatStatistic } from '$lib/components/app';
+
+	interface Props {
+		predictedTokens: number;
+		predictedMs: number;
+	}
+
+	let { predictedTokens, predictedMs }: Props = $props();
+
+	let tokensPerSecond = $derived((predictedTokens / predictedMs) * 1000);
+	let timeInSeconds = $derived((predictedMs / 1000).toFixed(2));
+</script>
+
+<BadgeChatStatistic icon={WholeWord} value="{predictedTokens} tokens" />
+
+<BadgeChatStatistic icon={Clock} value="{timeInSeconds}s" />
+
+<BadgeChatStatistic icon={Gauge} value="{tokensPerSecond.toFixed(2)} tokens/s" />

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageUser.svelte

@@ -5,7 +5,7 @@
 	import { ChatAttachmentsList, MarkdownContent } from '$lib/components/app';
 	import { INPUT_CLASSES } from '$lib/constants/input-classes';
 	import { config } from '$lib/stores/settings.svelte';
-	import autoResizeTextarea from '$lib/utils/autoresize-textarea';
+	import { autoResizeTextarea } from '$lib/utils';
 	import ChatMessageActions from './ChatMessageActions.svelte';
 
 	interface Props {

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessages.svelte

@@ -1,17 +1,9 @@
 <script lang="ts">
 	import { ChatMessage } from '$lib/components/app';
-	import { DatabaseStore } from '$lib/stores/database';
-	import {
-		activeConversation,
-		continueAssistantMessage,
-		deleteMessage,
-		editAssistantMessage,
-		editMessageWithBranching,
-		editUserMessagePreserveResponses,
-		navigateToSibling,
-		regenerateMessageWithBranching
-	} from '$lib/stores/chat.svelte';
-	import { getMessageSiblings } from '$lib/utils/branching';
+	import { DatabaseService } from '$lib/services/database';
+	import { chatStore } from '$lib/stores/chat.svelte';
+	import { conversationsStore, activeConversation } from '$lib/stores/conversations.svelte';
+	import { getMessageSiblings } from '$lib/utils';
 
 	interface Props {
 		class?: string;
@@ -27,7 +19,7 @@
 		const conversation = activeConversation();
 
 		if (conversation) {
-			DatabaseStore.getConversationMessages(conversation.id).then((messages) => {
+			DatabaseService.getConversationMessages(conversation.id).then((messages) => {
 				allConversationMessages = messages;
 			});
 		} else {
@@ -65,13 +57,13 @@
 	});
 
 	async function handleNavigateToSibling(siblingId: string) {
-		await navigateToSibling(siblingId);
+		await conversationsStore.navigateToSibling(siblingId);
 	}
 
 	async function handleEditWithBranching(message: DatabaseMessage, newContent: string) {
 		onUserAction?.();
 
-		await editMessageWithBranching(message.id, newContent);
+		await chatStore.editMessageWithBranching(message.id, newContent);
 
 		refreshAllMessages();
 	}
@@ -83,15 +75,15 @@
 	) {
 		onUserAction?.();
 
-		await editAssistantMessage(message.id, newContent, shouldBranch);
+		await chatStore.editAssistantMessage(message.id, newContent, shouldBranch);
 
 		refreshAllMessages();
 	}
 
-	async function handleRegenerateWithBranching(message: DatabaseMessage) {
+	async function handleRegenerateWithBranching(message: DatabaseMessage, modelOverride?: string) {
 		onUserAction?.();
 
-		await regenerateMessageWithBranching(message.id);
+		await chatStore.regenerateMessageWithBranching(message.id, modelOverride);
 
 		refreshAllMessages();
 	}
@@ -99,7 +91,7 @@
 	async function handleContinueAssistantMessage(message: DatabaseMessage) {
 		onUserAction?.();
 
-		await continueAssistantMessage(message.id);
+		await chatStore.continueAssistantMessage(message.id);
 
 		refreshAllMessages();
 	}
@@ -110,13 +102,13 @@
 	) {
 		onUserAction?.();
 
-		await editUserMessagePreserveResponses(message.id, newContent);
+		await chatStore.editUserMessagePreserveResponses(message.id, newContent);
 
 		refreshAllMessages();
 	}
 
 	async function handleDeleteMessage(message: DatabaseMessage) {
-		await deleteMessage(message.id);
+		await chatStore.deleteMessage(message.id);
 
 		refreshAllMessages();
 	}

tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte

@@ -3,47 +3,34 @@
 	import {
 		ChatForm,
 		ChatScreenHeader,
-		ChatScreenWarning,
 		ChatMessages,
 		ChatScreenProcessingInfo,
 		DialogEmptyFileAlert,
 		DialogChatError,
-		ServerErrorSplash,
-		ServerInfo,
 		ServerLoadingSplash,
 		DialogConfirmation
 	} from '$lib/components/app';
+	import * as Alert from '$lib/components/ui/alert';
 	import * as AlertDialog from '$lib/components/ui/alert-dialog';
 	import {
 		AUTO_SCROLL_AT_BOTTOM_THRESHOLD,
 		AUTO_SCROLL_INTERVAL,
 		INITIAL_SCROLL_DELAY
 	} from '$lib/constants/auto-scroll';
+	import { chatStore, errorDialog, isLoading } from '$lib/stores/chat.svelte';
 	import {
+		conversationsStore,
 		activeMessages,
-		activeConversation,
-		deleteConversation,
-		dismissErrorDialog,
-		errorDialog,
-		isLoading,
-		sendMessage,
-		stopGeneration
-	} from '$lib/stores/chat.svelte';
+		activeConversation
+	} from '$lib/stores/conversations.svelte';
 	import { config } from '$lib/stores/settings.svelte';
-	import {
-		supportsVision,
-		supportsAudio,
-		serverLoading,
-		serverWarning,
-		serverStore
-	} from '$lib/stores/server.svelte';
-	import { parseFilesToMessageExtras } from '$lib/utils/convert-files-to-extra';
-	import { isFileTypeSupported } from '$lib/utils/file-type';
-	import { filterFilesByModalities } from '$lib/utils/modality-file-validation';
-	import { processFilesToChatUploaded } from '$lib/utils/process-uploaded-files';
+	import { serverLoading, serverError, serverStore, isRouterMode } from '$lib/stores/server.svelte';
+	import { modelsStore, modelOptions, selectedModelId } from '$lib/stores/models.svelte';
+	import { isFileTypeSupported, filterFilesByModalities } from '$lib/utils';
+	import { parseFilesToMessageExtras, processFilesToChatUploaded } from '$lib/utils/browser-only';
 	import { onMount } from 'svelte';
 	import { fade, fly, slide } from 'svelte/transition';
-	import { Trash2 } from '@lucide/svelte';
+	import { Trash2, AlertTriangle, RefreshCw } from '@lucide/svelte';
 	import ChatScreenDragOverlay from './ChatScreenDragOverlay.svelte';
 
 	let { showCenteredEmpty = false } = $props();
@@ -84,20 +71,84 @@
 
 	let activeErrorDialog = $derived(errorDialog());
 	let isServerLoading = $derived(serverLoading());
+	let hasPropsError = $derived(!!serverError());
 
 	let isCurrentConversationLoading = $derived(isLoading());
 
+	let isRouter = $derived(isRouterMode());
+
+	let conversationModel = $derived(
+		chatStore.getConversationModel(activeMessages() as DatabaseMessage[])
+	);
+
+	let activeModelId = $derived.by(() => {
+		const options = modelOptions();
+
+		if (!isRouter) {
+			return options.length > 0 ? options[0].model : null;
+		}
+
+		const selectedId = selectedModelId();
+		if (selectedId) {
+			const model = options.find((m) => m.id === selectedId);
+			if (model) return model.model;
+		}
+
+		if (conversationModel) {
+			const model = options.find((m) => m.model === conversationModel);
+			if (model) return model.model;
+		}
+
+		return null;
+	});
+
+	let modelPropsVersion = $state(0);
+
+	$effect(() => {
+		if (activeModelId) {
+			const cached = modelsStore.getModelProps(activeModelId);
+			if (!cached) {
+				modelsStore.fetchModelProps(activeModelId).then(() => {
+					modelPropsVersion++;
+				});
+			}
+		}
+	});
+
+	let hasAudioModality = $derived.by(() => {
+		if (activeModelId) {
+			void modelPropsVersion;
+			return modelsStore.modelSupportsAudio(activeModelId);
+		}
+
+		return false;
+	});
+
+	let hasVisionModality = $derived.by(() => {
+		if (activeModelId) {
+			void modelPropsVersion;
+
+			return modelsStore.modelSupportsVision(activeModelId);
+		}
+
+		return false;
+	});
+
 	async function handleDeleteConfirm() {
 		const conversation = activeConversation();
+
 		if (conversation) {
-			await deleteConversation(conversation.id);
+			await conversationsStore.deleteConversation(conversation.id);
 		}
+
 		showDeleteDialog = false;
 	}
 
 	function handleDragEnter(event: DragEvent) {
 		event.preventDefault();
+
 		dragCounter++;
+
 		if (event.dataTransfer?.types.includes('Files')) {
 			isDragOver = true;
 		}
@@ -105,7 +156,9 @@
 
 	function handleDragLeave(event: DragEvent) {
 		event.preventDefault();
+
 		dragCounter--;
+
 		if (dragCounter === 0) {
 			isDragOver = false;
 		}
@@ -113,7 +166,7 @@
 
 	function handleErrorDialogOpenChange(open: boolean) {
 		if (!open) {
-			dismissErrorDialog();
+			chatStore.dismissErrorDialog();
 		}
 	}
 
@@ -123,6 +176,7 @@
 
 	function handleDrop(event: DragEvent) {
 		event.preventDefault();
+
 		isDragOver = false;
 		dragCounter = 0;
 
@@ -180,7 +234,9 @@
 	}
 
 	async function handleSendMessage(message: string, files?: ChatUploadedFile[]): Promise<boolean> {
-		const result = files ? await parseFilesToMessageExtras(files) : undefined;
+		const result = files
+			? await parseFilesToMessageExtras(files, activeModelId ?? undefined)
+			: undefined;
 
 		if (result?.emptyFiles && result.emptyFiles.length > 0) {
 			emptyFileNames = result.emptyFiles;
@@ -200,7 +256,7 @@
 			userScrolledUp = false;
 			autoScrollEnabled = true;
 		}
-		await sendMessage(message, extras);
+		await chatStore.sendMessage(message, extras);
 		scrollChatToBottom();
 
 		return true;
@@ -218,16 +274,20 @@
 			}
 		}
 
-		const { supportedFiles, unsupportedFiles, modalityReasons } =
-			filterFilesByModalities(generallySupported);
+		// Use model-specific capabilities for file validation
+		const capabilities = { hasVision: hasVisionModality, hasAudio: hasAudioModality };
+		const { supportedFiles, unsupportedFiles, modalityReasons } = filterFilesByModalities(
+			generallySupported,
+			capabilities
+		);
 
 		const allUnsupportedFiles = [...generallyUnsupported, ...unsupportedFiles];
 
 		if (allUnsupportedFiles.length > 0) {
 			const supportedTypes: string[] = ['text files', 'PDFs'];
 
-			if (supportsVision()) supportedTypes.push('images');
-			if (supportsAudio()) supportedTypes.push('audio files');
+			if (hasVisionModality) supportedTypes.push('images');
+			if (hasAudioModality) supportedTypes.push('audio files');
 
 			fileErrorData = {
 				generallyUnsupported,
@@ -239,7 +299,10 @@
 		}
 
 		if (supportedFiles.length > 0) {
-			const processed = await processFilesToChatUploaded(supportedFiles);
+			const processed = await processFilesToChatUploaded(
+				supportedFiles,
+				activeModelId ?? undefined
+			);
 			uploadedFiles = [...uploadedFiles, ...processed];
 		}
 	}
@@ -322,17 +385,37 @@
 		>
 			<ChatScreenProcessingInfo />
 
-			{#if serverWarning()}
-				<ChatScreenWarning class="pointer-events-auto mx-auto max-w-[48rem] px-4" />
+			{#if hasPropsError}
+				<div
+					class="pointer-events-auto mx-auto mb-4 max-w-[48rem] px-1"
+					in:fly={{ y: 10, duration: 250 }}
+				>
+					<Alert.Root variant="destructive">
+						<AlertTriangle class="h-4 w-4" />
+						<Alert.Title class="flex items-center justify-between">
+							<span>Server unavailable</span>
+							<button
+								onclick={() => serverStore.fetch()}
+								disabled={isServerLoading}
+								class="flex items-center gap-1.5 rounded-lg bg-destructive/20 px-2 py-1 text-xs font-medium hover:bg-destructive/30 disabled:opacity-50"
+							>
+								<RefreshCw class="h-3 w-3 {isServerLoading ? 'animate-spin' : ''}" />
+								{isServerLoading ? 'Retrying...' : 'Retry'}
+							</button>
+						</Alert.Title>
+						<Alert.Description>{serverError()}</Alert.Description>
+					</Alert.Root>
+				</div>
 			{/if}
 
 			<div class="conversation-chat-form pointer-events-auto rounded-t-3xl pb-4">
 				<ChatForm
+					disabled={hasPropsError}
 					isLoading={isCurrentConversationLoading}
 					onFileRemove={handleFileRemove}
 					onFileUpload={handleFileUpload}
 					onSend={handleSendMessage}
-					onStop={() => stopGeneration()}
+					onStop={() => chatStore.stopGeneration()}
 					showHelperText={false}
 					bind:uploadedFiles
 				/>
@@ -342,9 +425,7 @@
 {:else if isServerLoading}
 	<!-- Server Loading State -->
 	<ServerLoadingSplash />
-{:else if serverStore.error && !serverStore.modelName}
-	<ServerErrorSplash error={serverStore.error} />
-{:else if serverStore.modelName}
+{:else}
 	<div
 		aria-label="Welcome screen with file drop zone"
 		class="flex h-full items-center justify-center"
@@ -355,27 +436,44 @@
 		role="main"
 	>
 		<div class="w-full max-w-[48rem] px-4">
-			<div class="mb-8 text-center" in:fade={{ duration: 300 }}>
-				<h1 class="mb-2 text-3xl font-semibold tracking-tight">llama.cpp</h1>
+			<div class="mb-10 text-center" in:fade={{ duration: 300 }}>
+				<h1 class="mb-4 text-3xl font-semibold tracking-tight">llama.cpp</h1>
 
-				<p class="text-lg text-muted-foreground">How can I help you today?</p>
+				<p class="text-lg text-muted-foreground">
+					{serverStore.props?.modalities?.audio
+						? 'Record audio, type a message '
+						: 'Type a message'} or upload files to get started
+				</p>
 			</div>
 
-			<div class="mb-6 flex justify-center" in:fly={{ y: 10, duration: 300, delay: 200 }}>
-				<ServerInfo />
-			</div>
-
-			{#if serverWarning()}
-				<ChatScreenWarning />
+			{#if hasPropsError}
+				<div class="mb-4" in:fly={{ y: 10, duration: 250 }}>
+					<Alert.Root variant="destructive">
+						<AlertTriangle class="h-4 w-4" />
+						<Alert.Title class="flex items-center justify-between">
+							<span>Server unavailable</span>
+							<button
+								onclick={() => serverStore.fetch()}
+								disabled={isServerLoading}
+								class="flex items-center gap-1.5 rounded-lg bg-destructive/20 px-2 py-1 text-xs font-medium hover:bg-destructive/30 disabled:opacity-50"
+							>
+								<RefreshCw class="h-3 w-3 {isServerLoading ? 'animate-spin' : ''}" />
+								{isServerLoading ? 'Retrying...' : 'Retry'}
+							</button>
+						</Alert.Title>
+						<Alert.Description>{serverError()}</Alert.Description>
+					</Alert.Root>
+				</div>
 			{/if}
 
-			<div in:fly={{ y: 10, duration: 250, delay: 300 }}>
+			<div in:fly={{ y: 10, duration: 250, delay: hasPropsError ? 0 : 300 }}>
 				<ChatForm
+					disabled={hasPropsError}
 					isLoading={isCurrentConversationLoading}
 					onFileRemove={handleFileRemove}
 					onFileUpload={handleFileUpload}
 					onSend={handleSendMessage}
-					onStop={() => stopGeneration()}
+					onStop={() => chatStore.stopGeneration()}
 					showHelperText={true}
 					bind:uploadedFiles
 				/>
@@ -477,6 +575,7 @@
 
 <DialogChatError
 	message={activeErrorDialog?.message ?? ''}
+	contextInfo={activeErrorDialog?.contextInfo}
 	onOpenChange={handleErrorDialogOpenChange}
 	open={Boolean(activeErrorDialog)}
 	type={activeErrorDialog?.type ?? 'server'}

tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreenProcessingInfo.svelte

@@ -1,34 +1,47 @@
 <script lang="ts">
+	import { untrack } from 'svelte';
 	import { PROCESSING_INFO_TIMEOUT } from '$lib/constants/processing-info';
 	import { useProcessingState } from '$lib/hooks/use-processing-state.svelte';
-	import { slotsService } from '$lib/services/slots';
-	import { isLoading, activeMessages, activeConversation } from '$lib/stores/chat.svelte';
+	import { chatStore, isLoading, isChatStreaming } from '$lib/stores/chat.svelte';
+	import { activeMessages, activeConversation } from '$lib/stores/conversations.svelte';
 	import { config } from '$lib/stores/settings.svelte';
 
 	const processingState = useProcessingState();
 
 	let isCurrentConversationLoading = $derived(isLoading());
+	let isStreaming = $derived(isChatStreaming());
+	let hasProcessingData = $derived(processingState.processingState !== null);
 	let processingDetails = $derived(processingState.getProcessingDetails());
-	let showSlotsInfo = $derived(isCurrentConversationLoading || config().keepStatsVisible);
 
-	// Track loading state reactively by checking if conversation ID is in loading conversations array
+	let showProcessingInfo = $derived(
+		isCurrentConversationLoading || isStreaming || config().keepStatsVisible || hasProcessingData
+	);
+
+	$effect(() => {
+		const conversation = activeConversation();
+
+		untrack(() => chatStore.setActiveProcessingConversation(conversation?.id ?? null));
+	});
+
 	$effect(() => {
 		const keepStatsVisible = config().keepStatsVisible;
+		const shouldMonitor = keepStatsVisible || isCurrentConversationLoading || isStreaming;
 
-		if (keepStatsVisible || isCurrentConversationLoading) {
+		if (shouldMonitor) {
 			processingState.startMonitoring();
 		}
 
-		if (!isCurrentConversationLoading && !keepStatsVisible) {
-			setTimeout(() => {
-				if (!config().keepStatsVisible) {
+		if (!isCurrentConversationLoading && !isStreaming && !keepStatsVisible) {
+			const timeout = setTimeout(() => {
+				if (!config().keepStatsVisible && !isChatStreaming()) {
 					processingState.stopMonitoring();
 				}
 			}, PROCESSING_INFO_TIMEOUT);
+
+			return () => clearTimeout(timeout);
 		}
 	});
 
-	// Update processing state from stored timings
 	$effect(() => {
 		const conversation = activeConversation();
 		const messages = activeMessages() as DatabaseMessage[];
@@ -36,47 +49,18 @@
 
 		if (keepStatsVisible && conversation) {
 			if (messages.length === 0) {
-				slotsService.clearConversationState(conversation.id);
+				untrack(() => chatStore.clearProcessingState(conversation.id));
 				return;
 			}
 
-			// Search backwards through messages to find most recent assistant message with timing data
-			// Using reverse iteration for performance - avoids array copy and stops at first match
-			let foundTimingData = false;
-
-			for (let i = messages.length - 1; i >= 0; i--) {
-				const message = messages[i];
-				if (message.role === 'assistant' && message.timings) {
-					foundTimingData = true;
-
-					slotsService
-						.updateFromTimingData(
-							{
-								prompt_n: message.timings.prompt_n || 0,
-								predicted_n: message.timings.predicted_n || 0,
-								predicted_per_second:
-									message.timings.predicted_n && message.timings.predicted_ms
-										? (message.timings.predicted_n / message.timings.predicted_ms) * 1000
-										: 0,
-								cache_n: message.timings.cache_n || 0
-							},
-							conversation.id
-						)
-						.catch((error) => {
-							console.warn('Failed to update processing state from stored timings:', error);
-						});
-					break;
-				}
-			}
-
-			if (!foundTimingData) {
-				slotsService.clearConversationState(conversation.id);
+			if (!isCurrentConversationLoading && !isStreaming) {
+				untrack(() => chatStore.restoreProcessingStateFromMessages(messages, conversation.id));
 			}
 		}
 	});
 </script>
 
-<div class="chat-processing-info-container pointer-events-none" class:visible={showSlotsInfo}>
+<div class="chat-processing-info-container pointer-events-none" class:visible={showProcessingInfo}>
 	<div class="chat-processing-info-content">
 		{#each processingDetails as detail (detail)}
 			<span class="chat-processing-info-detail pointer-events-auto">{detail}</span>

tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreenWarning.svelte

@@ -1,38 +0,0 @@
-<script lang="ts">
-	import { AlertTriangle, RefreshCw } from '@lucide/svelte';
-	import { serverLoading, serverStore } from '$lib/stores/server.svelte';
-	import { fly } from 'svelte/transition';
-
-	interface Props {
-		class?: string;
-	}
-
-	let { class: className = '' }: Props = $props();
-
-	function handleRefreshServer() {
-		serverStore.fetchServerProps();
-	}
-</script>
-
-<div class="mb-3 {className}" in:fly={{ y: 10, duration: 250 }}>
-	<div
-		class="rounded-md border border-yellow-200 bg-yellow-50 px-3 py-2 dark:border-yellow-800 dark:bg-yellow-950"
-	>
-		<div class="flex items-center justify-between">
-			<div class="flex items-center">
-				<AlertTriangle class="h-4 w-4 text-yellow-600 dark:text-yellow-400" />
-				<p class="ml-2 text-sm text-yellow-800 dark:text-yellow-200">
-					Server `/props` endpoint not available - using cached data
-				</p>
-			</div>
-			<button
-				onclick={handleRefreshServer}
-				disabled={serverLoading()}
-				class="ml-3 flex items-center gap-1.5 rounded bg-yellow-100 px-2 py-1 text-xs font-medium text-yellow-800 hover:bg-yellow-200 disabled:opacity-50 dark:bg-yellow-900 dark:text-yellow-200 dark:hover:bg-yellow-800"
-			>
-				<RefreshCw class="h-3 w-3 {serverLoading() ? 'animate-spin' : ''}" />
-				{serverLoading() ? 'Checking...' : 'Retry'}
-			</button>
-		</div>
-	</div>
-</div>

tools/server/webui/src/lib/components/app/chat/ChatSettings/ChatSettings.svelte

@@ -17,7 +17,7 @@
 		ChatSettingsFields
 	} from '$lib/components/app';
 	import { ScrollArea } from '$lib/components/ui/scroll-area';
-	import { config, updateMultipleConfig } from '$lib/stores/settings.svelte';
+	import { config, settingsStore } from '$lib/stores/settings.svelte';
 	import { setMode } from 'mode-watcher';
 	import type { Component } from 'svelte';
 
@@ -79,19 +79,14 @@
 			title: 'Display',
 			icon: Monitor,
 			fields: [
-				{
-					key: 'showThoughtInProgress',
-					label: 'Show thought in progress',
-					type: 'checkbox'
-				},
 				{
 					key: 'showMessageStats',
 					label: 'Show message generation statistics',
 					type: 'checkbox'
 				},
 				{
-					key: 'showTokensPerSecond',
-					label: 'Show tokens per second',
+					key: 'showThoughtInProgress',
+					label: 'Show thought in progress',
 					type: 'checkbox'
 				},
 				{
@@ -100,19 +95,20 @@
 					type: 'checkbox'
 				},
 				{
-					key: 'showModelInfo',
-					label: 'Show model information',
+					key: 'autoMicOnEmpty',
+					label: 'Show microphone on empty input',
+					type: 'checkbox',
+					isExperimental: true
+				},
+				{
+					key: 'renderUserContentAsMarkdown',
+					label: 'Render user content as Markdown',
 					type: 'checkbox'
 				},
 				{
 					key: 'disableAutoScroll',
 					label: 'Disable automatic scroll',
 					type: 'checkbox'
-				},
-				{
-					key: 'renderUserContentAsMarkdown',
-					label: 'Render user content as Markdown',
-					type: 'checkbox'
 				}
 			]
 		},
@@ -232,11 +228,6 @@
 			title: 'Developer',
 			icon: Code,
 			fields: [
-				{
-					key: 'modelSelectorEnabled',
-					label: 'Enable model selector',
-					type: 'checkbox'
-				},
 				{
 					key: 'showToolCalls',
 					label: 'Show tool call labels',
@@ -342,7 +333,7 @@
 			}
 		}
 
-		updateMultipleConfig(processedConfig);
+		settingsStore.updateMultipleConfig(processedConfig);
 		onSave?.();
 	}
 

tools/server/webui/src/lib/components/app/chat/ChatSettings/ChatSettingsFields.svelte

@@ -6,8 +6,7 @@
 	import * as Select from '$lib/components/ui/select';
 	import { Textarea } from '$lib/components/ui/textarea';
 	import { SETTING_CONFIG_DEFAULT, SETTING_CONFIG_INFO } from '$lib/constants/settings-config';
-	import { supportsVision } from '$lib/stores/server.svelte';
-	import { getParameterInfo, resetParameterToServerDefault } from '$lib/stores/settings.svelte';
+	import { settingsStore } from '$lib/stores/settings.svelte';
 	import { ParameterSyncService } from '$lib/services/parameter-sync';
 	import { ChatSettingsParameterSourceIndicator } from '$lib/components/app';
 	import type { Component } from 'svelte';
@@ -27,7 +26,7 @@
 			return null;
 		}
 
-		return getParameterInfo(key);
+		return settingsStore.getParameterInfo(key);
 	}
 </script>
 
@@ -82,7 +81,7 @@
 					<button
 						type="button"
 						onclick={() => {
-							resetParameterToServerDefault(field.key);
+							settingsStore.resetParameterToServerDefault(field.key);
 							// Trigger UI update by calling onConfigChange with the default value
 							const defaultValue = propsDefault ?? SETTING_CONFIG_DEFAULT[field.key];
 							onConfigChange(field.key, String(defaultValue));
@@ -175,7 +174,7 @@
 						<button
 							type="button"
 							onclick={() => {
-								resetParameterToServerDefault(field.key);
+								settingsStore.resetParameterToServerDefault(field.key);
 								// Trigger UI update by calling onConfigChange with the default value
 								const defaultValue = propsDefault ?? SETTING_CONFIG_DEFAULT[field.key];
 								onConfigChange(field.key, String(defaultValue));
@@ -210,13 +209,10 @@
 				</p>
 			{/if}
 		{:else if field.type === 'checkbox'}
-			{@const isDisabled = field.key === 'pdfAsImage' && !supportsVision()}
-
 			<div class="flex items-start space-x-3">
 				<Checkbox
 					id={field.key}
 					checked={Boolean(localConfig[field.key])}
-					disabled={isDisabled}
 					onCheckedChange={(checked) => onConfigChange(field.key, checked)}
 					class="mt-1"
 				/>
@@ -224,9 +220,7 @@
 				<div class="space-y-1">
 					<label
 						for={field.key}
-						class="cursor-pointer text-sm leading-none font-medium {isDisabled
-							? 'text-muted-foreground'
-							: ''} flex items-center gap-1.5"
+						class="flex cursor-pointer items-center gap-1.5 pt-1 pb-0.5 text-sm leading-none font-medium"
 					>
 						{field.label}
 
@@ -239,11 +233,6 @@
 						<p class="text-xs text-muted-foreground">
 							{field.help || SETTING_CONFIG_INFO[field.key]}
 						</p>
-					{:else if field.key === 'pdfAsImage' && !supportsVision()}
-						<p class="text-xs text-muted-foreground">
-							PDF-to-image processing requires a vision-capable model. PDFs will be processed as
-							text.
-						</p>
 					{/if}
 				</div>
 			</div>

tools/server/webui/src/lib/components/app/chat/ChatSettings/ChatSettingsFooter.svelte

@@ -1,7 +1,7 @@
 <script lang="ts">
 	import { Button } from '$lib/components/ui/button';
 	import * as AlertDialog from '$lib/components/ui/alert-dialog';
-	import { forceSyncWithServerDefaults } from '$lib/stores/settings.svelte';
+	import { settingsStore } from '$lib/stores/settings.svelte';
 	import { RotateCcw } from '@lucide/svelte';
 
 	interface Props {
@@ -18,7 +18,7 @@
 	}
 
 	function handleConfirmReset() {
-		forceSyncWithServerDefaults();
+		settingsStore.forceSyncWithServerDefaults();
 		onReset?.();
 
 		showResetDialog = false;

tools/server/webui/src/lib/components/app/chat/ChatSettings/ChatSettingsImportExportTab.svelte

@@ -2,10 +2,9 @@
 	import { Download, Upload } from '@lucide/svelte';
 	import { Button } from '$lib/components/ui/button';
 	import { DialogConversationSelection } from '$lib/components/app';
-	import { DatabaseStore } from '$lib/stores/database';
-	import type { ExportedConversations } from '$lib/types/database';
-	import { createMessageCountMap } from '$lib/utils/conversation-utils';
-	import { chatStore } from '$lib/stores/chat.svelte';
+	import { DatabaseService } from '$lib/services/database';
+	import { createMessageCountMap } from '$lib/utils';
+	import { conversationsStore } from '$lib/stores/conversations.svelte';
 
 	let exportedConversations = $state<DatabaseConversation[]>([]);
 	let importedConversations = $state<DatabaseConversation[]>([]);
@@ -22,7 +21,7 @@
 
 	async function handleExportClick() {
 		try {
-			const allConversations = await DatabaseStore.getAllConversations();
+			const allConversations = await DatabaseService.getAllConversations();
 			if (allConversations.length === 0) {
 				alert('No conversations to export');
 				return;
@@ -30,7 +29,7 @@
 
 			const conversationsWithMessages = await Promise.all(
 				allConversations.map(async (conv) => {
-					const messages = await DatabaseStore.getConversationMessages(conv.id);
+					const messages = await DatabaseService.getConversationMessages(conv.id);
 					return { conv, messages };
 				})
 			);
@@ -48,7 +47,7 @@
 		try {
 			const allData: ExportedConversations = await Promise.all(
 				selectedConversations.map(async (conv) => {
-					const messages = await DatabaseStore.getConversationMessages(conv.id);
+					const messages = await DatabaseService.getConversationMessages(conv.id);
 					return { conv: $state.snapshot(conv), messages: $state.snapshot(messages) };
 				})
 			);
@@ -136,9 +135,9 @@
 				.snapshot(fullImportData)
 				.filter((item) => selectedIds.has(item.conv.id));
 
-			await DatabaseStore.importConversations(selectedData);
+			await DatabaseService.importConversations(selectedData);
 
-			await chatStore.loadConversations();
+			await conversationsStore.loadConversations();
 
 			importedConversations = selectedConversations;
 			showImportSummary = true;

tools/server/webui/src/lib/components/app/chat/ChatSidebar/ChatSidebar.svelte

@@ -7,11 +7,7 @@
 	import * as Sidebar from '$lib/components/ui/sidebar';
 	import * as AlertDialog from '$lib/components/ui/alert-dialog';
 	import Input from '$lib/components/ui/input/input.svelte';
-	import {
-		conversations,
-		deleteConversation,
-		updateConversationName
-	} from '$lib/stores/chat.svelte';
+	import { conversationsStore, conversations } from '$lib/stores/conversations.svelte';
 	import ChatSidebarActions from './ChatSidebarActions.svelte';
 
 	const sidebar = Sidebar.useSidebar();
@@ -56,7 +52,7 @@
 			showDeleteDialog = false;
 
 			setTimeout(() => {
-				deleteConversation(selectedConversation.id);
+				conversationsStore.deleteConversation(selectedConversation.id);
 				selectedConversation = null;
 			}, 100); // Wait for animation to finish
 		}
@@ -67,7 +63,7 @@
 
 		showEditDialog = false;
 
-		updateConversationName(selectedConversation.id, editedName);
+		conversationsStore.updateConversationName(selectedConversation.id, editedName);
 		selectedConversation = null;
 	}
 
@@ -105,7 +101,7 @@
 </script>
 
 <ScrollArea class="h-[100vh]">
-	<Sidebar.Header class=" top-0 z-10 gap-6 bg-sidebar/50 px-4 pt-4 pb-2 backdrop-blur-lg md:sticky">
+	<Sidebar.Header class=" top-0 z-10 gap-6 bg-sidebar/50 px-4 py-4 pb-2 backdrop-blur-lg md:sticky">
 		<a href="#/" onclick={handleMobileSidebarItemClick}>
 			<h1 class="inline-flex items-center gap-1 px-2 text-xl font-semibold">llama.cpp</h1>
 		</a>
@@ -154,8 +150,6 @@
 			</Sidebar.Menu>
 		</Sidebar.GroupContent>
 	</Sidebar.Group>
-
-	<div class="bottom-0 z-10 bg-sidebar bg-sidebar/50 px-4 py-4 backdrop-blur-lg md:sticky"></div>
 </ScrollArea>
 
 <DialogConfirmation

tools/server/webui/src/lib/components/app/chat/ChatSidebar/ChatSidebarConversationItem.svelte

@@ -1,7 +1,8 @@
 <script lang="ts">
 	import { Trash2, Pencil, MoreHorizontal, Download, Loader2 } from '@lucide/svelte';
 	import { ActionDropdown } from '$lib/components/app';
-	import { downloadConversation, getAllLoadingConversations } from '$lib/stores/chat.svelte';
+	import { getAllLoadingChats } from '$lib/stores/chat.svelte';
+	import { conversationsStore } from '$lib/stores/conversations.svelte';
 	import { onMount } from 'svelte';
 
 	interface Props {
@@ -25,7 +26,7 @@
 	let renderActionsDropdown = $state(false);
 	let dropdownOpen = $state(false);
 
-	let isLoading = $derived(getAllLoadingConversations().includes(conversation.id));
+	let isLoading = $derived(getAllLoadingChats().includes(conversation.id));
 
 	function handleEdit(event: Event) {
 		event.stopPropagation();
@@ -114,7 +115,7 @@
 						label: 'Export',
 						onclick: (e) => {
 							e.stopPropagation();
-							downloadConversation(conversation.id);
+							conversationsStore.downloadConversation(conversation.id);
 						},
 						shortcut: ['shift', 'cmd', 's']
 					},

tools/server/webui/src/lib/components/app/dialogs/DialogChatAttachmentPreview.svelte

@@ -1,49 +1,39 @@
 <script lang="ts">
 	import * as Dialog from '$lib/components/ui/dialog';
 	import { ChatAttachmentPreview } from '$lib/components/app';
-	import { formatFileSize } from '$lib/utils/file-preview';
+	import { formatFileSize } from '$lib/utils';
 
 	interface Props {
 		open: boolean;
+		onOpenChange?: (open: boolean) => void;
 		// Either an uploaded file or a stored attachment
 		uploadedFile?: ChatUploadedFile;
 		attachment?: DatabaseMessageExtra;
 		// For uploaded files
 		preview?: string;
 		name?: string;
-		type?: string;
 		size?: number;
 		textContent?: string;
+		// For vision modality check
+		activeModelId?: string;
 	}
 
 	let {
 		open = $bindable(),
+		onOpenChange,
 		uploadedFile,
 		attachment,
 		preview,
 		name,
-		type,
 		size,
-		textContent
+		textContent,
+		activeModelId
 	}: Props = $props();
 
 	let chatAttachmentPreviewRef: ChatAttachmentPreview | undefined = $state();
 
 	let displayName = $derived(uploadedFile?.name || attachment?.name || name || 'Unknown File');
 
-	let displayType = $derived(
-		uploadedFile?.type ||
-			(attachment?.type === 'imageFile'
-				? 'image'
-				: attachment?.type === 'textFile'
-					? 'text'
-					: attachment?.type === 'audioFile'
-						? attachment.mimeType || 'audio'
-						: attachment?.type === 'pdfFile'
-							? 'application/pdf'
-							: type || 'unknown')
-	);
-
 	let displaySize = $derived(uploadedFile?.size || size);
 
 	$effect(() => {
@@ -53,14 +43,13 @@
 	});
 </script>
 
-<Dialog.Root bind:open>
+<Dialog.Root bind:open {onOpenChange}>
 	<Dialog.Content class="grid max-h-[90vh] max-w-5xl overflow-hidden sm:w-auto sm:max-w-6xl">
 		<Dialog.Header>
-			<Dialog.Title>{displayName}</Dialog.Title>
+			<Dialog.Title class="pr-8">{displayName}</Dialog.Title>
 			<Dialog.Description>
-				{displayType}
 				{#if displaySize}
-					• {formatFileSize(displaySize)}
+					{formatFileSize(displaySize)}
 				{/if}
 			</Dialog.Description>
 		</Dialog.Header>
@@ -70,9 +59,9 @@
 			{uploadedFile}
 			{attachment}
 			{preview}
-			{name}
-			{type}
+			name={displayName}
 			{textContent}
+			{activeModelId}
 		/>
 	</Dialog.Content>
 </Dialog.Root>

tools/server/webui/src/lib/components/app/dialogs/DialogChatAttachmentsViewAll.svelte

@@ -11,6 +11,7 @@
 		imageHeight?: string;
 		imageWidth?: string;
 		imageClass?: string;
+		activeModelId?: string;
 	}
 
 	let {
@@ -21,7 +22,8 @@
 		onFileRemove,
 		imageHeight = 'h-24',
 		imageWidth = 'w-auto',
-		imageClass = ''
+		imageClass = '',
+		activeModelId
 	}: Props = $props();
 
 	let totalCount = $derived(uploadedFiles.length + attachments.length);
@@ -45,6 +47,7 @@
 				{imageHeight}
 				{imageWidth}
 				{imageClass}
+				{activeModelId}
 			/>
 		</Dialog.Content>
 	</Dialog.Portal>

tools/server/webui/src/lib/components/app/dialogs/DialogChatError.svelte

@@ -6,10 +6,11 @@
 		open: boolean;
 		type: 'timeout' | 'server';
 		message: string;
+		contextInfo?: { n_prompt_tokens: number; n_ctx: number };
 		onOpenChange?: (open: boolean) => void;
 	}
 
-	let { open = $bindable(), type, message, onOpenChange }: Props = $props();
+	let { open = $bindable(), type, message, contextInfo, onOpenChange }: Props = $props();
 
 	const isTimeout = $derived(type === 'timeout');
 	const title = $derived(isTimeout ? 'TCP Timeout' : 'Server Error');
@@ -51,6 +52,15 @@
 
 		<div class={`rounded-lg border px-4 py-3 text-sm ${badgeClass}`}>
 			<p class="font-medium">{message}</p>
+			{#if contextInfo}
+				<div class="mt-2 space-y-1 text-xs opacity-80">
+					<p>
+						<span class="font-medium">Prompt tokens:</span>
+						{contextInfo.n_prompt_tokens.toLocaleString()}
+					</p>
+					<p><span class="font-medium">Context size:</span> {contextInfo.n_ctx.toLocaleString()}</p>
+				</div>
+			{/if}
 		</div>
 
 		<AlertDialog.Footer>

tools/server/webui/src/lib/components/app/dialogs/DialogModelInformation.svelte

@@ -0,0 +1,226 @@
+<script lang="ts">
+	import * as Dialog from '$lib/components/ui/dialog';
+	import * as Table from '$lib/components/ui/table';
+	import { BadgeModality, CopyToClipboardIcon } from '$lib/components/app';
+	import { serverStore } from '$lib/stores/server.svelte';
+	import { modelsStore } from '$lib/stores/models.svelte';
+	import { ChatService } from '$lib/services/chat';
+	import { formatFileSize, formatParameters, formatNumber } from '$lib/utils';
+
+	interface Props {
+		open?: boolean;
+		onOpenChange?: (open: boolean) => void;
+	}
+
+	let { open = $bindable(), onOpenChange }: Props = $props();
+
+	let serverProps = $derived(serverStore.props);
+	let modelName = $derived(modelsStore.singleModelName);
+
+	// Get modalities from modelStore using the model ID from the first model
+	// For now it supports only for single-model mode, will be extended with further improvements for multi-model functioanlities
+	let modalities = $derived.by(() => {
+		if (!modelsData?.data?.[0]?.id) return [];
+
+		return modelsStore.getModelModalitiesArray(modelsData.data[0].id);
+	});
+
+	let modelsData = $state<ApiModelListResponse | null>(null);
+	let isLoadingModels = $state(false);
+
+	// Fetch models data when dialog opens
+	$effect(() => {
+		if (open && !modelsData) {
+			loadModelsData();
+		}
+	});
+
+	async function loadModelsData() {
+		isLoadingModels = true;
+
+		try {
+			modelsData = await ChatService.getModels();
+		} catch (error) {
+			console.error('Failed to load models data:', error);
+			// Set empty data to prevent infinite loading
+			modelsData = { object: 'list', data: [] };
+		} finally {
+			isLoadingModels = false;
+		}
+	}
+</script>
+
+<Dialog.Root bind:open {onOpenChange}>
+	<Dialog.Content class="@container z-9999 !max-w-[60rem] max-w-full">
+		<style>
+			@container (max-width: 56rem) {
+				.resizable-text-container {
+					max-width: calc(100vw - var(--threshold));
+				}
+			}
+		</style>
+
+		<Dialog.Header>
+			<Dialog.Title>Model Information</Dialog.Title>
+			<Dialog.Description>Current model details and capabilities</Dialog.Description>
+		</Dialog.Header>
+
+		<div class="space-y-6 py-4">
+			{#if isLoadingModels}
+				<div class="flex items-center justify-center py-8">
+					<div class="text-sm text-muted-foreground">Loading model information...</div>
+				</div>
+			{:else if modelsData && modelsData.data.length > 0}
+				{@const modelMeta = modelsData.data[0].meta}
+
+				{#if serverProps}
+					<Table.Root>
+						<Table.Header>
+							<Table.Row>
+								<Table.Head class="w-[10rem]">Model</Table.Head>
+
+								<Table.Head>
+									<div class="inline-flex items-center gap-2">
+										<span
+											class="resizable-text-container min-w-0 flex-1 truncate"
+											style:--threshold="12rem"
+										>
+											{modelName}
+										</span>
+
+										<CopyToClipboardIcon
+											text={modelName || ''}
+											canCopy={!!modelName}
+											ariaLabel="Copy model name to clipboard"
+										/>
+									</div>
+								</Table.Head>
+							</Table.Row>
+						</Table.Header>
+						<Table.Body>
+							<!-- Model Path -->
+							<Table.Row>
+								<Table.Cell class="h-10 align-middle font-medium">File Path</Table.Cell>
+
+								<Table.Cell
+									class="inline-flex h-10 items-center gap-2 align-middle font-mono text-xs"
+								>
+									<span
+										class="resizable-text-container min-w-0 flex-1 truncate"
+										style:--threshold="14rem"
+									>
+										{serverProps.model_path}
+									</span>
+
+									<CopyToClipboardIcon
+										text={serverProps.model_path}
+										ariaLabel="Copy model path to clipboard"
+									/>
+								</Table.Cell>
+							</Table.Row>
+
+							<!-- Context Size -->
+							<Table.Row>
+								<Table.Cell class="h-10 align-middle font-medium">Context Size</Table.Cell>
+								<Table.Cell
+									>{formatNumber(serverProps.default_generation_settings.n_ctx)} tokens</Table.Cell
+								>
+							</Table.Row>
+
+							<!-- Training Context -->
+							{#if modelMeta?.n_ctx_train}
+								<Table.Row>
+									<Table.Cell class="h-10 align-middle font-medium">Training Context</Table.Cell>
+									<Table.Cell>{formatNumber(modelMeta.n_ctx_train)} tokens</Table.Cell>
+								</Table.Row>
+							{/if}
+
+							<!-- Model Size -->
+							{#if modelMeta?.size}
+								<Table.Row>
+									<Table.Cell class="h-10 align-middle font-medium">Model Size</Table.Cell>
+									<Table.Cell>{formatFileSize(modelMeta.size)}</Table.Cell>
+								</Table.Row>
+							{/if}
+
+							<!-- Parameters -->
+							{#if modelMeta?.n_params}
+								<Table.Row>
+									<Table.Cell class="h-10 align-middle font-medium">Parameters</Table.Cell>
+									<Table.Cell>{formatParameters(modelMeta.n_params)}</Table.Cell>
+								</Table.Row>
+							{/if}
+
+							<!-- Embedding Size -->
+							{#if modelMeta?.n_embd}
+								<Table.Row>
+									<Table.Cell class="align-middle font-medium">Embedding Size</Table.Cell>
+									<Table.Cell>{formatNumber(modelMeta.n_embd)}</Table.Cell>
+								</Table.Row>
+							{/if}
+
+							<!-- Vocabulary Size -->
+							{#if modelMeta?.n_vocab}
+								<Table.Row>
+									<Table.Cell class="align-middle font-medium">Vocabulary Size</Table.Cell>
+									<Table.Cell>{formatNumber(modelMeta.n_vocab)} tokens</Table.Cell>
+								</Table.Row>
+							{/if}
+
+							<!-- Vocabulary Type -->
+							{#if modelMeta?.vocab_type}
+								<Table.Row>
+									<Table.Cell class="align-middle font-medium">Vocabulary Type</Table.Cell>
+									<Table.Cell class="align-middle capitalize">{modelMeta.vocab_type}</Table.Cell>
+								</Table.Row>
+							{/if}
+
+							<!-- Total Slots -->
+							<Table.Row>
+								<Table.Cell class="align-middle font-medium">Parallel Slots</Table.Cell>
+								<Table.Cell>{serverProps.total_slots}</Table.Cell>
+							</Table.Row>
+
+							<!-- Modalities -->
+							{#if modalities.length > 0}
+								<Table.Row>
+									<Table.Cell class="align-middle font-medium">Modalities</Table.Cell>
+									<Table.Cell>
+										<div class="flex flex-wrap gap-1">
+											<BadgeModality {modalities} />
+										</div>
+									</Table.Cell>
+								</Table.Row>
+							{/if}
+
+							<!-- Build Info -->
+							<Table.Row>
+								<Table.Cell class="align-middle font-medium">Build Info</Table.Cell>
+								<Table.Cell class="align-middle font-mono text-xs"
+									>{serverProps.build_info}</Table.Cell
+								>
+							</Table.Row>
+
+							<!-- Chat Template -->
+							{#if serverProps.chat_template}
+								<Table.Row>
+									<Table.Cell class="align-middle font-medium">Chat Template</Table.Cell>
+									<Table.Cell class="py-10">
+										<div class="max-h-120 overflow-y-auto rounded-md bg-muted p-4">
+											<pre
+												class="font-mono text-xs whitespace-pre-wrap">{serverProps.chat_template}</pre>
+										</div>
+									</Table.Cell>
+								</Table.Row>
+							{/if}
+						</Table.Body>
+					</Table.Root>
+				{/if}
+			{:else if !isLoadingModels}
+				<div class="flex items-center justify-center py-8">
+					<div class="text-sm text-muted-foreground">No model information available</div>
+				</div>
+			{/if}
+		</div>
+	</Dialog.Content>
+</Dialog.Root>

tools/server/webui/src/lib/components/app/dialogs/DialogModelNotAvailable.svelte

@@ -0,0 +1,76 @@
+<script lang="ts">
+	import * as AlertDialog from '$lib/components/ui/alert-dialog';
+	import { AlertTriangle, ArrowRight } from '@lucide/svelte';
+	import { goto } from '$app/navigation';
+	import { page } from '$app/state';
+
+	interface Props {
+		open: boolean;
+		modelName: string;
+		availableModels?: string[];
+		onOpenChange?: (open: boolean) => void;
+	}
+
+	let { open = $bindable(), modelName, availableModels = [], onOpenChange }: Props = $props();
+
+	function handleOpenChange(newOpen: boolean) {
+		open = newOpen;
+		onOpenChange?.(newOpen);
+	}
+
+	function handleSelectModel(model: string) {
+		// Build URL with selected model, preserving other params
+		const url = new URL(page.url);
+		url.searchParams.set('model', model);
+
+		handleOpenChange(false);
+		goto(url.toString());
+	}
+</script>
+
+<AlertDialog.Root {open} onOpenChange={handleOpenChange}>
+	<AlertDialog.Content class="max-w-lg">
+		<AlertDialog.Header>
+			<AlertDialog.Title class="flex items-center gap-2">
+				<AlertTriangle class="h-5 w-5 text-amber-500" />
+				Model Not Available
+			</AlertDialog.Title>
+
+			<AlertDialog.Description>
+				The requested model could not be found. Select an available model to continue.
+			</AlertDialog.Description>
+		</AlertDialog.Header>
+
+		<div class="space-y-3">
+			<div class="rounded-lg border border-amber-500/40 bg-amber-500/10 px-4 py-3 text-sm">
+				<p class="font-medium text-amber-600 dark:text-amber-400">
+					Requested: <code class="rounded bg-amber-500/20 px-1.5 py-0.5">{modelName}</code>
+				</p>
+			</div>
+
+			{#if availableModels.length > 0}
+				<div class="text-sm">
+					<p class="mb-2 font-medium text-muted-foreground">Select an available model:</p>
+					<div class="max-h-48 space-y-1 overflow-y-auto rounded-md border p-1">
+						{#each availableModels as model (model)}
+							<button
+								type="button"
+								class="group flex w-full items-center justify-between gap-2 rounded-sm px-3 py-2 text-left text-sm transition-colors hover:bg-accent hover:text-accent-foreground"
+								onclick={() => handleSelectModel(model)}
+							>
+								<span class="min-w-0 truncate font-mono text-xs">{model}</span>
+								<ArrowRight
+									class="h-4 w-4 shrink-0 text-muted-foreground opacity-0 transition-opacity group-hover:opacity-100"
+								/>
+							</button>
+						{/each}
+					</div>
+				</div>
+			{/if}
+		</div>
+
+		<AlertDialog.Footer>
+			<AlertDialog.Action onclick={() => handleOpenChange(false)}>Cancel</AlertDialog.Action>
+		</AlertDialog.Footer>
+	</AlertDialog.Content>
+</AlertDialog.Root>