use destructor instead

Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>

CODEOWNERS

@@ -10,7 +10,7 @@
 # ggml-org/ggml-rpc         : rgerganov
 # ggml-org/ggml-sycl        : arthw
 # ggml-org/ggml-vulkan      : 0cc4m, jeffbolznv
-# ggml-org/ggml-webgpu      : reeselevine
+# ggml-org/ggml-webgpu      : reeselevine, yomaytk
 # ggml-org/ggml-zdnn        : taronaeo
 # ggml-org/llama-common     : ggerganov, aldehir, angt, danbev, ngxson, pwilkin
 # ggml-org/llama-mtmd       : ngxson

common/arg.cpp

@@ -301,6 +301,8 @@ static handle_model_result common_params_handle_model(struct common_params_model
                                                       const common_download_opts & opts) {
     handle_model_result result;
 
+    // TODO @ngxson : refactor this into a new common_model_download_context
+
     if (!model.docker_repo.empty()) {
         model.path = common_docker_resolve_model(model.docker_repo);
     } else if (!model.hf_repo.empty()) {
@@ -396,7 +398,7 @@ static bool parse_bool_value(const std::string & value) {
 // CLI argument parsing functions
 //
 
-bool common_params_handle_models(common_params & params, llama_example curr_ex) {
+bool common_params_handle_models(common_params & params, llama_example curr_ex, const common_params_handle_models_params & handle_params) {
     const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
                                          params.speculative.types.end(),
                                          COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
@@ -407,6 +409,11 @@ bool common_params_handle_models(common_params & params, llama_example curr_ex)
     opts.skip_download   = params.skip_download;
     opts.download_mtp    = spec_type_draft_mtp;
     opts.download_mmproj = !params.no_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty();
+    opts.preset_only     = handle_params.preset_only;
+
+    if (handle_params.callback) {
+        opts.callback = handle_params.callback;
+    }
 
     // sub-models (draft, mmproj, vocoder) are explicitly specified by the user,
     // so we should not auto-discover mtp/mmproj siblings for them
@@ -584,19 +591,20 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
-    // export_graph_ops loads only metadata
-    const bool skip_model_download = ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;
+    const bool skip_model_download =
+        // server will call common_params_handle_models() later, so we skip it here
+        ctx_arg.ex == LLAMA_EXAMPLE_SERVER ||
+        // export_graph_ops loads only metadata
+        ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;
 
     if (!skip_model_download) {
         // handle model and download
-        common_params_handle_models(params, ctx_arg.ex);
+        common_params_handle_models(params, ctx_arg.ex, {});
 
         // 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
-                && !params.usage
-                && !params.completion) {
+        bool can_skip_model = params.usage || params.completion || !params.server_base.empty();
+        if (!can_skip_model && params.model.path.empty()) {
             throw std::invalid_argument("error: --model is required\n");
         }
     }
@@ -1110,6 +1118,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.completion = true;
         }
     ));
+    add_opt(common_arg(
+        {"--server-base"}, "URL",
+        string_format("connect to this server instead of starting a new one, example: 'http://localhost:8080' (default: none)"),
+        [](common_params & params, const std::string & value) {
+            params.server_base = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"--verbose-prompt"},
         string_format("print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false"),

common/arg.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "common.h"
+#include "download.h"
 
 #include <set>
 #include <map>
@@ -129,11 +130,19 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
 // see: https://github.com/ggml-org/llama.cpp/issues/18163
 void common_params_add_preset_options(std::vector<common_arg> & args);
 
+struct common_params_handle_models_params {
+    common_download_callback * callback = nullptr;
+    bool preset_only = false; // if true, only check & download remote preset (for router mode)
+};
+
 // populate model paths (main model, mmproj, etc) from -hf if necessary
 // return true if the model is ready to use
 // throw an exception if there is an error that prevents the model from being used (e.g. network error, model not found, etc)
 // if params.skip_download is true, no downloads will be attempted. return false if the model is invalid or missing (e.g. ETag check failed)
-bool common_params_handle_models(common_params & params, llama_example curr_ex);
+bool common_params_handle_models(
+    common_params & params,
+    llama_example curr_ex,
+    const common_params_handle_models_params & handle_params);
 
 // initialize argument parser context - used by test-arg-parser and preset
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/chat-auto-parser-generator.cpp

@@ -395,10 +395,11 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
                                            arguments.name_suffix) +
                            arguments.value_prefix +
                            (schema_info.resolves_to_string(param_schema) ?
-                                p.tool_arg_string_value(until_suffix) :
-                                p.tool_arg_json_value(p.schema(
-                                    p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false))) +
-                           p.tool_arg_close(p.literal(arguments.value_suffix)));
+                                p.ac(p.tool_arg_string_value(until_suffix) +
+                                    p.tool_arg_close(p.literal(arguments.value_suffix)), arguments.value_suffix) :
+                                (p.tool_arg_json_value(p.schema(
+                                    p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
+                                    p.tool_arg_close(p.literal(arguments.value_suffix)))));
 
             auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
             if (is_required) {

common/chat.cpp

@@ -90,41 +90,93 @@ std::string common_chat_msg::render_content(const std::string & delimiter) const
     return text;
 }
 
-std::vector<common_chat_msg_span> common_chat_split_by_role(const std::string & prompt, const std::vector<common_chat_msg_delimiter> & delims) {
-    if (delims.empty() || prompt.empty()) {
-        return {};
+common_chat_role common_chat_role_from_string(const std::string & role) {
+    if (role == "system")    { return COMMON_CHAT_ROLE_SYSTEM;    }
+    if (role == "assistant") { return COMMON_CHAT_ROLE_ASSISTANT; }
+    if (role == "user")      { return COMMON_CHAT_ROLE_USER;      }
+    if (role == "tool")      { return COMMON_CHAT_ROLE_TOOL;      }
+    return COMMON_CHAT_ROLE_UNKNOWN;
+}
+
+const char * common_chat_role_to_string(common_chat_role role) {
+    switch (role) {
+        case COMMON_CHAT_ROLE_SYSTEM:    return "system";
+        case COMMON_CHAT_ROLE_ASSISTANT: return "assistant";
+        case COMMON_CHAT_ROLE_USER:      return "user";
+        case COMMON_CHAT_ROLE_TOOL:      return "tool";
+        case COMMON_CHAT_ROLE_UNKNOWN:   return "";
+    }
+    return "";
+}
+
+json common_chat_msg_delimiters::to_json() const {
+    json result = json::array();
+    for (const auto & d : delimiters) {
+        result.push_back({
+            { "role",      common_chat_role_to_string(d.role) },
+            { "delimiter", d.delimiter                        },
+        });
+    }
+    return result;
+}
+
+common_chat_msg_delimiters common_chat_msg_delimiters_parse(const json & delimiters) {
+    common_chat_msg_delimiters result;
+
+    if (!delimiters.is_array()) {
+        return result;
     }
 
-    auto parser = build_peg_parser([&](common_peg_parser_builder & p) {
-        std::vector<std::string>       all_delims;
-        std::vector<common_peg_parser> tagged_messages;
-
-        all_delims.reserve(delims.size());
-        tagged_messages.reserve(delims.size());
-        for (const auto & d : delims) {
-            all_delims.push_back(d.delimiter);
+    result.delimiters.reserve(delimiters.size());
+    for (const auto & d : delimiters) {
+        if (!d.is_object()) {
+            continue;
         }
-
-        auto any_delim = p.until_one_of(all_delims);
-        for (const auto & d : delims) {
-            tagged_messages.push_back(p.tag(d.role, p.literal(d.delimiter) + any_delim));
-        }
-
-        return any_delim + p.zero_or_more(p.choice(tagged_messages)) + p.end();
-    });
-
-    common_peg_parse_context ctx(prompt);
-    const auto result = parser.parse(ctx);
-    if (!result.success()) {
-        return {};
+        result.delimiters.push_back({
+            common_chat_role_from_string(d.value("role", std::string())),
+            d.value("delimiter", std::string()),
+        });
     }
 
-    std::vector<common_chat_msg_span> spans;
-    ctx.ast.visit(result, [&](const common_peg_ast_node & node) {
-        if (!node.tag.empty()) {
-            spans.push_back({ node.tag, node.start, node.end - node.start });
+    return result;
+}
+
+void common_chat_msg_delimiters::tokenize(const llama_vocab * vocab) {
+    for (auto & d : delimiters) {
+        d.tokens = common_tokenize(vocab, d.delimiter, false, true);
+    }
+}
+
+common_chat_msg_spans common_chat_msg_delimiters::split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips) const {
+    std::vector<std::pair<common_chat_role, size_t>> matches;
+
+    auto skip = skips.begin();
+    for (size_t i = 0; i < tokens.size();) {
+        if (skip != skips.end() && i == skip->first) {
+            i += skip->second;
+            ++skip;
+            continue;
         }
-    });
+        for (const auto & d : delimiters) {
+            if (i + d.tokens.size() > tokens.size()) {
+                continue;
+            }
+            if (std::equal(d.tokens.begin(), d.tokens.end(), tokens.begin() + i)) {
+                matches.emplace_back(d.role, i);
+                break;
+            }
+        }
+        i++;
+    }
+
+    matches.emplace_back(COMMON_CHAT_ROLE_UNKNOWN, tokens.size());
+
+    common_chat_msg_spans spans;
+    for (size_t i = 0; i + 1 < matches.size(); i++) {
+        const auto & curr = matches[i];
+        const auto & next = matches[i + 1];
+        spans.add(curr.first, curr.second, next.second - curr.second);
+    }
 
     return spans;
 }
@@ -1081,13 +1133,13 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
 
     data.prompt            = prompt;
     data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
-    data.message_spans = common_chat_split_by_role(prompt, {
-        { "assistant", "<|start|>assistant" },
-        { "user",      "<|start|>user"      },
-        { "system",    "<|start|>developer" },
-        { "system",    "<|start|>system"    },
-        { "tool",      "<|start|>functions" },
-    });
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
+        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>developer" },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
+        { COMMON_CHAT_ROLE_TOOL,      "<|start|>functions" },
+    };
 
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.supports_thinking = true;
@@ -1228,10 +1280,10 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
         data.prompt += data.generation_prompt;
     }
 
-    data.message_spans = common_chat_split_by_role(data.prompt, {
-        { "user",      "<|turn>user\n"  },
-        { "assistant", "<|turn>model\n" },
-    });
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_USER,      "<|turn>user"  },
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|turn>model" },
+    };
 
     data.format            = COMMON_CHAT_FORMAT_PEG_GEMMA4;
     data.supports_thinking  = true;
@@ -2030,15 +2082,15 @@ static common_chat_params common_chat_params_init_cohere2moe(const common_chat_t
         RESULT_START, RESULT_END,
     };
 
-    // Split the rendered prompt into per-role message spans. Tool results are rendered with the
+    // Declare per-role message delimiters. Tool results are rendered with the
     // system token followed by <|START_TOOL_RESULT|>, so the "tool" delimiter must be listed before
     // the plain "system" one (it is a strict superset, and the role split tries delimiters in order).
-    data.message_spans = common_chat_split_by_role(data.prompt, {
-        { "assistant", GEN_PREFIX },
-        { "user",      TURN_START + USER },
-        { "tool",      TURN_START + SYSTEM + RESULT_START },
-        { "system",    TURN_START + SYSTEM },
-    });
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, GEN_PREFIX },
+        { COMMON_CHAT_ROLE_USER,      TURN_START + USER },
+        { COMMON_CHAT_ROLE_TOOL,      TURN_START + SYSTEM + RESULT_START },
+        { COMMON_CHAT_ROLE_SYSTEM,    TURN_START + SYSTEM },
+    };
 
     auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
     auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
@@ -2526,17 +2578,15 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
         autoparser.analyze_template(tmpl);
         auto auto_params = autoparser::peg_generator::generate_parser(tmpl, params, autoparser);
 
-        std::vector<common_chat_msg_delimiter> delimiters;
+        common_chat_msg_delimiters delimiters;
         if (!autoparser.assistant_start.empty()) {
-            delimiters.push_back({ "assistant", autoparser.assistant_start });
+            delimiters.add(COMMON_CHAT_ROLE_ASSISTANT, autoparser.assistant_start);
         }
         if (!autoparser.user_start.empty()) {
-            delimiters.push_back({ "user", autoparser.user_start });
+            delimiters.add(COMMON_CHAT_ROLE_USER, autoparser.user_start);
         }
 
-        if (!delimiters.empty()) {
-            auto_params.message_spans = common_chat_split_by_role(auto_params.prompt, delimiters);
-        }
+        auto_params.message_delimiters = std::move(delimiters);
 
         auto_params.supports_thinking = autoparser.reasoning.mode != autoparser::reasoning_mode::NONE;
         if (auto_params.supports_thinking) {

common/chat.h

@@ -143,15 +143,75 @@ struct common_chat_msg_diff {
     }
 };
 
+enum common_chat_role {
+    COMMON_CHAT_ROLE_UNKNOWN,
+    COMMON_CHAT_ROLE_SYSTEM,
+    COMMON_CHAT_ROLE_ASSISTANT,
+    COMMON_CHAT_ROLE_USER,
+    COMMON_CHAT_ROLE_TOOL
+};
+
+common_chat_role common_chat_role_from_string(const std::string & role);
+const char *     common_chat_role_to_string(common_chat_role role);
+
 struct common_chat_msg_span {
-    std::string role;
+    common_chat_role role = COMMON_CHAT_ROLE_UNKNOWN;
     std::size_t pos = 0;
     std::size_t len = 0;
+
+    bool valid() const {
+        return role != COMMON_CHAT_ROLE_UNKNOWN;
+    }
+};
+
+struct common_chat_msg_spans {
+    std::vector<common_chat_msg_span> spans;
+
+    void add(common_chat_role role, size_t pos, size_t len) {
+        spans.push_back({ role, pos, len });
+    }
+
+    bool is_user_start(int32_t pos) const {
+        for (auto it = spans.begin(); it != spans.end(); ++it) {
+            if (it->role == COMMON_CHAT_ROLE_USER && pos == (int32_t) it->pos) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    int32_t last_user_message_pos() const {
+        for (auto it = spans.rbegin(); it != spans.rend(); ++it) {
+            if (it->role == COMMON_CHAT_ROLE_USER) {
+                return (int32_t) it->pos;
+            }
+        }
+        return -1;
+    }
 };
 
 struct common_chat_msg_delimiter {
-    std::string role;
-    std::string delimiter;
+    common_chat_role role = COMMON_CHAT_ROLE_UNKNOWN;
+    std::string      delimiter;
+    llama_tokens     tokens = {};
+};
+
+struct common_chat_msg_delimiters {
+    std::vector<common_chat_msg_delimiter> delimiters;
+
+    common_chat_msg_delimiters() = default;
+    common_chat_msg_delimiters(std::initializer_list<common_chat_msg_delimiter> delims) : delimiters(delims) {}
+
+    void add(common_chat_role role, const std::string & delimiter) {
+        delimiters.push_back({ role, delimiter });
+    }
+
+    void tokenize(const llama_vocab * vocab);
+
+    // split tokens into message spans. skips maps a start index to a length of a region to jump over without matching
+    common_chat_msg_spans split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips = {}) const;
+
+    nlohmann::ordered_json to_json() const;
 };
 
 struct common_chat_tool {
@@ -219,7 +279,7 @@ struct common_chat_params {
     std::vector<std::string>            preserved_tokens;
     std::vector<std::string>            additional_stops;
     std::string                         parser;
-    std::vector<common_chat_msg_span>   message_spans;
+    common_chat_msg_delimiters          message_delimiters;
 };
 
 // per-message parsing syntax
@@ -325,5 +385,4 @@ struct common_chat_prompt_preset {
 
 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);
 
-std::vector<common_chat_msg_span> common_chat_split_by_role(const std::string & prompt, const std::vector<common_chat_msg_delimiter> & delims);
-
+common_chat_msg_delimiters common_chat_msg_delimiters_parse(const nlohmann::ordered_json & delimiters);

common/common.h

@@ -609,7 +609,7 @@ struct common_params {
     bool    cache_prompt        = true;  // whether to enable prompt caching
     bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
     int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
-    int32_t checkpoint_min_step = 256;   // minimum spacing between context checkpoints
+    int32_t checkpoint_min_step = 8192;  // minimum spacing between context checkpoints
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
     std::string hostname      = "127.0.0.1";
@@ -631,6 +631,9 @@ struct common_params {
 
     std::map<std::string, std::string> default_template_kwargs;
 
+    // CLI params
+    std::string server_base; // if set, connect to this server instead of starting a new one
+
     // UI configs
     bool ui = true;
     bool ui_mcp_proxy = false;

common/download.cpp

@@ -799,6 +799,7 @@ common_download_model_result common_download_model(const common_params_model  &
 
     bool download_mmproj = opts.download_mmproj;
     bool download_mtp = opts.download_mtp;
+    bool preset_only = opts.preset_only;
     bool is_hf = !model.hf_repo.empty();
 
     if (is_hf) {
@@ -806,7 +807,8 @@ common_download_model_result common_download_model(const common_params_model  &
         if (!hf.preset.path.empty()) {
             // if preset.ini exists, only download that file alone
             tasks.push_back({hf.preset.url, hf.preset.local_path});
-        } else {
+        } else if (!preset_only) {
+            // only add other files if we're NOT in preset-only mode (normal run, non-router)
             for (const auto & f : hf.model_files) {
                 tasks.push_back({f.url, f.local_path});
             }

common/download.h

@@ -55,6 +55,7 @@ struct common_download_opts {
     bool skip_download = false; // if true, only validation is performed, common_skip_download_exception may be thrown if the file is missing or invalid
     bool download_mmproj = false;
     bool download_mtp = false;
+    bool preset_only = false; // if true, only check & download remote preset (for router mode)
     common_download_callback * callback = nullptr;
 };
 

common/http.h

@@ -2,6 +2,16 @@
 
 #include <cpp-httplib/httplib.h>
 
+#ifdef _WIN32
+#include <winsock2.h>
+#include <windows.h>
+#else
+#include <sys/socket.h>
+#include <netinet/in.h>
+#include <arpa/inet.h>
+#include <unistd.h>
+#endif
+
 struct common_http_url {
     std::string scheme;
     std::string user;
@@ -97,3 +107,63 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
 static std::string common_http_show_masked_url(const common_http_url & parts) {
     return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + parts.host + parts.path;
 }
+
+static int common_http_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;
+}

common/peg-parser.cpp

@@ -921,6 +921,10 @@ struct parser_executor {
     common_peg_parse_result operator()(const common_peg_gbnf_parser & p) {
         return arena.parse(p.child, ctx, start_pos);
     }
+
+    common_peg_parse_result operator()(const common_peg_ac_parser & p) {
+        return arena.parse(p.child, ctx, start_pos);
+    }
 };
 
 common_peg_parse_result common_peg_arena::parse(common_peg_parse_context & ctx, size_t start) const {
@@ -989,7 +993,8 @@ void common_peg_arena::resolve_refs() {
                                  std::is_same_v<T, common_peg_not_parser> ||
                                  std::is_same_v<T, common_peg_tag_parser> ||
                                  std::is_same_v<T, common_peg_atomic_parser> ||
-                                 std::is_same_v<T, common_peg_gbnf_parser>) {
+                                 std::is_same_v<T, common_peg_gbnf_parser> ||
+                                 std::is_same_v<T, common_peg_ac_parser>) {
                 p.child = resolve_ref(p.child);
             } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
                 p.child = resolve_ref(p.child);
@@ -1070,6 +1075,8 @@ std::string common_peg_arena::dump_impl(common_peg_parser_id
             return "Atomic(" + dump_impl(p.child, visited) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
             return "Gbnf(" + p.grammar + ", " + dump_impl(p.child, visited) + ")";
+        } else if constexpr (std::is_same_v<T, common_peg_ac_parser>) {
+            return "Ac(" + string_join(p.delimiters, " | ") + ", " + dump_impl(p.child, visited) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_any_parser>) {
             return "Any";
         } else if constexpr (std::is_same_v<T, common_peg_space_parser>) {
@@ -1479,6 +1486,13 @@ common_peg_parser common_peg_parser_builder::json_member(const std::string & key
     });
 }
 
+common_peg_parser common_peg_parser_builder::ac(const common_peg_parser & p, const std::vector<std::string> & delimiters) {
+    if (delimiters.empty()) {
+        throw std::runtime_error("ac parser requires at least one delimiter");
+    }
+    return add(common_peg_ac_parser{p, delimiters});
+}
+
 static std::string gbnf_escape_char_class(uint32_t c) {
     if (c == '-' || c == ']' || c == '[' || c == '\\') {
         return "\\" + std::string(1, (char) c);
@@ -1529,14 +1543,22 @@ static std::string gbnf_escape_char_class(uint32_t c) {
     return std::string(buf);
 }
 
-// GBNF grammar matching strings that contain no string in `strings` as a
-// substring. Emits the complement of an Aho-Corasick automaton DFA and returns
-// the start state rule name.
-//
-// ref: https://github.com/ggml-org/llama.cpp/pull/24839
-static std::string gbnf_excluding_grammar(const common_grammar_builder & builder,
-                                          const std::string &            prefix,
-                                          const std::vector<std::string> & strings) {
+static std::string gbnf_char_class(const std::vector<uint32_t> & chars, bool negate) {
+    std::string s = negate ? "[^" : "[";
+    for (uint32_t ch : chars) {
+        s += gbnf_escape_char_class(ch);
+    }
+    return s + "]";
+}
+
+static std::string gbnf_ac_grammar(
+    const common_grammar_builder &   builder,
+    const std::string &              prefix,
+    const std::vector<std::string> & strings,
+    const std::function<std::string(const std::vector<uint32_t> &,
+                                    const std::map<size_t, std::vector<uint32_t>> &,
+                                    const std::vector<uint32_t> &,
+                                    const std::function<std::string(size_t)> &)> & build_rule) {
     aho_corasick ac(strings);
 
     auto state_name = [&](size_t s) -> std::string {
@@ -1548,42 +1570,30 @@ static std::string gbnf_excluding_grammar(const common_grammar_builder & builder
         return prefix + "-" + num;
     };
 
-    auto char_class = [](const std::vector<uint32_t> & chars, bool negate) {
-        std::string s = negate ? "[^" : "[";
-        for (uint32_t ch : chars) {
-            s += gbnf_escape_char_class(ch);
-        }
-        return s + "]";
-    };
-
     for (size_t q = 0; q < ac.num_states(); q++) {
         if (ac.is_terminal(q)) {
-            continue; // match states are dropped
+            continue; // match states
         }
 
         std::map<size_t, std::vector<uint32_t>> buckets;
-        std::vector<uint32_t> excluded;
+        std::vector<uint32_t> completing;  // chars that complete a delimiter
+        std::vector<uint32_t> specific;    // chars with an explicit transition
         for (uint32_t c : ac.alphabet) {
             size_t d = ac.next(q, c);
             if (ac.is_terminal(d)) {
-                excluded.push_back(c); // completes a forbidden string -> omit
+                completing.push_back(c);
+                specific.push_back(c);
             } else if (d != 0) {
                 buckets[d].push_back(c); // specific non-root destination
-                excluded.push_back(c);
+                specific.push_back(c);
             }
         }
 
-        std::string rhs = "|"; // every state is accepting
-        for (const auto & [d, chars] : buckets) {
-            rhs += " " + char_class(chars, false) + " " + state_name(d) + " |";
-        }
-        rhs += " " + char_class(excluded, true) + " " + state_name(0);
-
-        builder.add_rule(state_name(q), rhs);
+        builder.add_rule(state_name(q), build_rule(completing, buckets, specific, state_name));
     }
 
     // An empty delimiter makes the start state terminal. Emit an entry rule
-    // that matches nothing so the returned reference stays valid.
+    // that matches the empty string so the returned reference stays valid.
     if (ac.is_terminal(0)) {
         builder.add_rule(prefix, "|");
     }
@@ -1591,6 +1601,54 @@ static std::string gbnf_excluding_grammar(const common_grammar_builder & builder
     return state_name(0);
 }
 
+// GBNF grammar matching strings that contain no string in `strings` as a
+// substring. Emits the complement of an Aho-Corasick automaton DFA and returns
+// the start state rule name.
+//
+// ref: https://github.com/ggml-org/llama.cpp/pull/24839
+static std::string gbnf_excluding_grammar(const common_grammar_builder & builder,
+                                          const std::string &            prefix,
+                                          const std::vector<std::string> & strings) {
+    return gbnf_ac_grammar(builder, prefix, strings,
+        [](const std::vector<uint32_t> & /*completing*/,
+           const std::map<size_t, std::vector<uint32_t>> & buckets,
+           const std::vector<uint32_t> & specific,
+           const std::function<std::string(size_t)> & state_name) {
+            // every state is accepting and completing chars get no
+            // alternative, so a forbidden string can never be matched
+            std::string rhs = "|";
+            for (const auto & [d, chars] : buckets) {
+                rhs += " " + gbnf_char_class(chars, false) + " " + state_name(d) + " |";
+            }
+            rhs += " " + gbnf_char_class(specific, true) + " " + state_name(0);
+            return rhs;
+        });
+}
+
+// GBNF grammar matching everything up to and including the first occurrence of
+// any string in `strings`. Emits the Aho-Corasick automaton DFA and returns
+// the start state rule name.
+static std::string gbnf_including_grammar(const common_grammar_builder & builder,
+                                          const std::string &            prefix,
+                                          const std::vector<std::string> & strings) {
+    return gbnf_ac_grammar(builder, prefix, strings,
+        [](const std::vector<uint32_t> & completing,
+           const std::map<size_t, std::vector<uint32_t>> & buckets,
+           const std::vector<uint32_t> & specific,
+           const std::function<std::string(size_t)> & state_name) {
+            std::vector<std::string> alts;
+            if (!completing.empty()) {
+                alts.push_back(gbnf_char_class(completing, false)); // terminate on match
+            }
+            for (const auto & [d, chars] : buckets) {
+                alts.push_back(gbnf_char_class(chars, false) + " " + state_name(d));
+            }
+            // every other character keeps scanning from the start state
+            alts.push_back(gbnf_char_class(specific, true) + " " + state_name(0));
+            return string_join(alts, " | ");
+        });
+}
+
 static std::set<std::string> collect_reachable_rules(
     const common_peg_arena & arena,
     const common_peg_parser_id & rule
@@ -1628,6 +1686,7 @@ static std::set<std::string> collect_reachable_rules(
                                  std::is_same_v<T, common_peg_tag_parser> ||
                                  std::is_same_v<T, common_peg_atomic_parser> ||
                                  std::is_same_v<T, common_peg_gbnf_parser> ||
+                                 std::is_same_v<T, common_peg_ac_parser> ||
                                  std::is_same_v<T, common_peg_schema_parser>) {
                 visit(p.child);
             } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1822,6 +1881,8 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
                 return to_gbnf(p.child);
             } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
                 return p.grammar;
+            } else if constexpr (std::is_same_v<T, common_peg_ac_parser>) {
+                return gbnf_including_grammar(builder, "ac-" + std::to_string(id), p.delimiters);
             } else {
                 static_assert(is_always_false_v<T>);
             }
@@ -1958,6 +2019,8 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
             };
         } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
             return json{{"type", "gbnf"}, {"child", p.child}, {"grammar", p.grammar}};
+        } else if constexpr (std::is_same_v<T, common_peg_ac_parser>) {
+            return json{{"type", "ac"}, {"child", p.child}, {"delimiters", p.delimiters}};
         }
     }, variant);
 }
@@ -2130,6 +2193,16 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         };
     }
 
+    if (type == "ac") {
+        if (!j.contains("child") || !j.contains("delimiters") || !j["delimiters"].is_array() || j["delimiters"].empty()) {
+            throw std::runtime_error("ac parser requires 'child' and a non-empty 'delimiters' array");
+        }
+        return common_peg_ac_parser{
+            j["child"].get<common_peg_parser_id>(),
+            j["delimiters"].get<std::vector<std::string>>(),
+        };
+    }
+
     throw std::runtime_error("Unknown parser type: " + type);
 }
 

common/peg-parser.h

@@ -275,6 +275,11 @@ struct common_peg_gbnf_parser {
     std::string grammar;
 };
 
+struct common_peg_ac_parser {
+    common_peg_parser_id child;
+    std::vector<std::string> delimiters;
+};
+
 // Variant holding all parser types
 using common_peg_parser_variant = std::variant<
     common_peg_epsilon_parser,
@@ -296,7 +301,8 @@ using common_peg_parser_variant = std::variant<
     common_peg_ref_parser,
     common_peg_atomic_parser,
     common_peg_tag_parser,
-    common_peg_gbnf_parser
+    common_peg_gbnf_parser,
+    common_peg_ac_parser
 >;
 
 class common_peg_arena {
@@ -514,6 +520,13 @@ class common_peg_parser_builder {
     // the child's grammar. Parsing delegates entirely to the child.
     common_peg_parser gbnf(const common_peg_parser & p, const std::string & grammar) { return add(common_peg_gbnf_parser{p, grammar}); }
 
+    // Wraps a child parser but emits a GBNF grammar built from the Aho-Corasick
+    // automaton of `delimiters`, matching everything up to and including the
+    // first delimiter. Parsing delegates entirely to the child, which is
+    // responsible for consuming the delimiter (e.g. until(D) + literal(D)).
+    common_peg_parser ac(const common_peg_parser & p, const std::vector<std::string> & delimiters);
+    common_peg_parser ac(const common_peg_parser & p, const std::string & delimiter) { return ac(p, std::vector<std::string>{delimiter}); }
+
     void set_root(const common_peg_parser & p);
 
     common_peg_arena build();

conversion/__init__.py

@@ -96,6 +96,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "GraniteMoeHybridForCausalLM": "granite",
     "GraniteMoeSharedForCausalLM": "granite",
     "GraniteSpeechForConditionalGeneration": "granite",
+    "GraniteSpeechPlusForConditionalGeneration": "granite",
     "Grok1ForCausalLM": "grok",
     "GrokForCausalLM": "grok",
     "GroveMoeForCausalLM": "grovemoe",
@@ -261,6 +262,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "GlmasrModel": "ultravox",
     "Granite4VisionForConditionalGeneration": "granite",
     "GraniteSpeechForConditionalGeneration": "granite",
+    "GraniteSpeechPlusForConditionalGeneration": "granite",
     "HunYuanVLForConditionalGeneration": "hunyuan",
     "Idefics3ForConditionalGeneration": "smolvlm",
     "InternVisionModel": "internvl",

conversion/granite.py

@@ -348,6 +348,34 @@ class GraniteSpeechMmprojModel(MmprojModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("GraniteSpeechPlusForConditionalGeneration")
+class GraniteSpeechPlusMmprojModel(GraniteSpeechMmprojModel):
+    """Conversion for GraniteSpeechPlus - extends GraniteSpeech with feature layer concatenation"""
+    has_vision_encoder = False
+    has_audio_encoder = True
+
+    def set_gguf_parameters(self):
+        assert self.hparams_audio is not None
+        super().set_gguf_parameters()
+
+        # Add feature_layer if present in encoder config
+        if feature_layers := self.hparams_audio.get("cat_hidden_layers"):
+            self.gguf_writer.add_audio_feature_layers(feature_layers)
+            logger.info(f"gguf: audio feature_layers = {feature_layers}")
+
+            # Validate projector dimension matches concatenated encoder output
+            hidden_dim = self.hparams_audio["hidden_dim"]
+            expected_dim = hidden_dim * (len(feature_layers) + 1)
+            projector_dim = self.global_config["projector_config"]["encoder_hidden_size"]
+
+            if projector_dim != expected_dim:
+                raise ValueError(
+                    f"Projector encoder_hidden_size ({projector_dim}) does not match "
+                    f"expected concatenated dimension ({expected_dim}). "
+                    f"Expected: hidden_dim ({hidden_dim}) * (len(feature_layers) + 1) = {expected_dim}"
+                )
+
+
 @ModelBase.register("Granite4VisionForConditionalGeneration")
 class Granite4VisionMmprojModel(MmprojModel):
     has_vision_encoder = True

docs/android.md

@@ -29,7 +29,7 @@ With Termux, you can install and run `llama.cpp` as if the environment were Linu
 
 ```
 $ apt update && apt upgrade -y
-$ apt install git cmake
+$ apt install git cmake libandroid-spawn
 ```
 
 Then, follow the [build instructions](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md), specifically for CMake.

ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl

@@ -174,7 +174,7 @@ __kernel void kernel_gemv_noshuffle_q8_0_f32(
         regA.s6 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
         regA.s7 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
 
-        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, regS, regB);
+        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, convert_float(regS), regB);
     }
 
     // reduction in local memory, assumes #wave=4

ggml/src/ggml-sycl/binbcast.cpp

@@ -293,6 +293,11 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t
              (sycl::ext::oneapi::bfloat16 *) dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2,
              ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3, ggml_is_contiguous(src0),
              ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1), main_stream);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_BF16) {
+        op()((const sycl::ext::oneapi::bfloat16 *) src0->data, (const float *) src1->data,
+             (sycl::ext::oneapi::bfloat16 *) dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2,
+             ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3, ggml_is_contiguous(src0),
+             ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1), main_stream);
 #endif
     } else {
         fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, ggml_type_name(dst->type),

ggml/src/ggml-sycl/element_wise.cpp

@@ -43,14 +43,44 @@ static __dpct_inline__ T op_sgn(T x) {
     return x > static_cast<T>(0.f) ? static_cast<T>(1.f) : ((x < static_cast<T>(0.f) ? static_cast<T>(-1.f) : static_cast<T>(0.f)));
 }
 
+
 template<typename T>
 static __dpct_inline__ T op_abs(T x) {
-    return sycl::fabs(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fabs(x);  // or experimental namespace if needed
+    } else {
+        return sycl::fabs(x);
+    }
+}
+
+template<typename T>
+static __dpct_inline__ T op_expm1(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::expm1(static_cast<float>(x))
+        );
+    } else {
+        return sycl::expm1(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_elu(T x) {
-    return (x > static_cast<T>(0.f)) ? x : sycl::expm1(x);
+    return (x > static_cast<T>(0.f)) ? x : op_expm1(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_tanh(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        constexpr int ver = __INTEL_LLVM_COMPILER;
+#if defined(__INTEL_LLVM_COMPILER) && (__INTEL_LLVM_COMPILER >= 20260000)
+            return sycl::ext::oneapi::experimental::tanh(x);
+#else
+            return static_cast<T>(sycl::tanh(static_cast<float>(x)));
+#endif
+    } else {
+        return sycl::tanh(x);
+    }
 }
 
 template<typename T>
@@ -59,74 +89,106 @@ static __dpct_inline__ T op_gelu(T x) {
     const T SQRT_2_OVER_PI = static_cast<T>(0.79788456080286535587989211986876f);
     return static_cast<T>(0.5f) * x *
            (static_cast<T>(1.0f) +
-            sycl::tanh(SQRT_2_OVER_PI * x * (static_cast<T>(1.0f) + GELU_COEF_A * x * x)));
+            op_tanh(SQRT_2_OVER_PI * x * (static_cast<T>(1.0f) + GELU_COEF_A * x * x)));
+}
+
+template<typename T>
+static __dpct_inline__ T op_exp(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::exp(x);
+    } else {
+        return sycl::exp(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_silu(T x) {
-    return x / (static_cast<T>(1.0f) + sycl::native::exp(-x));
+    return x / (static_cast<T>(1.0f) + op_exp(-x));
 }
 
 template<typename T>
-static __dpct_inline__ T op_gelu_quick(T x) {
-    const T GELU_QUICK_COEF_LOCAL = static_cast<T>(-1.702f);
-    return x * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(GELU_QUICK_COEF_LOCAL * x)));
+static __dpct_inline__ T op_erf(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::erf(static_cast<float>(x))
+        );
+    } else {
+        return sycl::erf(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_gelu_erf(T x) {
     const T SQRT_2_INV = static_cast<T>(0.70710678118654752440084436210484f);
-    return static_cast<T>(0.5f) * x * (static_cast<T>(1.0f) + sycl::erf(x * SQRT_2_INV));
+    return static_cast<T>(0.5f) * x * (static_cast<T>(1.0f) + op_erf(x * SQRT_2_INV));
 }
 
 template<typename T>
-static __dpct_inline__ T op_tanh(T x) {
-    return sycl::tanh(x);
+static __dpct_inline__ T op_gelu_quick(T x) {
+    const T GELU_QUICK_COEF_LOCAL = static_cast<T>(-1.702f);
+    return x * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + op_exp(GELU_QUICK_COEF_LOCAL * x)));
 }
 
 template<typename T>
 static __dpct_inline__ T op_relu(T x) {
-    return sycl::fmax(x, static_cast<T>(0));
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmax(x, static_cast<T>(0));
+    } else {
+        return sycl::fmax(x, static_cast<T>(0));
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_sigmoid(T x) {
-    return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(-x));
+    return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + op_exp(-x));
 }
 
 template<typename T>
 static __dpct_inline__ T op_sqrt(T x) {
-    return sycl::sqrt(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::sqrt(x);
+    } else {
+        return sycl::sqrt(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_sin(T x) {
-    return sycl::sin(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::sin(x);
+    } else {
+        return sycl::sin(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_cos(T x) {
-    return sycl::cos(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::cos(x);
+    } else {
+        return sycl::cos(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_hardsigmoid(T x) {
-    return sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmin(
+            static_cast<T>(1.0f), sycl::ext::oneapi::experimental::fmax(
+                                      static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    } else {
+        return sycl::fmin(static_cast<T>(1.0f),
+                          sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_hardswish(T x) {
-    return x * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
-}
-
-template<typename T>
-static __dpct_inline__ T op_exp(T x) {
-    return sycl::exp(x);
-}
-
-template<typename T>
-static __dpct_inline__ T op_expm1(T x) {
-    return sycl::expm1(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return x * sycl::ext::oneapi::experimental::fmin(static_cast<T>(1.0f), sycl::ext::oneapi::experimental::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    } else {
+        return x * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    }
 }
 
 template<typename T>
@@ -134,13 +196,17 @@ static __dpct_inline__ T op_log(T x) {
     if (x <= static_cast<T>(0)) {
         return neg_infinity<T>();
     }
-    return sycl::log(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::log(x);
+    } else {
+        return sycl::log(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_softplus(T x) {
     const float xf = (float) x;
-    const float ax = sycl::fabs(xf);
+    const float ax = op_abs(xf);
     const float m  = sycl::fmax(xf, 0.0f);
     const float y  = m + sycl::log1p(sycl::exp(-ax));
     return (T) y;
@@ -159,8 +225,14 @@ static __dpct_inline__ T op_step(T x) {
 template<typename T>
 static __dpct_inline__ T op_leaky_relu(T x, float negative_slope) {
     T neg_slope_T = static_cast<T>(negative_slope);
-    return sycl::fmax(x, static_cast<T>(0)) +
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmax(x, static_cast<T>(0)) +
+           sycl::ext::oneapi::experimental::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
+
+    } else {
+        return sycl::fmax(x, static_cast<T>(0)) +
            sycl::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
+    }
 }
 
 template<typename T>
@@ -175,22 +247,40 @@ static __dpct_inline__ T op_clamp(T x, float min_val, float max_val) {
 
 template<typename T>
 static __dpct_inline__ T op_floor(T x) {
-    return sycl::floor(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::floor(x);
+    } else {
+        return sycl::floor(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_ceil(T x) {
-    return sycl::ceil(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::ceil(x);
+    } else {
+        return sycl::ceil(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_round(T x) {
-    return sycl::round(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::round(static_cast<float>(x))
+        );
+    } else {
+        return sycl::round(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_trunc(T x) {
-    return sycl::trunc(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::trunc(x);
+    } else {
+        return sycl::trunc(x);
+    }
 }
 
 template<typename T, typename F>
@@ -339,7 +429,7 @@ static void acc_f32_sycl(const float *x, const float *y, float *dst,
     const int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
                                            sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
-                         [=](sycl::nd_item<3> /*item_ct1*/) {
+                         [=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                              acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, ne13, s1, s2, s3, offset);
                          });
 }
@@ -354,8 +444,8 @@ static void arange_kernel(T * dst, const int k, T start, T step,
 
 template<typename KernelInvoker, typename... Args>
 static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16 || dst->src[0]->type == GGML_TYPE_BF16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_BF16);
     GGML_ASSERT(dst->src[0]->type == dst->type);
 
     dpct::queue_ptr main_stream = ctx.stream();
@@ -367,6 +457,14 @@ static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx,
                 kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
                 break;
             }
+#ifdef GGML_SYCL_HAS_BF16
+        case GGML_TYPE_BF16:
+            {
+                auto data_pts = cast_data<sycl::ext::oneapi::bfloat16>(dst);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
+                break;
+            }
+#endif
         case GGML_TYPE_F32:
             {
                 auto data_pts = cast_data<float>(dst);
@@ -480,7 +578,7 @@ static inline void ggml_sycl_op_unary(
             stream->parallel_for(
                 sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
                                   sycl::range<1>(256)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_generic_kernel(
                         src, dst_ptr, k_elements,
                         ne0, ne1, ne2, ne3,
@@ -508,7 +606,7 @@ static inline void ggml_sycl_op_arange(ggml_backend_sycl_context & ctx, ggml_ten
     stream->parallel_for(
         sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_ARANGE_BLOCK_SIZE),
                           sycl::range<1>(SYCL_ARANGE_BLOCK_SIZE)),
-        [=](sycl::nd_item<1> item_ct1) {
+        [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             arange_kernel(dst_ptr, k, start, step, item_ct1);
         });
 }
@@ -602,7 +700,7 @@ static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor
             stream->parallel_for(
                 sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
                                   sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_log_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -640,7 +738,7 @@ static inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tenso
             stream->parallel_for(
                 sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQRT_BLOCK_SIZE),
                                   sycl::range<1>(SYCL_SQRT_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_sqrt_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -653,7 +751,7 @@ static inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor
             stream->parallel_for(
                 sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
                                   sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_sin_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -666,7 +764,7 @@ static inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor
             stream->parallel_for(
                 sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
                                   sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_cos_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -681,7 +779,7 @@ static inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml
             stream->parallel_for(
                 sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
                                   sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_leaky_relu_kernel(src, dst_ptr, k_elements, slope, item_ct1);
                 });
         }, negative_slope);
@@ -694,7 +792,7 @@ static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor
             stream->parallel_for(
                 sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQR_BLOCK_SIZE),
                                   sycl::range<1>(SYCL_SQR_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_sqr_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -711,7 +809,7 @@ static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tens
             stream->parallel_for(
                 sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE),
                                   sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     clamp(src, dst_ptr, min_arg, max_arg, k_elements, item_ct1);
                 });
         }, min_val, max_val);
@@ -774,7 +872,8 @@ static inline void ggml_sycl_op_geglu(ggml_backend_sycl_context & ctx, ggml_tens
         [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
             const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
             main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_geglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -785,7 +884,8 @@ static inline void ggml_sycl_op_reglu(ggml_backend_sycl_context & ctx, ggml_tens
         [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
             const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_RELU_BLOCK_SIZE); // Using RELU block size for reglu
             main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_reglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -796,7 +896,8 @@ static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_ten
         [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
             const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_SILU_BLOCK_SIZE); // Using SILU block size for swiglu
             main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_SILU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_swiglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -811,7 +912,6 @@ __dpct_inline__ float ggml_sycl_op_swiglu_oai_single(float x, float g, float alp
     return out_glu;
 }
 
-
 template <typename T>
 static void swiglu_oai_kernel(const T * x, const T * g, T * dst, const int64_t k,
                               const int64_t n, const int64_t o0, const int64_t o1,
@@ -845,7 +945,7 @@ static void swiglu_oai_sycl(const T *       x,
     const int64_t num_blocks = (k + SYCL_GLU_BLOCK_SIZE - 1) / SYCL_GLU_BLOCK_SIZE;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE),
                                            sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                              swiglu_oai_kernel(x, g, dst, k, n, o0, o1, alpha, limit, item_ct1);
                          });
 }
@@ -899,7 +999,8 @@ static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_
         [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
             const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
             main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_geglu_erf(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -910,7 +1011,8 @@ static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggm
         [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
             const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
             main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_geglu_quick(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -108,6 +108,9 @@ if (Vulkan_FOUND)
 
     if (GGML_VULKAN_CHECK_RESULTS)
         add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+        # the result-checking path computes a CPU reference graph via
+        # ggml_graph_compute_with_ctx(), which is defined in ggml-cpu
+        target_link_libraries(ggml-vulkan PRIVATE ggml-cpu)
     endif()
 
     if (GGML_VULKAN_DEBUG)
@@ -129,6 +132,8 @@ if (Vulkan_FOUND)
 
     if (GGML_VULKAN_RUN_TESTS)
         add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+        # the test path also calls ggml_graph_compute_with_ctx() (ggml-cpu)
+        target_link_libraries(ggml-vulkan PRIVATE ggml-cpu)
     endif()
 
     # Set up toolchain for host compilation whether cross-compiling or not

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -905,11 +905,12 @@ struct ggml_webgpu_mul_mat_vec_pipeline_key {
     ggml_type src0_type;
     ggml_type src1_type;
     int       vectorized;
+    uint32_t  num_cols;
     bool      use_mmvq;
 
     bool operator==(const ggml_webgpu_mul_mat_vec_pipeline_key & other) const {
         return src0_type == other.src0_type && src1_type == other.src1_type && vectorized == other.vectorized &&
-               use_mmvq == other.use_mmvq;
+               num_cols == other.num_cols && use_mmvq == other.use_mmvq;
     }
 };
 
@@ -919,6 +920,7 @@ struct ggml_webgpu_mul_mat_vec_pipeline_key_hash {
         ggml_webgpu_hash_combine(seed, key.src0_type);
         ggml_webgpu_hash_combine(seed, key.src1_type);
         ggml_webgpu_hash_combine(seed, key.vectorized);
+        ggml_webgpu_hash_combine(seed, key.num_cols);
         ggml_webgpu_hash_combine(seed, key.use_mmvq);
         return seed;
     }
@@ -993,11 +995,12 @@ struct ggml_webgpu_mul_mat_id_pipeline_key {
     ggml_type src0_type;
     ggml_type src1_type;
     uint32_t  n_experts;
+    uint32_t  num_cols;
     int       vectorized;
 
     bool operator==(const ggml_webgpu_mul_mat_id_pipeline_key & other) const {
         return src0_type == other.src0_type && src1_type == other.src1_type && n_experts == other.n_experts &&
-               vectorized == other.vectorized;
+               num_cols == other.num_cols && vectorized == other.vectorized;
     }
 };
 
@@ -1007,6 +1010,7 @@ struct ggml_webgpu_mul_mat_id_pipeline_key_hash {
         ggml_webgpu_hash_combine(seed, key.src0_type);
         ggml_webgpu_hash_combine(seed, key.src1_type);
         ggml_webgpu_hash_combine(seed, key.n_experts);
+        ggml_webgpu_hash_combine(seed, key.num_cols);
         ggml_webgpu_hash_combine(seed, key.vectorized);
         return seed;
     }
@@ -1107,7 +1111,7 @@ inline bool ggml_webgpu_can_use_mmvq(const ggml_tensor * src0,
                                      const ggml_tensor * src1,
                                      bool                supports_dot_product,
                                      const std::string & vendor) {
-    if (src1->ne[1] == 1) {
+    if (src1->ne[1] <= 4) {
         bool supports_dp4a = vendor == "amd" || vendor == "intel" || vendor == "nvidia";
         if (supports_dp4a && supports_dot_product) {
             switch (src1->type) {
@@ -1889,6 +1893,7 @@ class ggml_webgpu_shader_lib {
                           (context.src0->type == GGML_TYPE_F32 || context.src0->type == GGML_TYPE_F16)) ?
                              1 :
                              0;
+        key.num_cols   = context.dst->ne[1];
         key.use_mmvq =
             ggml_webgpu_can_use_mmvq(context.src0, context.src1, context.supports_dot_product, context.vendor);
 
@@ -2004,6 +2009,7 @@ class ggml_webgpu_shader_lib {
         if (key.vectorized) {
             variant += "_vectorized";
         }
+        defines.push_back(std::string("NUM_COLS=") + std::to_string(key.num_cols));
 
         auto processed            = preprocessor.preprocess(shader_src, defines);
         auto decisions            = std::make_shared<ggml_webgpu_mul_mat_vec_shader_decisions>();
@@ -2421,6 +2427,7 @@ class ggml_webgpu_shader_lib {
         if (key.vectorized) {
             variant += "_vectorized";
         }
+        defines.push_back(std::string("NUM_COLS=1"));
 
         defines.push_back(std::string("N_EXPERTS=") + std::to_string(key.n_experts));
 

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

@@ -1418,15 +1418,17 @@ static void ggml_webgpu_quantize_q8_dispatch(webgpu_context &
     const size_t dst_offset           = ggml_webgpu_tensor_offset(dst);
     const size_t q8_src1_align_offset = ROUNDUP_POW2(
         dst_offset + ggml_nbytes(dst), ctx->global_ctx->capabilities.limits.minStorageBufferOffsetAlignment);
-    const size_t q8_src1_binding_size =
-        ROUNDUP_POW2(src1->ne[3] * src1->ne[2] * (36 /* sizeof(q8_1) */ * (src1->ne[0] / /* block_size */ 32)),
-                     WEBGPU_STORAGE_BUF_BINDING_MULT);
+    const size_t q8_src1_binding_size = ROUNDUP_POW2(
+        src1->ne[3] * src1->ne[2] * src1->ne[1] * (36 /* sizeof(q8_1) */ * (src1->ne[0] / /* block_size */ 32)),
+        WEBGPU_STORAGE_BUF_BINDING_MULT);
 
     std::vector<uint32_t> q8_params = {
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
+        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
         (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
         (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
         (uint32_t) src1->ne[0],
+        (uint32_t) src1->ne[1],
         (uint32_t) src1->ne[2],
         (uint32_t) src1->ne[3],
     };
@@ -1442,7 +1444,7 @@ static void ggml_webgpu_quantize_q8_dispatch(webgpu_context &
     uint32_t       q8_wg_x        = 1;
     uint32_t       q8_wg_y        = 1;
     const uint32_t wg_per_vec     = (src0->ne[0] / 4 + (q8_wg_size - 1)) / q8_wg_size;
-    const uint32_t q8_total_wg    = src1->ne[2] * src1->ne[3] * wg_per_vec;
+    const uint32_t q8_total_wg    = src1->ne[1] * src1->ne[2] * src1->ne[3] * wg_per_vec;
     const uint32_t max_wg_per_dim = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
     compute_2d_workgroups(q8_total_wg, max_wg_per_dim, q8_wg_x, q8_wg_y);
 
@@ -1456,7 +1458,7 @@ static webgpu_encoded_op ggml_webgpu_mul_mat(webgpu_context & ctx,
                                              ggml_tensor *    src1,
                                              ggml_tensor *    dst) {
     // Determine if this is a mat-vec operation
-    bool is_vec = (dst->ne[1] == 1);
+    bool use_mat_vec = (dst->ne[1] <= 4);
 
     // use MMVQ path for mat-vec
     bool use_mmvq = ggml_webgpu_can_use_mmvq(src0, src1, ctx->global_ctx->capabilities.supports_dot_product,
@@ -1482,7 +1484,7 @@ static webgpu_encoded_op ggml_webgpu_mul_mat(webgpu_context & ctx,
     webgpu_pipeline                   pipeline;
     std::vector<webgpu_dispatch_desc> dispatches;
 
-    if (is_vec) {
+    if (use_mat_vec) {
         if (use_mmvq) {
             ggml_webgpu_quantize_q8_dispatch(ctx, src0, src1, dst, dispatches);
         }
@@ -1529,7 +1531,7 @@ static webgpu_encoded_op ggml_webgpu_mul_mat(webgpu_context & ctx,
     uint32_t       wg_y           = 1;
     const uint32_t max_wg_per_dim = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
 
-    if (is_vec) {
+    if (use_mat_vec) {
         auto * decisions = static_cast<ggml_webgpu_mul_mat_vec_shader_decisions *>(pipeline.context.get());
 
         uint32_t batches       = dst->ne[2] * dst->ne[3];
@@ -3691,8 +3693,8 @@ static size_t ggml_backend_webgpu_buffer_type_get_alloc_size(ggml_backend_buffer
                     ggml_webgpu_can_use_mmvq(src0, src1, ctx->webgpu_global_ctx->capabilities.supports_dot_product,
                                              ctx->webgpu_global_ctx->vendor);
                 if (use_mmvq) {
-                    const size_t q8_src1_size =
-                        src1->ne[3] * src1->ne[2] * (36 /* sizeof(q8_1) */ * (src1->ne[0] / /* block_size */ 32));
+                    const size_t q8_src1_size = src1->ne[3] * src1->ne[2] * src1->ne[1] *
+                                                (36 /* sizeof(q8_1) */ * (src1->ne[0] / /* block_size */ 32));
                     res = ROUNDUP_POW2(res + q8_src1_size +
                                            ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment,
                                        WEBGPU_STORAGE_BUF_BINDING_MULT);

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl

@@ -103,7 +103,7 @@ fn main(
 
 #ifdef USE_SUBGROUP_REDUCTION
     for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
-        let subgroup_total = subgroupAdd(acc[row]);
+        let subgroup_total = subgroupAdd(acc[0][row]);
         if (subgroup_invocation_id == 0u) {
             partial_sums[partial_index(row, subgroup_id)] = subgroup_total;
         }
@@ -126,7 +126,7 @@ fn main(
 
 #ifdef USE_WORKGROUP_REDUCTION
     for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
-        partial_sums[partial_index(row, thread_id)] = acc[row];
+        partial_sums[partial_index(row, thread_id)] = acc[0][row];
     }
 
     workgroupBarrier();

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl

@@ -91,61 +91,67 @@ fn main(
     let dst_idx_base = params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row_base;
 
 #ifdef MMVQ
-    let src1q_idx_base = (src13_idx * params.bs02 * params.broadcast2 + src12_idx) * (params.k / 32u);
+    let src1q_idx_base = (src13_idx * params.bs02 * params.broadcast2 + src12_idx) * params.n * (params.k / 32u);
     let acc = accumulate_vec_q_dot(thread_id, row_base, src0_batch_offset, src1q_idx_base);
 #else
     let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12;
     let acc = accumulate_vec_dot(thread_id, row_base, src0_batch_offset, src1_idx_base);
 #endif
 
+    for (var col = 0u;col < NUM_COLS;col += 1) {
+
 #ifdef USE_SUBGROUP_REDUCTION
-    for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
-        let subgroup_total = subgroupAdd(acc[row]);
-        if (subgroup_invocation_id == 0u) {
-            partial_sums[partial_index(row, subgroup_id)] = subgroup_total;
-        }
-    }
+            for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
+                let subgroup_total = subgroupAdd(acc[col][row]);
+                if (subgroup_invocation_id == 0u) {
+                    partial_sums[partial_index(row, subgroup_id)] = subgroup_total;
+                }
+            }
 
-    workgroupBarrier();
+            workgroupBarrier();
 
-    for (var row = subgroup_id; (row < OUTPUTS_PER_WG) && (row_base + row < params.m); row += num_subgroups) {
-        let output_row = row_base + row;
-        var row_acc = 0.0f;
-        for (var k = subgroup_invocation_id; k < num_subgroups; k += subgroup_size) {
-            row_acc += partial_sums[partial_index(row, k)];
-        }
-        let row_total = subgroupAdd(row_acc);
-        if (subgroup_invocation_id == 0) {
-            dst[dst_idx_base + row] = row_total;
-        }
-    }
+            for (var row = subgroup_id; (row < OUTPUTS_PER_WG) && (row_base + row < params.m); row += num_subgroups) {
+                let output_row = row_base + row;
+                var row_acc = 0.0f;
+                for (var k = subgroup_invocation_id; k < num_subgroups; k += subgroup_size) {
+                    row_acc += partial_sums[partial_index(row, k)];
+                }
+                let row_total = subgroupAdd(row_acc);
+                if (subgroup_invocation_id == 0) {
+                    dst[dst_idx_base + col * params.m + row] = row_total;
+                }
+            }
 #endif
 
 #ifdef USE_WORKGROUP_REDUCTION
-    for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
-        partial_sums[partial_index(row, thread_id)] = acc[row];
-    }
+            for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
+                partial_sums[partial_index(row, thread_id)] = acc[col][row];
+            }
+
+            workgroupBarrier();
+
+            var stride = WG_SIZE / 2u;
+
+            while (stride > 0) {
+                if (thread_id < stride) {
+                    for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
+                        partial_sums[partial_index(row, thread_id)] += partial_sums[partial_index(row, thread_id + stride)];
+                    }
+                }
+
+                workgroupBarrier();
+                stride = stride / 2;
+            }
+
+            if (thread_id < OUTPUTS_PER_WG) {
+                let output_row = row_base + thread_id;
+                if (output_row < params.m) {
+                    dst[dst_idx_base + col * params.m + thread_id] = partial_sums[partial_index(thread_id, 0)];
+                }
+            }
+#endif
 
     workgroupBarrier();
 
-    var stride = WG_SIZE / 2u;
-
-    while (stride > 0) {
-        if (thread_id < stride) {
-            for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
-                partial_sums[partial_index(row, thread_id)] += partial_sums[partial_index(row, thread_id + stride)];
-            }
-        }
-
-        workgroupBarrier();
-        stride = stride / 2;
     }
-
-    if (thread_id < OUTPUTS_PER_WG) {
-        let output_row = row_base + thread_id;
-        if (output_row < params.m) {
-            dst[dst_idx_base + thread_id] = partial_sums[partial_index(thread_id, 0)];
-        }
-    }
-#endif
 }

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl

@@ -51,10 +51,7 @@ fn repack_b_dm(block: u32) -> B_DS_TYPE {
 fn get_dm(block_byte_base: u32) -> f32 {
     return f32(load_f16_at_src0(block_byte_base));
 }
-fn mul_q8_1(row_sum: i32, da: f32, b_ds: B_DS_TYPE) -> f32 {
-    return f32(row_sum) * (da * b_ds.x) - 8.0 * da * b_ds.y / THREADS_PER_BLOCK;
-}
-#endif
+#endif // MUL_ACC_Q4_0
 
 #ifdef MUL_ACC_Q4_1
 #define BLOCK_SIZE_BYTES 20
@@ -85,10 +82,7 @@ fn get_dm(block_byte_base: u32) -> vec2<f32> {
         f32(load_f16_at_src0(block_byte_base + 2u))
     );
 }
-fn mul_q8_1(row_sum: i32, dma: vec2<f32>, b_ds: B_DS_TYPE) -> f32 {
-    return f32(row_sum) * (dma.x * b_ds.x) + dma.y * b_ds.y / THREADS_PER_BLOCK;
-}
-#endif
+#endif // MUL_ACC_Q4_1
 
 #ifdef MUL_ACC_Q8_0
 #define BLOCK_SIZE_BYTES 34
@@ -111,46 +105,48 @@ fn repack_b_dm(block: u32) -> B_DS_TYPE {
 fn get_dm(block_byte_base: u32) -> f32 {
     return f32(load_f16_at_src0(block_byte_base));
 }
-fn mul_q8_1(row_sum: i32, da: f32, b_ds: B_DS_TYPE) -> f32 {
-    return f32(row_sum) * (da * b_ds);
-}
-#endif
+#endif // MUL_ACC_Q8_0
 
-#ifdef LEGACY_QUANTS
-fn mmvq_dot_product(a_byte_base: u32, b_inner_id: u32, b_repacked: vec2<u32>, b_ds: B_DS_TYPE) -> f32 {
-    var row_sum = 0;
-    let a_repacked = repack_a(a_byte_base, b_inner_id);
-
-    row_sum += dot4I8Packed(a_repacked[0], b_repacked[0]);
-    row_sum += dot4I8Packed(a_repacked[1], b_repacked[1]);
-
-    return mul_q8_1(row_sum, get_dm(a_byte_base), b_ds);
-}
-
-fn accumulate_vec_q_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1q_idx_base: u32) -> array<f32, OUTPUTS_PER_WG> {
-    var acc: array<f32, OUTPUTS_PER_WG>;
+#if defined(LEGACY_QUANTS)
+fn accumulate_vec_q_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1q_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
+    var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
 
     let num_blocks = params.k / BLOCK_SIZE;
 
     for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
-        let b_inner_id = thread_id % THREADS_PER_BLOCK;
-        let b_block_idx = src1q_idx_base + block;
-
-        let b_repacked = repack_b_qs(b_block_idx, b_inner_id);
-        let b_ds = repack_b_dm(b_block_idx);
-
+        let inner_id = thread_id % THREADS_PER_BLOCK;
         for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
             let output_row = row_base + row;
             if (output_row < params.m) {
                 let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
-                acc[row] += mmvq_dot_product(block_byte_base, b_inner_id, b_repacked, b_ds);
+                let a_repacked = repack_a(block_byte_base, inner_id);
+                let da = get_dm(block_byte_base);
+                for (var col = 0u;col < NUM_COLS;col += 1) {
+                    let src1q_idx = src1q_idx_base + col * (params.k / Q8_BLOCK_SIZE) + block;
+                    let b_repacked = repack_b_qs(src1q_idx, inner_id);
+                    let b_ds = repack_b_dm(src1q_idx);
+
+                    let row_sum = dot4I8Packed(a_repacked[0], b_repacked[0]) + dot4I8Packed(a_repacked[1], b_repacked[1]);
+
+#if defined(MUL_ACC_Q4_0)
+                    acc[col][row] += f32(row_sum) * (da * b_ds.x) - 8.0 * da * b_ds.y / THREADS_PER_BLOCK;
+#endif // MUL_ACC_Q4_0
+
+#if defined(MUL_ACC_Q4_1)
+                    acc[col][row] += f32(row_sum) * (da.x * b_ds.x) + da.y * b_ds.y / THREADS_PER_BLOCK;
+#endif // MUL_ACC_Q4_1
+
+#if defined(MUL_ACC_Q8_0)
+                    acc[col][row] += f32(row_sum) * (da * b_ds);
+#endif // MUL_ACC_Q8_0
+                }
             }
         }
     }
 
     return acc;
 }
-#endif
+#endif // LEGACY_QUANTS
 
 #ifdef MUL_ACC_Q2_K
 #define BLOCK_SIZE_BYTES 84
@@ -191,22 +187,7 @@ fn get_scale_min(block_byte_base: u32, tid: u32) -> vec2<f32> {
     let scale = byte_of(load_u32_at_src0_aligned(scale_byte), scale_byte & 3u);
     return vec2<f32>(f32(scale & 0xFu), f32(scale >> 4u));
 }
-fn mmvq_dot_product(a_byte_base: u32, tid: u32, b_repacked: vec4<u32>, b_ds: B_DS_TYPE) -> f32 {
-    let a_repacked = repack_a(a_byte_base, tid);
-    let dm = get_dm(a_byte_base);
-    let scale_min = get_scale_min(a_byte_base, tid);
-
-    let scale_q = i32(scale_min.x);
-    let scale_m_i8x4 = u32(scale_min.y) * 0x01010101u;
-
-    let row_sum_d = (dot4I8Packed(b_repacked[0], a_repacked[0]) + dot4I8Packed(b_repacked[1], a_repacked[1])
-                   + dot4I8Packed(b_repacked[2], a_repacked[2]) + dot4I8Packed(b_repacked[3], a_repacked[3])) * scale_q;
-    let row_sum_m = dot4I8Packed(b_repacked[0], scale_m_i8x4) + dot4I8Packed(b_repacked[1], scale_m_i8x4)
-                  + dot4I8Packed(b_repacked[2], scale_m_i8x4) + dot4I8Packed(b_repacked[3], scale_m_i8x4);
-
-    return b_ds * (dm.x * f32(row_sum_d) - dm.y * f32(row_sum_m));
-}
-#endif
+#endif // MUL_ACC_Q2_K
 
 #ifdef MUL_ACC_Q4_K
 #define BLOCK_SIZE_BYTES 144
@@ -265,39 +246,52 @@ fn get_scale_min(block_byte_base: u32, tid: u32) -> vec2<f32> {
 
     return vec2<f32>(scale, min_val);
 }
-fn mmvq_dot_product(a_byte_base: u32, tid: u32, b_repacked: vec4<u32>, b_ds: B_DS_TYPE) -> f32 {
-    let a_repacked = repack_a(a_byte_base, tid);
-    let dm = get_dm(a_byte_base);
-    let scale_min = get_scale_min(a_byte_base, tid);
-
-    let row_sum = dot4I8Packed(a_repacked[0], b_repacked[0]) + dot4I8Packed(a_repacked[1], b_repacked[1])
-                + dot4I8Packed(a_repacked[2], b_repacked[2]) + dot4I8Packed(a_repacked[3], b_repacked[3]);
-
-    // Each thread covers half of the Q8_1 block, so add only b_ds.y/2.
-    return b_ds.x * dm.x * scale_min.x * f32(row_sum) - dm.y * scale_min.y * (b_ds.y / (Q8_BLOCK_SIZE / ELEMS_PER_THREAD));
-}
-#endif
+#endif // MUL_ACC_Q4_K
 
 #ifdef K_QUANTS
-fn accumulate_vec_q_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1q_idx_base: u32) -> array<f32, OUTPUTS_PER_WG> {
-    var acc: array<f32, OUTPUTS_PER_WG>;
+fn accumulate_vec_q_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1q_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
+    var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
 
     let tid = thread_id % THREADS_PER_BLOCK;
 
     for (var block = thread_id / THREADS_PER_BLOCK; block < params.k / BLOCK_SIZE; block += WG_SIZE / THREADS_PER_BLOCK) {
-        let src1q_idx = src1q_idx_base + (block * BLOCK_SIZE + ELEMS_PER_THREAD * tid) / Q8_BLOCK_SIZE;
-        let b_repacked = repack_b_qs(src1q_idx, tid);
-        let b_ds = repack_b_dm(src1q_idx);
-
         for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
             let output_row = row_base + row;
             if (output_row < params.m) {
                 let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
-                acc[row] += mmvq_dot_product(block_byte_base, tid, b_repacked, b_ds);
+                let a_repacked = repack_a(block_byte_base, tid);
+                let dm = get_dm(block_byte_base);
+                let scale_min = get_scale_min(block_byte_base, tid);
+                for (var col = 0u;col < NUM_COLS;col += 1) {
+                    let src1q_idx = src1q_idx_base + col * (params.k / Q8_BLOCK_SIZE) + (block * BLOCK_SIZE + ELEMS_PER_THREAD * tid) / Q8_BLOCK_SIZE;
+                    let b_repacked = repack_b_qs(src1q_idx, tid);
+                    let b_ds = repack_b_dm(src1q_idx);
+
+#if defined(MUL_ACC_Q2_K)
+                    let scale_q = i32(scale_min.x);
+                    let scale_m_i8x4 = u32(scale_min.y) * 0x01010101u;
+
+                    let row_sum_d = (dot4I8Packed(b_repacked[0], a_repacked[0]) + dot4I8Packed(b_repacked[1], a_repacked[1])
+                                        + dot4I8Packed(b_repacked[2], a_repacked[2]) + dot4I8Packed(b_repacked[3], a_repacked[3])) * scale_q;
+                    let row_sum_m = dot4I8Packed(b_repacked[0], scale_m_i8x4) + dot4I8Packed(b_repacked[1], scale_m_i8x4)
+                                        + dot4I8Packed(b_repacked[2], scale_m_i8x4) + dot4I8Packed(b_repacked[3], scale_m_i8x4);
+
+                    acc[col][row] += b_ds * (dm.x * f32(row_sum_d) - dm.y * f32(row_sum_m));
+#endif // MUL_ACC_Q2_K
+
+#if defined(MUL_ACC_Q4_K)
+                    let row_sum = dot4I8Packed(a_repacked[0], b_repacked[0]) + dot4I8Packed(a_repacked[1], b_repacked[1])
+                                    + dot4I8Packed(a_repacked[2], b_repacked[2]) + dot4I8Packed(a_repacked[3], b_repacked[3]);
+
+                    // Each thread covers half of the Q8_1 block, so add only b_ds.y/2.
+                    acc[col][row] += b_ds.x * dm.x * scale_min.x * f32(row_sum) - dm.y * scale_min.y * (b_ds.y / (Q8_BLOCK_SIZE / ELEMS_PER_THREAD));
+#endif // MUL_ACC_Q4_K
+
+                }
             }
         }
     }
 
     return acc;
 }
-#endif
+#endif // K_QUANTS

ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl

@@ -9,9 +9,11 @@ requires packed_4x8_integer_dot_product;
 
 struct Params {
     offset_src1: u32,
+    stride_11: u32,
     stride_12: u32,
     stride_13: u32,
     ne0: u32,
+    ne1: u32,
     ne2: u32,
     ne3: u32,
 };
@@ -57,25 +59,28 @@ fn main(
     @builtin(num_workgroups) num_wg: vec3<u32>
 ) {
     let thread_id = local_id.x;
-    let num_vec4 = params.ne0 / 4u;
+    let ne0_vec4 = params.ne0 / 4u;
 
-    let wg_per_vec = (num_vec4 + (WG_SIZE - 1u)) / WG_SIZE;
-    let total_batches = wg_per_vec * params.ne2 * params.ne3;
+    let wg_per_vec = (ne0_vec4 + (WG_SIZE - 1u)) / WG_SIZE;
+    let total_batches = wg_per_vec * params.ne1 * params.ne2 * params.ne3;
 
     let wg_linear = wg_id.y * num_wg.x + wg_id.x;
     if (wg_linear >= total_batches) {
         return;
     }
 
-    let src13_idx = wg_linear / (params.ne2 * wg_per_vec);
-    let src12_idx = (wg_linear - src13_idx * (params.ne2 * wg_per_vec)) / wg_per_vec;
-    let src11_wg_idx = wg_linear % wg_per_vec;
-    let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12;
+    let vec_idx = wg_linear / wg_per_vec;
+    let src13_idx = vec_idx / (params.ne2 * params.ne1);
+    let vec_ne12_num       = vec_idx % (params.ne2 * params.ne1);
+    let src12_idx = vec_ne12_num / params.ne1;
+    let src11_idx = vec_ne12_num % params.ne1;
+    let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12 + src11_idx * params.stride_11;
     let src1_idx_vec4_base = src1_idx_base / 4u;
 
     let blocks_per_row = params.ne0 / 32u;
     let blocks_per_wg = (WG_SIZE * 4u) / 32u;
-    let src1q_idx_base = (src13_idx * params.ne2 + src12_idx) * blocks_per_row;
+    let src1q_idx_base = ((src13_idx * params.ne2 + src12_idx) * params.ne1 + src11_idx) * blocks_per_row;
+    let src11_wg_idx = wg_linear % wg_per_vec;
     let src1q_idx = src1q_idx_base + src11_wg_idx * blocks_per_wg + thread_id / 8u;
     let qs_idx = thread_id % 8u;
 
@@ -85,7 +90,7 @@ fn main(
     var thread_amax = 0.0;
 
     let src11_vec4_idx = src11_wg_idx * WG_SIZE + thread_id;
-    let is_valid = src11_vec4_idx < num_vec4;
+    let is_valid = src11_vec4_idx < ne0_vec4;
 
 #ifdef USE_SUBGROUP_REDUCTION
 

gguf-py/gguf/constants.py

@@ -359,6 +359,7 @@ class Keys:
         CHUNK_SIZE          = "clip.audio.chunk_size"
         CONV_KERNEL_SIZE    = "clip.audio.conv_kernel_size"
         MAX_POS_EMB         = "clip.audio.max_pos_emb"
+        FEATURE_LAYERS      = "clip.audio.feature_layer" # Granite Speech Plus
 
         class Attention:
             HEAD_COUNT      = "clip.audio.attention.head_count"

gguf-py/gguf/gguf_writer.py

@@ -1310,6 +1310,9 @@ class GGUFWriter:
     def add_audio_max_pos_emb(self, value: int) -> None:
         self.add_uint32(Keys.ClipAudio.MAX_POS_EMB, value)
 
+    def add_audio_feature_layers(self, layers: Sequence[int]) -> None:
+        self.add_array(Keys.ClipAudio.FEATURE_LAYERS, layers)
+
     def add_audio_projector_window_size(self, value: int) -> None:
         self.add_uint32(Keys.ClipAudio.Projector.WINDOW_SIZE, value)
 

src/llama-sampler.cpp

@@ -2813,8 +2813,6 @@ static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_t
             cur_p->data[i].logit = -INFINITY;
         }
     }
-
-    llama_sampler_softmax_impl(cur_p, true);
 }
 
 static struct llama_sampler * llama_sampler_top_n_sigma_clone(const struct llama_sampler * smpl) {

tests/peg-parser/test-gbnf-generation.cpp

@@ -212,6 +212,75 @@ void test_gbnf_generation(testing &t) {
         )""", gbnf);
     });
 
+    t.test("ac grammar", [](testing &t) {
+        auto parser = build_peg_parser([](common_peg_parser_builder & p)  {
+            return p.ac(p.until("</tag>") + p.literal("</tag>"), "</tag>");
+        });
+
+        auto gbnf = build_grammar([&](const common_grammar_builder & builder) {
+            parser.build_grammar(builder);
+        });
+
+        assert_gbnf_equal(t, R"""(
+            ac-3 ::= [<] ac-3-01 | [^<] ac-3
+            ac-3-01 ::= [<] ac-3-01 | [/] ac-3-02 | [^/<] ac-3
+            ac-3-02 ::= [<] ac-3-01 | [t] ac-3-03 | [^<t] ac-3
+            ac-3-03 ::= [<] ac-3-01 | [a] ac-3-04 | [^<a] ac-3
+            ac-3-04 ::= [<] ac-3-01 | [g] ac-3-05 | [^<g] ac-3
+            ac-3-05 ::= [>] | [<] ac-3-01 | [^<>] ac-3
+            root ::= ac-3
+            space ::= | " " | "\n"{1,2} [ \t]{0,20}
+        )""", gbnf);
+    });
+
+    t.test("ac grammar terminates at first delimiter", [](testing &t) {
+        auto parser = build_peg_parser([](common_peg_parser_builder & p)  {
+            return p.ac(p.until("\n</parameter>\n") + p.literal("\n</parameter>\n"), "\n</parameter>\n");
+        });
+
+        auto gbnf = build_grammar([&](const common_grammar_builder & builder) {
+            parser.build_grammar(builder);
+        });
+
+        assert_gbnf_equal(t, R"""(
+            ac-3 ::= [\n] ac-3-01 | [^\n] ac-3
+            ac-3-01 ::= [\n] ac-3-01 | [<] ac-3-02 | [^\n<] ac-3
+            ac-3-02 ::= [\n] ac-3-01 | [/] ac-3-03 | [^\n/] ac-3
+            ac-3-03 ::= [\n] ac-3-01 | [p] ac-3-04 | [^\np] ac-3
+            ac-3-04 ::= [\n] ac-3-01 | [a] ac-3-05 | [^\na] ac-3
+            ac-3-05 ::= [\n] ac-3-01 | [r] ac-3-06 | [^\nr] ac-3
+            ac-3-06 ::= [\n] ac-3-01 | [a] ac-3-07 | [^\na] ac-3
+            ac-3-07 ::= [\n] ac-3-01 | [m] ac-3-08 | [^\nm] ac-3
+            ac-3-08 ::= [\n] ac-3-01 | [e] ac-3-09 | [^\ne] ac-3
+            ac-3-09 ::= [\n] ac-3-01 | [t] ac-3-10 | [^\nt] ac-3
+            ac-3-10 ::= [\n] ac-3-01 | [e] ac-3-11 | [^\ne] ac-3
+            ac-3-11 ::= [\n] ac-3-01 | [r] ac-3-12 | [^\nr] ac-3
+            ac-3-12 ::= [\n] ac-3-01 | [>] ac-3-13 | [^\n>] ac-3
+            ac-3-13 ::= [\n] | [^\n] ac-3
+            root ::= ac-3
+            space ::= | " " | "\n"{1,2} [ \t]{0,20}
+        )""", gbnf);
+    });
+
+    t.test("ac grammar multiple delimiters", [](testing &t) {
+        auto parser = build_peg_parser([](common_peg_parser_builder & p)  {
+            return p.ac(p.eps(), std::vector<std::string>{"ab", "cd", "ef"});
+        });
+
+        auto gbnf = build_grammar([&](const common_grammar_builder & builder) {
+            parser.build_grammar(builder);
+        });
+
+        assert_gbnf_equal(t, R"""(
+            ac-1 ::= [a] ac-1-01 | [c] ac-1-03 | [e] ac-1-05 | [^ace] ac-1
+            ac-1-01 ::= [b] | [a] ac-1-01 | [c] ac-1-03 | [e] ac-1-05 | [^abce] ac-1
+            ac-1-03 ::= [d] | [a] ac-1-01 | [c] ac-1-03 | [e] ac-1-05 | [^acde] ac-1
+            ac-1-05 ::= [f] | [a] ac-1-01 | [c] ac-1-03 | [e] ac-1-05 | [^acef] ac-1
+            root ::= ac-1
+            space ::= | " " | "\n"{1,2} [ \t]{0,20}
+        )""", gbnf);
+    });
+
     t.test("complex expressions with parentheses", [](testing &t) {
         auto parser = build_peg_parser([](common_peg_parser_builder & p) {
             return p.one_or_more(p.literal("a") | p.literal("b"));

tests/test-backend-ops.cpp

@@ -8433,6 +8433,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 2}, {2, 1}));
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 2}, {1, 2}));
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 2}, {2, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  4, k, {3, 2}, {2, 2}));
 
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {1, 1}));
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {2, 1}));
@@ -8449,6 +8450,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {2, 3}, {1, 1}, {0, 1, 3, 2}));
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));
 
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  4, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, k, {2, 3}, {1, 1}, {0, 2, 1, 3}));
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, k, {2, 3}, {1, 1}, {0, 1, 3, 2}));
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));

tests/test-chat.cpp

@@ -1562,37 +1562,112 @@ static void test_msgs_oaicompat_json_conversion() {
     }
 }
 
-static void test_split_by_role() {
+static void test_msg_token_delimiters_split() {
     LOG_DBG("%s\n", __func__);
 
+    // Delimiters that share a leading token, distinguished by the second token,
+    // to exercise the per-position token matching.
+    const common_chat_msg_delimiters delims = {
+        { { COMMON_CHAT_ROLE_USER,      "", { 10, 11 } },
+          { COMMON_CHAT_ROLE_ASSISTANT, "", { 10, 12 } } }
+    };
+
     // Empty inputs
-    assert_equals<size_t>(0, common_chat_split_by_role("", {}).size());
-    assert_equals<size_t>(0, common_chat_split_by_role("hello", {}).size());
-    assert_equals<size_t>(0, common_chat_split_by_role("", { { "user", "<|user|>" } }).size());
+    assert_equals<size_t>(0, common_chat_msg_delimiters{}.split({}).spans.size());
+    assert_equals<size_t>(0, common_chat_msg_delimiters{}.split({ 10, 11 }).spans.size());
+    assert_equals<size_t>(0, delims.split({}).spans.size());
 
-    // Multi-role conversation, no leading/trailing content
+    // No delimiters match -> no spans
+    assert_equals<size_t>(0, delims.split({ 100, 101, 102 }).spans.size());
+
+    // Multi-role conversation: <user>Hi<assistant>Hello<user>Bye
     {
-        const std::string prompt = "<|user|>Hi<|assistant|>Hello<|user|>Bye";
-        const auto splits = common_chat_split_by_role(prompt, {
-            { "user",      "<|user|>"      },
-            { "assistant", "<|assistant|>" },
-        });
-        assert_equals<size_t>(3, splits.size());
+        const llama_tokens tokens = {
+            10, 11,            // <user>
+            100, 101,          // Hi
+            10, 12,            // <assistant>
+            200, 201, 202,     // Hello
+            10, 11,            // <user>
+            300, 301,          // Bye
+        };
 
-        assert_equals<std::string>("user", splits[0].role);
-        assert_equals<size_t>(0, splits[0].pos);
-        assert_equals<size_t>(10, splits[0].len);
-        assert_equals<std::string>("<|user|>Hi", prompt.substr(splits[0].pos, splits[0].len));
+        const auto result = delims.split(tokens);
+        const auto & spans = result.spans;
+        assert_equals<size_t>(3, spans.size());
 
-        assert_equals<std::string>("assistant", splits[1].role);
-        assert_equals<size_t>(10, splits[1].pos);
-        assert_equals<size_t>(18, splits[1].len);
-        assert_equals<std::string>("<|assistant|>Hello", prompt.substr(splits[1].pos, splits[1].len));
+        assert_equals(COMMON_CHAT_ROLE_USER, spans[0].role);
+        assert_equals<size_t>(0, spans[0].pos);
+        assert_equals<size_t>(4, spans[0].len);
 
-        assert_equals<std::string>("user", splits[2].role);
-        assert_equals<size_t>(28, splits[2].pos);
-        assert_equals<size_t>(11, splits[2].len);
-        assert_equals<std::string>("<|user|>Bye", prompt.substr(splits[2].pos, splits[2].len));
+        assert_equals(COMMON_CHAT_ROLE_ASSISTANT, spans[1].role);
+        assert_equals<size_t>(4, spans[1].pos);
+        assert_equals<size_t>(5, spans[1].len);
+
+        assert_equals(COMMON_CHAT_ROLE_USER, spans[2].role);
+        assert_equals<size_t>(9, spans[2].pos);
+        assert_equals<size_t>(4, spans[2].len);
+
+        // is_user_start() is true at the token position where a user span begins
+        assert_equals(true,  result.is_user_start(0));
+        assert_equals(false, result.is_user_start(4));  // assistant span
+        assert_equals(true,  result.is_user_start(9));
+    }
+
+    // Content before the first delimiter is not captured as a span
+    {
+        const llama_tokens tokens = {
+            500, 501,    // leading content (dropped)
+            10, 11,      // <user>
+            100,         // Hi
+        };
+
+        const auto spans = delims.split(tokens).spans;
+        assert_equals<size_t>(1, spans.size());
+        assert_equals(COMMON_CHAT_ROLE_USER, spans[0].role);
+        assert_equals<size_t>(2, spans[0].pos);
+        assert_equals<size_t>(3, spans[0].len);
+    }
+
+    // Skipped regions (media chunks) are jumped over but still count as span content
+    {
+        const llama_tokens tokens = {
+            10, 11,             // <user>
+            LLAMA_TOKEN_NULL,   // media chunk (3 tokens)
+            LLAMA_TOKEN_NULL,
+            LLAMA_TOKEN_NULL,
+            100,                // Hi
+            10, 12,             // <assistant>
+        };
+
+        const std::map<size_t, size_t> skips = { { 2, 3 } };
+
+        const auto spans = delims.split(tokens, skips).spans;
+        assert_equals<size_t>(2, spans.size());
+
+        assert_equals(COMMON_CHAT_ROLE_USER, spans[0].role);
+        assert_equals<size_t>(0, spans[0].pos);
+        assert_equals<size_t>(6, spans[0].len);
+
+        assert_equals(COMMON_CHAT_ROLE_ASSISTANT, spans[1].role);
+        assert_equals<size_t>(6, spans[1].pos);
+        assert_equals<size_t>(2, spans[1].len);
+    }
+
+    // A delimiter sequence inside a skipped region is not matched
+    {
+        const llama_tokens tokens = {
+            10, 11,      // <user>
+            10, 12,      // skipped region that happens to contain delimiter tokens
+            100,         // Hi
+        };
+
+        const std::map<size_t, size_t> skips = { { 2, 2 } };
+
+        const auto spans = delims.split(tokens, skips).spans;
+        assert_equals<size_t>(1, spans.size());
+        assert_equals(COMMON_CHAT_ROLE_USER, spans[0].role);
+        assert_equals<size_t>(0, spans[0].pos);
+        assert_equals<size_t>(5, spans[0].len);
     }
 }
 
@@ -5857,7 +5932,7 @@ int main(int argc, char ** argv) {
     {
         test_msg_diffs_compute();
         test_msgs_oaicompat_json_conversion();
-        test_split_by_role();
+        test_msg_token_delimiters_split();
         test_tools_oaicompat_json_conversion();
         test_convert_responses_to_chatcmpl();
         test_developer_role_to_system_workaround();

tests/test-sampling.cpp

@@ -360,9 +360,9 @@ int main(void) {
     test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.032727f, 0.241818f, 0.241818f}, 2.0f, 1.1f, 2, 5, {});
     test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 4, 7, {});
 
-    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f, 0.0f, 0.0f}, 1.00f);
+    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.0f, 0.0f, 0.428571f, 0.571429f}, 1.00f);
     test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 0.00f); // top_n_sigma == 0 now represents a no-op rather than greedy decoding as of PR#13345
-    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 3.00f);
+    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 3.00f);
 
     test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
     test_sampler_queue(10000, "k",     1, 1.0f, 1.0f);

tools/cli/CMakeLists.txt

@@ -2,11 +2,13 @@
 
 set(TARGET llama-cli-impl)
 
-add_library(${TARGET} cli.cpp)
+add_library(${TARGET} cli.cpp
+                      cli-client.cpp
+                      cli-context.cpp)
 set_target_properties(${TARGET} PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS ON)
 
 target_include_directories(${TARGET} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR} ../server)
-target_link_libraries(${TARGET} PUBLIC server-context llama-common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PUBLIC llama-server-impl llama-common ${CMAKE_THREAD_LIBS_INIT})
 
 if(LLAMA_TOOLS_INSTALL)
     install(TARGETS ${TARGET} LIBRARY)

tools/cli/cli-client.cpp

@@ -0,0 +1,164 @@
+#include "cli-client.h"
+
+#include "http.h"
+
+#include <algorithm>
+#include <chrono>
+#include <thread>
+
+// generation can stall for a long time during prompt processing, so the
+// read timeout must be generous
+static constexpr time_t CLI_HTTP_READ_TIMEOUT_SEC = 3600;
+
+// upper bound for the accumulated response body kept for error reporting
+static constexpr size_t CLI_HTTP_MAX_ERROR_BODY = 1024 * 1024;
+
+// returns the path with the base url's path prefix prepended (if any)
+static std::string join_path(const common_http_url & parts, const std::string & path) {
+    if (parts.path.empty() || parts.path == "/") {
+        return path;
+    }
+    std::string prefix = parts.path;
+    if (prefix.back() == '/') {
+        prefix.pop_back();
+    }
+    return prefix + path;
+}
+
+json cli_client::get(const std::string & path) {
+    auto [cli, parts] = common_http_client(server_base);
+    cli.set_read_timeout(CLI_HTTP_READ_TIMEOUT_SEC, 0);
+    auto path_with_model = path + (model.empty() ? "" : ("?model=" + model));
+    auto res = cli.Get(join_path(parts, path_with_model));
+    if (!res) {
+        throw std::runtime_error("failed to connect to " + server_base + ": " + httplib::to_string(res.error()));
+    }
+    if (res->status < 200 || res->status >= 300) {
+        throw std::runtime_error("GET " + path + " failed with status " + std::to_string(res->status) + ": " + res->body);
+    }
+    json result = json::parse(res->body, nullptr, false);
+    if (result.is_discarded()) {
+        throw std::runtime_error("GET " + path + " returned invalid JSON");
+    }
+    return result;
+}
+
+json cli_client::post(const std::string & path, const json & body) {
+    auto [cli, parts] = common_http_client(server_base);
+    cli.set_read_timeout(CLI_HTTP_READ_TIMEOUT_SEC, 0);
+    auto body_with_model = body;
+    if (!model.empty()) {
+        body_with_model["model"] = model;
+    }
+    auto res = cli.Post(join_path(parts, path), body_with_model.dump(), "application/json");
+    if (!res) {
+        throw std::runtime_error("failed to connect to " + server_base + ": " + httplib::to_string(res.error()));
+    }
+    if (res->status < 200 || res->status >= 300) {
+        throw std::runtime_error("POST " + path + " failed with status " + std::to_string(res->status) + ": " + res->body);
+    }
+    json result = json::parse(res->body, nullptr, false);
+    if (result.is_discarded()) {
+        throw std::runtime_error("POST " + path + " returned invalid JSON");
+    }
+    return result;
+}
+
+json cli_client::post_sse(const std::string & path,
+                          const json & body,
+                          const std::function<bool()> & should_stop,
+                          const std::function<void(const json &)> & on_data) {
+    auto [cli, parts] = common_http_client(server_base);
+    cli.set_read_timeout(CLI_HTTP_READ_TIMEOUT_SEC, 0);
+
+    std::string pending;  // buffer for incomplete SSE lines
+    std::string raw_body; // accumulated body, used only for error reporting
+
+    auto receiver = [&](const char * data, size_t len) -> bool {
+        if (should_stop()) {
+            return false; // aborts the request
+        }
+        if (raw_body.size() < CLI_HTTP_MAX_ERROR_BODY) {
+            raw_body.append(data, std::min(len, CLI_HTTP_MAX_ERROR_BODY - raw_body.size()));
+        }
+        pending.append(data, len);
+        size_t pos;
+        while ((pos = pending.find('\n')) != std::string::npos) {
+            std::string line = pending.substr(0, pos);
+            pending.erase(0, pos + 1);
+            if (!line.empty() && line.back() == '\r') {
+                line.pop_back();
+            }
+            if (line.rfind("data: ", 0) != 0) {
+                continue;
+            }
+            std::string payload = line.substr(6);
+            if (payload == "[DONE]") {
+                continue;
+            }
+            json event = json::parse(payload, nullptr, false);
+            if (!event.is_discarded()) {
+                on_data(event);
+            }
+        }
+        return true;
+    };
+
+    httplib::Headers headers = {{"Accept", "text/event-stream"}};
+    auto body_with_model = body;
+    if (!model.empty()) {
+        body_with_model["model"] = model;
+    }
+    auto res = cli.Post(join_path(parts, path), headers, body_with_model.dump(), "application/json", receiver);
+
+    if (!res) {
+        if (res.error() == httplib::Error::Canceled && should_stop()) {
+            return json(); // cancelled by the user
+        }
+        return json {{"error", {{"message", "failed to connect to " + server_base + ": " + httplib::to_string(res.error())}}}};
+    }
+    if (res->status < 200 || res->status >= 300) {
+        json error_body = json::parse(raw_body, nullptr, false);
+        if (!error_body.is_discarded() && error_body.contains("error")) {
+            return error_body;
+        }
+        return json {{"error", {{"message", "request failed with status " + std::to_string(res->status)}}}};
+    }
+    return json();
+}
+
+bool cli_client::wait_health(const std::function<bool()> & is_aborted) {
+    int connect_attempts = 0;
+    while (!is_aborted()) {
+        auto [cli, parts] = common_http_client(server_base);
+        cli.set_connection_timeout(1, 0);
+        auto res = cli.Get(join_path(parts, "/health"));
+        if (res) {
+            if (res->status == 200) {
+                return true;
+            }
+            // any other status means the server is up but not ready yet
+            // (e.g. 503 while the model is still loading)
+        } else if (++connect_attempts >= 10) {
+            last_error = "failed to connect to " + server_base + ": " + httplib::to_string(res.error());
+            return false;
+        }
+        std::this_thread::sleep_for(std::chrono::milliseconds(300));
+    }
+    last_error = "aborted while waiting for the server to become ready";
+    return false;
+}
+
+std::vector<std::string> cli_client::list_models() {
+    json resp = get("/v1/models");
+    if (!resp.contains("data") || !resp.at("data").is_array()) {
+        throw std::runtime_error("invalid response from /v1/models");
+    }
+    std::vector<std::string> models;
+    for (const auto & m : resp.at("data")) {
+        if (m.contains("id") && m.at("id").is_string()) {
+            models.push_back(m.at("id").get<std::string>());
+        }
+    }
+    return models;
+}

tools/cli/cli-client.h

@@ -0,0 +1,56 @@
+#pragma once
+
+#include "ggml.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
+#include <functional>
+#include <string>
+
+using json = nlohmann::ordered_json;
+
+// openai-like client for CLI
+struct cli_client {
+    std::string server_base; // base url, for example "http://127.0.0.1:8080"
+    std::string last_error;  // set when wait_health() fails
+
+    std::string model; // optional, set when the server has multiple models (router mode)
+
+    // simple GET request, returns the response json
+    // throws std::runtime_error on transport error or non-2xx status
+    json get(const std::string & path);
+
+    // simple POST request, returns the response json
+    // throws std::runtime_error on transport error or non-2xx status
+    json post(const std::string & path, const json & body);
+
+    // POST request with an SSE streaming response; on_data is invoked once
+    // per "data:" event; the function returns after the stream is finished:
+    // a null json on graceful exit (incl. cancellation via should_stop),
+    // the error response json otherwise
+    json post_sse(const std::string & path,
+                  const json & body,
+                  const std::function<bool()> & should_stop,
+                  const std::function<void(const json &)> & on_data);
+
+    // poll /health until the server is ready to accept requests
+    // returns false if is_aborted returned true or the server is unreachable
+    bool wait_health(const std::function<bool()> & is_aborted);
+
+    //
+    // higher-level wrappers
+    //
+
+    json create_chat_completion(const json & request,
+                                const std::function<bool()> & should_stop,
+                                const std::function<void(const json &)> & on_data) {
+        return post_sse("/v1/chat/completions", request, should_stop, on_data);
+    }
+
+    json get_props() {
+        return get("/props");
+    }
+
+    std::vector<std::string> list_models();
+};

tools/cli/cli-context.cpp

@@ -0,0 +1,559 @@
+#include "cli-context.h"
+#include "cli-view.h"
+
+#include "arg.h"
+#include "base64.hpp"
+#include "log.h"
+#include "console.h"
+
+#include <algorithm>
+#include <filesystem>
+#include <fstream>
+#include <map>
+#include <set>
+
+std::atomic<bool> g_cli_interrupted = false;
+
+static bool should_stop() {
+    return g_cli_interrupted.load();
+}
+
+static constexpr size_t FILE_GLOB_MAX_RESULTS = 100;
+
+const char * LLAMA_ASCII_LOGO = R"(
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+)";
+
+// number of values an arg consumes on the command line
+static int arg_num_values(const common_arg & opt) {
+    if (opt.value_hint_2 != nullptr) {
+        return 2;
+    }
+    if (opt.value_hint != nullptr) {
+        return 1;
+    }
+    return 0;
+}
+
+static std::string format_error_message(const json & err) {
+    if (err.contains("error") && err.at("error").is_object()) {
+        const auto & e = err.at("error");
+        if (e.contains("message") && e.at("message").is_string()) {
+            return e.at("message").get<std::string>();
+        }
+    }
+    return err.dump();
+}
+
+static std::string media_type_from_ext(const std::string & fname) {
+    std::string ext = std::filesystem::path(fname).extension().string();
+    std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
+    if (ext == ".wav" || ext == ".mp3") {
+        return "audio";
+    }
+    if (ext == ".mp4" || ext == ".avi" || ext == ".mkv" || ext == ".mov" || ext == ".webm") {
+        return "video";
+    }
+    return "image";
+}
+
+bool cli_context::init() {
+    view::init(params);
+
+    std::optional<view::spinner> spinner;
+
+    bool use_external_server = !params.server_base.empty();
+    if (use_external_server) {
+        std::string base = params.server_base;
+        while (!base.empty() && base.back() == '/') {
+            base.pop_back();
+        }
+        client.server_base = base;
+
+        spinner.emplace("Connecting to server at " + base);
+    } else {
+        if (params.model.path.empty() && params.model.url.empty() &&
+                params.model.hf_repo.empty() && params.model.docker_repo.empty()) {
+            view::show_error(
+                "no model specified",
+                "use -m <file.gguf> or -hf <user/repo> to run a local model,\n"
+                "or --server-base <url> to connect to a running llama-server"
+            );
+            return false;
+        }
+
+        spinner.emplace("\n\nLoading model...");
+
+        server.emplace();
+        if (!server->start(params)) {
+            view::show_error("server start failed");
+            return false;
+        }
+        if (!server->wait_ready(should_stop)) {
+            if (!should_stop()) {
+                view::show_error("the server exited before becoming ready");
+            }
+            return false;
+        }
+        client.server_base = server->address();
+    }
+
+    // for --server-base this is the main availability check; for a spawned
+    // server it is a cheap sanity check on top of the ready signal
+    auto is_aborted = [this]() {
+        return should_stop() || (server && !server->alive());
+    };
+    bool healthy = false;
+    try {
+        healthy = client.wait_health(is_aborted);
+    } catch (const std::exception & e) {
+        client.last_error = e.what();
+    }
+    if (!healthy) {
+        if (!should_stop()) {
+            view::show_error(client.last_error);
+        }
+        return false;
+    }
+
+    if (use_external_server) {
+        spinner.reset();
+        if (!list_and_ask_models()) {
+            return false;
+        }
+        // restore the spinner for the next step
+        spinner.emplace("Waiting for server...");
+    }
+
+    fetch_server_props();
+
+    return true;
+}
+
+void cli_context::fetch_server_props() {
+    try {
+        json props = client.get_props();
+        model_name = props.value("model_alias", "");
+        if (model_name.empty()) {
+            const std::string path = props.value("model_path", "");
+            if (!path.empty()) {
+                model_name = std::filesystem::path(path).filename().string();
+            }
+        }
+        build_info = props.value("build_info", "");
+        if (props.contains("modalities") && props.at("modalities").is_object()) {
+            const auto & modalities = props.at("modalities");
+            has_vision = modalities.value("vision", false);
+            has_audio  = modalities.value("audio", false);
+            has_video  = modalities.value("video", false);
+        }
+    } catch (const std::exception & e) {
+        // /props can be disabled on remote servers; not fatal
+        LOG_DBG("failed to fetch /props: %s\n", e.what());
+    }
+}
+
+bool cli_context::list_and_ask_models() {
+    auto models = client.list_models();
+
+    // only one model: use it without asking
+    if (models.size() == 1) {
+        model_name = models[0];
+        client.model = model_name;
+        return true;
+    }
+
+    std::string message = "\nAvailable models:";
+    if (!models.empty()) {
+        for (size_t i = 0; i < models.size(); ++i) {
+            message += "\n  " + std::to_string(i + 1) + ". " + models[i];
+        }
+    }
+    message += "\n";
+    view::show_message(message);
+    std::string selection;
+    while (selection.empty()) {
+        if (should_stop()) {
+            return false;
+        }
+        view::user_turn user_turn;
+        selection = user_turn.read_input(false, "Select model by number: ");
+        if (selection.empty()) {
+            continue;
+        }
+        try {
+            size_t idx = std::stoul(selection);
+            if (idx > 0 && idx <= models.size()) {
+                model_name = models[idx - 1];
+                client.model = model_name;
+                view::show_message("Selected model: " + model_name);
+                break;
+            }
+        } catch (...) {
+            // ignore
+        }
+        view::show_error("Invalid selection. Please enter a valid number.");
+        selection.clear();
+        continue;
+    }
+    return true;
+}
+
+void cli_context::add_system_prompt() {
+    if (!params.system_prompt.empty()) {
+        messages.push_back({
+            {"role",    "system"},
+            {"content", params.system_prompt}
+        });
+    }
+}
+
+void cli_context::push_user_message(const std::string & text) {
+    json content;
+    if (pending_media.empty()) {
+        content = text;
+    } else {
+        // multimodal message: media parts first, then the text
+        content = pending_media;
+        content.push_back({
+            {"type", "text"},
+            {"text", text}
+        });
+        pending_media = json::array();
+    }
+    messages.push_back({
+        {"role",    "user"},
+        {"content", content}
+    });
+}
+
+bool cli_context::stage_media_file(const std::string & fname, const std::string & type) {
+    std::ifstream file(fname, std::ios::binary);
+    if (!file) {
+        return false;
+    }
+    std::string data((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+    std::string encoded = base64::encode(data);
+
+    if (type == "audio") {
+        std::string ext = std::filesystem::path(fname).extension().string();
+        std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
+        pending_media.push_back({
+            {"type", "input_audio"},
+            {"input_audio", {
+                {"data",   encoded},
+                {"format", ext == ".mp3" ? "mp3" : "wav"}
+            }}
+        });
+    } else if (type == "video") {
+        pending_media.push_back({
+            {"type", "input_video"},
+            {"input_video", {
+                {"data", encoded}
+            }}
+        });
+    } else {
+        // the server detects the actual image type from the data
+        pending_media.push_back({
+            {"type", "image_url"},
+            {"image_url", {
+                {"url", "data:image/unknown;base64," + encoded}
+            }}
+        });
+    }
+    return true;
+}
+
+bool cli_context::generate_completion(std::string & assistant_content, cli_timings & timings) {
+    json body = {
+        {"messages",          messages},
+        {"stream",            true},
+        // in order to get timings even when we cancel mid-way
+        {"timings_per_token", true},
+    };
+
+    bool stream_error = false;
+
+    view::assistant_turn a;
+
+    json err = client.create_chat_completion(body, should_stop, [&](const json & chunk) {
+        if (chunk.contains("error")) {
+            stream_error = true;
+            view::show_error(format_error_message(chunk));
+            return;
+        }
+        if (chunk.contains("timings")) {
+            const auto & t = chunk.at("timings");
+            timings.prompt_per_second    = t.value("prompt_per_second",    0.0);
+            timings.predicted_per_second = t.value("predicted_per_second", 0.0);
+        }
+        if (!chunk.contains("choices") || !chunk.at("choices").is_array() || chunk.at("choices").empty()) {
+            return;
+        }
+        const auto & choice = chunk.at("choices").at(0);
+        if (!choice.contains("delta")) {
+            return;
+        }
+        const auto & delta = choice.at("delta");
+        if (delta.contains("reasoning_content") && delta.at("reasoning_content").is_string()) {
+            const std::string text = delta.at("reasoning_content").get<std::string>();
+            if (!text.empty()) {
+                a.push(view::ASSISTANT_DISPLAY_MODE_REASONING, text);
+            }
+        }
+        if (delta.contains("content") && delta.at("content").is_string()) {
+            const std::string text = delta.at("content").get<std::string>();
+            if (!text.empty()) {
+                assistant_content += text;
+                a.push(view::ASSISTANT_DISPLAY_MODE_CONTENT, text);
+            }
+        }
+    });
+
+    g_cli_interrupted.store(false);
+
+    if (!err.is_null()) {
+        view::show_error(format_error_message(err));
+        return false;
+    }
+    return !stream_error;
+}
+
+int cli_context::run() {
+    add_system_prompt();
+
+    std::string modalities = "text";
+    if (has_vision) {
+        modalities += ", vision";
+    }
+    if (has_audio) {
+        modalities += ", audio";
+    }
+    if (has_video) {
+        modalities += ", video";
+    }
+
+    std::string banner;
+    banner += "\n";
+    banner += LLAMA_ASCII_LOGO;
+    banner += "\n";
+    banner += "build      : " + build_info + "\n";
+    banner += "model      : " + model_name + "\n";
+    banner += "modalities : " + modalities + "\n";
+    if (!params.system_prompt.empty()) {
+        banner += "using custom system prompt\n";
+    }
+    banner += "\n";
+    banner += "available commands:\n";
+    banner += "  /exit or Ctrl+C     stop or exit\n";
+    banner += "  /regen              regenerate the last response\n";
+    banner += "  /clear              clear the chat history\n";
+    banner += "  /read <file>        add a text file\n";
+    banner += "  /glob <pattern>     add text files using globbing pattern\n";
+    if (has_vision) {
+        banner += "  /image <file>       add an image file\n";
+    }
+    if (has_audio) {
+        banner += "  /audio <file>       add an audio file\n";
+    }
+    if (has_video) {
+        banner += "  /video <file>       add a video file\n";
+    }
+    banner += "\n";
+
+    view::show_message(banner);
+
+    // interactive loop
+    std::string cur_msg;
+
+    auto add_text_file = [&](const std::string & fname) -> bool {
+        std::ifstream file(fname, std::ios::binary);
+        if (!file) {
+            view::show_error(string_format("file does not exist or cannot be opened: '%s'", fname.c_str()));
+            return false;
+        }
+        std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+        cur_msg += "--- File: ";
+        cur_msg += fname;
+        cur_msg += " ---\n";
+        cur_msg += content;
+        view::show_message(string_format("Loaded text from '%s'", fname.c_str()));
+        return true;
+    };
+
+    while (true) {
+        std::string buffer;
+        {
+            view::user_turn user_turn;
+
+            if (params.prompt.empty()) {
+                buffer = user_turn.read_input(params.multiline_input);
+            } else {
+                // process input prompt from args
+                for (auto & fname : params.image) {
+                    if (!stage_media_file(fname, media_type_from_ext(fname))) {
+                        view::show_error(string_format("file does not exist or cannot be opened: '%s'", fname.c_str()));
+                        break;
+                    }
+                    view::show_message(string_format("Loaded media from '%s'", fname.c_str()));
+                }
+                buffer = params.prompt;
+                user_turn.echo(buffer);
+                params.prompt.clear(); // only use it once
+            }
+        }
+
+        if (should_stop()) {
+            g_cli_interrupted.store(false);
+            break;
+        }
+
+        // remove trailing newline
+        if (!buffer.empty() && buffer.back() == '\n') {
+            buffer.pop_back();
+        }
+
+        // skip empty messages
+        if (buffer.empty()) {
+            continue;
+        }
+
+        bool add_user_msg = true;
+
+        // process commands
+        if (string_starts_with(buffer, "/exit")) {
+            break;
+        } else if (string_starts_with(buffer, "/regen")) {
+            if (messages.size() >= 2) {
+                size_t last_idx = messages.size() - 1;
+                messages.erase(last_idx);
+                add_user_msg = false;
+            } else {
+                view::show_error("No message to regenerate.");
+                continue;
+            }
+        } else if (string_starts_with(buffer, "/clear")) {
+            messages.clear();
+            add_system_prompt();
+
+            pending_media = json::array();
+            view::show_message("Chat history cleared.");
+            continue;
+        } else if (
+                (string_starts_with(buffer, "/image ") && has_vision) ||
+                (string_starts_with(buffer, "/audio ") && has_audio) ||
+                (string_starts_with(buffer, "/video ") && has_video)) {
+            std::string type = buffer.substr(1, 5);
+            // just in case (bad copy-paste for example), we strip all trailing/leading spaces
+            std::string fname = string_strip(buffer.substr(7));
+            if (!stage_media_file(fname, type)) {
+                view::show_error(string_format("file does not exist or cannot be opened: '%s'", fname.c_str()));
+                continue;
+            }
+            view::show_message(string_format("Loaded media from '%s'", fname.c_str()));
+            continue;
+        } else if (string_starts_with(buffer, "/read ")) {
+            std::string fname = string_strip(buffer.substr(6));
+            add_text_file(fname);
+            continue;
+        } else if (string_starts_with(buffer, "/glob ")) {
+            std::error_code ec;
+            size_t count = 0;
+            auto curdir = std::filesystem::current_path();
+            std::string pattern = string_strip(buffer.substr(6));
+            std::filesystem::path rel_path;
+
+            auto startglob = pattern.find_first_of("![*?");
+            if (startglob != std::string::npos && startglob != 0) {
+                auto endpath = pattern.substr(0, startglob).find_last_of('/');
+                if (endpath != std::string::npos) {
+                    std::string rel_pattern = pattern.substr(0, endpath);
+#if !defined(_WIN32)
+                    if (string_starts_with(rel_pattern, '~')) {
+                        const char * home = std::getenv("HOME");
+                        if (home && home[0]) {
+                            rel_pattern = home + rel_pattern.substr(1);
+                        }
+                    }
+#endif
+                    rel_path = rel_pattern;
+                    pattern.erase(0, endpath + 1);
+                    curdir /= rel_path;
+                }
+            }
+
+            for (const auto & entry : std::filesystem::recursive_directory_iterator(curdir,
+                    std::filesystem::directory_options::skip_permission_denied, ec)) {
+                if (!entry.is_regular_file()) {
+                    continue;
+                }
+
+                std::string rel = std::filesystem::relative(entry.path(), curdir, ec).string();
+                if (ec) {
+                    ec.clear();
+                    continue;
+                }
+                std::replace(rel.begin(), rel.end(), '\\', '/');
+
+                if (!glob_match(pattern, rel)) {
+                    continue;
+                }
+
+                if (!add_text_file((rel_path / rel).string())) {
+                    continue;
+                }
+
+                if (++count >= FILE_GLOB_MAX_RESULTS) {
+                    view::show_error(string_format("Maximum number of globbed files allowed (%zu) reached.", FILE_GLOB_MAX_RESULTS));
+                    break;
+                }
+            }
+            continue;
+        } else {
+            // not a command
+            cur_msg += buffer;
+        }
+
+        // generate response
+        if (add_user_msg) {
+            push_user_message(cur_msg);
+            cur_msg.clear();
+        }
+        cli_timings timings;
+        std::string assistant_content;
+        generate_completion(assistant_content, timings);
+        messages.push_back({
+            {"role",    "assistant"},
+            {"content", assistant_content}
+        });
+
+        if (params.show_timings) {
+            view::show_info(string_format(
+                "\n[ Prompt: %.1f t/s | Generation: %.1f t/s ]",
+                timings.prompt_per_second,
+                timings.predicted_per_second
+            ));
+        }
+
+        if (params.single_turn) {
+            break;
+        }
+    }
+
+    view::show_message("\n\nExiting...");
+
+    return 0;
+}
+
+void cli_context::shutdown() {
+    if (server) {
+        server->stop();
+        server.reset();
+    }
+}

tools/cli/cli-context.h

@@ -0,0 +1,65 @@
+#pragma once
+
+#include "common.h"
+
+#include "cli-client.h"
+#include "cli-server.h"
+
+#include <atomic>
+#include <optional>
+#include <string>
+
+struct cli_timings {
+    double prompt_per_second    = 0.0;
+    double predicted_per_second = 0.0;
+};
+
+// set by the SIGINT handler; cleared once the interrupt has been handled
+extern std::atomic<bool> g_cli_interrupted;
+
+struct cli_context {
+    common_params params;
+
+    cli_client client;                // always initialized
+    std::optional<cli_server> server; // only set when no --server-base is given
+
+    json messages      = json::array();
+    json pending_media = json::array(); // staged multimodal content parts
+
+    // properties of the connected server
+    // will be populated by fetch_server_props()
+    std::string model_name;
+    std::string build_info;
+    bool has_vision = false;
+    bool has_audio  = false;
+    bool has_video  = false;
+
+    cli_context(const common_params & params) : params(params) {}
+    ~cli_context() {
+        shutdown();
+    }
+
+    // connect to --server-base or spawn a local llama-server child;
+    // argc/argv are needed to forward the server-relevant args to the child
+    bool init();
+
+    // run the interactive chat loop, returns the process exit code
+    int run();
+
+    // stop the local server child (if any)
+    void shutdown();
+
+private:
+    bool generate_completion(std::string & assistant_content, cli_timings & timings);
+    void fetch_server_props();
+    void add_system_prompt();
+    void push_user_message(const std::string & text);
+
+    // check if server have multiple models (router mode)
+    // if yes, list them then ask; do nothing otherwise
+    bool list_and_ask_models();
+
+    // read a file and stage it as a multimodal content part; type is one of
+    // "image", "audio", "video"; returns false if the file cannot be read
+    bool stage_media_file(const std::string & fname, const std::string & type);
+};

tools/cli/cli-server.h

@@ -0,0 +1,83 @@
+#pragma once
+
+#include <thread>
+
+#include "http.h"
+
+// llama_server will be available as a dynamic library symbol
+int llama_server(common_params & params, int argc, char ** argv);
+void llama_server_terminate();
+
+struct cli_server {
+    std::thread th;
+    int port = -1;
+    std::atomic<bool> is_alive = false;
+    std::atomic<bool> is_stopping = false;
+
+    ~cli_server() {
+        stop();
+    }
+
+    void stop() {
+        if (alive() && !is_stopping.exchange(true)) {
+            llama_server_terminate();
+            th.join();
+        }
+    }
+
+    // spawn llama-server in a thread and interact with it via a random port
+    bool start(common_params & params) {
+        port = common_http_get_free_port();
+        if (port <= 0) {
+            fprintf(stderr, "failed to get a free port\n");
+            exit(1);
+        }
+
+        is_alive.store(true, std::memory_order_release);
+
+        th = std::thread([&]() {
+            common_params server_params = params; // copy
+            server_params.port = port;
+            // argc / argv are only used in router mode, we can skip them for now
+            int res = llama_server(server_params, 0, nullptr);
+            if (res != 0) {
+                fprintf(stderr, "llama_server exited with code %d\n", res);
+            }
+            is_alive.store(false, std::memory_order_release);
+        });
+
+        return true;
+    }
+
+    std::string address() const {
+        return "http://127.0.0.1:" + std::to_string(port);
+    }
+
+    bool wait_ready(std::function<bool()> should_stop) {
+        if (!alive()) {
+            return false;
+        }
+        while (!should_stop()) {
+            auto [cli, parts] = common_http_client(address());
+            cli.set_connection_timeout(1, 0);
+            auto res = cli.Get("/health");
+            if (res) {
+                if (res->status == 200) {
+                    return true;
+                }
+                // any other status means the server is up but not ready yet
+                // (e.g. 503 while the model is still loading)
+            }
+            if (!alive()) {
+                // in case server die permanently
+                return false;
+            }
+            std::this_thread::sleep_for(std::chrono::milliseconds(200));
+        }
+        return true;
+    }
+
+    bool alive() const {
+        return is_alive.load(std::memory_order_acquire);
+    }
+};

tools/cli/cli-view.h

@@ -0,0 +1,250 @@
+#pragma once
+
+#include "common.h"
+#include "console.h"
+
+#include <array>
+#include <algorithm>
+#include <filesystem>
+#include <string_view>
+
+// TODO?: Make this reusable, enums, docs
+static const std::array<std::string_view, 8> cmds = {
+    "/audio ",
+    "/clear",
+    "/exit",
+    "/glob ",
+    "/image ",
+    "/read ",
+    "/regen",
+    "/video ",
+};
+
+static std::vector<std::pair<std::string, size_t>> auto_completion_callback(std::string_view line, size_t cursor_byte_pos) {
+    std::vector<std::pair<std::string, size_t>> matches;
+    std::string cmd;
+
+    if (line.length() > 1 && line.front() == '/' && !std::any_of(cmds.begin(), cmds.end(), [line](std::string_view prefix) {
+        return string_starts_with(line, prefix);
+    })) {
+        auto it = cmds.begin();
+
+        while ((it = std::find_if(it, cmds.end(), [line](std::string_view cmd_line) {
+            return string_starts_with(cmd_line, line);
+        })) != cmds.end()) {
+            matches.emplace_back(*it, it->length());
+            ++it;
+        }
+    } else {
+        auto it = std::find_if(cmds.begin(), cmds.end(), [line](std::string_view prefix) {
+            return prefix.back() == ' ' && string_starts_with(line, prefix);
+        });
+
+        if (it != cmds.end()) {
+            cmd = *it;
+        }
+    }
+
+    if (!cmd.empty() && cmd != "/glob " && line.length() >= cmd.length() && cursor_byte_pos >= cmd.length()) {
+        const std::string path_prefix  = std::string(line.substr(cmd.length(), cursor_byte_pos - cmd.length()));
+        const std::string path_postfix = std::string(line.substr(cursor_byte_pos));
+        auto cur_dir = std::filesystem::current_path();
+        std::string cur_dir_str = cur_dir.string();
+        std::string expanded_prefix = path_prefix;
+
+#if !defined(_WIN32)
+        if (string_starts_with(path_prefix, '~')) {
+            const char * home = std::getenv("HOME");
+            if (home && home[0]) {
+                expanded_prefix = home + path_prefix.substr(1);
+            }
+        }
+        if (string_starts_with(expanded_prefix, '/')) {
+#else
+        if (std::isalpha(expanded_prefix[0]) && expanded_prefix.find(':') == 1) {
+#endif
+            cur_dir = std::filesystem::path(expanded_prefix).parent_path();
+            cur_dir_str.clear();
+        } else if (!path_prefix.empty()) {
+            cur_dir /= std::filesystem::path(path_prefix).parent_path();
+        }
+
+        std::error_code ec;
+        for (const auto & entry : std::filesystem::directory_iterator(cur_dir, ec)) {
+            if (ec) {
+                break;
+            }
+            if (!entry.exists(ec)) {
+                ec.clear();
+                continue;
+            }
+
+            const std::string path_full = entry.path().string();
+            std::string path_entry = !cur_dir_str.empty() && string_starts_with(path_full, cur_dir_str) ? path_full.substr(cur_dir_str.length() + 1) : path_full;
+
+            if (entry.is_directory(ec)) {
+                path_entry.push_back(std::filesystem::path::preferred_separator);
+            }
+
+            if (expanded_prefix.empty() || string_starts_with(path_entry, expanded_prefix)) {
+                const std::string updated_line = cmd + path_entry;
+                matches.emplace_back(updated_line + path_postfix, updated_line.length());
+            }
+
+            if (ec) {
+                ec.clear();
+            }
+        }
+
+        if (matches.empty()) {
+            const std::string updated_line = cmd + path_prefix;
+            matches.emplace_back(updated_line + path_postfix, updated_line.length());
+        }
+
+        // Add the longest common prefix
+        if (!expanded_prefix.empty() && matches.size() > 1) {
+            const std::string_view match0(matches[0].first);
+            const std::string_view match1(matches[1].first);
+            auto it = std::mismatch(match0.begin(), match0.end(), match1.begin(), match1.end());
+            size_t len = it.first - match0.begin();
+
+            for (size_t i = 2; i < matches.size(); ++i) {
+                const std::string_view matchi(matches[i].first);
+                auto cmp = std::mismatch(match0.begin(), match0.end(), matchi.begin(), matchi.end());
+                len = std::min(len, static_cast<size_t>(cmp.first - match0.begin()));
+            }
+
+            const std::string updated_line = std::string(match0.substr(0, len));
+            matches.emplace_back(updated_line + path_postfix, updated_line.length());
+        }
+
+        std::sort(matches.begin(), matches.end(), [](const auto & a, const auto & b) {
+            return a.first.compare(0, a.second, b.first, 0, b.second) < 0;
+        });
+    }
+
+    return matches;
+}
+
+// note: make this view implementation generic, so that we can move to TUI in the future if we want to
+namespace view {
+    static void init(const common_params & params) {
+        // TODO: avoid using atexit() here by making `console` a singleton
+        console::init(params.simple_io, params.use_color);
+        atexit([]() { console::cleanup(); });
+
+        console::set_completion_callback(auto_completion_callback);
+    }
+
+    struct spinner {
+        spinner(const std::string & message) {
+            if (!message.empty()) {
+                console::log("%s ", message.c_str());
+            }
+            console::spinner::start();
+        }
+        ~spinner() {
+            console::spinner::stop();
+        }
+    };
+
+    struct user_turn {
+        user_turn() {
+            console::set_display(DISPLAY_TYPE_USER_INPUT);
+        }
+        ~user_turn() {
+            console::set_display(DISPLAY_TYPE_RESET);
+        }
+        void echo(const std::string & buffer) {
+            if (buffer.size() > 500) {
+                console::log("\n> %s ... (truncated)\n", buffer.substr(0, 500).c_str());
+            } else {
+                console::log("\n> %s\n", buffer.c_str());
+            }
+        }
+        std::string read_input(bool multiline_input, const char * prompt = nullptr) {
+            if (prompt) {
+                console::log("%s", prompt);
+            } else {
+                console::log("\n> ");
+            }
+            std::string buffer;
+            std::string line;
+            bool another_line = true;
+            do {
+                another_line = console::readline(line, multiline_input);
+                buffer += line;
+            } while (another_line);
+            return buffer;
+        }
+    };
+
+    enum assistant_display_mode {
+        ASSISTANT_DISPLAY_MODE_REASONING,
+        ASSISTANT_DISPLAY_MODE_CONTENT,
+    };
+    struct assistant_turn {
+        assistant_display_mode mode = ASSISTANT_DISPLAY_MODE_CONTENT;
+        bool trailing_newline = true;
+        bool is_inside_reasoning = false;
+        assistant_turn() {
+            console::set_display(DISPLAY_TYPE_RESET);
+        }
+        ~assistant_turn() {
+            console::set_display(DISPLAY_TYPE_RESET);
+            add_newline_if_needed();
+        }
+        void push(assistant_display_mode m, const std::string & buffer) {
+            if (m != mode) {
+                add_newline_if_needed();
+                switch (m) {
+                    case ASSISTANT_DISPLAY_MODE_CONTENT:
+                        {
+                            if (is_inside_reasoning) {
+                                console::log("[End thinking]\n\n");
+                                is_inside_reasoning = false;
+                            }
+                            console::set_display(DISPLAY_TYPE_RESET);
+                        } break;
+                    case ASSISTANT_DISPLAY_MODE_REASONING:
+                        {
+                            console::set_display(DISPLAY_TYPE_REASONING);
+                            is_inside_reasoning = true;
+                            console::log("\n[Start thinking]\n\n");
+                        } break;
+                }
+            }
+            mode = m;
+            if (buffer.empty()) {
+                return;
+            }
+            trailing_newline = buffer.back() == '\n';
+            console::log("%s", buffer.c_str());
+            console::flush();
+        }
+        void add_newline_if_needed() {
+            if (!trailing_newline) {
+                console::log("\n");
+                console::flush();
+            }
+        }
+    };
+
+    static void show_error(const std::string & title, const std::string & message = "") {
+        console::spinner::stop();
+        console::error("Error: %s\n", title.c_str());
+        if (!message.empty()) {
+            console::log("%s\n", message.c_str());
+        }
+    }
+
+    static void show_message(const std::string & message) {
+        console::log("%s\n", message.c_str());
+    }
+
+    static void show_info(const std::string & message) {
+        console::set_display(DISPLAY_TYPE_INFO);
+        console::log("%s\n", message.c_str());
+        console::set_display(DISPLAY_TYPE_RESET);
+    }
+}

tools/cli/cli.cpp

@@ -1,20 +1,10 @@
-#include "chat.h"
-#include "common.h"
 #include "arg.h"
-#include "console.h"
-#include "fit.h"
-// #include "log.h"
+#include "common.h"
+#include "log.h"
 
-#include "server-common.h"
-#include "server-context.h"
-#include "server-task.h"
+#include "cli-context.h"
+#include "cli-view.h"
 
-#include <array>
-#include <atomic>
-#include <algorithm>
-#include <filesystem>
-#include <fstream>
-#include <thread>
 #include <signal.h>
 
 #if defined(_WIN32)
@@ -25,342 +15,19 @@
 #include <windows.h>
 #endif
 
-const char * LLAMA_ASCII_LOGO = R"(
-▄▄ ▄▄
-██ ██
-██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
-██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
-██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
-                                    ██    ██
-                                    ▀▀    ▀▀
-)";
-
-static std::atomic<bool> g_is_interrupted = false;
-static bool should_stop() {
-    return g_is_interrupted.load();
-}
-
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void signal_handler(int) {
-    if (g_is_interrupted.load()) {
+    if (g_cli_interrupted.load()) {
         // second Ctrl+C - exit immediately
         // make sure to clear colors before exiting (not using LOG or console.cpp here to avoid deadlock)
         fprintf(stdout, "\033[0m\n");
         fflush(stdout);
         std::exit(130);
     }
-    g_is_interrupted.store(true);
+    g_cli_interrupted.store(true);
 }
 #endif
 
-struct cli_context {
-    server_context ctx_server;
-    json messages = json::array();
-    std::vector<raw_buffer> input_files;
-    task_params defaults;
-    bool verbose_prompt;
-
-    // thread for showing "loading" animation
-    std::atomic<bool> loading_show;
-
-    cli_context(const common_params & params) {
-        defaults.sampling    = params.sampling;
-        defaults.speculative = params.speculative;
-        defaults.n_keep      = params.n_keep;
-        defaults.n_predict   = params.n_predict;
-        defaults.antiprompt  = params.antiprompt;
-
-        defaults.stream = true; // make sure we always use streaming mode
-        defaults.timings_per_token = true; // in order to get timings even when we cancel mid-way
-        // defaults.return_progress = true; // TODO: show progress
-
-        verbose_prompt = params.verbose_prompt;
-    }
-
-    std::string generate_completion(result_timings & out_timings) {
-        server_response_reader rd = ctx_server.get_response_reader();
-        auto chat_params = format_chat();
-        {
-            // TODO: reduce some copies here in the future
-            server_task task = server_task(SERVER_TASK_TYPE_COMPLETION);
-            task.id         = rd.get_new_id();
-            task.index      = 0;
-            task.params     = defaults;           // copy
-            task.cli_prompt = chat_params.prompt; // copy
-            task.cli_files  = input_files;        // copy
-            task.cli        = true;
-
-            // chat template settings
-            task.params.chat_parser_params = common_chat_parser_params(chat_params);
-            task.params.chat_parser_params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
-            if (!chat_params.parser.empty()) {
-                task.params.chat_parser_params.parser.load(chat_params.parser);
-            }
-
-            // Copy the preserved tokens into the sampling params
-            const llama_vocab * vocab = llama_model_get_vocab(
-                llama_get_model(ctx_server.get_llama_context()));
-            for (const auto & token : chat_params.preserved_tokens) {
-                auto ids = common_tokenize(vocab, token, false, true);
-                if (ids.size() == 1) {
-                    task.params.sampling.preserved_tokens.insert(ids[0]);
-                }
-            }
-
-            // reasoning budget sampler
-            if (!chat_params.thinking_end_tag.empty()) {
-                task.params.sampling.reasoning_budget_tokens = defaults.sampling.reasoning_budget_tokens;
-                task.params.sampling.generation_prompt = chat_params.generation_prompt;
-
-                if (!chat_params.thinking_start_tag.empty()) {
-                    task.params.sampling.reasoning_budget_start =
-                        common_tokenize(vocab, chat_params.thinking_start_tag, false, true);
-                }
-                task.params.sampling.reasoning_budget_end =
-                    common_tokenize(vocab, chat_params.thinking_end_tag, false, true);
-                task.params.sampling.reasoning_budget_forced =
-                    common_tokenize(vocab, defaults.sampling.reasoning_budget_message + chat_params.thinking_end_tag, false, true);
-            }
-
-            rd.post_task({std::move(task)});
-        }
-
-        if (verbose_prompt) {
-            console::set_display(DISPLAY_TYPE_PROMPT);
-            console::log("%s\n\n", chat_params.prompt.c_str());
-            console::set_display(DISPLAY_TYPE_RESET);
-        }
-
-        // wait for first result
-        console::spinner::start();
-        server_task_result_ptr result = rd.next(should_stop);
-
-        while (true) {
-            auto res_partial = dynamic_cast<server_task_result_cmpl_partial *>(result.get());
-            if (res_partial && res_partial->is_begin) {
-                // this is the "send 200 status to client" signal in streaming mode
-                // skip, do not stop the spinner
-                result = rd.next(should_stop);
-            } else {
-                console::spinner::stop();
-                break;
-            }
-        }
-
-        std::string curr_content;
-        bool is_thinking = false;
-
-        while (result) {
-            if (should_stop()) {
-                break;
-            }
-            if (result->is_error()) {
-                json err_data = result->to_json();
-                if (err_data.contains("message")) {
-                    console::error("Error: %s\n", err_data["message"].get<std::string>().c_str());
-                } else {
-                    console::error("Error: %s\n", err_data.dump().c_str());
-                }
-                return curr_content;
-            }
-            auto res_partial = dynamic_cast<server_task_result_cmpl_partial *>(result.get());
-            if (res_partial) {
-                out_timings = std::move(res_partial->timings);
-                for (const auto & diff : res_partial->oaicompat_msg_diffs) {
-                    if (!diff.content_delta.empty()) {
-                        if (is_thinking) {
-                            console::log("\n[End thinking]\n\n");
-                            console::set_display(DISPLAY_TYPE_RESET);
-                            is_thinking = false;
-                        }
-                        curr_content += diff.content_delta;
-                        console::log("%s", diff.content_delta.c_str());
-                        console::flush();
-                    }
-                    if (!diff.reasoning_content_delta.empty()) {
-                        console::set_display(DISPLAY_TYPE_REASONING);
-                        if (!is_thinking) {
-                            console::log("[Start thinking]\n");
-                        }
-                        is_thinking = true;
-                        console::log("%s", diff.reasoning_content_delta.c_str());
-                        console::flush();
-                    }
-                }
-            }
-            auto res_final = dynamic_cast<server_task_result_cmpl_final *>(result.get());
-            if (res_final) {
-                out_timings = std::move(res_final->timings);
-                break;
-            }
-            result = rd.next(should_stop);
-        }
-        g_is_interrupted.store(false);
-        // server_response_reader automatically cancels pending tasks upon destruction
-        return curr_content;
-    }
-
-    // TODO: support remote files in the future (http, https, etc)
-    std::string load_input_file(const std::string & fname, bool is_media) {
-        std::ifstream file = fs_open_ifstream(fname, std::ios::binary);
-        if (!file) {
-            return "";
-        }
-        if (is_media) {
-            raw_buffer buf;
-            buf.assign((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
-            input_files.push_back(std::move(buf));
-            return get_media_marker();
-        } else {
-            std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
-            return content;
-        }
-    }
-
-    common_chat_params format_chat() {
-        auto meta = ctx_server.get_meta();
-        auto & chat_params = meta.chat_params;
-
-        auto caps = common_chat_templates_get_caps(chat_params.tmpls.get());
-
-        common_chat_templates_inputs inputs;
-        inputs.messages              = common_chat_msgs_parse_oaicompat(messages);
-        inputs.tools                 = {}; // TODO
-        inputs.tool_choice           = COMMON_CHAT_TOOL_CHOICE_NONE;
-        inputs.json_schema           = ""; // TODO
-        inputs.grammar               = ""; // TODO
-        inputs.use_jinja             = chat_params.use_jinja;
-        inputs.parallel_tool_calls   = caps["supports_parallel_tool_calls"];
-        inputs.add_generation_prompt = true;
-        inputs.reasoning_format      = COMMON_REASONING_FORMAT_DEEPSEEK;
-        inputs.force_pure_content    = chat_params.force_pure_content;
-        inputs.enable_thinking       = chat_params.enable_thinking ? common_chat_templates_support_enable_thinking(chat_params.tmpls.get()) : false;
-
-        // Apply chat template to the list of messages
-        return common_chat_templates_apply(chat_params.tmpls.get(), inputs);
-    }
-};
-
-// TODO?: Make this reusable, enums, docs
-static const std::array<std::string_view, 8> cmds = {
-    "/audio ",
-    "/clear",
-    "/exit",
-    "/glob ",
-    "/image ",
-    "/read ",
-    "/regen",
-    "/video ",
-};
-
-static std::vector<std::pair<std::string, size_t>> auto_completion_callback(std::string_view line, size_t cursor_byte_pos) {
-    std::vector<std::pair<std::string, size_t>> matches;
-    std::string cmd;
-
-    if (line.length() > 1 && line.front() == '/' && !std::any_of(cmds.begin(), cmds.end(), [line](std::string_view prefix) {
-        return string_starts_with(line, prefix);
-    })) {
-        auto it = cmds.begin();
-
-        while ((it = std::find_if(it, cmds.end(), [line](std::string_view cmd_line) {
-            return string_starts_with(cmd_line, line);
-        })) != cmds.end()) {
-            matches.emplace_back(*it, it->length());
-            ++it;
-        }
-    } else {
-        auto it = std::find_if(cmds.begin(), cmds.end(), [line](std::string_view prefix) {
-            return prefix.back() == ' ' && string_starts_with(line, prefix);
-        });
-
-        if (it != cmds.end()) {
-            cmd = *it;
-        }
-    }
-
-    if (!cmd.empty() && cmd != "/glob " && line.length() >= cmd.length() && cursor_byte_pos >= cmd.length()) {
-        const std::string path_prefix  = std::string(line.substr(cmd.length(), cursor_byte_pos - cmd.length()));
-        const std::string path_postfix = std::string(line.substr(cursor_byte_pos));
-        auto cur_dir = std::filesystem::current_path();
-        std::string cur_dir_str = cur_dir.string();
-        std::string expanded_prefix = path_prefix;
-
-#if !defined(_WIN32)
-        if (string_starts_with(path_prefix, '~')) {
-            const char * home = std::getenv("HOME");
-            if (home && home[0]) {
-                expanded_prefix = home + path_prefix.substr(1);
-            }
-        }
-        if (string_starts_with(expanded_prefix, '/')) {
-#else
-        if (std::isalpha(expanded_prefix[0]) && expanded_prefix.find(':') == 1) {
-#endif
-            cur_dir = std::filesystem::path(expanded_prefix).parent_path();
-            cur_dir_str.clear();
-        } else if (!path_prefix.empty()) {
-            cur_dir /= std::filesystem::path(path_prefix).parent_path();
-        }
-
-        std::error_code ec;
-        for (const auto & entry : std::filesystem::directory_iterator(cur_dir, ec)) {
-            if (ec) {
-                break;
-            }
-            if (!entry.exists(ec)) {
-                ec.clear();
-                continue;
-            }
-
-            const std::string path_full = entry.path().string();
-            std::string path_entry = !cur_dir_str.empty() && string_starts_with(path_full, cur_dir_str) ? path_full.substr(cur_dir_str.length() + 1) : path_full;
-
-            if (entry.is_directory(ec)) {
-                path_entry.push_back(std::filesystem::path::preferred_separator);
-            }
-
-            if (expanded_prefix.empty() || string_starts_with(path_entry, expanded_prefix)) {
-                const std::string updated_line = cmd + path_entry;
-                matches.emplace_back(updated_line + path_postfix, updated_line.length());
-            }
-
-            if (ec) {
-                ec.clear();
-            }
-        }
-
-        if (matches.empty()) {
-            const std::string updated_line = cmd + path_prefix;
-            matches.emplace_back(updated_line + path_postfix, updated_line.length());
-        }
-
-        // Add the longest common prefix
-        if (!expanded_prefix.empty() && matches.size() > 1) {
-            const std::string_view match0(matches[0].first);
-            const std::string_view match1(matches[1].first);
-            auto it = std::mismatch(match0.begin(), match0.end(), match1.begin(), match1.end());
-            size_t len = it.first - match0.begin();
-
-            for (size_t i = 2; i < matches.size(); ++i) {
-                const std::string_view matchi(matches[i].first);
-                auto cmp = std::mismatch(match0.begin(), match0.end(), matchi.begin(), matchi.end());
-                len = std::min(len, static_cast<size_t>(cmp.first - match0.begin()));
-            }
-
-            const std::string updated_line = std::string(match0.substr(0, len));
-            matches.emplace_back(updated_line + path_postfix, updated_line.length());
-        }
-
-        std::sort(matches.begin(), matches.end(), [](const auto & a, const auto & b) {
-            return a.first.compare(0, a.second, b.first, 0, b.second) < 0;
-        });
-    }
-
-    return matches;
-}
-
-static constexpr size_t FILE_GLOB_MAX_RESULTS = 100;
-
 // satisfies -Wmissing-declarations
 int llama_cli(int argc, char ** argv);
 
@@ -375,25 +42,6 @@ int llama_cli(int argc, char ** argv) {
         return 1;
     }
 
-    // TODO: maybe support it later?
-    if (params.conversation_mode == COMMON_CONVERSATION_MODE_DISABLED) {
-        console::error("--no-conversation is not supported by llama-cli\n");
-        console::error("please use llama-completion instead\n");
-    }
-
-    // struct that contains llama context and inference
-    cli_context ctx_cli(params);
-
-    llama_backend_init();
-    llama_numa_init(params.numa);
-
-    // TODO: avoid using atexit() here by making `console` a singleton
-    console::init(params.simple_io, params.use_color);
-    atexit([]() { console::cleanup(); });
-
-    console::set_display(DISPLAY_TYPE_RESET);
-    console::set_completion_callback(auto_completion_callback);
-
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
     struct sigaction sigint_action;
     sigint_action.sa_handler = signal_handler;
@@ -408,273 +56,11 @@ int llama_cli(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-    console::log("\nLoading model... "); // followed by loading animation
-    console::spinner::start();
-    if (!ctx_cli.ctx_server.load_model(params)) {
-        console::spinner::stop();
-        console::error("\nFailed to load the model\n");
+    cli_context ctx_cli(params);
+
+    if (!ctx_cli.init()) {
         return 1;
     }
 
-    ctx_cli.defaults.sampling = params.sampling;
-
-    console::spinner::stop();
-    console::log("\n");
-
-    std::thread inference_thread([&ctx_cli]() {
-        ctx_cli.ctx_server.start_loop();
-    });
-
-    auto inf = ctx_cli.ctx_server.get_meta();
-    std::string modalities = "text";
-    if (inf.has_inp_image) {
-        modalities += ", vision";
-    }
-    if (inf.has_inp_audio) {
-        modalities += ", audio";
-    }
-
-    auto add_system_prompt = [&]() {
-        if (!params.system_prompt.empty()) {
-            ctx_cli.messages.push_back({
-                {"role",    "system"},
-                {"content", params.system_prompt}
-            });
-        }
-    };
-    add_system_prompt();
-
-    console::log("\n");
-    console::log("%s\n", LLAMA_ASCII_LOGO);
-    console::log("build      : %s\n", inf.build_info.c_str());
-    console::log("model      : %s\n", inf.model_name.c_str());
-    console::log("modalities : %s\n", modalities.c_str());
-    if (!params.system_prompt.empty()) {
-        console::log("using custom system prompt\n");
-    }
-    console::log("\n");
-    console::log("available commands:\n");
-    console::log("  /exit or Ctrl+C     stop or exit\n");
-    console::log("  /regen              regenerate the last response\n");
-    console::log("  /clear              clear the chat history\n");
-    console::log("  /read <file>        add a text file\n");
-    console::log("  /glob <pattern>     add text files using globbing pattern\n");
-    if (inf.has_inp_image) {
-        console::log("  /image <file>       add an image file\n");
-    }
-    if (inf.has_inp_audio) {
-        console::log("  /audio <file>       add an audio file\n");
-    }
-    if (inf.has_inp_video) {
-        console::log("  /video <file>       add a video file\n");
-    }
-    console::log("\n");
-
-    // interactive loop
-    std::string cur_msg;
-
-    auto add_text_file = [&](const std::string & fname) -> bool {
-        std::string marker = ctx_cli.load_input_file(fname, false);
-        if (marker.empty()) {
-            console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
-            return false;
-        }
-        if (inf.fim_sep_token != LLAMA_TOKEN_NULL) {
-            cur_msg += common_token_to_piece(ctx_cli.ctx_server.get_llama_context(), inf.fim_sep_token, true);
-            cur_msg += fname;
-            cur_msg.push_back('\n');
-        } else {
-            cur_msg += "--- File: ";
-            cur_msg += fname;
-            cur_msg += " ---\n";
-        }
-        cur_msg += marker;
-        console::log("Loaded text from '%s'\n", fname.c_str());
-        return true;
-    };
-
-    while (true) {
-        std::string buffer;
-        console::set_display(DISPLAY_TYPE_USER_INPUT);
-        if (params.prompt.empty()) {
-            console::log("\n> ");
-            std::string line;
-            bool another_line = true;
-            do {
-                another_line = console::readline(line, params.multiline_input);
-                buffer += line;
-            } while (another_line);
-        } else {
-            // process input prompt from args
-            for (auto & fname : params.image) {
-                std::string marker = ctx_cli.load_input_file(fname, true);
-                if (marker.empty()) {
-                    console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
-                    break;
-                }
-                console::log("Loaded media from '%s'\n", fname.c_str());
-                cur_msg += marker;
-            }
-            buffer = params.prompt;
-            if (buffer.size() > 500) {
-                console::log("\n> %s ... (truncated)\n", buffer.substr(0, 500).c_str());
-            } else {
-                console::log("\n> %s\n", buffer.c_str());
-            }
-            params.prompt.clear(); // only use it once
-        }
-        console::set_display(DISPLAY_TYPE_RESET);
-        console::log("\n");
-
-        if (should_stop()) {
-            g_is_interrupted.store(false);
-            break;
-        }
-
-        // remove trailing newline
-        if (!buffer.empty() &&buffer.back() == '\n') {
-            buffer.pop_back();
-        }
-
-        // skip empty messages
-        if (buffer.empty()) {
-            continue;
-        }
-
-        bool add_user_msg = true;
-
-        // process commands
-        if (string_starts_with(buffer, "/exit")) {
-            break;
-        } else if (string_starts_with(buffer, "/regen")) {
-            if (ctx_cli.messages.size() >= 2) {
-                size_t last_idx = ctx_cli.messages.size() - 1;
-                ctx_cli.messages.erase(last_idx);
-                add_user_msg = false;
-            } else {
-                console::error("No message to regenerate.\n");
-                continue;
-            }
-        } else if (string_starts_with(buffer, "/clear")) {
-            ctx_cli.messages.clear();
-            add_system_prompt();
-
-            ctx_cli.input_files.clear();
-            console::log("Chat history cleared.\n");
-            continue;
-        } else if (
-                (string_starts_with(buffer, "/image ") && inf.has_inp_image) ||
-                (string_starts_with(buffer, "/audio ") && inf.has_inp_audio) ||
-                (string_starts_with(buffer, "/video ") && inf.has_inp_video)) {
-            // just in case (bad copy-paste for example), we strip all trailing/leading spaces
-            std::string fname = string_strip(buffer.substr(7));
-            std::string marker = ctx_cli.load_input_file(fname, true);
-            if (marker.empty()) {
-                console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
-                continue;
-            }
-            cur_msg += marker;
-            console::log("Loaded media from '%s'\n", fname.c_str());
-            continue;
-        } else if (string_starts_with(buffer, "/read ")) {
-            std::string fname = string_strip(buffer.substr(6));
-            add_text_file(fname);
-            continue;
-        } else if (string_starts_with(buffer, "/glob ")) {
-            std::error_code ec;
-            size_t count = 0;
-            auto curdir = std::filesystem::current_path();
-            std::string pattern = string_strip(buffer.substr(6));
-            std::filesystem::path rel_path;
-
-            auto startglob = pattern.find_first_of("![*?");
-            if (startglob != std::string::npos && startglob != 0) {
-                auto endpath = pattern.substr(0, startglob).find_last_of('/');
-                if (endpath != std::string::npos) {
-                    std::string rel_pattern = pattern.substr(0, endpath);
-#if !defined(_WIN32)
-                    if (string_starts_with(rel_pattern, '~')) {
-                        const char * home = std::getenv("HOME");
-                        if (home && home[0]) {
-                            rel_pattern = home + rel_pattern.substr(1);
-                        }
-                    }
-#endif
-                    rel_path = rel_pattern;
-                    pattern.erase(0, endpath + 1);
-                    curdir /= rel_path;
-                }
-            }
-
-            for (const auto & entry : std::filesystem::recursive_directory_iterator(curdir,
-                    std::filesystem::directory_options::skip_permission_denied, ec)) {
-                if (!entry.is_regular_file()) {
-                    continue;
-                }
-
-                std::string rel = std::filesystem::relative(entry.path(), curdir, ec).string();
-                if (ec) {
-                    ec.clear();
-                    continue;
-                }
-                std::replace(rel.begin(), rel.end(), '\\', '/');
-
-                if (!glob_match(pattern, rel)) {
-                    continue;
-                }
-
-                if (!add_text_file((rel_path / rel).string())) {
-                    continue;
-                }
-
-                if (++count >= FILE_GLOB_MAX_RESULTS) {
-                    console::error("Maximum number of globbed files allowed (%zu) reached.\n", FILE_GLOB_MAX_RESULTS);
-                    break;
-                }
-            }
-            continue;
-        } else {
-            // not a command
-            cur_msg += buffer;
-        }
-
-        // generate response
-        if (add_user_msg) {
-            ctx_cli.messages.push_back({
-                {"role",    "user"},
-                {"content", cur_msg}
-            });
-            cur_msg.clear();
-        }
-        result_timings timings;
-        std::string assistant_content = ctx_cli.generate_completion(timings);
-        ctx_cli.messages.push_back({
-            {"role",    "assistant"},
-            {"content", assistant_content}
-        });
-        console::log("\n");
-
-        if (params.show_timings) {
-            console::set_display(DISPLAY_TYPE_INFO);
-            console::log("\n");
-            console::log("[ Prompt: %.1f t/s | Generation: %.1f t/s ]\n", timings.prompt_per_second, timings.predicted_per_second);
-            console::set_display(DISPLAY_TYPE_RESET);
-        }
-
-        if (params.single_turn) {
-            break;
-        }
-    }
-
-    console::set_display(DISPLAY_TYPE_RESET);
-
-    console::log("\nExiting...\n");
-    ctx_cli.ctx_server.terminate();
-    inference_thread.join();
-
-    // bump the log level to display timings
-    common_log_set_verbosity_thold(LOG_LEVEL_INFO);
-    common_memory_breakdown_print(ctx_cli.ctx_server.get_llama_context());
-
-    return 0;
+    return ctx_cli.run();
 }

tools/mtmd/clip-impl.h

@@ -42,6 +42,7 @@
 #define KEY_N_HEAD              "clip.%s.attention.head_count"
 #define KEY_N_HEAD_KV           "clip.%s.attention.head_count_kv"
 #define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+#define KEY_FEATURE_LAYERS      "clip.%s.feature_layer"
 
 // vision-specific
 #define KEY_VISION_PROJ_TYPE        "clip.vision.projector_type" // for models with mixed modalities
@@ -54,7 +55,6 @@
 #define KEY_PATCH_SIZE              "clip.vision.patch_size"
 #define KEY_IMAGE_MEAN              "clip.vision.image_mean"
 #define KEY_IMAGE_STD               "clip.vision.image_std"
-#define KEY_FEATURE_LAYER           "clip.vision.feature_layer"
 #define KEY_PROJ_SCALE_FACTOR       "clip.vision.projector.scale_factor"
 #define KEY_PROJ_SAMPLE_QUERY_SIDE  "clip.vision.projector.query_side"
 #define KEY_PROJ_SAMPLE_WINDOW_SIDE "clip.vision.projector.window_side"

tools/mtmd/clip-model.h

@@ -91,7 +91,7 @@ struct clip_hparams {
 
     float eps = 1e-6;
     float rope_theta = 0.0;
-    std::vector<int32_t> vision_feature_layer;
+    std::vector<int32_t> feature_layers;
     int32_t attn_window_size = 0;
     int32_t n_wa_pattern = 0;
     std::unordered_set<int32_t> wa_layer_indexes; // explicit layer indexes that use full attention (for irregular patterns like YoutuVL)
@@ -165,8 +165,8 @@ struct clip_hparams {
         return false;
     }
 
-    bool is_vision_feature_layer(int32_t layer) const {
-        return std::find(vision_feature_layer.begin(), vision_feature_layer.end(), layer) != vision_feature_layer.end();
+    bool is_feature_layer(int32_t layer) const {
+        return std::find(feature_layers.begin(), feature_layers.end(), layer) != feature_layers.end();
     }
 };
 

tools/mtmd/clip.cpp

@@ -1264,12 +1264,10 @@ struct clip_model_loader {
                 }
             }
 
-            // Load the vision feature layer indices if they are explicitly provided;
-            // if multiple vision feature layers are present, the values will be concatenated
-            // to form the final visual features.
+            // Load the vision/audio feature layer indices if they are explicitly provided
             // NOTE: gguf conversions should standardize the values of the vision feature layer to
             // be non-negative, since we use -1 to mark values as unset here.
-            get_arr_int(KEY_FEATURE_LAYER, hparams.vision_feature_layer, false);
+            get_arr_int(string_format(KEY_FEATURE_LAYERS, prefix), hparams.feature_layers, false);
 
             // model-specific params
             switch (model.proj_type) {
@@ -1651,6 +1649,7 @@ struct clip_model_loader {
                         get_u32(KEY_A_PROJ_WINDOW_SIZE,     hparams.audio_proj_window_size);
                         get_u32(KEY_A_PROJ_DOWNSAMPLE_RATE, hparams.audio_proj_downsample_rate);
                         get_u32(KEY_A_PROJ_HEAD_COUNT,      hparams.audio_proj_head_count);
+                        // NOTE: feature layers loaded above in common path
                     } break;
                 case PROJECTOR_TYPE_JANUS_PRO:
                     {
@@ -1663,11 +1662,11 @@ struct clip_model_loader {
                         hparams.image_resize_algo = RESIZE_ALGO_BICUBIC_PILLOW;
                         hparams.image_resize_pad = PAD_CEIL;
 
-                        get_arr_int(KEY_FEATURE_LAYER, hparams.vision_feature_layer);
+                        // NOTE: feature_layers loaded in common path as optional
                         get_arr_int(KEY_PROJ_SPATIAL_OFFSETS, hparams.proj_spatial_offsets);
-                        if (hparams.vision_feature_layer.size() != hparams.proj_spatial_offsets.size()) {
-                            throw std::runtime_error(string_format("%s: vision_feature_layer.size() %d != proj_spatial_offsets.size() %d",
-                                                                   hparams.vision_feature_layer.size(), hparams.proj_spatial_offsets.size()));
+                        if (hparams.feature_layers.size() != hparams.proj_spatial_offsets.size()) {
+                            throw std::runtime_error(string_format("%s: feature_layers.size() %d != proj_spatial_offsets.size() %d",
+                                                                   hparams.feature_layers.size(), hparams.proj_spatial_offsets.size()));
                         }
 
                         get_u32(KEY_PROJ_SAMPLE_QUERY_SIDE,  hparams.downsample_query_side);
@@ -2740,7 +2739,7 @@ struct clip_model_loader {
                     model.image_newline = get_tensor(TN_IMAGE_NEWLINE);
 
                     // Load separate layerwise and spatial projector tensors
-                    const auto projector_count = hparams.vision_feature_layer.size();
+                    const auto projector_count = hparams.feature_layers.size();
                     model.qf_proj_blocks.resize(projector_count);
                     for (size_t bid = 0; bid < projector_count; ++bid) {
                         auto & b = model.qf_proj_blocks[bid];
@@ -4388,7 +4387,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
 
                 // Stage 1b only uses block 0's permutations; future stages
                 // will upload all blocks.
-                for (size_t bid = 0; bid < hparams.vision_feature_layer.size(); ++bid) {
+                for (size_t bid = 0; bid < hparams.feature_layers.size(); ++bid) {
                     const std::string prefix = "g4v_blk" + std::to_string(bid) + "_";
                     upload(prefix + "win_idx",     make_win_idx(image_side, window_side));
                     upload(prefix + "qwin_idx",    make_win_idx(new_side, query_side));

tools/mtmd/models/granite-speech.cpp

@@ -1,5 +1,7 @@
 #include "models.h"
 
+#include <algorithm>
+
 ggml_cgraph * clip_graph_granite_speech::build() {
     const int n_frames     = img.nx();
     const int context_size = hparams.audio_chunk_size;
@@ -11,6 +13,10 @@ ggml_cgraph * clip_graph_granite_speech::build() {
     const int padded_len   = num_blocks * context_size;
     const int remainder    = n_frames % context_size;
 
+    // Calculate projector input dimension based on feature layers
+    const int proj_input_dim = n_embd * (hparams.feature_layers.size() + 1);
+    const bool use_feature_concat = !hparams.feature_layers.empty();
+
     ggml_tensor * attn_dists = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, context_size * context_size);
     ggml_set_name(attn_dists, "attn_dists");
     ggml_set_input(attn_dists);
@@ -31,6 +37,15 @@ ggml_cgraph * clip_graph_granite_speech::build() {
     cur = ggml_add(ctx0, cur, model.inp_proj_b);
     cb(cur, "inp_linear", -1);
 
+    // Capture layer 0 if requested (after input_linear)
+    ggml_tensor * concat_result = nullptr;
+    if (use_feature_concat) {
+        if (std::find(hparams.feature_layers.begin(), hparams.feature_layers.end(), 0) != hparams.feature_layers.end()) {
+            concat_result = cur;
+            cb(concat_result, "feature_layer_0", -1);
+        }
+    }
+
     for (int il = 0; il < n_layer; il++) {
         const auto & layer = model.layers[il];
         auto * residual = cur;
@@ -168,6 +183,18 @@ ggml_cgraph * clip_graph_granite_speech::build() {
                          NORM_TYPE_NORMAL, eps, il);
         cb(cur, "layer_out", il);
 
+        // Capture intermediate layer (il + 1) if requested
+        if (use_feature_concat) {
+            if (hparams.is_feature_layer(il + 1)) {
+                if (concat_result == nullptr) {
+                    concat_result = cur;
+                } else {
+                    concat_result = ggml_concat(ctx0, concat_result, cur, 0);
+                }
+                cb(concat_result, string_format("feature_layer_%d", il + 1).c_str(), il);
+            }
+        }
+
         // CTC branch
         if (il + 1 == ctc_layer) {
             auto * mid = build_mm(model.ctc_out_w, cur);
@@ -180,6 +207,13 @@ ggml_cgraph * clip_graph_granite_speech::build() {
         }
     }
 
+    // Append final output to concatenated features if using feature concatenation
+    if (use_feature_concat && concat_result != nullptr) {
+        concat_result = ggml_concat(ctx0, concat_result, cur, 0);
+        cb(concat_result, "concat_final", -1);
+        cur = concat_result;
+    }
+
     cb(cur, "encoder_out", -1);
 
     // QFormer projector
@@ -197,7 +231,7 @@ ggml_cgraph * clip_graph_granite_speech::build() {
             cur = ggml_pad(ctx0, cur, 0, padded_proj - n_frames, 0, 0);
         }
 
-        ggml_tensor * enc_windows = ggml_reshape_3d(ctx0, cur, n_embd, window_size, nblocks_proj);
+        ggml_tensor * enc_windows = ggml_reshape_3d(ctx0, cur, proj_input_dim, window_size, nblocks_proj);
 
         ggml_tensor * queries = build_norm(model.qf_proj_blocks[0].qf_proj_query,
             model.qf_proj_blocks[0].qf_proj_norm_w, model.qf_proj_blocks[0].qf_proj_norm_b,

tools/mtmd/models/granite4-vision.cpp

@@ -304,14 +304,14 @@ ggml_cgraph * clip_graph_granite4_vision::build() {
     }
 
     // --- Stage 1b/1c: WindowQFormer blocks ---
-    const int projector_count = hparams.vision_feature_layer.size();
+    const int projector_count = hparams.feature_layers.size();
     const float qformer_eps = 1e-12f;
 
     ggml_tensor * mmproj = nullptr;
     for (int bid = 0; bid < projector_count; ++bid) {
         const auto & blk = model.qf_proj_blocks[bid];
 
-        int vlayer = hparams.vision_feature_layer[bid];
+        int vlayer = hparams.feature_layers[bid];
         GGML_ASSERT(vlayer >= 0 && vlayer < n_layer);
         ggml_tensor * h = layer_outs[vlayer];
 

tools/mtmd/models/llava.cpp

@@ -21,7 +21,7 @@ ggml_cgraph * clip_graph_llava::build() {
 
         // If we set explicit vision feature layers, only go up to the deepest one
         // NOTE: only used by granite-vision models for now
-        for (const auto & feature_layer : hparams.vision_feature_layer) {
+        for (const auto & feature_layer : hparams.feature_layers) {
             if (feature_layer > deepest_feature_layer) {
                 deepest_feature_layer = feature_layer;
             }
@@ -59,7 +59,7 @@ ggml_cgraph * clip_graph_llava::build() {
 
         // If this is an embedding feature layer, save the output.
         // NOTE: 0 index here refers to the input to the encoder.
-        if (hparams.is_vision_feature_layer(il)) {
+        if (hparams.is_feature_layer(il)) {
             embedding_stack.push_back(cur);
         }
 
@@ -134,7 +134,7 @@ ggml_cgraph * clip_graph_llava::build() {
     // process vision feature layers (used by granite)
     {
         // final layer is a vision feature layer
-        if (hparams.is_vision_feature_layer(max_feature_layer)) {
+        if (hparams.is_feature_layer(max_feature_layer)) {
             embedding_stack.push_back(inpL);
         }
 

tools/server/README-dev.md

@@ -204,9 +204,9 @@ Instead of building everything from the ground up (like what most AI agents will
 
 The flow for downloading a new model:
 - POST request comes in --> `post_router_models` --> validation
-- `server_models::download()` is called
-    - Sets up a new thread `inst.th` and runs the download inside
-- If a stop request comes in, set `stop_download` to `true`
+- A new `llama-server` subprocess will be spawned with special `SERVER_CHILD_MODE_DOWNLOAD`
+- Child process runs the download and report status back to router via stdin/out
+- If a stop request comes in, the router asks the child process to stop (same mechanism as running a model in child process)
 - Otherwise, upon completion, we call `load_models()` to refresh the list of models
 
 ### Notable Related PRs

tools/server/README.md

@@ -1230,8 +1230,6 @@ print(completion.choices[0].text)
 
 Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
-If model supports multimodal, you can input the media file via `image_url` content part. We support both base64 and remote URL as input. See OAI documentation for more.
-
 *Options:*
 
 See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). llama.cpp `/completion`-specific features such as `mirostat` are also supported.
@@ -1250,9 +1248,18 @@ The `response_format` parameter supports both plain JSON output (e.g. `{"type":
 
 `parallel_tool_calls` : Whether to enable parallel/multiple tool calls (only supported on some models, verification is based on jinja template).
 
-For multimodal input:
-- Content type `image_url` and `input_audio` are the same as OAI schema
-- Content type `input_video` is an extension from OAI schema. For now, it only accepts base64 input
+For multimodal input (typed content, `messages[i].content[j]`):
+- If `type == "image_url"`:
+    - `image_url.url` can be a remote URL, base64 (raw or URI-encoded via `data:image/...;base64`) or path to local file
+    - Accepts formats supported by `stb_image` (jpeg, png, tga, bmp, gif, ...)
+- If `type == "input_audio"`:
+    - Either `input_audio.data` or `input_audio.url` can be specified, can be a remote URL, raw base64 or path to local file
+    - Accepts formats supported by `miniaudio` (mp3, wav, flac)
+    - `input_audio.format` will be ignored, the file format will be determined automatically
+- If `type == "input_video"`:
+    - Either `input_video.data` or `input_video.url` can be specified, can be a remote URL, raw base64 or path to local file
+    - Accepts formats supported by `ffmpeg`
+- Note: for local file, make sure to set `--media-path`. File path must be prefixed by `file://`
 
 *Examples:*
 

tools/server/server-common.cpp

@@ -518,6 +518,14 @@ size_t server_tokens::get_common_prefix(const server_tokens & b) const {
     return max_idx; // all tokens are equal
 }
 
+common_chat_msg_spans server_tokens::find_message_spans(const common_chat_msg_delimiters & delims) const {
+    std::map<size_t, size_t> skips;
+    for (const auto & it : map_idx_to_media) {
+        skips[it.first] = mtmd_input_chunk_get_n_tokens(it.second.get());
+    }
+    return delims.split(tokens, skips);
+}
+
 bool server_tokens::validate(const struct llama_context * ctx) const {
     const llama_model * model = llama_get_model(ctx);
     const llama_vocab * vocab = llama_model_get_vocab(model);
@@ -817,12 +825,21 @@ json oaicompat_completion_params_parse(const json & body) {
     return llama_params;
 }
 
-// media_path always end with '/', see arg.cpp
+// url can be
+// - http(s):// for remote files
+// - file:// for local files (only allowed if media_path is set)
+// - data: for base64 encoded data with uri scheme (e.g. data:image/png;base64,...)
+// - raw base64 encoded data
 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());
+        const std::string & url,
+        const std::string & media_path,
+        bool accept_base64_uri) {
+    if (!media_path.empty()) {
+        // should already be enforced by arg.cpp, but checking just in case
+        GGML_ASSERT(media_path.back() == DIRECTORY_SEPARATOR);
+    }
+
     if (string_starts_with(url, "http")) {
         // download remote image
         // TODO @ngxson : maybe make these params configurable
@@ -858,20 +875,28 @@ static void handle_media(
         data.assign((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
         out_files.push_back(data);
 
-    } else {
+    } else if (accept_base64_uri && string_starts_with(url, "data:")) {
         // 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");
+            throw std::runtime_error("Invalid uri-encoded base64 value");
         } else if (!string_starts_with(parts[0], "data:image/")) {
-            throw std::runtime_error("Invalid url format: " + parts[0]);
+            throw std::runtime_error("Invalid uri format: " + parts[0]);
         } else if (!string_ends_with(parts[0], "base64")) {
-            throw std::runtime_error("url must be base64 encoded");
+            throw std::runtime_error("uri must be base64 encoded");
         } else {
             auto base64_data = parts[1];
             auto decoded_data = base64_decode(base64_data);
             out_files.push_back(decoded_data);
         }
+
+    } else {
+        // try as raw base64 string
+        auto decoded_data = base64_decode(url);
+        if (decoded_data.empty()) {
+            throw std::runtime_error("Invalid base64 value");
+        }
+        out_files.push_back(decoded_data);
     }
 }
 
@@ -957,14 +982,15 @@ json oaicompat_chat_params_parse(
         }
 
         for (auto & p : content) {
-            std::string type      = json_value(p, "type", std::string());
+            std::string type = json_value(p, "type", std::string());
             if (type == "image_url") {
                 if (!opt.allow_image) {
                     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());
-                handle_media(out_files, image_url, opt.media_path);
+                std::string url = json_value(image_url, "url", std::string());
+                handle_media(out_files, url, opt.media_path, true);
 
                 p["type"] = "media_marker";
                 p["text"] = get_media_marker();
@@ -975,17 +1001,11 @@ json oaicompat_chat_params_parse(
                     throw std::runtime_error("audio input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
                 }
 
-                json input_audio   = json_value(p, "input_audio", json::object());
-                std::string data   = json_value(input_audio, "data", std::string());
-                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::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()
+                // note: don't need to validate "format", it's redundant
+                json input_audio = json_value(p, "input_audio", json::object());
+                std::string url  = json_value(input_audio, "data",
+                                        json_value(input_audio, "url", std::string()));
+                handle_media(out_files, url, opt.media_path, false);
 
                 p["type"] = "media_marker";
                 p["text"] = get_media_marker();
@@ -996,10 +1016,10 @@ json oaicompat_chat_params_parse(
                     throw std::runtime_error("video input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
                 }
 
-                json input_video  = json_value(p, "input_video", json::object());
-                std::string data  = json_value(input_video, "data", std::string());
-                auto decoded_data = base64_decode(data); // expected to be base64 encoded
-                out_files.push_back(decoded_data);
+                json input_video = json_value(p, "input_video", json::object());
+                std::string url  = json_value(input_video, "data",
+                                        json_value(input_video, "url", std::string()));
+                handle_media(out_files, url, opt.media_path, false);
 
                 p["type"] = "media_marker";
                 p["text"] = get_media_marker();
@@ -1092,15 +1112,7 @@ json oaicompat_chat_params_parse(
         llama_params["chat_parser"] = chat_params.parser;
     }
 
-    llama_params["message_spans"] = json::array();
-
-    for (const auto & span : chat_params.message_spans) {
-        llama_params["message_spans"].push_back({
-            { "role", span.role },
-            { "pos",  span.pos  },
-            { "len",  span.len  },
-        });
-    }
+    llama_params["message_delimiters"] = chat_params.message_delimiters.to_json();
 
     // Reasoning budget: pass parameters through to sampling layer
     {

tools/server/server-common.h

@@ -218,6 +218,9 @@ public:
 
     size_t get_common_prefix(const server_tokens & b) const;
 
+    // split the tokens into message spans, skipping over media chunks
+    common_chat_msg_spans find_message_spans(const common_chat_msg_delimiters & delims) const;
+
     // make sure all text tokens are within the vocab range
     bool validate(const struct llama_context * ctx) const;
 

tools/server/server-context.cpp

@@ -931,6 +931,8 @@ private:
 
     bool sleeping = false;
 
+    int64_t t_last_load_progress_ms = 0;
+
     void destroy() {
         spec.reset();
         ctx_dft.reset();
@@ -1244,6 +1246,10 @@ private:
         }
 
         if (has_mmproj) {
+            if (callback_state) {
+                callback_state(SERVER_STATE_LOADING, {{"stage", "mmproj_model"}});
+            }
+
             if (!is_resume) {
                 mtmd_helper_log_set(common_log_default_callback, nullptr);
             }
@@ -3430,8 +3436,8 @@ private:
                         has_mtmd = true;
                     }
 
-                    const int32_t n_before_user = slot.task->params.n_before_user;
-                    const bool n_before_user_known = n_before_user > 0;
+                    const auto & spans = slot.task->params.message_spans;
+                    const auto last_user_pos = spans.last_user_message_pos();
 
                     // add prompt tokens for processing in the current batch
                     while (slot.prompt.n_tokens() < slot.task->n_tokens() && batch.size() < n_batch) {
@@ -3460,10 +3466,8 @@ private:
 
                         slot.n_prompt_tokens_processed++;
 
-                        // stop the prompt batch exactly before the latest user input, so a checkpoint
-                        // can be created after the previous messages
-                        if (n_before_user_known &&
-                            slot.prompt.n_tokens() == n_before_user) {
+                        // stop the prompt batch exactly before a user message
+                        if (spans.is_user_start(slot.prompt.n_tokens())) {
                             break;
                         }
 
@@ -3492,8 +3496,13 @@ private:
                     // the number of tokens added to the batch for the current slot
                     const auto n_tokens_cur = batch.size() - n_tokens_prev;
 
+                    const auto n_tokens_start = slot.prompt.n_tokens() - n_tokens_cur;
+
                     const bool near_prompt_end = slot.task->n_tokens() < slot.prompt.n_tokens() + n_ubatch;
 
+                    const bool is_user_start = spans.is_user_start(n_tokens_start);
+                    const bool is_last_user_message = n_tokens_start == last_user_pos;
+
                     // entire prompt has been processed
                     if (slot.prompt.n_tokens() == slot.task->n_tokens()) {
                         slot.state = SLOT_STATE_DONE_PROMPT;
@@ -3508,8 +3517,9 @@ private:
 
                         slot.init_sampler();
                     } else {
-                        // skip ordinary mid-prompt checkpoints
-                        if (!n_before_user_known && !near_prompt_end) {
+                        // skip ordinary mid-prompt checkpoints, unless the batch starts a user
+                        // message or we are near the end of the prompt
+                        if (!is_user_start && !near_prompt_end) {
                             do_checkpoint = false;
                         }
                     }
@@ -3517,29 +3527,6 @@ private:
                     const auto pos_min = llama_memory_seq_pos_min(llama_get_memory(ctx_tgt), slot.id);
                     const auto pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_tgt), slot.id);
 
-                    // checkpoints are created before the current batch is decoded, so
-                    // their token position is the batch start rather than the prompt end
-                    const int32_t n_tokens_start = slot.prompt.n_tokens() - n_tokens_cur;
-
-                    {
-                        const bool is_on_user =
-                            n_before_user_known &&
-                            n_tokens_start == n_before_user;
-
-                        const bool is_after_user =
-                            n_before_user_known &&
-                            n_tokens_start > n_before_user;
-
-                        const bool is_allowed =
-                            !n_before_user_known ||
-                            is_on_user ||
-                            (is_after_user && near_prompt_end);
-
-                        if (do_checkpoint && !is_allowed) {
-                            do_checkpoint = false;
-                        }
-                    }
-
                     // nothing to checkpoint yet
                     // TODO: is this check needed?
                     if (do_checkpoint && pos_min < 0) {
@@ -3549,8 +3536,8 @@ private:
                     // do not checkpoint after mtmd chunks
                     do_checkpoint = do_checkpoint && !has_mtmd;
 
-                    // no need to create checkpoints that are too close together
-                    do_checkpoint = do_checkpoint && (slot.prompt.checkpoints.empty() || n_tokens_start > slot.prompt.checkpoints.back().n_tokens + params_base.checkpoint_min_step);
+                    // no need to create checkpoints that are too close together, unless it's the last user message
+                    do_checkpoint = do_checkpoint && (slot.prompt.checkpoints.empty() || is_last_user_message || n_tokens_start > slot.prompt.checkpoints.back().n_tokens + params_base.checkpoint_min_step);
                     SLT_DBG(slot, "main/do_checkpoint = %s, pos_min = %d, pos_max = %d\n", do_checkpoint ? "yes" : "no", pos_min, pos_max);
 
                     // note: we create the checkpoint before calling llama_decode(), so the current batch is not
@@ -4049,54 +4036,6 @@ void server_context::set_state_callback(server_state_callback_t callback) {
     });
 }
 
-// compute the number of tokens before the last user message in the prompt
-static int32_t prompt_get_n_before_user(
-        const json & message_spans,
-        const std::string & prompt,
-        const std::vector<raw_buffer> & files,
-        const llama_vocab * vocab,
-        mtmd_context * mctx) {
-    int32_t result = -1;
-    int32_t byte_pos = -1;
-
-    for (const auto & span : message_spans) {
-        const std::string role = json_value(span, "role", std::string());
-
-        if (role == "user") {
-            byte_pos = json_value(span, "pos", -1);
-        }
-    }
-
-    if (byte_pos >= 0) {
-        GGML_ASSERT((size_t) byte_pos <= prompt.size());
-
-        const std::string prefix = prompt.substr(0, (size_t) byte_pos);
-
-        const std::string marker = get_media_marker();
-        size_t n_prefix_media = 0;
-        for (size_t pos = 0; (pos = prefix.find(marker, pos)) != std::string::npos; pos += marker.size()) {
-            n_prefix_media++;
-        }
-
-        GGML_ASSERT(n_prefix_media <= files.size());
-
-        if (mctx != nullptr && n_prefix_media > 0) {
-            // TODO: this makes a copy - avoid it
-            std::vector<raw_buffer> prefix_files(files.begin(), files.begin() + n_prefix_media);
-
-            result = (int32_t) process_mtmd_prompt(mctx, prefix, prefix_files).size();
-        } else {
-            result = (int32_t) tokenize_input_prompts(vocab, nullptr, prefix, true, true)[0].size();
-        }
-
-        SRV_TRC("message_spans: last user message: byte_pos=%d, media=%zu, n_before_user=%d\n",
-                byte_pos, n_prefix_media, result);
-    }
-
-    return result;
-}
-
-
 //
 // server_routes
 //
@@ -4144,6 +4083,10 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
 
         // tasks.reserve(inputs.size()); // TODO: this is inaccurate due to child tasks
 
+        // message delimiters for checkpointing
+        auto delimiters = common_chat_msg_delimiters_parse(json_value(data, "message_delimiters", json::array()));
+        delimiters.tokenize(ctx_server.vocab);
+
         for (size_t i = 0; i < inputs.size(); i++) {
             server_task task = server_task(type);
 
@@ -4157,16 +4100,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
                     meta->logit_bias_eog,
                     data);
 
-            const auto message_spans = json_value(data, "message_spans", json::array());
-            if (prompt.is_string() && message_spans.is_array()) {
-                task.params.n_before_user =
-                    prompt_get_n_before_user(
-                        message_spans,
-                        prompt.get<std::string>(),
-                        files,
-                        ctx_server.vocab,
-                        ctx_server.mctx);
-            }
+            task.params.message_spans = task.tokens.find_message_spans(delimiters);
 
             task.id_slot = json_value(data, "id_slot", -1);
 

tools/server/server-context.h

@@ -53,7 +53,7 @@ struct server_context_meta {
 };
 
 enum server_state {
-    // SERVER_STATE_DOWNLOADING,
+    SERVER_STATE_DOWNLOADING,
     SERVER_STATE_LOADING,
     SERVER_STATE_READY,
     SERVER_STATE_SLEEPING,
@@ -61,6 +61,7 @@ enum server_state {
 
 static std::string server_state_to_str(server_state state) {
     switch (state) {
+        case SERVER_STATE_DOWNLOADING: return "downloading";
         case SERVER_STATE_LOADING:     return "loading";
         case SERVER_STATE_READY:       return "ready";
         case SERVER_STATE_SLEEPING:    return "sleeping";
@@ -69,6 +70,7 @@ static std::string server_state_to_str(server_state state) {
 }
 
 static server_state server_state_from_str(const std::string & str) {
+    if (str == "downloading") return SERVER_STATE_DOWNLOADING;
     if (str == "loading")     return SERVER_STATE_LOADING;
     if (str == "ready")       return SERVER_STATE_READY;
     if (str == "sleeping")    return SERVER_STATE_SLEEPING;

tools/server/server-models.cpp

@@ -5,6 +5,7 @@
 #include "build-info.h"
 #include "preset.h"
 #include "download.h"
+#include "http.h"
 
 #include <cpp-httplib/httplib.h> // TODO: remove this once we use HTTP client from download.h
 #include <sheredom/subprocess.h>
@@ -25,14 +26,7 @@
 #include <sstream>
 #include <cstring>
 
-#ifdef _WIN32
-#include <winsock2.h>
-#include <windows.h>
-#else
-#include <sys/socket.h>
-#include <netinet/in.h>
-#include <arpa/inet.h>
-#include <unistd.h>
+#ifndef _WIN32
 extern char **environ;
 #endif
 
@@ -64,6 +58,17 @@ struct server_subproc {
         return sproc.has_value() && subprocess_alive(&sproc.value());
     }
 
+    void request_exit() {
+        if (sproc.has_value()) {
+            FILE * stdin_file = subprocess_stdin(&sproc.value());
+            if (stdin_file) {
+                fprintf(stdin_file, "%s\n", CMD_ROUTER_TO_CHILD_EXIT);
+                fflush(stdin_file);
+            }
+        }
+        stopped.store(true, std::memory_order_relaxed);
+    }
+
     void terminate() {
         if (!sproc.has_value()) {
             return;
@@ -213,7 +218,7 @@ void server_model_meta::update_caps() {
         });
         params.offline = true;
         // params.skip_download = true; // TODO: ideally, we should validate the model here, but it takes too much time
-        common_params_handle_models(params, LLAMA_EXAMPLE_SERVER);
+        common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, {});
         if (params.mmproj.path.empty()) {
             multimodal = { false, false };
         } else {
@@ -323,7 +328,7 @@ void server_models::notify_sse(const std::string & event, const std::string & mo
 }
 
 void server_models::load_models() {
-    // Phase 1: load presets from all sources — pure I/O, no lock needed
+    // Phase 1: load presets from all sources - pure I/O, no lock needed
     // 1. cached models
     common_presets cached_models = ctx_preset.load_from_cache();
     SRV_INF("Loaded %zu cached model presets\n", cached_models.size());
@@ -376,7 +381,7 @@ void server_models::load_models() {
         return source_map.count(name) ? source_map.at(name) : SERVER_MODEL_SOURCE_PRESET;
     };
 
-    // Helpers that read `mapping` — must be called while holding the lock.
+    // Helpers that read `mapping` - must be called while holding the lock.
     std::unordered_set<std::string> custom_names;
     for (const auto & [name, preset] : custom_presets) custom_names.insert(name);
     auto join_set = [](const std::set<std::string> & s) {
@@ -523,7 +528,7 @@ void server_models::load_models() {
             }
         }
 
-        // join outside the lock — monitoring thread calls update_status (needs lock)
+        // join outside the lock - monitoring thread calls update_status (needs lock)
         lk.unlock();
         for (auto & th : threads_to_join) th.join();
         lk.lock();
@@ -622,7 +627,7 @@ void server_models::load_models() {
 
         apply_stop_timeout();
 
-        // clear reload flag before unlocking for autoload — load() blocks on !is_reloading,
+        // clear reload flag before unlocking for autoload - load() blocks on !is_reloading,
         // so clearing it here (while still locked) prevents a deadlock in the autoload calls below
         is_reloading = false;
         cv.notify_all();
@@ -693,66 +698,6 @@ std::optional<server_model_meta> server_models::get_meta(const std::string & nam
     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) {
@@ -815,17 +760,23 @@ void server_models::unload_lru() {
 }
 
 void server_models::load(const std::string & name) {
-    if (!has_model(name)) {
-        throw std::runtime_error("model name=" + name + " is not found");
+    load(name, load_options{});
+}
+
+void server_models::load(const std::string & name, const load_options & opts) {
+    if (!opts.custom_meta.has_value()) {
+        if (!has_model(name)) {
+            throw std::runtime_error("model name=" + name + " is not found");
+        }
+        unload_lru();
     }
-    unload_lru();
 
     std::unique_lock<std::mutex> lk(mutex);
     // edge case: block until any in-progress reload has finished so we always load
     // against the freshest preset and a consistent mapping state
     cv.wait(lk, [this]() { return !is_reloading; });
 
-    auto meta = mapping[name].meta;
+    auto meta = opts.custom_meta.has_value() ? *opts.custom_meta : mapping[name].meta;
     if (meta.status != SERVER_MODEL_STATUS_UNLOADED) {
         SRV_INF("model %s is not ready\n", name.c_str());
         return;
@@ -850,7 +801,7 @@ void server_models::load(const std::string & name) {
     // prepare new instance info
     instance_t inst;
     inst.meta             = meta;
-    inst.meta.port        = get_free_port();
+    inst.meta.port        = common_http_get_free_port();
     inst.meta.status      = SERVER_MODEL_STATUS_LOADING;
     inst.meta.loaded_info = json{};
     inst.meta.last_used   = ggml_time_ms();
@@ -869,6 +820,12 @@ void server_models::load(const std::string & name) {
         std::vector<std::string> child_env  = base_env; // copy
         child_env.push_back("LLAMA_SERVER_ROUTER_PORT=" + std::to_string(base_params.port));
 
+        if (opts.mode == SERVER_CHILD_MODE_DOWNLOAD) {
+            inst.meta.status = SERVER_MODEL_STATUS_DOWNLOADING;
+            child_env.push_back("LLAMA_SERVER_CHILD_MODE=download");
+            child_env.push_back("LLAMA_ARG_HF_REPO=" + name);
+        }
+
         SRV_INF("%s", "spawning server instance with args:\n");
         for (const auto & arg : child_args) {
             SRV_INF("  %s\n", arg.c_str());
@@ -886,13 +843,17 @@ void server_models::load(const std::string & name) {
         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, stop_timeout = inst.meta.stop_timeout]() {
+    inst.th = std::thread([
+        this, name,
+        child_proc = inst.subproc,
+        port = inst.meta.port,
+        stop_timeout = inst.meta.stop_timeout,
+        child_mode = opts.mode
+    ]() {
         FILE * stdin_file = subprocess_stdin(&child_proc->get());
         FILE * stdout_file = subprocess_stdout(&child_proc->get()); // combined stdout/stderr
 
@@ -925,7 +886,7 @@ void server_models::load(const std::string & name) {
                     return is_stopping() || child_proc->stopped.load(std::memory_order_acquire);
                 });
             }
-            // child crashed or finished on its own — skip graceful shutdown sequence
+            // child crashed or finished on its own, skip graceful shutdown sequence
             if (child_proc->stopped.load(std::memory_order_acquire)) {
                 return;
             }
@@ -973,10 +934,14 @@ void server_models::load(const std::string & name) {
         subprocess_destroy(&child_proc->get());
 
         // update status and exit code
-        this->update_status(name, {
-            SERVER_MODEL_STATUS_UNLOADED,
-            exit_code
-        });
+        if (child_mode == SERVER_CHILD_MODE_DOWNLOAD) {
+            // instance will be cleaned up on next load_models() call
+        } else {
+            this->update_status(name, {
+                SERVER_MODEL_STATUS_UNLOADED,
+                exit_code
+            });
+        }
         SRV_INF("instance name=%s exited with status %d\n", name.c_str(), exit_code);
     });
 
@@ -984,7 +949,7 @@ void server_models::load(const std::string & name) {
     {
         auto & old_instance = mapping[name];
         // old process should have exited already, but just in case, we clean it up here
-        if (old_instance.subproc->is_alive()) {
+        if (old_instance.subproc && old_instance.subproc->is_alive()) {
             SRV_WRN("old process for model name=%s is still alive, this is unexpected\n", name.c_str());
             old_instance.subproc->terminate(); // force kill
         }
@@ -1001,92 +966,13 @@ void server_models::load(const std::string & name) {
     cv.notify_all();
 }
 
-// callback for model downloading functionality
-struct server_models_download_res : public common_download_callback {
-    common_params_model model;
-    common_download_opts opts;
-
-    std::function<bool()> should_stop;
-    std::function<void(const common_download_progress & p)> on_progress;
-
-    bool is_ok = false;
-
-    bool run() {
-        try {
-            common_download_model(model, opts);
-            is_ok = true;
-        } catch (const std::exception & e) {
-            auto model_name = model.get_name();
-            SRV_ERR("download failed for model name=%s: %s\n", model_name.c_str(), e.what());
-            is_ok = false;
-        }
-        return is_ok;
-    }
-    void on_start(const common_download_progress & p) override {
-        on_progress(p);
-    }
-    void on_update(const common_download_progress & p) override {
-        on_progress(p);
-    }
-    void on_done(const common_download_progress &, bool ok) override {
-        is_ok = ok;
-    }
-    bool is_cancelled() const override {
-        return should_stop();
-    }
-};
-
-void server_models::download(common_params_model && model, common_download_opts && opts) {
-    std::string name = model.get_name();
-    GGML_ASSERT(name == model.hf_repo);
-
-    std::unique_lock<std::mutex> lk(mutex);
-    if (mapping.find(name) != mapping.end()) {
-        throw std::runtime_error("model name=" + name + " already exists");
-    }
-
-    instance_t inst;
-    inst.meta.name   = name;
-    inst.meta.status = SERVER_MODEL_STATUS_DOWNLOADING;
-    inst.subproc     = std::make_shared<server_subproc>();
-
-    auto dl = std::make_unique<server_models_download_res>();
-    dl->model = model; // copy
-    dl->opts  = opts;  // copy
-
-    dl->should_stop = [sp = inst.subproc]() {
-        return sp->stopped.load(std::memory_order_relaxed);
-    };
-
-    dl->on_progress = [this, name](const common_download_progress & p) {
-        update_download_progress(name, p, false);
-    };
-
-    inst.th = std::thread([this, dl = std::move(dl)]() {
-        dl->opts.callback = dl.get();
-        bool ok = dl->run();
-        auto model_name = dl->model.get_name();
-        SRV_INF("download finished for model name=%s with status=%s\n",
-                    model_name.c_str(), ok ? "success" : "failure");
-        update_download_progress(model_name, {}, true, ok);
-        // need_reload is set inside update_download_progress under the mutex;
-        // the next load_models() call will clean up this instance
-    });
-
-    mapping[name] = std::move(inst);
-    notify_sse("status_update", name, {
-        {"status", server_model_status_to_string(SERVER_MODEL_STATUS_DOWNLOADING)},
-    });
-    cv.notify_all();
-}
-
 void server_models::unload(const std::string & name) {
     std::unique_lock<std::mutex> lk(mutex);
     auto it = mapping.find(name);
     if (it != mapping.end()) {
         if (it->second.meta.status == SERVER_MODEL_STATUS_DOWNLOADING) {
             SRV_INF("cancelling download for model name=%s\n", name.c_str());
-            it->second.subproc->stopped.store(true, std::memory_order_relaxed);
+            it->second.subproc->request_exit();
             // for convenience, we wait the status change here
             wait(lk, name, [](const server_model_meta & new_meta) {
                 return new_meta.status != SERVER_MODEL_STATUS_DOWNLOADING;
@@ -1198,37 +1084,65 @@ void server_models::update_download_progress(const std::string & name, const com
 }
 
 bool server_models::remove(const std::string & name) {
-    auto meta = get_meta(name);
+    // do everything under one lock acquisition; avoid get_meta() /
+    // unload() because they can trigger load_models() which erases
+    // transient DOWNLOADING / DOWNLOADED entries as a side-effect
+    std::unique_lock<std::mutex> lk(mutex);
 
-    if (!meta.has_value()) {
+    auto it = mapping.find(name);
+    if (it == mapping.end()) {
         throw std::runtime_error("model name=" + name + " is not found");
     }
-    if (meta->source != SERVER_MODEL_SOURCE_CACHE) {
+    if (it->second.meta.source != SERVER_MODEL_SOURCE_CACHE) {
         throw std::runtime_error("model name=" + name + " is not removable (not from cache)");
     }
 
-    unload(name); // cancel download or stop running instance
-    {
-        std::unique_lock<std::mutex> lk(mutex);
-        // a cancelled download lands on DOWNLOADED; a stopped instance lands on UNLOADED
-        wait(lk, name, [](const server_model_meta & new_meta) {
-            return new_meta.status == SERVER_MODEL_STATUS_UNLOADED
-                || new_meta.status == SERVER_MODEL_STATUS_DOWNLOADED;
-        });
-        // join before erasing - after status reaches UNLOADED/DOWNLOADED the thread no
-        // longer acquires this mutex, so joining while holding it is safe
-        if (mapping[name].th.joinable()) {
-            mapping[name].th.join();
+    if (it->second.meta.status == SERVER_MODEL_STATUS_DOWNLOADING) {
+        // cancel in-flight download
+        SRV_INF("cancelling download for model name=%s\n", name.c_str());
+        it->second.subproc->request_exit();
+    } else if (it->second.meta.is_running()) {
+        // stop running instance
+        SRV_INF("stopping model instance name=%s\n", name.c_str());
+        stopping_models.insert(name);
+        if (it->second.meta.status == SERVER_MODEL_STATUS_LOADING) {
+            it->second.subproc->terminate();
         }
-        // remove the model from disk (hold lock to prevent concurrent load)
-        bool ok = common_download_remove(name);
-        if (ok) {
-            mapping.erase(name);
-        }
-        SRV_INF("removing model name=%s from cache (%s)\n", name.c_str(), ok ? "succeeded" : "failed");
-        notify_sse("model_remove", name, {});
-        return ok;
+        cv_stop.notify_all();
     }
+
+    // wait until the monitoring thread finishes
+    wait(lk, name, [](const server_model_meta & meta) {
+        return meta.status == SERVER_MODEL_STATUS_UNLOADED
+            || meta.status == SERVER_MODEL_STATUS_DOWNLOADED;
+    });
+
+    // re-find after wait - load_models() may have erased the entry during the wait
+    it = mapping.find(name);
+    if (it == mapping.end()) {
+        // load_models() already joined the thread and erased the entry;
+        // we just need to clean up the cached files on disk
+        lk.unlock();
+        bool ok = common_download_remove(name);
+        SRV_INF("removing model name=%s from cache (%s)\n", name.c_str(), ok ? "succeeded" : "partial");
+        notify_sse("model_remove", name, {});
+        return true;
+    }
+
+    // join before erasing - thread no longer acquires this mutex
+    if (it->second.th.joinable()) {
+        it->second.th.join();
+    }
+
+    // remove from disk (best-effort: cancelled downloads may have no cached files)
+    bool ok = common_download_remove(name);
+    mapping.erase(name);
+    if (!ok) {
+        SRV_WRN("removing model name=%s from disk returned false (no cached files?)\n", name.c_str());
+    }
+    SRV_INF("removing model name=%s from cache (%s)\n", name.c_str(), ok ? "succeeded" : "partial");
+    notify_sse("model_remove", name, {});
+    return true;
 }
 
 void server_models::wait(const std::string & name, std::function<bool(const server_model_meta &)> predicate) {
@@ -1243,7 +1157,9 @@ void server_models::wait(std::unique_lock<std::mutex> & lk, const std::string &
             return predicate(it->second.meta);
 
         }
-        return false;
+        // model was removed from mapping by another code path (e.g. load_models()).
+        // nothing left to wait for - tell the caller to proceed.
+        return true;
     });
 }
 
@@ -1328,6 +1244,31 @@ void server_models::handle_child_state(const std::string & name, const std::stri
     }
 
     switch (state) {
+        case SERVER_STATE_DOWNLOADING:
+            {
+                std::string result = json_value(payload, "result", std::string());
+                std::string url    = json_value(payload, "url",    std::string());
+                auto request_exit = [&]() {
+                    std::lock_guard<std::mutex> lk(mutex);
+                    auto it = mapping.find(name);
+                    if (it != mapping.end()) {
+                        return it->second.subproc->request_exit();
+                    }
+                };
+                if (result == "download_finished") {
+                    update_download_progress(name, {}, true, true);
+                    request_exit();
+                } else if (result == "download_failed") {
+                    update_download_progress(name, {}, true, false);
+                    request_exit();
+                } else if (!url.empty()) {
+                    common_download_progress p;
+                    p.url        = url;
+                    p.downloaded = json_value(payload, "downloaded", (size_t)0);
+                    p.total      = json_value(payload, "total", (size_t)0);
+                    update_download_progress(name, p, false);
+                }
+            } break;
         case SERVER_STATE_LOADING:
             {
                 update_status(name, {
@@ -1366,6 +1307,92 @@ bool server_child::is_child() {
     return router_port != nullptr;
 }
 
+server_child_mode server_child::get_mode() {
+    const char * mode = std::getenv("LLAMA_SERVER_CHILD_MODE");
+    std::string mode_str(mode ? mode : "");
+    if (mode_str == "download") {
+        return SERVER_CHILD_MODE_DOWNLOAD;
+    } else {
+        return SERVER_CHILD_MODE_NORMAL;
+    }
+}
+
+struct server_download_state : public common_download_callback {
+    server_child * self;
+    std::function<bool()> should_stop;
+    std::atomic<int64_t> last_progress_time{0}; // multiple files downloading in different threads
+    bool is_ok = false;
+
+    server_download_state(server_child * s) : self(s) {}
+
+    bool run(common_params & params) {
+        try {
+            common_params_handle_models_params p;
+            p.callback = this;
+            common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, p);
+            is_ok = true;
+        } catch (const std::exception & e) {
+            auto model_name = params.model.get_name();
+            SRV_ERR("download failed for model name=%s: %s\n", model_name.c_str(), e.what());
+            is_ok = false;
+        }
+        return is_ok;
+    }
+    void on_progress(const common_download_progress & p) {
+        json data = {
+            {"url", p.url},
+            {"downloaded", p.downloaded},
+            {"total", p.total},
+        };
+        self->notify_to_router(server_state_to_str(SERVER_STATE_DOWNLOADING), data);
+    }
+    void on_start(const common_download_progress & p) override {
+        on_progress(p);
+    }
+    void on_update(const common_download_progress & p) override {
+        int64_t now = ggml_time_ms();
+        // throttle progress updates to avoid flooding logs
+        if (now - last_progress_time.load(std::memory_order_relaxed) >= 100) {
+            on_progress(p);
+            last_progress_time.store(now, std::memory_order_relaxed);
+        }
+    }
+    void on_done(const common_download_progress & p, bool) override {
+        on_progress(p);
+    }
+    bool is_cancelled() const override {
+        return should_stop ? should_stop() : false;
+    }
+};
+
+int server_child::run_download(common_params & params) {
+    auto cancelled = std::make_shared<std::atomic<bool>>(false);
+
+    // monitor stdin for cancellation command from the router
+    std::thread signal_thread = setup([cancelled](int) {
+        cancelled->store(true, std::memory_order_relaxed);
+    });
+
+    server_download_state dl(this);
+    dl.should_stop = [cancelled]() {
+        return cancelled->load(std::memory_order_relaxed);
+    };
+
+    bool ok = dl.run(params);
+
+    notify_to_router(server_state_to_str(SERVER_STATE_DOWNLOADING), {
+        {"result", ok ? "download_finished" : "download_failed"},
+    });
+
+    // router should send CMD_ROUTER_TO_CHILD_EXIT after receiving the result
+    if (signal_thread.joinable()) {
+        signal_thread.join();
+    }
+
+    SRV_INF("download completed %s\n", ok ? "successfully" : "with errors");
+    return 0;
+}
+
 std::thread server_child::setup(const std::function<void(int)> & shutdown_handler) {
     // setup thread for monitoring stdin
     return std::thread([shutdown_handler]() {
@@ -1639,7 +1666,7 @@ void server_models_routes::init_routes() {
             res_err(res, format_error_response("model is not found", ERROR_TYPE_INVALID_REQUEST));
             return res;
         }
-        if (!model->is_running()) {
+        if (!model->is_running() && model->status != SERVER_MODEL_STATUS_DOWNLOADING) {
             res_err(res, format_error_response("model is not running", ERROR_TYPE_INVALID_REQUEST));
             return res;
         }
@@ -1680,8 +1707,9 @@ void server_models_routes::init_routes() {
 
         model.hf_repo        = name;
         opts.bearer_token    = params.hf_token;
-        opts.download_mmproj = true;
-        opts.download_mtp    = true;
+        // note: we only check main model, no need sidecar here
+        opts.download_mmproj = false;
+        opts.download_mtp    = false;
 
         // first, only check if the model is valid and can be downloaded
         opts.skip_download = true;
@@ -1702,10 +1730,21 @@ void server_models_routes::init_routes() {
             throw std::invalid_argument("model validation failed, unable to download");
         }
 
+        // reject if model already exists
+        if (models.has_model(name)) {
+            throw std::invalid_argument("model '" + name + "' already exists");
+        }
+
         // then, proceed with the actual download
-        opts.skip_download = false;
         SRV_INF("starting download for model '%s'\n", name.c_str());
-        models.download(std::move(model), std::move(opts));
+        {
+            server_models::load_options load_opts;
+            load_opts.mode = SERVER_CHILD_MODE_DOWNLOAD;
+            load_opts.custom_meta = server_model_meta{};
+            load_opts.custom_meta->source = SERVER_MODEL_SOURCE_CACHE;
+            load_opts.custom_meta->name   = name;
+            models.load(name, load_opts);
+        }
 
         res_ok(res, {{"success", true}});
         return res;
@@ -1719,10 +1758,7 @@ void server_models_routes::init_routes() {
             throw std::invalid_argument("model must be a non-empty string");
         }
 
-        bool ok = models.remove(name);
-        if (!ok) {
-            throw std::runtime_error("failed to remove model '" + name + "'");
-        }
+        models.remove(name); // throws on error
 
         res_ok(res, {{"success", true}});
         return res;

tools/server/server-models.h

@@ -40,6 +40,11 @@ enum server_model_source {
     SERVER_MODEL_SOURCE_CACHE,
 };
 
+enum server_child_mode {
+    SERVER_CHILD_MODE_NORMAL,   // load the model and run normally
+    SERVER_CHILD_MODE_DOWNLOAD, // download the model and exit
+};
+
 static std::string server_model_status_to_string(server_model_status status) {
     switch (status) {
         case SERVER_MODEL_STATUS_DOWNLOADING: return "downloading";
@@ -105,7 +110,6 @@ private:
         std::shared_ptr<server_subproc> subproc; // shared between main thread and monitoring thread
         std::thread th;
         server_model_meta meta;
-        FILE * stdin_file = nullptr;
     };
 
     std::mutex mutex;
@@ -161,16 +165,19 @@ public:
     // return a copy of all model metadata (thread-safe)
     std::vector<server_model_meta> get_all_meta();
 
+    struct load_options {
+        server_child_mode mode = SERVER_CHILD_MODE_NORMAL;
+        // used for spawning a downloading child process
+        std::optional<server_model_meta> custom_meta = std::nullopt;
+    };
+
     // load and unload model instances
     // these functions are thread-safe
     void load(const std::string & name);
+    void load(const std::string & name, const load_options & opts);
     void unload(const std::string & name);
     void unload_all();
 
-    // download a new model, progress is reported via SSE
-    // to stop the download, call unload()
-    void download(common_params_model && model, common_download_opts && opts);
-
     struct update_status_args {
         server_model_status status;
         int exit_code = 0; // only valid if status == UNLOADED
@@ -213,9 +220,12 @@ public:
 struct server_child {
     // serializes the notify_to_router writes
     std::mutex mtx_stdout;
+    std::atomic<bool> is_finished_downloading = false; // set by run_download
 
     // return true if the current process is a child server instance
     bool is_child();
+    server_child_mode get_mode();
+    int run_download(common_params & params);
 
     // register the shutdown_handler to be called by the router
     // return the monitoring thread (to be joined by the caller)

tools/server/server-task.cpp

@@ -591,10 +591,11 @@ json server_task_result_cmpl_final::to_json_oaicompat_resp() {
 
     for (const common_chat_tool_call & tool_call : oaicompat_msg.tool_calls) {
         output.push_back(json {
+            {"id",        "fc_" + tool_call.id},
             {"type",      "function_call"},
             {"status",    "completed"},
             {"arguments", tool_call.arguments},
-            {"call_id",   "fc_" + tool_call.id},
+            {"call_id",   "call_" + tool_call.id},
             {"name",      tool_call.name},
         });
     }
@@ -690,10 +691,11 @@ json server_task_result_cmpl_final::to_json_oaicompat_resp_stream() {
 
     for (const common_chat_tool_call & tool_call : oaicompat_msg.tool_calls) {
         const json output_item = {
+            {"id",        "fc_" + tool_call.id},
             {"type",      "function_call"},
             {"status",    "completed"},
             {"arguments", tool_call.arguments},
-            {"call_id",   "fc_" + tool_call.id},
+            {"call_id",   "call_" + tool_call.id},
             {"name",      tool_call.name}
         };
         server_sent_events.push_back(json {
@@ -1277,8 +1279,9 @@ json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
                 {"data", json {
                     {"type",  "response.output_item.added"},
                     {"item", json {
+                        {"id",        "fc_" + diff.tool_call_delta.id},
                         {"arguments", ""},
-                        {"call_id",   "fc_" + diff.tool_call_delta.id},
+                        {"call_id",   "call_" + diff.tool_call_delta.id},
                         {"name",      diff.tool_call_delta.name},
                         {"type",      "function_call"},
                         {"status",    "in_progress"},

tools/server/server-task.h

@@ -62,9 +62,6 @@ struct task_params {
 
     int32_t n_cache_reuse = 0; // min chunk size to attempt reusing from the cache via KV shifting (0 = disabled)
 
-    // number of prompt tokens before the latest user message
-    int32_t n_before_user = -1;
-
     int64_t t_max_prompt_ms  = -1; // TODO: implement
     int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
 
@@ -92,6 +89,9 @@ struct task_params {
     // per-request parameters for chat parsing
     common_chat_parser_params chat_parser_params;
 
+    // message spans for checkpointing
+    common_chat_msg_spans message_spans;
+
     // Embeddings
     int32_t embd_normalize = 2; // (-1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm)
 

tools/server/server-tools.cpp

@@ -569,9 +569,13 @@ struct server_tool_edit_file : server_tool {
             }
             int n = (int) lines.size();
             if (e.line_start == -1) {
-                // -1 means end of file; line_end is ignored — normalize to point past last line
-                e.line_start = n + 1;
-                e.line_end   = n + 1;
+                // -1 targets end of file -> valid for append only; line_end is ignored
+                if (e.mode != "append") {
+                    return {{"error", "line_start -1 (end of file) is only valid for append mode"}};
+                }
+                // append at end of file: insert position is the current line count
+                e.line_start = n;
+                e.line_end   = n;
             } else {
                 if (e.line_start < 1 || e.line_end < e.line_start) {
                     return {{"error", string_format("invalid line range [%d, %d]", e.line_start, e.line_end)}};
@@ -612,8 +616,8 @@ struct server_tool_edit_file : server_tool {
             } else if (e.mode == "delete") {
                 lines.erase(lines.begin() + idx_start, lines.begin() + idx_end + 1);
             } else { // append
-                // idx_end + 1 may equal lines.size() when line_start == -1 (end of file)
-                lines.insert(lines.begin() + idx_end + 1, new_lines.begin(), new_lines.end());
+                // insert after idx_end; idx_end + 1 == lines.size() for end-of-file append
+                lines.insert(lines.begin() + (idx_end + 1), new_lines.begin(), new_lines.end());
             }
         }
 

tools/server/server.cpp

@@ -35,6 +35,19 @@ static inline void signal_handler(int signal) {
     shutdown_handler(signal);
 }
 
+// satisfies -Wmissing-declarations (used by llama command)
+int llama_server(int argc, char ** argv);
+
+// to be used via CLI (argc / argv are used by router mode only)
+int llama_server(common_params & params, int argc, char ** argv);
+void llama_server_terminate();
+void llama_server_terminate() {
+    if (shutdown_handler) {
+        shutdown_handler(0);
+    }
+}
+
+
 // wrapper function that handles exceptions and logs errors
 // this is to make sure handler_t never throws exceptions; instead, it returns an error response
 static server_http_context::handler_t ex_wrapper(server_http_context::handler_t func) {
@@ -71,9 +84,6 @@ static server_http_context::handler_t ex_wrapper(server_http_context::handler_t
     };
 }
 
-// satisfies -Wmissing-declarations
-int llama_server(int argc, char ** argv);
-
 int llama_server(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
 
@@ -89,6 +99,23 @@ int llama_server(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
+    return llama_server(params, argc, argv);
+}
+
+int llama_server(common_params & params, int argc, char ** argv) {
+    bool is_run_by_cli = (argv == nullptr);
+
+    // note: router mode also accepts -hf remote-preset, so we need to check that first
+    if (!is_run_by_cli && !params.model.hf_repo.empty()) {
+        try {
+            common_params_handle_models_params handle_params;
+            handle_params.preset_only = true;
+            common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, handle_params);
+        } catch (const std::exception & e) {
+            // ignored for now
+        }
+    }
+
     // router server never loads a model and must not touch the GPU
     const bool is_router_server = params.model.path.empty()
                                && params.model.hf_repo.empty();
@@ -134,6 +161,7 @@ int llama_server(int argc, char ** argv) {
     //
 
     // register API routes
+    server_child child; // only used in non-router mode
     server_routes routes(params, ctx_server);
     server_tools tools;
 
@@ -254,11 +282,22 @@ int llama_server(int argc, char ** argv) {
         ctx_http.post("/tools",           ex_wrapper(tools.handle_post));
     }
 
+    //
+    // Handle downloading model
+    //
+
+    if (child.is_child() && child.get_mode() == SERVER_CHILD_MODE_DOWNLOAD) {
+        return child.run_download(params);
+    } else if (!is_router_server && !is_run_by_cli) {
+        // single-model mode (NOT spawned by router)
+        // if this is invoked by CLI, model downloading should be already handled
+        common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, {});
+    }
+
     //
     // Start the server
     //
 
-    server_child child; // only used in non-router mode
     std::function<void()> clean_up;
 
     if (is_router_server) {
@@ -334,20 +373,22 @@ int llama_server(int argc, char ** argv) {
         };
     }
 
-    // TODO: refactor in common/console
+    // register signal handler if not running by CLI
+    if (!is_run_by_cli) {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-    struct sigaction sigint_action;
-    sigint_action.sa_handler = signal_handler;
-    sigemptyset (&sigint_action.sa_mask);
-    sigint_action.sa_flags = 0;
-    sigaction(SIGINT, &sigint_action, NULL);
-    sigaction(SIGTERM, &sigint_action, NULL);
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = signal_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+        sigaction(SIGTERM, &sigint_action, NULL);
 #elif defined (_WIN32)
-    auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-        return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-    };
-    SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
+    }
 
     if (is_router_server) {
         SRV_INF("router server is listening on %s\n", ctx_http.listening_address.c_str());

tools/server/tests/unit/test_router.py

@@ -256,15 +256,45 @@ def test_router_reload_models():
         os.remove(preset_path)
 
 
+def test_router_remote_preset():
+    global server
+    server.model_hf_repo = "ggml-org/test-preset-ci"
+    server.model_hf_file = None
+    server.offline = False
+    server.start()
+
+    # Should see preset models in GET /models
+    res = server.make_request("GET", "/models")
+    assert res.status_code == 200
+    ids = {item["id"] for item in res.body.get("data", [])}
+    assert "tinygemma3-preset" in ids
+    assert "stories260K-test" in ids
+
+    # Should be able to load a preset model
+    model_id = "tinygemma3-preset"
+    _load_model_and_wait(model_id)
+
+
 MODEL_DOWNLOAD_ID = "ggml-org/test-model-router-download:F16"
-MODEL_DOWNLOAD_TIMEOUT = 300
+MODEL_DOWNLOAD_TIMEOUT = 30
 
 
-def _listen_sse(server: ServerProcess, collected: list, stop: threading.Event):
-    """Collect /models/sse events into `collected` until `stop` is set."""
+def _listen_sse(
+    server: ServerProcess, collected: list, stop: threading.Event, ready: threading.Event | None = None
+):
+    """Collect /models/sse events into `collected` until `stop` is set.
+
+    When `ready` is provided, it is set once the streaming response is open,
+    i.e. the server has accepted the connection and registered us as a
+    subscriber. Callers that trigger one-shot events (e.g. download_finished)
+    must wait on `ready` before acting, otherwise the event can be broadcast
+    before this client is subscribed and be lost.
+    """
     url = f"http://{server.server_host}:{server.server_port}/models/sse"
     try:
         with requests.get(url, stream=True, timeout=MODEL_DOWNLOAD_TIMEOUT) as resp:
+            if ready is not None:
+                ready.set()
             for line_bytes in resp.iter_lines():
                 if stop.is_set():
                     break
@@ -294,11 +324,17 @@ def test_router_download_model():
 
     sse_events: list = []
     stop = threading.Event()
+    sse_ready = threading.Event()
     sse_thread = threading.Thread(
-        target=_listen_sse, args=(server, sse_events, stop), daemon=True
+        target=_listen_sse, args=(server, sse_events, stop, sse_ready), daemon=True
     )
     sse_thread.start()
 
+    # wait for the SSE client to be subscribed before triggering the download,
+    # otherwise the one-shot download_finished event can be broadcast before
+    # this client is registered and be lost
+    assert sse_ready.wait(10), "SSE client failed to connect"
+
     # Trigger the download
     res = server.make_request("POST", "/models", data={"model": MODEL_DOWNLOAD_ID})
     assert res.status_code == 200
@@ -328,13 +364,17 @@ def test_router_delete_model():
 
     # Ensure the model exists (download it if needed)
     if MODEL_DOWNLOAD_ID not in _get_model_ids(is_reload=False):
-        res = server.make_request("POST", "/models", data={"model": MODEL_DOWNLOAD_ID})
-        assert res.status_code == 200
         sse_events: list = []
         stop = threading.Event()
+        sse_ready = threading.Event()
         threading.Thread(
-            target=_listen_sse, args=(server, sse_events, stop), daemon=True
+            target=_listen_sse, args=(server, sse_events, stop, sse_ready), daemon=True
         ).start()
+        # subscribe before triggering the download so the one-shot
+        # download_finished event is not lost (see test_router_download_model)
+        assert sse_ready.wait(10), "SSE client failed to connect"
+        res = server.make_request("POST", "/models", data={"model": MODEL_DOWNLOAD_ID})
+        assert res.status_code == 200
         finished = _wait_for_sse_event(
             sse_events, "download_finished", MODEL_DOWNLOAD_ID, MODEL_DOWNLOAD_TIMEOUT
         )

tools/ui/src/app.d.ts

@@ -19,6 +19,10 @@ import type {
 	ApiErrorResponse,
 	ApiLlamaCppServerProps,
 	ApiModelDataEntry,
+	ApiModelLoadStage,
+	ApiModelsSseProgress,
+	ApiModelsSseData,
+	ApiModelsSseEvent,
 	ApiModelListResponse,
 	ApiProcessingState,
 	ApiRouterModelMeta,
@@ -52,6 +56,7 @@ import type {
 	// Model types
 	ModelModalities,
 	ModelOption,
+	ModelLoadProgress,
 	// Settings types
 	SettingsChatServiceOptions,
 	SettingsConfigValue,
@@ -83,6 +88,10 @@ declare global {
 		ApiErrorResponse,
 		ApiLlamaCppServerProps,
 		ApiModelDataEntry,
+		ApiModelLoadStage,
+		ApiModelsSseProgress,
+		ApiModelsSseData,
+		ApiModelsSseEvent,
 		ApiModelListResponse,
 		ApiProcessingState,
 		ApiRouterModelMeta,
@@ -120,6 +129,7 @@ declare global {
 		// Model types
 		ModelModalities,
 		ModelOption,
+		ModelLoadProgress,
 		// Settings types
 		SettingsChatServiceOptions,
 		SettingsConfigValue,

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageAssistant/ChatMessageAssistant.svelte

@@ -10,7 +10,7 @@
 	import { getMessageEditContext } from '$lib/contexts';
 	import { useProcessingState } from '$lib/hooks/use-processing-state.svelte';
 	import { isLoading, isChatStreaming } from '$lib/stores/chat.svelte';
-	import { copyToClipboard, deriveAgenticSections } from '$lib/utils';
+	import { copyToClipboard, deriveAgenticSections, modelLoadProgressText } from '$lib/utils';
 	import { AgenticSectionType } from '$lib/enums';
 	import { REASONING_TAGS } from '$lib/constants/agentic';
 	import { tick } from 'svelte';
@@ -185,6 +185,13 @@
 	let hasNoContent = $derived(!message?.content?.trim());
 	let isActivelyProcessing = $derived(isCurrentlyLoading || isStreaming);
 
+	// during a router auto-load the message has no model yet, so target the selected one
+	let loadTargetModel = $derived(message.model ?? modelsStore.selectedModelName);
+	let modelLoadProgress = $derived(
+		isRouter && loadTargetModel ? modelsStore.getLoadProgress(loadTargetModel) : null
+	);
+	let modelLoadingText = $derived(modelLoadProgressText(modelLoadProgress));
+
 	let showProcessingInfoTop = $derived(
 		message?.role === MessageRole.ASSISTANT &&
 			isActivelyProcessing &&
@@ -220,7 +227,8 @@
 		<div class="mt-6 w-full max-w-[48rem]" in:fade>
 			<div class="processing-container">
 				<span class="processing-text">
-					{processingState.getPromptProgressText() ??
+					{modelLoadingText ??
+						processingState.getPromptProgressText() ??
 						processingState.getProcessingMessage() ??
 						'Processing...'}
 				</span>
@@ -252,7 +260,8 @@
 		<div class="mt-4 w-full max-w-[48rem]" in:fade>
 			<div class="processing-container">
 				<span class="processing-text">
-					{processingState.getPromptProgressText() ??
+					{modelLoadingText ??
+						processingState.getPromptProgressText() ??
 						processingState.getProcessingMessage() ??
 						'Processing...'}
 				</span>

tools/ui/src/lib/components/app/models/ModelsSelectorOption.svelte

@@ -13,6 +13,7 @@
 	import type { ModelOption } from '$lib/types/models';
 	import { ServerModelStatus } from '$lib/enums';
 	import { modelsStore, routerModels } from '$lib/stores/models.svelte';
+	import { modelLoadFraction, modelLoadProgressText } from '$lib/utils';
 
 	interface Props {
 		option: ModelOption;
@@ -50,11 +51,15 @@
 		(serverStatus === ServerModelStatus.LOADED || isSleeping) && !isOperationInProgress
 	);
 	let isLoading = $derived(serverStatus === ServerModelStatus.LOADING || isOperationInProgress);
+
+	let loadProgress = $derived(isLoading ? modelsStore.getLoadProgress(option.model) : null);
+	let loadPercent = $derived(Math.round(modelLoadFraction(loadProgress) * 100));
+	let loadTitle = $derived(modelLoadProgressText(loadProgress));
 </script>
 
 <div
 	class={[
-		'group flex w-full items-center gap-2 rounded-sm p-2 text-left text-sm transition focus:outline-none',
+		'group relative flex w-full items-center gap-2 rounded-sm p-2 text-left text-sm transition focus:outline-none',
 		'cursor-pointer hover:bg-muted focus:bg-muted',
 		(isSelected || isHighlighted) && 'bg-accent text-accent-foreground',
 		!(isSelected || isHighlighted) && 'hover:bg-accent hover:text-accent-foreground',
@@ -62,6 +67,7 @@
 	]}
 	role="option"
 	aria-selected={isSelected || isHighlighted}
+	title={loadTitle}
 	tabindex="0"
 	onclick={() => onSelect(option.id)}
 	onmouseenter={onMouseEnter}
@@ -188,4 +194,15 @@
 			</div>
 		{/if}
 	</div>
+
+	{#if isLoading}
+		<div
+			class="pointer-events-none absolute inset-x-0 bottom-0 h-0.5 overflow-hidden rounded-b-sm bg-muted"
+		>
+			<div
+				class="h-full bg-primary transition-[width] duration-200 ease-out"
+				style="width: {loadPercent}%"
+			></div>
+		</div>
+	{/if}
 </div>

tools/ui/src/lib/constants/api-endpoints.ts

@@ -1,7 +1,8 @@
 export const API_MODELS = {
 	LIST: '/v1/models',
 	LOAD: '/models/load',
-	UNLOAD: '/models/unload'
+	UNLOAD: '/models/unload',
+	SSE: '/models/sse'
 };
 
 // chat completion routes, the control route drives realtime inference (e.g. end reasoning)

tools/ui/src/lib/constants/index.ts

@@ -37,6 +37,8 @@ export * from './mcp-form';
 export * from './mcp-resource';
 export * from './message-export';
 export * from './model-id';
+export * from './model-loading';
+export * from './sse';
 export * from './precision';
 export * from './processing-info';
 export * from './pwa';

tools/ui/src/lib/constants/model-loading.ts

@@ -0,0 +1,14 @@
+/**
+ * Labels shown while a model loads, keyed by the stage reported on /models/sse.
+ */
+export const MODEL_LOAD_STAGE_LABELS: Record<ApiModelLoadStage, string> = {
+	text_model: 'Loading weights',
+	spec_model: 'Loading draft',
+	mmproj_model: 'Loading projector'
+};
+
+/**
+ * Share of the bar reserved for each load phase after text_model.
+ * text_model fills the rest, so a plain model reaches 100% on its own.
+ */
+export const MODEL_LOAD_TAIL_SHARE = 0.1;

tools/ui/src/lib/constants/sse.ts

@@ -0,0 +1,16 @@
+/**
+ * Server-sent events wire format, shared by the chat stream and the
+ * /models/sse status feed (text/event-stream).
+ */
+
+// blank line between two events
+export const SSE_RECORD_SEPARATOR = '\n\n';
+
+// line break inside an event
+export const SSE_LINE_SEPARATOR = '\n';
+
+// data field prefix, the value follows after an optional space
+export const SSE_DATA_PREFIX = 'data:';
+
+// end-of-stream marker on the chat completion stream
+export const SSE_DONE_MARKER = '[DONE]';

tools/ui/src/lib/enums/index.ts

@@ -54,7 +54,7 @@ export {
 
 export { ModelModality } from './model.enums';
 
-export { ServerRole, ServerModelStatus } from './server.enums';
+export { ServerRole, ServerModelStatus, ServerModelsSseEventType } from './server.enums';
 
 export { ParameterSource, SyncableParameterType, SettingsFieldType } from './settings.enums';
 

tools/ui/src/lib/enums/server.enums.ts

@@ -19,3 +19,17 @@ export enum ServerModelStatus {
 	SLEEPING = 'sleeping',
 	FAILED = 'failed'
 }
+
+/**
+ * /models/sse event type enum - discriminates the records broadcast on the
+ * model status feed in ROUTER mode. Matches the event names emitted by
+ * tools/server/server-models.cpp from the C++ server.
+ */
+export enum ServerModelsSseEventType {
+	STATUS_CHANGE = 'status_change',
+	MODEL_STATUS = 'model_status',
+	STATUS_UPDATE = 'status_update',
+	MODELS_RELOAD = 'models_reload',
+	MODEL_REMOVE = 'model_remove',
+	DOWNLOAD_PROGRESS = 'download_progress'
+}

tools/ui/src/lib/services/chat.service.ts

@@ -10,7 +10,10 @@ import {
 	SETTINGS_KEYS,
 	API_CHAT,
 	API_SLOTS,
-	CONTROL_ACTION
+	CONTROL_ACTION,
+	SSE_LINE_SEPARATOR,
+	SSE_DATA_PREFIX,
+	SSE_DONE_MARKER
 } from '$lib/constants';
 import {
 	AttachmentType,
@@ -18,8 +21,7 @@ import {
 	FileTypeAudio,
 	MessageRole,
 	MimeTypeAudio,
-	ReasoningFormat,
-	UrlProtocol
+	ReasoningFormat
 } from '$lib/enums';
 import type {
 	ApiChatMessageContentPart,
@@ -642,15 +644,15 @@ export class ChatService {
 				if (abortSignal?.aborted) break;
 
 				chunk += decoder.decode(value, { stream: true });
-				const lines = chunk.split('\n');
+				const lines = chunk.split(SSE_LINE_SEPARATOR);
 				chunk = lines.pop() || '';
 
 				for (const line of lines) {
 					if (abortSignal?.aborted) break;
 
-					if (line.startsWith(UrlProtocol.DATA)) {
-						const data = line.slice(6);
-						if (data === '[DONE]') {
+					if (line.startsWith(SSE_DATA_PREFIX)) {
+						const data = line.slice(SSE_DATA_PREFIX.length).trim();
+						if (data === SSE_DONE_MARKER) {
 							streamFinished = true;
 
 							continue;

tools/ui/src/lib/stores/models.svelte.ts

@@ -1,6 +1,7 @@
+import { base } from '$app/paths';
 import { SvelteMap, SvelteSet } from 'svelte/reactivity';
 import { toast } from 'svelte-sonner';
-import { ServerModelStatus, ModelModality } from '$lib/enums';
+import { ServerModelStatus, ServerModelsSseEventType, ModelModality } from '$lib/enums';
 import { ModelsService } from '$lib/services/models.service';
 import { PropsService } from '$lib/services/props.service';
 import { serverStore, isRouterMode } from '$lib/stores/server.svelte';
@@ -8,11 +9,15 @@ import {
 	detectThinkingSupport,
 	detectThinkingSupportWithReason
 } from '$lib/utils/chat-template-thinking-detector';
-import { TTLCache } from '$lib/utils';
+import { TTLCache, getAuthHeaders } from '$lib/utils';
 import {
 	MODEL_PROPS_CACHE_TTL_MS,
 	MODEL_PROPS_CACHE_MAX_ENTRIES,
-	FAVORITE_MODELS_LOCALSTORAGE_KEY
+	FAVORITE_MODELS_LOCALSTORAGE_KEY,
+	API_MODELS,
+	SSE_RECORD_SEPARATOR,
+	SSE_LINE_SEPARATOR,
+	SSE_DATA_PREFIX
 } from '$lib/constants';
 
 import { conversationsStore } from '$lib/stores/conversations.svelte';
@@ -55,6 +60,15 @@ class ModelsStore {
 	private modelUsage = $state<Map<string, SvelteSet<string>>>(new Map());
 	private modelLoadingStates = new SvelteMap<string, boolean>();
 
+	// /models/sse feed state, the single source of truth for status and load progress
+	private statusAbort: AbortController | null = null;
+	private statusReaderActive = false;
+	private loadProgress = new SvelteMap<string, ModelLoadProgress>();
+	private statusWaiters = new Map<
+		string,
+		{ target: ServerModelStatus; resolve: () => void; reject: (e: Error) => void }
+	>();
+
 	favoriteModelIds = $state<Set<string>>(this.loadFavoritesFromStorage());
 
 	/**
@@ -531,7 +545,8 @@ class ModelsStore {
 	 * 1. Model from active conversation's last assistant response (if loaded)
 	 * 2. Model from active conversation's last assistant response (if not loaded)
 	 * 3. First loaded model (not from active conversation)
-	 * 4. First available model
+	 * 4. A favorite model
+	 * 5. First available model
 	 */
 	async ensureFirstModelSelected(): Promise<void> {
 		if (this.selectedModelName) return;
@@ -560,6 +575,13 @@ class ModelsStore {
 			return;
 		}
 
+		// Try loading a favorite model
+		const favorite = this.favoriteModelIds.values().next()?.value
+		if (favorite) {
+			await this.selectModelById(favorite);
+			return;
+		}
+
 		// Fall back to the first available model
 		await this.selectModelById(availableModels[0].id);
 	}
@@ -626,49 +648,218 @@ class ModelsStore {
 	 *
 	 */
 
-	/**
-	 * WORKAROUND: Polling for model status after load/unload operations.
-	 *
-	 * Currently, `/models/load` and `/models/unload` return success before
-	 * the operation actually completes on the server.
-	 *
-	 * TODO: Remove polling once llama-server properly waits for the operation
-	 * to complete before returning success.
-	 */
-
-	private static readonly STATUS_POLL_INTERVAL = 500;
+	// reconnect delay after the feed drops or the server is not ready yet
+	private static readonly SSE_RECONNECT_MS = 1000;
 
 	/**
-	 * Poll for expected model status after load/unload operation.
-	 * Keeps polling until the model reaches the expected status or fails.
+	 * Open the /models/sse feed and keep it live with auto reconnect.
+	 * Idempotent and router mode only. The feed drives status and progress,
+	 * so it replaces any post-operation polling.
 	 */
-	private async pollForModelStatus(
-		modelId: string,
-		expectedStatus: ServerModelStatus
-	): Promise<void> {
-		let attempt = 0;
-		while (true) {
-			await this.fetchRouterModels();
+	subscribeStatus(): void {
+		if (this.statusReaderActive) return;
+		if (!isRouterMode()) return;
 
-			const currentStatus = this.getModelStatus(modelId);
-			if (currentStatus === expectedStatus) return;
+		this.statusReaderActive = true;
+		this.statusAbort = new AbortController();
+		void this.runStatusReader(this.statusAbort.signal);
+	}
 
-			if (currentStatus === ServerModelStatus.FAILED) {
-				throw new Error(
-					`Model failed to ${expectedStatus === ServerModelStatus.LOADED ? 'load' : 'unload'}`
-				);
+	/**
+	 * Close the /models/sse feed and drop transient progress.
+	 */
+	unsubscribeStatus(): void {
+		this.statusReaderActive = false;
+		this.statusAbort?.abort();
+		this.statusAbort = null;
+		this.loadProgress.clear();
+	}
+
+	/**
+	 * Current load progress for a model, or null when not loading.
+	 */
+	getLoadProgress(modelId: string): ModelLoadProgress | null {
+		return this.loadProgress.get(modelId) ?? null;
+	}
+
+	/**
+	 * Read the feed and reconnect until unsubscribed. Splits the byte stream
+	 * into SSE records on the blank line boundary.
+	 */
+	private async runStatusReader(signal: AbortSignal): Promise<void> {
+		const decoder = new TextDecoder();
+
+		while (!signal.aborted) {
+			try {
+				const response = await fetch(`${base}${API_MODELS.SSE}`, {
+					headers: getAuthHeaders(),
+					signal
+				});
+
+				if (response.ok && response.body) {
+					const reader = response.body.getReader();
+					let buffer = '';
+
+					while (!signal.aborted) {
+						const { value, done } = await reader.read();
+						if (done) break;
+
+						buffer += decoder.decode(value, { stream: true });
+
+						let boundary = buffer.indexOf(SSE_RECORD_SEPARATOR);
+						while (boundary !== -1) {
+							this.handleStatusRecord(buffer.slice(0, boundary));
+							buffer = buffer.slice(boundary + SSE_RECORD_SEPARATOR.length);
+							boundary = buffer.indexOf(SSE_RECORD_SEPARATOR);
+						}
+					}
+				}
+			} catch {
+				// network drop or abort falls through to the reconnect delay
 			}
 
-			if (
-				expectedStatus === ServerModelStatus.LOADED &&
-				currentStatus === ServerModelStatus.UNLOADED &&
-				attempt > 2
-			) {
-				throw new Error('Model was unloaded unexpectedly during loading');
-			}
+			if (signal.aborted) return;
 
-			attempt++;
-			await new Promise((resolve) => setTimeout(resolve, ModelsStore.STATUS_POLL_INTERVAL));
+			await new Promise((resolve) => setTimeout(resolve, ModelsStore.SSE_RECONNECT_MS));
+		}
+	}
+
+	/**
+	 * Parse one SSE record. The payload rides in the data lines as a JSON
+	 * envelope that carries its own model, event and data fields.
+	 */
+	private handleStatusRecord(record: string): void {
+		const payload = record
+			.split(SSE_LINE_SEPARATOR)
+			.filter((line) => line.startsWith(SSE_DATA_PREFIX))
+			.map((line) => line.slice(SSE_DATA_PREFIX.length).trim())
+			.join(SSE_LINE_SEPARATOR);
+
+		if (payload.length === 0) return;
+
+		let envelope: ApiModelsSseEvent;
+		try {
+			envelope = JSON.parse(payload);
+		} catch {
+			return;
+		}
+
+		this.applyStatusEvent(envelope);
+	}
+
+	/**
+	 * Route one feed record by event kind. Only the status_* events carry a
+	 * status payload, models_reload triggers a list refresh, model_remove drops
+	 * the row, download_* belong to the download surface, not here.
+	 */
+	private applyStatusEvent(event: ApiModelsSseEvent): void {
+		switch (event.event) {
+			case ServerModelsSseEventType.STATUS_CHANGE:
+			case ServerModelsSseEventType.MODEL_STATUS:
+			case ServerModelsSseEventType.STATUS_UPDATE:
+				this.applyModelStatus(event);
+				break;
+			case ServerModelsSseEventType.MODELS_RELOAD:
+				void this.fetchRouterModels();
+				break;
+			case ServerModelsSseEventType.MODEL_REMOVE:
+				this.removeRouterModel(event.model);
+				break;
+			case ServerModelsSseEventType.DOWNLOAD_PROGRESS:
+				break;
+		}
+	}
+
+	/**
+	 * Apply a status envelope: update the model row, track or clear progress,
+	 * settle any pending load or unload awaiter.
+	 */
+	private applyModelStatus(event: ApiModelsSseEvent): void {
+		const model = event.model;
+		const data = event.data;
+		if (!model || !data?.status) return;
+
+		const status = data.status;
+
+		this.setRouterModelStatus(model, status);
+
+		if (status === ServerModelStatus.LOADING) {
+			if (data.progress) this.loadProgress.set(model, data.progress);
+		} else {
+			this.loadProgress.delete(model);
+		}
+
+		if (status === ServerModelStatus.LOADED) {
+			void this.updateModelModalities(model);
+		}
+
+		const failed =
+			status === ServerModelStatus.FAILED ||
+			(status === ServerModelStatus.UNLOADED && (data.exit_code ?? 0) !== 0);
+
+		if (failed) {
+			this.rejectStatus(model, new Error(`Model failed: ${this.toDisplayName(model)}`));
+			return;
+		}
+
+		this.settleStatus(model, status);
+	}
+
+	/**
+	 * Drop a model row reported gone by the feed and settle its awaiters.
+	 */
+	private removeRouterModel(modelId: string): void {
+		if (this.routerModels.findIndex((m) => m.id === modelId) === -1) return;
+
+		this.routerModels = this.routerModels.filter((m) => m.id !== modelId);
+		this.loadProgress.delete(modelId);
+		this.rejectStatus(modelId, new Error(`Model removed: ${this.toDisplayName(modelId)}`));
+	}
+
+	/**
+	 * Update one model row status in place, reassigning to trigger reactivity.
+	 */
+	private setRouterModelStatus(modelId: string, status: ServerModelStatus): void {
+		const idx = this.routerModels.findIndex((m) => m.id === modelId);
+		if (idx === -1) return;
+
+		const current = this.routerModels[idx];
+		if (current.status.value === status) return;
+
+		const next = [...this.routerModels];
+		next[idx] = { ...current, status: { ...current.status, value: status } };
+		this.routerModels = next;
+	}
+
+	/**
+	 * Register an awaiter that resolves when the feed reports target status.
+	 * One operation runs per model at a time, so one awaiter per model is kept.
+	 */
+	private waitForStatus(modelId: string, target: ServerModelStatus): Promise<void> {
+		return new Promise((resolve, reject) => {
+			this.statusWaiters.set(modelId, { target, resolve, reject });
+		});
+	}
+
+	/**
+	 * Resolve and drop the awaiter when the model reaches its target status.
+	 */
+	private settleStatus(modelId: string, status: ServerModelStatus): void {
+		const waiter = this.statusWaiters.get(modelId);
+		if (waiter && waiter.target === status) {
+			this.statusWaiters.delete(modelId);
+			waiter.resolve();
+		}
+	}
+
+	/**
+	 * Reject and drop the awaiter for a model.
+	 */
+	private rejectStatus(modelId: string, error: Error): void {
+		const waiter = this.statusWaiters.get(modelId);
+		if (waiter) {
+			this.statusWaiters.delete(modelId);
+			waiter.reject(error);
 		}
 	}
 
@@ -679,12 +870,18 @@ class ModelsStore {
 		this.modelLoadingStates.set(modelId, true);
 		this.error = null;
 
+		// the feed drives completion, so it must be live before the request
+		this.subscribeStatus();
+
+		const reachedLoaded = this.waitForStatus(modelId, ServerModelStatus.LOADED);
+		reachedLoaded.catch(() => {});
+
 		try {
 			await ModelsService.load(modelId);
-			await this.pollForModelStatus(modelId, ServerModelStatus.LOADED);
-			await this.updateModelModalities(modelId);
+			await reachedLoaded;
 			toast.success(`Model loaded: ${this.toDisplayName(modelId)}`);
 		} catch (error) {
+			this.rejectStatus(modelId, error instanceof Error ? error : new Error('load failed'));
 			this.error = error instanceof Error ? error.message : 'Failed to load model';
 			toast.error(`Failed to load model: ${this.toDisplayName(modelId)}`);
 			throw error;
@@ -700,11 +897,17 @@ class ModelsStore {
 		this.modelLoadingStates.set(modelId, true);
 		this.error = null;
 
+		this.subscribeStatus();
+
+		const reachedUnloaded = this.waitForStatus(modelId, ServerModelStatus.UNLOADED);
+		reachedUnloaded.catch(() => {});
+
 		try {
 			await ModelsService.unload(modelId);
-			await this.pollForModelStatus(modelId, ServerModelStatus.UNLOADED);
+			await reachedUnloaded;
 			toast.info(`Model unloaded: ${this.toDisplayName(modelId)}`);
 		} catch (error) {
+			this.rejectStatus(modelId, error instanceof Error ? error : new Error('unload failed'));
 			this.error = error instanceof Error ? error.message : 'Failed to unload model';
 			toast.error(`Failed to unload model: ${this.toDisplayName(modelId)}`);
 			throw error;
@@ -783,6 +986,9 @@ class ModelsStore {
 	}
 
 	clear(): void {
+		this.unsubscribeStatus();
+		this.statusWaiters.forEach((waiter) => waiter.reject(new Error('Models store cleared')));
+		this.statusWaiters.clear();
 		this.models = [];
 		this.routerModels = [];
 		this.loading = false;

tools/ui/src/lib/types/api.d.ts

@@ -1,4 +1,10 @@
-import type { ContentPartType, FileTypeAudio, ServerModelStatus, ServerRole } from '$lib/enums';
+import type {
+	ContentPartType,
+	FileTypeAudio,
+	ServerModelStatus,
+	ServerModelsSseEventType,
+	ServerRole
+} from '$lib/enums';
 import type { ChatMessagePromptProgress, ChatRole } from './chat';
 
 export type AudioInputFormat = FileTypeAudio.WAV | FileTypeAudio.MP3;
@@ -96,6 +102,46 @@ export interface ApiModelDataEntry {
 	meta?: Record<string, unknown> | null;
 }
 
+/**
+ * Load stage reported by the /models/sse feed, in load order.
+ */
+export type ApiModelLoadStage = 'text_model' | 'spec_model' | 'mmproj_model';
+
+/**
+ * Load progress snapshot: the full ordered stage plan, the active stage,
+ * and its fractional value (0.0 -> 1.0).
+ */
+export interface ApiModelsSseProgress {
+	stages: ApiModelLoadStage[];
+	current: ApiModelLoadStage;
+	value: number;
+}
+
+/**
+ * Status payload carried by a /models/sse envelope.
+ * exit_code appears on unload.
+ */
+export interface ApiModelsSseData {
+	status: ServerModelStatus;
+	progress?: ApiModelsSseProgress;
+	exit_code?: number;
+}
+
+/**
+ * Event kind multiplexed on the /models/sse feed.
+ * Only the status_* events carry a status payload, models_reload signals a
+ * full list refresh, model_remove drops a row, download_* drive download UI.
+ */
+/**
+ * One /models/sse record. event discriminates the kind, model names the
+ * target instance, data carries the status payload when present.
+ */
+export interface ApiModelsSseEvent {
+	model: string;
+	event: ServerModelsSseEventType;
+	data: ApiModelsSseData;
+}
+
 export interface ApiModelDetails {
 	name: string;
 	model: string;

tools/ui/src/lib/types/index.ts

@@ -11,6 +11,10 @@ export type {
 	ApiChatMessageData,
 	ApiModelStatus,
 	ApiModelDataEntry,
+	ApiModelLoadStage,
+	ApiModelsSseProgress,
+	ApiModelsSseData,
+	ApiModelsSseEvent,
 	ApiModelDetails,
 	ApiModelListResponse,
 	ApiLlamaCppServerProps,
@@ -70,7 +74,12 @@ export type {
 } from './database';
 
 // Model types
-export type { ModelModalities, ModelOption, ModalityCapabilities } from './models';
+export type {
+	ModelModalities,
+	ModelOption,
+	ModelLoadProgress,
+	ModalityCapabilities
+} from './models';
 
 // Settings types
 export type {

tools/ui/src/lib/types/models.d.ts

@@ -1,4 +1,4 @@
-import type { ApiModelDataEntry, ApiModelDetails } from '$lib/types/api';
+import type { ApiModelDataEntry, ApiModelDetails, ApiModelLoadStage } from '$lib/types/api';
 
 export interface ModelModalities {
 	vision: boolean;
@@ -20,6 +20,17 @@ export interface ModelOption {
 	tags?: string[];
 }
 
+/**
+ * Ephemeral UI-only load progress for one model instance.
+ * Lives only while a load runs, driven by the /models/sse feed.
+ * stage is absent until the feed reports its first stage.
+ */
+export interface ModelLoadProgress {
+	stages: ApiModelLoadStage[];
+	current: ApiModelLoadStage;
+	value: number;
+}
+
 export interface ParsedModelId {
 	raw: string;
 	orgName: string | null;

tools/ui/src/lib/utils/index.ts

@@ -44,6 +44,9 @@ export { buildProxiedUrl, buildProxiedHeaders } from './cors-proxy';
 // URL utilities
 export { extractRootDomain, sanitizeExternalUrl } from './url';
 
+// Progress helpers
+export { modelLoadFraction, modelLoadProgressText } from './progress';
+
 // Conversation utilities
 export { createMessageCountMap, getMessageCount } from './conversation-utils';
 

tools/ui/src/lib/utils/progress.ts

@@ -0,0 +1,43 @@
+/**
+ * Model load progress helpers for the /models/sse surfaces
+ * (selector row and chat message).
+ */
+
+import { MODEL_LOAD_STAGE_LABELS, MODEL_LOAD_TAIL_SHARE } from '$lib/constants';
+
+/**
+ * Human label for a model load stage.
+ */
+export function modelLoadStageLabel(stage: ApiModelLoadStage): string {
+	return MODEL_LOAD_STAGE_LABELS[stage];
+}
+
+/**
+ * Overall load fraction (0.0 -> 1.0) across the declared stage plan.
+ * text_model fills [0, 1 - tail], each later phase owns one tail slice.
+ */
+export function modelLoadFraction(progress: ModelLoadProgress | null): number {
+	if (!progress) return 0;
+
+	const { stages, current, value } = progress;
+	const tailCount = Math.max(stages.length - 1, 0);
+	const textCeiling = 1 - tailCount * MODEL_LOAD_TAIL_SHARE;
+	const idx = stages.indexOf(current);
+
+	if (idx <= 0) {
+		return value * textCeiling;
+	}
+
+	return textCeiling + (idx - 1 + value) * MODEL_LOAD_TAIL_SHARE;
+}
+
+/**
+ * Single line describing load progress: active stage label and overall percent.
+ * Returns null when there is no progress to show.
+ */
+export function modelLoadProgressText(progress: ModelLoadProgress | null): string | null {
+	if (!progress) return null;
+
+	const label = modelLoadStageLabel(progress.current);
+	return `${label} ${Math.round(modelLoadFraction(progress) * 100)}%`;
+}

tools/ui/src/routes/+layout.svelte

@@ -230,6 +230,20 @@
 		}
 	});
 
+	// Live model status and load progress via the /models/sse feed (router mode)
+	$effect(() => {
+		if (!browser) return;
+		if (!isRouterMode()) return;
+
+		untrack(() => {
+			modelsStore.subscribeStatus();
+		});
+
+		return () => {
+			modelsStore.unsubscribeStatus();
+		};
+	});
+
 	// Background MCP server health checks on app load
 	// Fetch enabled servers from settings and run health checks in background
 	$effect(() => {