server : check draft context creation error (#24922)

* vulkan: make SQR/SQRT/SIN/COS/CLAMP/LEAKY_RELU use unary.comp

* vulkan: make NORM support noncontig

* add noncontiguous row test cases for norm/l2_norm, handle this in the CPU backend and l2_norm.comp

* fix supports_op for cuda and webgpu

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
@@ -584,17 +591,19 @@ 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) {
             throw std::invalid_argument("error: --model is required\n");

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";

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/jinja/runtime.cpp

@@ -686,59 +686,62 @@ value set_statement::execute_impl(context & ctx) {
     return mk_val<value_undefined>();
 }
 
+static inline void bind_parameters(const std::string & name, const statements & this_args, const func_args & args, context & ctx) {
+    const size_t expected_count = this_args.size();
+    const size_t input_count = args.count();
+
+    JJ_DEBUG("Invoking '%s' with %zu input arguments (expected %zu)", name.c_str(), input_count, expected_count);
+    for (size_t i = 0; i < expected_count; ++i) {
+        if (i < input_count) {
+            if (is_stmt<identifier>(this_args[i])) {
+                // normal parameter
+                std::string param_name = cast_stmt<identifier>(this_args[i])->val;
+                value param_value = args.get_kwarg_or_pos(param_name, i);
+                JJ_DEBUG("  Binding parameter '%s' to argument of type %s", param_name.c_str(), param_value->type().c_str());
+                ctx.set_val(param_name, param_value);
+            } else if (is_stmt<keyword_argument_expression>(this_args[i])) {
+                // default argument used as normal parameter
+                auto kwarg = cast_stmt<keyword_argument_expression>(this_args[i]);
+                if (!is_stmt<identifier>(kwarg->key)) {
+                    throw std::runtime_error("Keyword argument key must be an identifier in '" + name + "'");
+                }
+                std::string param_name = cast_stmt<identifier>(kwarg->key)->val;
+                value param_value = args.get_kwarg_or_pos(param_name, i);
+                JJ_DEBUG("  Binding parameter '%s' to argument of type %s", param_name.c_str(), param_value->type().c_str());
+                ctx.set_val(param_name, param_value);
+            } else {
+                throw std::runtime_error("Invalid parameter type in '" + name + "'");
+            }
+        } else {
+            auto & default_arg = this_args[i];
+            if (is_stmt<keyword_argument_expression>(default_arg)) {
+                auto kwarg = cast_stmt<keyword_argument_expression>(default_arg);
+                if (!is_stmt<identifier>(kwarg->key)) {
+                    throw std::runtime_error("Keyword argument key must be an identifier in '" + name + "'");
+                }
+                std::string param_name = cast_stmt<identifier>(kwarg->key)->val;
+                JJ_DEBUG("  Binding parameter '%s' to default argument of type %s", param_name.c_str(), kwarg->val->type().c_str());
+                ctx.set_val(param_name, kwarg->val->execute(args.ctx));
+            } else {
+                throw std::runtime_error("Not enough arguments provided to '" + name + "'");
+            }
+            //std::string param_name = cast_stmt<identifier>(default_args[i])->val;
+            //JJ_DEBUG("  Binding parameter '%s' to default", param_name.c_str());
+            //ctx.var[param_name] = default_args[i]->execute(ctx);
+        }
+    }
+}
+
 value macro_statement::execute_impl(context & ctx) {
     if (!is_stmt<identifier>(this->name)) {
         throw std::runtime_error("Macro name must be an identifier");
     }
     std::string name = cast_stmt<identifier>(this->name)->val;
 
-    const func_handler func = [this, name, &ctx](const func_args & args) -> value {
-        size_t expected_count = this->args.size();
-        size_t input_count = args.count();
+    const func_handler func = [this, name](const func_args & args) -> value {
+        context macro_ctx(args.ctx); // new scope for macro execution
 
-        JJ_DEBUG("Invoking macro '%s' with %zu input arguments (expected %zu)", name.c_str(), input_count, expected_count);
-        context macro_ctx(ctx); // new scope for macro execution
-
-        // bind parameters
-        for (size_t i = 0; i < expected_count; ++i) {
-            if (i < input_count) {
-                if (is_stmt<identifier>(this->args[i])) {
-                    // normal parameter
-                    std::string param_name = cast_stmt<identifier>(this->args[i])->val;
-                    value param_value = args.get_kwarg_or_pos(param_name, i);
-                    JJ_DEBUG("  Binding parameter '%s' to argument of type %s", param_name.c_str(), param_value->type().c_str());
-                    macro_ctx.set_val(param_name, param_value);
-                } else if (is_stmt<keyword_argument_expression>(this->args[i])) {
-                    // default argument used as normal parameter
-                    auto kwarg = cast_stmt<keyword_argument_expression>(this->args[i]);
-                    if (!is_stmt<identifier>(kwarg->key)) {
-                        throw std::runtime_error("Keyword argument key must be an identifier in macro '" + name + "'");
-                    }
-                    std::string param_name = cast_stmt<identifier>(kwarg->key)->val;
-                    value param_value = args.get_kwarg_or_pos(param_name, i);
-                    JJ_DEBUG("  Binding parameter '%s' to argument of type %s", param_name.c_str(), param_value->type().c_str());
-                    macro_ctx.set_val(param_name, param_value);
-                } else {
-                    throw std::runtime_error("Invalid parameter type in macro '" + name + "'");
-                }
-            } else {
-                auto & default_arg = this->args[i];
-                if (is_stmt<keyword_argument_expression>(default_arg)) {
-                    auto kwarg = cast_stmt<keyword_argument_expression>(default_arg);
-                    if (!is_stmt<identifier>(kwarg->key)) {
-                        throw std::runtime_error("Keyword argument key must be an identifier in macro '" + name + "'");
-                    }
-                    std::string param_name = cast_stmt<identifier>(kwarg->key)->val;
-                    JJ_DEBUG("  Binding parameter '%s' to default argument of type %s", param_name.c_str(), kwarg->val->type().c_str());
-                    macro_ctx.set_val(param_name, kwarg->val->execute(ctx));
-                } else {
-                    throw std::runtime_error("Not enough arguments provided to macro '" + name + "'");
-                }
-                //std::string param_name = cast_stmt<identifier>(default_args[i])->val;
-                //JJ_DEBUG("  Binding parameter '%s' to default", param_name.c_str());
-                //macro_ctx.var[param_name] = default_args[i]->execute(ctx);
-            }
-        }
+        bind_parameters(name, this->args, args, macro_ctx);
 
         // execute macro body
         JJ_DEBUG("Executing macro '%s' body with %zu statements", name.c_str(), this->body.size());
@@ -752,6 +755,46 @@ value macro_statement::execute_impl(context & ctx) {
     return mk_val<value_undefined>();
 }
 
+value call_statement::execute_impl(context & ctx) {
+    auto call_expr = cast_stmt<call_expression>(this->call);
+    if (!call_expr) {
+        throw std::runtime_error("Call statement requires a valid call expression");
+    }
+
+    value callee_val = call_expr->callee->execute(ctx);
+    if (!is_val<value_func>(callee_val)) {
+        throw std::runtime_error("Callee is not a function: got " + callee_val->type());
+    }
+    auto * callee_func = cast_val<value_func>(callee_val);
+
+    context caller_ctx(ctx); // new scope for caller execution
+
+    const func_handler func = [this, caller_ctx = std::move(caller_ctx)](const func_args & args) -> value {
+        context block_ctx(caller_ctx); // new scope for block execution
+
+        bind_parameters("caller", this->caller_args, args, block_ctx);
+
+        JJ_DEBUG("Executing call body with %zu statements", this->body.size());
+        auto res = exec_statements(this->body, block_ctx);
+        JJ_DEBUG("Call body execution complete, result: %s", res->val_str.str().c_str());
+        return res;
+    };
+
+    context call_ctx(ctx);
+    call_ctx.set_val("caller", mk_val<value_func>("caller", func));
+
+    func_args args(call_ctx);
+
+    for (const auto & arg_expr : call_expr->args) {
+        auto arg_val = arg_expr->execute(ctx);
+        JJ_DEBUG("  Argument type: %s", arg_val->type().c_str());
+        args.push_back(arg_val);
+    }
+
+    JJ_DEBUG("Calling macro '%s' with %zu arguments", callee_func->name.c_str(), args.count());
+    return callee_func->invoke(args);
+}
+
 value member_expression::execute_impl(context & ctx) {
     value object = this->object->execute(ctx);
 

common/jinja/runtime.h

@@ -552,6 +552,7 @@ struct call_statement : public statement {
         for (const auto & arg : this->caller_args) chk_type<expression>(arg);
     }
     std::string type() const override { return "CallStatement"; }
+    value execute_impl(context & ctx) override;
 };
 
 struct ternary_expression : public expression {

common/peg-parser.cpp

@@ -6,13 +6,14 @@
 #include "unicode.h"
 
 #include <algorithm>
+#include <deque>
 #include <initializer_list>
 #include <map>
 #include <memory>
 #include <nlohmann/json.hpp>
 #include <regex>
+#include <set>
 #include <stdexcept>
-#include <unordered_set>
 
 // Trick to catch missing branches
 template <typename T>
@@ -88,40 +89,7 @@ struct trie {
         return match_result{match_result::NO_MATCH};
     }
 
-    struct prefix_and_next {
-        std::vector<uint32_t> prefix;
-        std::vector<uint32_t> next_chars;
-    };
-
-    std::vector<prefix_and_next> collect_prefix_and_next() {
-        std::vector<uint32_t>        prefix;
-        std::vector<prefix_and_next> result;
-        collect_prefix_and_next(0, prefix, result);
-        return result;
-    }
-
   private:
-    void collect_prefix_and_next(size_t index, std::vector<uint32_t> & prefix, std::vector<prefix_and_next> & out) {
-        if (!nodes[index].is_word) {
-            if (!nodes[index].children.empty()) {
-                std::vector<uint32_t> chars;
-                chars.reserve(nodes[index].children.size());
-                for (const auto & p : nodes[index].children) {
-                    chars.push_back(p.first);
-                }
-                out.emplace_back(prefix_and_next{prefix, chars});
-            }
-        }
-
-        for (const auto & p : nodes[index].children) {
-            uint32_t ch = p.first;
-            auto child = p.second;
-            prefix.push_back(ch);
-            collect_prefix_and_next(child, prefix, out);
-            prefix.pop_back();
-        }
-    }
-
     size_t create_node() {
         size_t index = nodes.size();
         nodes.emplace_back();
@@ -153,6 +121,65 @@ struct trie {
     }
 };
 
+// Aho-Corasick automaton
+struct aho_corasick {
+    trie                t;
+    std::vector<size_t> fail;      // failure links
+    std::vector<size_t> order;     // states in BFS order
+    std::vector<bool>   terminal;  // match states (directly or via a suffix link)
+    std::set<uint32_t>  alphabet;  // every character with a transition
+
+    aho_corasick(const std::vector<std::string> & strings) : t(strings) {
+        const auto & nodes = t.nodes;
+        const size_t n = nodes.size();
+
+        fail.assign(n, 0);
+        order.reserve(n);
+
+        std::deque<size_t> queue{ 0 };
+        while (!queue.empty()) {
+            size_t u = queue.front();
+            queue.pop_front();
+            order.push_back(u);
+            for (const auto & [ch, v] : nodes[u].children) {
+                if (u != 0) {
+                    size_t f = fail[u];
+                    while (f && nodes[f].children.find(ch) == nodes[f].children.end()) {
+                        f = fail[f];
+                    }
+                    auto it = nodes[f].children.find(ch);
+                    fail[v] = (it != nodes[f].children.end() && it->second != v) ? it->second : 0;
+                }
+                queue.push_back(v);
+            }
+        }
+
+        terminal.assign(n, false);
+        for (size_t u : order) {
+            terminal[u] = nodes[u].is_word || (u != 0 && terminal[fail[u]]);
+        }
+
+        for (const auto & node : nodes) {
+            for (const auto & [ch, v] : node.children) {
+                alphabet.insert(ch);
+            }
+        }
+    }
+
+    size_t num_states()          const { return t.nodes.size(); }
+    bool   is_terminal(size_t s) const { return terminal[s]; }
+
+    // follow failure links until a transition on `ch` exists.
+    size_t next(size_t state, uint32_t ch) const {
+        const auto & nodes = t.nodes;
+        while (state && nodes[state].children.find(ch) == nodes[state].children.end()) {
+            state = fail[state];
+        }
+        auto it = nodes[state].children.find(ch);
+        return it != nodes[state].children.end() ? it->second : 0;
+    }
+};
+
 static std::pair<uint32_t, size_t> parse_hex_escape(const std::string & str, size_t pos, int hex_count) {
     if (pos + hex_count > str.length()) {
         return {0, 0};
@@ -894,6 +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 {
@@ -962,7 +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);
@@ -992,12 +1024,12 @@ void common_peg_arena::resolve_refs() {
 }
 
 std::string common_peg_arena::dump(common_peg_parser_id id) const {
-    std::unordered_set<common_peg_parser_id> visited;
+    std::set<common_peg_parser_id> visited;
     return dump_impl(id, visited);
 }
 
 std::string common_peg_arena::dump_impl(common_peg_parser_id                       id,
-                                        std::unordered_set<common_peg_parser_id> & visited) const {
+                                        std::set<common_peg_parser_id> & visited) const {
     // Check for cycles
     if (visited.count(id)) {
         return "[cycle]";
@@ -1043,6 +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>) {
@@ -1452,6 +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);
@@ -1502,61 +1543,118 @@ static std::string gbnf_escape_char_class(uint32_t c) {
     return std::string(buf);
 }
 
-static std::string gbnf_excluding_pattern(const std::vector<std::string> & strings) {
-    trie matcher(strings);
-    auto pieces = matcher.collect_prefix_and_next();
-
-    std::string pattern;
-    std::string trailing;  // optional proper-prefix of a delimiter, allowed only at the very end
-    for (size_t i = 0; i < pieces.size(); ++i) {
-        if (i > 0) {
-            pattern += " | ";
-        }
-
-        const auto & pre = pieces[i].prefix;
-        const auto & chars = pieces[i].next_chars;
-
-        std::string cls;
-        cls.reserve(chars.size());
-        for (uint32_t ch : chars) {
-            cls += gbnf_escape_char_class(ch);
-        }
-
-        if (!pre.empty()) {
-            std::string pre_literal = gbnf_format_literal(common_unicode_cpts_to_utf8(pre));
-            pattern += pre_literal + " [^" + cls + "]";
-            // Each interior alternative consumes a delimiter-prefix plus a disambiguating
-            // char, so the repetition alone cannot match a value that *ends* on a proper
-            // prefix of a delimiter (e.g. a trailing "\n" when the delimiter is
-            // "\n</parameter>\n"). The runtime until() (greedy first-match) accepts such
-            // values, so without this the grammar would reject input the parser accepts.
-            // Allow the value to terminate on any proper prefix as an optional tail.
-            // This makes the grammar a slight superset of the runtime language (a value
-            // may end on the longest prefix, which greedy first-match would not itself
-            // produce); harmless for constrained generation, which only needs to admit
-            // every runtime-valid string.
-            if (!trailing.empty()) {
-                trailing += " | ";
-            }
-            trailing += pre_literal;
-        } else {
-            pattern += "[^" + cls + "]";
-        }
+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);
     }
-
-    std::string result = "(" + pattern + ")*";
-    if (!trailing.empty()) {
-        result += " (" + trailing + ")?";
-    }
-    return result;
+    return s + "]";
 }
 
-static std::unordered_set<std::string> collect_reachable_rules(
+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 {
+        if (s == 0) {
+            return prefix;
+        }
+        std::string num = std::to_string(s);
+        num = num.size() == 1 ? ("0" + num) : num;
+        return prefix + "-" + num;
+    };
+
+    for (size_t q = 0; q < ac.num_states(); q++) {
+        if (ac.is_terminal(q)) {
+            continue; // match states
+        }
+
+        std::map<size_t, std::vector<uint32_t>> buckets;
+        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)) {
+                completing.push_back(c);
+                specific.push_back(c);
+            } else if (d != 0) {
+                buckets[d].push_back(c); // specific non-root destination
+                specific.push_back(c);
+            }
+        }
+
+        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 the empty string so the returned reference stays valid.
+    if (ac.is_terminal(0)) {
+        builder.add_rule(prefix, "|");
+    }
+
+    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
 ) {
-    std::unordered_set<std::string> reachable;
-    std::unordered_set<std::string> visited;
+    std::set<std::string> reachable;
+    std::set<std::string> visited;
 
     std::function<void(common_peg_parser_id)> visit = [&](common_peg_parser_id id) {
         const auto & parser = arena.get(id);
@@ -1588,6 +1686,7 @@ static std::unordered_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>) {
@@ -1765,7 +1864,7 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
                 if (p.delimiters.empty()) {
                     return ".*";
                 }
-                return gbnf_excluding_pattern(p.delimiters);
+                return gbnf_excluding_grammar(builder, "until-" + std::to_string(id), p.delimiters);
             } else if constexpr (std::is_same_v<T, common_peg_schema_parser>) {
                 if (schema_delegates(p)) {
                     return to_gbnf(p.child);
@@ -1782,6 +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>);
             }
@@ -1789,7 +1890,7 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
     };
 
     // Collect reachable rules
-    std::unordered_set<std::string> reachable_rules;
+    std::set<std::string> reachable_rules;
 
     if (lazy) {
         // Collect rules reachable from trigger rules
@@ -1918,6 +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);
 }
@@ -2090,6 +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

@@ -3,8 +3,8 @@
 #include <nlohmann/json_fwd.hpp>
 
 #include <memory>
+#include <set>
 #include <unordered_map>
-#include <unordered_set>
 #include <string>
 #include <string_view>
 #include <functional>
@@ -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 {
@@ -335,7 +341,7 @@ class common_peg_arena {
     friend class common_peg_parser_builder;
 
   private:
-    std::string dump_impl(common_peg_parser_id id, std::unordered_set<common_peg_parser_id> & visited) const;
+    std::string dump_impl(common_peg_parser_id id, std::set<common_peg_parser_id> & visited) const;
 
     common_peg_parser_id add_parser(common_peg_parser_variant parser);
     void add_rule(const std::string & name, common_peg_parser_id id);
@@ -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();

common/speculative.cpp

@@ -905,7 +905,13 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
     int32_t n_embd = 0;
 
-    bool is_mem_shared = false;
+    // One MTP draft driver, three modes (set once in the ctor):
+    //   is_mem_shared (gemma4): shares the target KV, runs all heads in one graph.
+    //   chain_heads (step35): n_mtp_layers trained heads, one per draft step.
+    //   neither (qwen35 / qwen35moe): a single trained MTP head.
+    int32_t n_mtp_layers  = 1;
+    bool    is_mem_shared = false;   // gemma4
+    bool    chain_heads   = false;   // derived in the ctor: n_mtp_layers > 1 && !is_mem_shared
 
     // Per-sequence cross-batch carryover: pair (h_p, x_{p+1}) at MTP pos p+1.
     // The last h-row of one process() call needs the first token of the NEXT
@@ -920,10 +926,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
     std::vector<std::vector<float>> verify_h;
     std::vector<int32_t> verify_h_rows;
 
-    // Per-seq draft length from the last draft() call, used in accept() to
-    // roll back ctx_dft's recurrent state past the AR draft's redundant
-    // pre-advancement before process() mirrored the verify batch.
-    std::vector<uint16_t> last_n_drafted;
+    std::vector<int>                i_last;
+    std::vector<std::vector<float>> chain_h;
 
     common_speculative_impl_draft_mtp(const common_params_speculative & params, uint32_t n_seq)
         : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq)
@@ -936,6 +940,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         n_embd = llama_model_n_embd_out(llama_get_model(ctx_dft));
         GGML_ASSERT(n_embd == llama_model_n_embd(llama_get_model(ctx_tgt)) &&
                 "MTP input row width must match the target h_nextn width");
+        n_mtp_layers = std::max(1, (int) llama_model_n_layer_nextn(llama_get_model(ctx_dft)));
 
         LOG_INF("%s: adding speculative implementation 'draft-mtp'\n", __func__);
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling);
@@ -982,16 +987,25 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
 
         is_mem_shared = llama_get_ctx_other(ctx_dft) == ctx_tgt;
+        chain_heads   = n_mtp_layers > 1 && !is_mem_shared;
+
+        if (chain_heads) {
+            this->params.n_max = std::min(this->params.n_max, n_mtp_layers);
+
+            chain_h.assign(n_seq, {});
+            for (auto & c : chain_h) {
+                c.reserve((size_t) (this->params.n_max + 1) * n_embd);
+            }
+        }
 
         pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
 
+        i_last.assign(n_seq, -1);
         i_batch_beg.assign(n_seq, -1);
         i_batch_end.assign(n_seq, -1);
 
         verify_h.assign(n_seq, {});
         verify_h_rows.assign(n_seq, 0);
-
-        last_n_drafted.assign(n_seq, 0);
     }
 
     ~common_speculative_impl_draft_mtp() override {
@@ -1097,9 +1111,34 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                 set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
             }
 
-            const int32_t rc = llama_decode(ctx_dft, batch);
-            if (rc != 0) {
-                LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (pos=%d)\n", __func__, (int) rc, (int) batch_in.pos[0]);
+            auto * mem_dft = llama_get_memory(ctx_dft);
+
+            bool ok = true;
+            for (int head = 0; head < n_mtp_layers; ++head) {
+                if (chain_heads) {
+                    // ref: https://github.com/ggml-org/llama.cpp/pull/24340/changes#r3413498544
+                    for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                        if (i_batch_beg[seq_id] < 0) {
+                            continue;
+                        }
+                        llama_memory_seq_rm(mem_dft, seq_id, batch_in.pos[i_batch_beg[seq_id]], -1);
+                    }
+                    llama_set_nextn_layer_offset(ctx_dft, head);
+                }
+
+                const int32_t rc = llama_decode(ctx_dft, batch);
+                if (rc != 0) {
+                    LOG_ERR("%s: llama_decode(ctx_dft) head=%d failed rc=%d (pos=%d)\n",
+                            __func__, head, (int) rc, (int) batch_in.pos[0]);
+                    ok = false;
+                    break;
+                }
+            }
+
+            if (chain_heads) {
+                llama_set_nextn_layer_offset(ctx_dft, 0); // restore default for non-draft decodes
+            }
+            if (!ok) {
                 return false;
             }
         }
@@ -1134,7 +1173,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         int n_drafting = 0;
         std::vector<bool> drafting(n_seq);
 
-        const float * h_row = nullptr;
         const size_t row_bytes = (size_t) n_embd * sizeof(float);
 
         for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
@@ -1149,22 +1187,43 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
             common_sampler_reset(smpls[seq_id].get());
 
             common_batch_add(batch, dp.id_last, dp.n_past, { seq_id }, true);
+            std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, pending_h[seq_id].data(), row_bytes);
 
-            h_row = pending_h[seq_id].data();
-            std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
-        }
+            i_last[seq_id] = batch.n_tokens - 1;
 
-        int ret = llama_decode(ctx_dft, batch);
-        if (ret != 0) {
-            LOG_WRN("%s: llama_decode returned %d\n", __func__, ret);
-            return;
+            if (chain_heads) {
+                chain_h[seq_id].assign(pending_h[seq_id].begin(), pending_h[seq_id].end());
+            }
         }
 
         int i = 0;
 
         while (n_drafting > 0) {
-            int i_batch = 0;
+            // each step decodes under a different head, i.e. a different decoder layer, and
+            // KV is per layer. process() filled this layer's KV only for positions < n_past
+            // (prompt + accepted prefix) — nothing in the draft region yet. so reset the
+            // draft region (the seq_rm lower bound is n_past, leaving the prompt KV intact)
+            // and select head i so it rebuilds its own layer's KV there; decoding just the
+            // latest token would leave its attention reading cells only another head wrote.
+            if (chain_heads) {
+                auto * mem_dft = llama_get_memory(ctx_dft);
+                for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                    if (drafting[seq_id]) {
+                        llama_memory_seq_rm(mem_dft, seq_id, dparams[seq_id].n_past, -1);
+                    }
+                }
+                llama_set_nextn_layer_offset(ctx_dft, i);
+            }
 
+            int ret = llama_decode(ctx_dft, batch);
+            if (ret != 0) {
+                LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret);
+                break;
+            }
+
+            // rebuild the batch for the next step: the growing-KV paths re-add only the
+            // new token (the KV already holds the prefix), while chained heads re-add the
+            // whole prefix at the next head. dropped sequences are simply not re-added.
             common_batch_clear(batch);
 
             for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
@@ -1174,9 +1233,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
                 auto * smpl = smpls[seq_id].get();
 
-                common_sampler_sample(smpl, ctx_dft, i_batch, true);
-                h_row = llama_get_embeddings_nextn_ith(ctx_dft, i_batch);
-                ++i_batch;
+                common_sampler_sample(smpl, ctx_dft, i_last[seq_id], true);
+                const float * h_row = llama_get_embeddings_nextn_ith(ctx_dft, i_last[seq_id]);
 
                 const auto * cur_p = common_sampler_get_candidates(smpl, true);
 
@@ -1210,30 +1268,41 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                     continue;
                 }
 
-                if (is_mem_shared) {
+                if (chain_heads) {
+                    // ref: https://github.com/ggml-org/llama.cpp/pull/24340#discussion_r3448031546
+                    chain_h[seq_id].insert(chain_h[seq_id].end(), h_row, h_row + n_embd);
+
+                    const int n_rows = (int) result.size() + 1; // id_last + tokens drafted so far
+                    for (int t = 0; t < n_rows; ++t) {
+                        const llama_token tok = (t == 0) ? dp.id_last : result[t - 1];
+                        common_batch_add(batch, tok, dp.n_past + t, { seq_id }, t == n_rows - 1);
+                        std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd,
+                                    chain_h[seq_id].data() + (size_t) t * n_embd, row_bytes);
+                    }
+                } else if (is_mem_shared) {
                     // note: with shared memory (e.g. Gemma4 assistants) we use the same position for all draft tokens
                     // ref: https://github.com/huggingface/transformers/blob/effde20942e3f82a1b97449f60b3a48c5ff96145/docs/source/en/model_doc/gemma4_assistant.md?plain=1#L36-L37
                     common_batch_add(batch, id, dp.n_past, { seq_id }, true);
+                    std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
                 } else {
                     common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
+                    std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
                 }
-                std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
+
+                i_last[seq_id] = batch.n_tokens - 1;
             }
 
             if (batch.n_tokens == 0) {
                 break;
             }
 
-            // evaluate the drafted tokens on the draft model
-            ret = llama_decode(ctx_dft, batch);
-            if (ret != 0) {
-                LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret);
-                break;
-            }
-
             ++i;
         }
 
+        if (chain_heads) {
+            llama_set_nextn_layer_offset(ctx_dft, 0); // restore default for non-draft decodes
+        }
+
         for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
             auto & dp = dparams[seq_id];
             if (!dp.drafting) {
@@ -1243,8 +1312,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
             if (dp.result->size() < (size_t) params.n_min) {
                 dp.result->clear();
             }
-
-            last_n_drafted[seq_id] = (uint16_t) dp.result->size();
         }
     }
 
@@ -1857,7 +1924,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
 
         bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
         bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr;
-        bool has_mtp = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP)) && params.draft.ctx_dft != nullptr;
+        bool has_draft_mtp    = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP))    && params.draft.ctx_dft != nullptr;
 
 
 
@@ -1895,7 +1962,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
         if (has_draft_eagle3) {
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params));
         }
-        if (has_mtp) {
+        if (has_draft_mtp) {
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, params));
         }
     }

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-cpu/ops.cpp

@@ -3688,8 +3688,6 @@ static void ggml_compute_forward_norm_f32(
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -3703,25 +3701,49 @@ static void ggml_compute_forward_norm_f32(
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
 
-                float sum = 0.0;
-                ggml_vec_sum_f32(ne00, &sum, x);
-                float mean = sum/ne00;
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    const float * xf = (const float *) x;
 
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-                float variance = 0;
+                    float sum = 0.0;
+                    ggml_vec_sum_f32(ne00, &sum, xf);
+                    float mean = sum/ne00;
+
+                    float * yf = (float *) y;
+                    float variance = 0;
 
 #ifdef GGML_USE_ACCELERATE
-                mean = -mean;
-                vDSP_vsadd(x, 1, &mean, y, 1, ne00);
-                vDSP_measqv(y, 1, &variance, ne00);
+                    mean = -mean;
+                    vDSP_vsadd(xf, 1, &mean, yf, 1, ne00);
+                    vDSP_measqv(yf, 1, &variance, ne00);
 #else
-                variance = ggml_vec_cvar_f32(ne00, y, x, mean);
+                    variance = ggml_vec_cvar_f32(ne00, yf, xf, mean);
 #endif //GGML_USE_ACCELERATE
 
-                const float scale = 1.0f/sqrtf(variance + eps);
-                ggml_vec_scale_f32(ne00, y, scale);
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    ggml_vec_scale_f32(ne00, yf, scale);
+                } else {
+                    float sum = 0.0;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        sum += *(const float *) (x + i00*nb00);
+                    }
+                    const float mean = sum/ne00;
+
+                    float variance = 0.0f;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float v = *(const float *) (x + i00*nb00) - mean;
+                        *(float *) (y + i00*nb0) = v;
+                        variance += v * v;
+                    }
+                    variance /= ne00;
+
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        *(float *) (y + i00*nb0) *= scale;
+                    }
+                }
             }
         }
     }
@@ -4142,8 +4164,6 @@ static void ggml_compute_forward_l2_norm_f32(
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -4158,20 +4178,27 @@ static void ggml_compute_forward_l2_norm_f32(
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
 
                 ggml_float sum = 0.0;
                 for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    sum += (ggml_float)(x[i00] * x[i00]);
+                    const float xi = *(const float *) (x + i00*nb00);
+                    sum += (ggml_float)(xi * xi);
                 }
 
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-
-                memcpy(y, x, ne00 * sizeof(float));
-
                 const float scale = 1.0f/fmaxf(sqrtf(sum), eps);
 
-                ggml_vec_scale_f32(ne00, y, scale);
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
+
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    memcpy(y, x, ne00 * sizeof(float));
+                    ggml_vec_scale_f32(ne00, (float *) y, scale);
+                } else {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float xi = *(const float *) (x + i00*nb00);
+                        *(float *) (y + i00*nb0) = xi * scale;
+                    }
+                }
             }
         }
     }

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

@@ -5334,7 +5334,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_L2_NORM:
-            return true;
+            return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_RMS_NORM_BACK:
             return ggml_is_contiguous(op->src[0]);
             break;

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-vulkan/ggml-vulkan.cpp

@@ -493,6 +493,20 @@ struct vk_conv2d_pipeline_state {
     }
 };
 
+struct vk_conv3d_pipeline_state {
+    vk_conv3d_pipeline_state(uint32_t s0, uint32_t s1, uint32_t s2, uint32_t p0, uint32_t p1, uint32_t p2,
+                             uint32_t d0, uint32_t d1, uint32_t d2, uint32_t KW, uint32_t KH, uint32_t KD, uint32_t aligned)
+        : s0(s0), s1(s1), s2(s2), p0(p0), p1(p1), p2(p2), d0(d0), d1(d1), d2(d2), KW(KW), KH(KH), KD(KD), aligned(aligned) {}
+
+    uint32_t s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD;
+    uint32_t aligned;
+
+    bool operator<(const vk_conv3d_pipeline_state &b) const {
+        return std::tie(s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD, aligned) <
+               std::tie(b.s0, b.s1, b.s2, b.p0, b.p1, b.p2, b.d0, b.d1, b.d2, b.KW, b.KH, b.KD, b.aligned);
+    }
+};
+
 struct vk_solve_tri_pipeline_state {
     vk_solve_tri_pipeline_state(uint32_t N, uint32_t K)
         : N(N), K(K) {}
@@ -777,6 +791,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
     vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_get_rows_back_f32;
     vk_pipeline pipeline_acc_f32;
     vk_pipeline pipeline_set_f32;
 
@@ -801,14 +816,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_concat_i8, pipeline_concat_i16, pipeline_concat_i32, pipeline_concat_i64;
     vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32, pipeline_upscale_bilinear_antialias_f32;
     vk_pipeline pipeline_scale_f32;
-    vk_pipeline pipeline_sqr_f32;
-    vk_pipeline pipeline_sqrt_f32;
-    vk_pipeline pipeline_sin_f32;
-    vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_log[2];
     vk_pipeline pipeline_tri[2];
     vk_pipeline pipeline_diag[2];
-    vk_pipeline pipeline_clamp_f32;
+    vk_pipeline pipeline_clamp[2];
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
@@ -840,6 +851,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_gelu_quick[2];
     vk_pipeline pipeline_silu[2];
     vk_pipeline pipeline_relu[2];
+    vk_pipeline pipeline_sqr[2];
+    vk_pipeline pipeline_sqrt[2];
+    vk_pipeline pipeline_sin[2];
+    vk_pipeline pipeline_cos[2];
     vk_pipeline pipeline_xielu[2];
     vk_pipeline pipeline_neg[2];
     vk_pipeline pipeline_tanh[2];
@@ -871,7 +886,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_geglu_erf[2];
     vk_pipeline pipeline_geglu_quick[2];
 
-    vk_pipeline pipeline_leaky_relu_f32;
+    vk_pipeline pipeline_leaky_relu[2];
     vk_pipeline pipeline_silu_back_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
@@ -924,6 +939,8 @@ struct vk_device_struct {
     std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
     std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv_transpose_2d_f32[CONV_SHAPE_COUNT];
     std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv_transpose_2d_f16_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv3d_pipeline_state, vk_pipeline> pipeline_conv3d_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv3d_pipeline_state, vk_pipeline> pipeline_conv3d_f16_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_dw_whcn_f32, pipeline_conv2d_dw_whcn_f16_f32;
     vk_pipeline pipeline_conv2d_dw_cwhn_f32, pipeline_conv2d_dw_cwhn_f16_f32;
 
@@ -1669,6 +1686,41 @@ template <> void init_pushconst_fastdiv(vk_op_conv2d_push_constants &p) {
     init_fastdiv_values(p.OW*p.OH,  p.OWOHmp,  p.OWOHL);
 }
 
+struct vk_op_conv3d_push_constants {
+    uint32_t OC;
+    uint32_t IC;
+    uint32_t N;
+
+    uint32_t IW;
+    uint32_t IH;
+    uint32_t ID;
+    uint32_t OW;
+    uint32_t OH;
+    uint32_t OD;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+
+    uint32_t nb1;
+    uint32_t nb2;
+    uint32_t nb3;
+
+    uint32_t OWmp;     uint32_t OWL;
+    uint32_t OWOHmp;   uint32_t OWOHL;
+    uint32_t OWOHODmp; uint32_t OWOHODL;
+};
+
+template <> void init_pushconst_fastdiv(vk_op_conv3d_push_constants &p) {
+    init_fastdiv_values(p.OW,             p.OWmp,     p.OWL);
+    init_fastdiv_values(p.OW*p.OH,        p.OWOHmp,   p.OWOHL);
+    init_fastdiv_values(p.OW*p.OH*p.OD,   p.OWOHODmp, p.OWOHODL);
+}
+
 struct vk_op_conv2d_dw_push_constants {
     uint32_t ne;
     uint32_t batches;
@@ -4074,19 +4126,35 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     }
 #endif
 
+    auto const &ggml_vk_mul_mm_spec = [](std::vector<uint32_t> spec, bool aligned) {
+        spec.push_back(aligned ? 1u : 0u);
+        return spec;
+    };
+
     const int mul_mat_id_param_count = 5;
 
 #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
     if (device->coopmat2) {
+        auto const &ggml_vk_mul_mm_cm2_spec = [](std::vector<uint32_t> spec, bool aligned, bool mul_mat_id) {
+            if (mul_mat_id && spec.size() > 5) {
+                spec.insert(spec.begin() + 5, aligned ? 1u : 0u);
+            } else {
+                spec.push_back(aligned ? 1u : 0u);
+            }
+            if (mul_mat_id && spec.size() == 6) {
+                spec.push_back(32);
+            }
+            return spec;
+        };
 
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(l_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(m_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(s_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(l_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), l_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(m_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), m_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(s_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), s_align, true);   \
 
         // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
@@ -4161,17 +4229,17 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, true);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, true);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, true);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, true);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, true);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, true);   \
 
         // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
@@ -4284,32 +4352,32 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         // Selects dot2 SPIR-V variant at runtime when device->dot2_f16 is true
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
 
         // bf16 scalar path promotes to f32, no dot2 variant
 #define CREATE_MM_NODOT2(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
 
 #define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l_int[TYPE]) { \
@@ -4474,17 +4542,17 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
 
 #define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
         if (device->mul_mat ## ID ## _l_int[TYPE]) \
@@ -4879,6 +4947,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL],  "get_rows_iq4_nl_f32",  get_rows_iq4_nl_f32_len,  get_rows_iq4_nl_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4],   "get_rows_mxfp4_f32",   get_rows_mxfp4_f32_len,   get_rows_mxfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_NVFP4],   "get_rows_nvfp4_f32",   get_rows_nvfp4_f32_len,   get_rows_nvfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_back_f32, "get_rows_back_f32", get_rows_back_f32_len, get_rows_back_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {256, 1, 1}, {}, 1, true);
 
     ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, sizeof(vk_op_flash_attn_split_k_reduce_push_constants), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
@@ -4903,7 +4972,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     }
     ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_nc_push_constants), {1, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1, true);
@@ -5023,11 +5092,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
 
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_sqrt_f32, "sqrt_f32", sqrt_f32_len, sqrt_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
     ggml_vk_create_pipeline(device, device->pipeline_log[0], "log_f32", log_f32_len, log_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_log[1], "log_f16", log_f16_len, log_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -5037,8 +5101,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_diag[0], "diag_f32", diag_f32_len, diag_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -5058,6 +5120,12 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     CREATE_UNARY(gelu_quick)
     CREATE_UNARY(silu)
     CREATE_UNARY(relu)
+    CREATE_UNARY(sqr)
+    CREATE_UNARY(sqrt)
+    CREATE_UNARY(sin)
+    CREATE_UNARY(cos)
+    CREATE_UNARY(clamp)
+    CREATE_UNARY(leaky_relu)
     CREATE_UNARY(xielu)
     CREATE_UNARY(neg)
     CREATE_UNARY(tanh)
@@ -5097,7 +5165,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     CREATE_GLU(geglu_quick)
 #undef CREATE_GLU
 
-    ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true);
@@ -5314,7 +5381,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
 
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_sgd_f32, "opt_step_sgd_f32", opt_step_sgd_f32_len, opt_step_sgd_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
-    // conv2d, conv_transpose_2d
+    // conv2d, conv_transpose_2d, conv3d
     for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
         // smaller WG for the small-tile fallback gives more concurrent WGs per SM
         uint32_t conv2d_WG_SIZE  = (s == CONV_SHAPE_64x32) ? 128 : 256;
@@ -5377,8 +5444,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
             return (conv2d_BS.K * (conv2d_BS.CRS + pad) + conv2d_BS.CRS * (conv2d_BS.NPQ + pad) + csh_elems) * elem_size;
         };
 
-        // coopmat1 needs to store the output through shared memory, so check up front
-        // whether it'll fit and disable it before applying coopmat1 parameters.
+        // 2D, transpose-2D, and 3D conv use the same KxCRS @ CRSxNPQ shmem
+        // layout. cm1 needs Csh for output, so check before applying cm1 params.
         if (conv2d_use_cm1 && device->properties.limits.maxComputeSharedMemorySize < shmem_req(conv2d_cm1_shmem_pad, true, true)) {
             conv2d_use_cm1 = false;
         }
@@ -5470,6 +5537,53 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         }
 #undef CREATE_CONV
 #undef CREATE_CONVS
+
+        std::vector<uint32_t> conv3d_spec_constants = { conv2d_WG_SIZE, conv2d_BS.K, conv2d_BS.CRS, conv2d_BS.NPQ, conv2d_TS_K, conv2d_SHMEM_PAD };
+#define CREATE_CONV3D(type_suffix, spv_suffix) \
+        for (auto &c : device->pipeline_conv3d##type_suffix[s]) { \
+            const vk_conv3d_pipeline_state &state = c.first; \
+            std::vector<uint32_t> spec_constants_cpy = conv3d_spec_constants; \
+            spec_constants_cpy.push_back(state.s0); \
+            spec_constants_cpy.push_back(state.s1); \
+            spec_constants_cpy.push_back(state.s2); \
+            spec_constants_cpy.push_back(state.p0); \
+            spec_constants_cpy.push_back(state.p1); \
+            spec_constants_cpy.push_back(state.p2); \
+            spec_constants_cpy.push_back(state.d0); \
+            spec_constants_cpy.push_back(state.d1); \
+            spec_constants_cpy.push_back(state.d2); \
+            spec_constants_cpy.push_back(state.KW); \
+            spec_constants_cpy.push_back(state.KH); \
+            spec_constants_cpy.push_back(state.KD); \
+            spec_constants_cpy.push_back(state.aligned); \
+            spec_constants_cpy.push_back(conv2d_csh_store); \
+            spec_constants_cpy.push_back(conv2d_WM); \
+            spec_constants_cpy.push_back(conv2d_WN); \
+            ggml_vk_create_pipeline( \
+                device, c.second, "conv3d" #type_suffix, \
+                conv3d##type_suffix##spv_suffix##_len, conv3d##type_suffix##spv_suffix##_data, "main", 3, \
+                sizeof(vk_op_conv3d_push_constants), wg_denoms, spec_constants_cpy, 1, true, conv2d_required_subgroup_size != 0, conv2d_required_subgroup_size); \
+        }
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        if (device->coopmat2) {
+            CREATE_CONV3D(_f32, _cm2)
+            CREATE_CONV3D(_f16_f32, _cm2)
+        } else
+#endif
+#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+        if (conv2d_use_cm1) {
+            CREATE_CONV3D(_f32, _cm1)
+            CREATE_CONV3D(_f16_f32, _cm1)
+        } else
+#endif
+        if (conv2d_UNROLL) {
+            CREATE_CONV3D(_f32, _unroll)
+            CREATE_CONV3D(_f16_f32, _unroll)
+        } else {
+            CREATE_CONV3D(_f32, )
+            CREATE_CONV3D(_f16_f32, )
+        }
+#undef CREATE_CONV3D
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
@@ -10294,6 +10408,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_get_rows_f32[src0->type];
         }
         return nullptr;
+    case GGML_OP_GET_ROWS_BACK:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_get_rows_back_f32;
+        }
+        return nullptr;
     case GGML_OP_ACC:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_acc_f32;
@@ -10400,23 +10519,27 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_SQR:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sqr_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sqr[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_SQRT:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sqrt_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sqrt[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_SIN:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sin_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sin[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_COS:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_cos_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_cos[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_LOG:
@@ -10438,8 +10561,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_CLAMP:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_clamp_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_clamp[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_PAD:
@@ -10807,8 +10931,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_LEAKY_RELU:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_leaky_relu_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_leaky_relu[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_CONV_2D:
@@ -10885,6 +11010,61 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             }
         }
         return nullptr;
+    case GGML_OP_CONV_3D:
+        if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            const uint32_t OC = (uint32_t)ggml_get_op_params_i32(dst, 11);
+            const uint32_t IC = (uint32_t)ggml_get_op_params_i32(dst, 9);
+            const uint32_t N  = (uint32_t)ggml_get_op_params_i32(dst, 10);
+            const uint32_t NPQ = N * dst->ne[2] * dst->ne[1] * dst->ne[0];
+            const vk_conv_shapes shape = ggml_vk_conv_select_shape(ctx, OC, NPQ);
+
+            const uint32_t KW = (uint32_t)src0->ne[0];
+            const uint32_t KH = (uint32_t)src0->ne[1];
+            const uint32_t KD = (uint32_t)src0->ne[2];
+            const uint32_t s0 = (uint32_t)ggml_get_op_params_i32(dst, 0);
+            const uint32_t s1 = (uint32_t)ggml_get_op_params_i32(dst, 1);
+            const uint32_t s2 = (uint32_t)ggml_get_op_params_i32(dst, 2);
+            const uint32_t p0 = (uint32_t)ggml_get_op_params_i32(dst, 3);
+            const uint32_t p1 = (uint32_t)ggml_get_op_params_i32(dst, 4);
+            const uint32_t p2 = (uint32_t)ggml_get_op_params_i32(dst, 5);
+            const uint32_t d0 = (uint32_t)ggml_get_op_params_i32(dst, 6);
+            const uint32_t d1 = (uint32_t)ggml_get_op_params_i32(dst, 7);
+            const uint32_t d2 = (uint32_t)ggml_get_op_params_i32(dst, 8);
+
+            const uint32_t CRS = IC * KW * KH * KD;
+            const uint32_t BS_K   = vk_conv_block_sizes[shape].K;
+            const uint32_t BS_CRS = vk_conv_block_sizes[shape].CRS;
+            const uint32_t BS_NPQ = vk_conv_block_sizes[shape].NPQ;
+            const uint32_t aligned = ((OC  % BS_K   == 0) &&
+                                      (CRS % BS_CRS == 0) &&
+                                      (NPQ % BS_NPQ == 0)) ? 1u : 0u;
+
+            vk_conv3d_pipeline_state conv3d_pipeline_state(s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD, aligned);
+
+            std::map<vk_conv3d_pipeline_state, vk_pipeline> *pipelines = nullptr;
+            if (src0->type == GGML_TYPE_F32) {
+                pipelines = &ctx->device->pipeline_conv3d_f32[shape];
+            } else if (src0->type == GGML_TYPE_F16) {
+                pipelines = &ctx->device->pipeline_conv3d_f16_f32[shape];
+            } else {
+                return nullptr;
+            }
+
+            vk_pipeline pipeline = nullptr;
+
+            {
+                std::lock_guard<std::mutex> guard(ctx->device->compile_mutex);
+                auto it = pipelines->find(conv3d_pipeline_state);
+                if (it != pipelines->end()) {
+                    pipeline = it->second;
+                } else {
+                    (*pipelines)[conv3d_pipeline_state] = pipeline = std::make_shared<vk_pipeline_struct>();
+                }
+            }
+
+            return pipeline;
+        }
+        return nullptr;
     case GGML_OP_ADD1:
         if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
             return ctx->device->pipeline_add1_f16_f16;
@@ -11135,6 +11315,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
         elements[2] = std::min(elements[2], ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
         break;
+    case GGML_OP_GET_ROWS_BACK:
+        elements = { (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], 1 };
+        elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+        break;
     case GGML_OP_ARGSORT:
         GGML_ASSERT(0);
         break;
@@ -11220,6 +11404,21 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             GGML_ABORT("invalid push constant type for CONV_2D");
         }
         break;
+    case GGML_OP_CONV_3D:
+        if constexpr (std::is_same_v<PC, vk_op_conv3d_push_constants>) {
+            const uint32_t NPQ = pc.N * pc.OD * pc.OH * pc.OW;
+            const vk_conv_shapes shape = ggml_vk_conv_select_shape(ctx, pc.OC, NPQ);
+            const uint32_t NPQ_blocks = CEIL_DIV(NPQ, vk_conv_block_sizes[shape].NPQ);
+
+            elements = { pc.OC, NPQ_blocks, 1 };
+            if (elements[1] > 512) {
+                elements[2] = CEIL_DIV(elements[1], 512);
+                elements[1] = 512;
+            }
+        } else {
+            GGML_ABORT("invalid push constant type for CONV_3D");
+        }
+        break;
     case GGML_OP_ADD:
     case GGML_OP_SUB:
     case GGML_OP_DIV:
@@ -11236,6 +11435,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_TRI:
     case GGML_OP_DIAG:
     case GGML_OP_CLAMP:
+    case GGML_OP_LEAKY_RELU:
     case GGML_OP_PAD:
     case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
@@ -11380,6 +11580,21 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
     });
 }
 
+static void ggml_vk_get_rows_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_GET_ROWS_BACK, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2], (uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    });
+}
+
 static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
@@ -12087,8 +12302,10 @@ static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx,
 
 static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f, 0.0f, 0.0f });
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, std::move(p));
 }
 
 static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -13118,6 +13335,51 @@ static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx,
     ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, dst->op, std::move(p));
 }
 
+static void ggml_vk_conv_3d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
+                            const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_ASSERT(nb00 == sizeof(float) || nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb10 == sizeof(float));
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    vk_op_conv3d_push_constants p{};
+    p.IC = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 9));
+    p.N  = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 10));
+    p.OC = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 11));
+    GGML_ASSERT(src0->ne[3] == (int64_t)p.IC * p.OC);
+    GGML_ASSERT(src1->ne[3] == (int64_t)p.IC * p.N);
+    GGML_ASSERT(dst->ne[3] == (int64_t)p.OC * p.N);
+
+    p.IW = static_cast<uint32_t>(ne10);
+    p.IH = static_cast<uint32_t>(ne11);
+    p.ID = static_cast<uint32_t>(ne12);
+    p.OW = static_cast<uint32_t>(ne0);
+    p.OH = static_cast<uint32_t>(ne1);
+    p.OD = static_cast<uint32_t>(ne2);
+
+    // the shader clamps src addresses to p.IC * p.N * p.IW * p.IH * p.ID - 1 in uint32, so the
+    // total input element count must fit in a uint32.
+    GGML_ASSERT((uint64_t)p.IC * p.N * p.IW * p.IH * p.ID <= 0xFFFFFFFFull);
+
+    p.nb01 = static_cast<uint32_t>(nb01 / nb00);
+    p.nb02 = static_cast<uint32_t>(nb02 / nb00);
+    p.nb03 = static_cast<uint32_t>(nb03 / nb00);
+
+    p.nb11 = static_cast<uint32_t>(nb11 / nb10);
+    p.nb12 = static_cast<uint32_t>(nb12 / nb10);
+    p.nb13 = static_cast<uint32_t>(nb13 / nb10);
+
+    p.nb1 = static_cast<uint32_t>(nb1 / nb0);
+    p.nb2 = static_cast<uint32_t>(nb2 / nb0);
+    p.nb3 = static_cast<uint32_t>(nb3 / nb0);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_3D, std::move(p));
+}
+
 static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     vk_op_conv2d_dw_push_constants p{};
     p.ne = ggml_nelements(dst);
@@ -13144,7 +13406,10 @@ static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx
 
 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const float * op_params = (const float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f, 0.0f, 0.0f });
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, std::move(p));
 }
 
 #ifdef GGML_VULKAN_RUN_TESTS
@@ -14247,6 +14512,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_GET_ROWS:
         ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node);
 
+        break;
+    case GGML_OP_GET_ROWS_BACK:
+        ggml_vk_get_rows_back(ctx, compute_ctx, src0, src1, node);
+
         break;
     case GGML_OP_ADD:
         if (ctx->num_additional_fused_ops) {
@@ -14515,6 +14784,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_CONV_TRANSPOSE_2D:
         ggml_vk_conv_2d(ctx, compute_ctx, src0, src1, node);
 
+        break;
+    case GGML_OP_CONV_3D:
+        ggml_vk_conv_3d(ctx, compute_ctx, src0, src1, node);
+
         break;
     case GGML_OP_CONV_2D_DW:
         ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node);
@@ -16964,6 +17237,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                         return false;
                 }
             }
+        case GGML_OP_GET_ROWS_BACK:
+            return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SET_ROWS:
             {
                 switch (op->type) {
@@ -17060,12 +17335,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_TRANSPOSE:
         case GGML_OP_RMS_NORM:
             return true;
-        case GGML_OP_NORM:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_NORM:
         case GGML_OP_L2_NORM:
-            return ggml_is_contiguous_rows(op->src[0]) &&
-                   op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_ADD:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
@@ -17084,8 +17358,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
-            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_LEAKY_RELU:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                   op->type == op->src[0]->type;
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
@@ -17285,6 +17560,13 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     ggml_is_contiguous(op->src[1]) &&
                     ggml_is_contiguous(op));
             }
+        case GGML_OP_CONV_3D:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->type == GGML_TYPE_F32 &&
+                ggml_is_contiguous(op->src[0]) &&
+                ggml_is_contiguous(op->src[1]) &&
+                ggml_is_contiguous(op);
         default:
             return false;
     }
@@ -18128,6 +18410,20 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             const int32_t d0 = tensor->op_params[4];
             const int32_t d1 = tensor->op_params[5];
             tensor_clone = ggml_conv_2d(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1);
+        } else if (tensor->op == GGML_OP_CONV_3D) {
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t s1 = tensor->op_params[1];
+            const int32_t s2 = tensor->op_params[2];
+            const int32_t p0 = tensor->op_params[3];
+            const int32_t p1 = tensor->op_params[4];
+            const int32_t p2 = tensor->op_params[5];
+            const int32_t d0 = tensor->op_params[6];
+            const int32_t d1 = tensor->op_params[7];
+            const int32_t d2 = tensor->op_params[8];
+            const int32_t IC = tensor->op_params[9];
+            const int32_t N  = tensor->op_params[10];
+            const int32_t OC = tensor->op_params[11];
+            tensor_clone = ggml_conv_3d_direct(ggml_ctx, src_clone[0], src_clone[1], s0, s1, s2, p0, p1, p2, d0, d1, d2, IC, N, OC);
         } else if (tensor->op == GGML_OP_CONV_2D_DW) {
             const int32_t s0 = tensor->op_params[0];
             const int32_t s1 = tensor->op_params[1];

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
-}

ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp

@@ -0,0 +1,431 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#ifdef COOPMAT2
+#extension GL_NV_cooperative_matrix2 : enable
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_KHR_memory_scope_semantics : enable
+#endif
+
+#ifdef COOPMAT
+#extension GL_KHR_cooperative_matrix : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_KHR_memory_scope_semantics : enable
+#endif
+
+#include "types.glsl"
+
+// shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
+layout(binding = 0) readonly buffer A {
+    A_TYPE knl_data[];
+};  // src0 - kernel:   [KW, KH, KD, IC*OC]
+
+layout(binding = 1) readonly buffer B {
+    B_TYPE src_data[];
+};  // src1 - input:    [IW, IH, ID, IC*N] -- channel_first format
+
+layout(binding = 2) writeonly buffer D {
+    D_TYPE dst_data[];
+};  // dst - result:    [OW, OH, OD, OC*N]
+
+layout(push_constant) uniform parameter {
+    // I/O channels, batch size
+    uint32_t OC;
+    uint32_t IC;
+    uint32_t N;
+
+    // Tensor spatial sizes: input, output
+    uint32_t IW;
+    uint32_t IH;
+    uint32_t ID;
+    uint32_t OW;
+    uint32_t OH;
+    uint32_t OD;
+
+    // Strides in elements
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+
+    uint32_t nb1;
+    uint32_t nb2;
+    uint32_t nb3;
+
+    // fastdiv helper values
+    uint32_t OWmp;   uint32_t OWL;
+    uint32_t OWOHmp; uint32_t OWOHL;
+    uint32_t OWOHODmp; uint32_t OWOHODL;
+}
+
+p;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+// Blocktile sizes
+layout(constant_id = 1) const uint BS_K            = 128;
+layout(constant_id = 2) const uint BS_CRS          = 16;
+layout(constant_id = 3) const uint BS_NPQ          = 128;
+// Thread-tile sizes
+layout(constant_id = 4) const uint TS_K            = 8;
+layout(constant_id = 5) const uint SHMEM_PAD       = 4;
+// Stride, padding, dilation
+layout(constant_id = 6)  const uint s0             = 1;
+layout(constant_id = 7)  const uint s1             = 1;
+layout(constant_id = 8)  const uint s2             = 1;
+layout(constant_id = 9)  const uint p0             = 0;
+layout(constant_id = 10) const uint p1             = 0;
+layout(constant_id = 11) const uint p2             = 0;
+layout(constant_id = 12) const uint d0             = 1;
+layout(constant_id = 13) const uint d1             = 1;
+layout(constant_id = 14) const uint d2             = 1;
+// Kernel spatial sizes
+layout(constant_id = 15) const uint KW             = 1;
+layout(constant_id = 16) const uint KH             = 1;
+layout(constant_id = 17) const uint KD             = 1;
+// when set, skip bounds checks and address clamps (K/CRS/NPQ are tile-aligned)
+layout(constant_id = 18) const uint aligned        = 0;
+// stage cm2 result through shmem (Csh) for coalesced stores. cm1 always does this.
+layout(constant_id = 19) const uint csh_store      = 0;
+
+#ifdef COOPMAT
+// cm1 subgroup tile: each subgroup computes a WM x WN region as a grid of
+// TM x TN x TK fragments. Requires WM%TM == WN%TN == BS_K%WM == BS_NPQ%WN ==
+// BS_CRS%TK == 0, and WG_SIZE == (BS_K/WM) * (BS_NPQ/WN) * subgroup_size.
+layout(constant_id = 20) const uint WM             = 32;
+layout(constant_id = 21) const uint WN             = 32;
+const uint TM = 16;
+const uint TN = 16;
+const uint TK = 16;
+const uint cms_per_row = WM / TM;
+const uint cms_per_col = WN / TN;
+const uint warps_M     = BS_K / WM;
+const uint warps_N     = BS_NPQ / WN;
+#endif
+
+// without padding, ID_idx/IH_idx/IW_idx are in bounds by construction
+const bool dhw_in_bounds = (p0 == 0) && (p1 == 0) && (p2 == 0);
+
+uint32_t       tid     = gl_LocalInvocationID.x;
+const uint32_t WG_SIZE = gl_WorkGroupSize.x;
+
+uint splitWork(uint work_size, uint block_size) {
+    return (block_size + work_size - 1) / block_size;
+}
+
+uint32_t K   = p.OC;
+uint32_t CRS = p.IC * KD * KH * KW;
+uint32_t NPQ = p.N * p.OD * p.OH * p.OW;
+
+// Number of blocktiles per input
+uint32_t NB_CRS = splitWork(CRS, BS_CRS);
+
+#if defined(COOPMAT2) || defined(COOPMAT)
+#define SHMEM_TYPE float16_t
+#else
+#define SHMEM_TYPE float
+#endif
+
+const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
+const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
+
+const uint32_t Ash_len = BS_K * Ash_stride;
+const uint32_t Bsh_len = BS_CRS * Bsh_stride;
+
+shared SHMEM_TYPE Ash[Ash_len];  // K x CRS
+shared SHMEM_TYPE Bsh[Bsh_len];  // CRS x NPQ
+
+#if defined(COOPMAT2) || defined(COOPMAT)
+// stage matC through shmem so global stores are row-major (NPQ-contiguous)
+const uint32_t Csh_stride = BS_NPQ;
+#ifdef COOPMAT
+const uint32_t Csh_len    = BS_K * Csh_stride;
+#else
+const uint32_t Csh_len    = csh_store != 0 ? BS_K * Csh_stride : 1;
+#endif
+shared SHMEM_TYPE Csh[Csh_len];  // K x NPQ
+#endif
+
+// Threadtile sizes
+const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
+
+// Number of threadtiles per blocktile
+const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
+
+/*
+Compute
+KxCRS @ CRSxNPQ = K x NPQ
+K=OC
+C=IC
+D,R,S=KD,KH,KW
+Z,P,Q=OD,OH,OW
+*/
+
+uint32_t B_idx_K   = gl_WorkGroupID.x;
+uint32_t B_idx_NPQ = gl_WorkGroupID.y + gl_WorkGroupID.z * 512;
+
+uint32_t T_y = tid / NT_NPQ;
+uint32_t T_x = tid % NT_NPQ;
+
+uint32_t       Ar    = tid / BS_CRS;
+uint32_t       Ac    = tid % BS_CRS;
+const uint32_t ArpWg = WG_SIZE / BS_CRS;
+
+uint32_t       Br    = tid / BS_NPQ;
+uint32_t       Bc    = tid % BS_NPQ;
+const uint32_t BrpWg = WG_SIZE / BS_NPQ;
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
+void split_crs(uint32_t crs_idx, out uint32_t ic, out uint32_t kd, out uint32_t kh, out uint32_t kw) {
+    const uint32_t KHKW = KH * KW;
+    const uint32_t KDKHKW = KD * KHKW;
+    ic = crs_idx / KDKHKW;
+    uint32_t rem = crs_idx - ic * KDKHKW;
+    kd = rem / KHKW;
+    rem = rem - kd * KHKW;
+    kh = rem / KW;
+    kw = rem - kh * KW;
+}
+
+void split_npq(uint32_t npq_idx, out uint32_t n, out uint32_t od, out uint32_t oh, out uint32_t ow) {
+    const uint32_t OWOH = p.OW * p.OH;
+    n = fastdiv(npq_idx, p.OWOHODmp, p.OWOHODL);
+    uint32_t rem = npq_idx - n * p.OD * OWOH;
+    od = fastdiv(rem, p.OWOHmp, p.OWOHL);
+    rem = rem - od * OWOH;
+    oh = fastdiv(rem, p.OWmp, p.OWL);
+    ow = rem - oh * p.OW;
+}
+
+#ifdef COOPMAT2
+#define ACC_TYPE float16_t
+
+ACC_TYPE perElemOpStore(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem)
+{
+    uint32_t K_idx   = B_idx_K * BS_K + r;
+    uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + c;
+    uint32_t N_idx;
+    uint32_t OD_idx;
+    uint32_t OH_idx;
+    uint32_t OW_idx;
+    split_npq(NPQ_idx, N_idx, OD_idx, OH_idx, OW_idx);
+    uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + OD_idx * p.nb2 + (N_idx * p.OC + K_idx) * p.nb3;
+    if (aligned != 0 || (K_idx < K && NPQ_idx < NPQ)) {
+        dst_data[dst_idx] = D_TYPE(elem);
+    }
+    return elem;
+}
+#endif
+
+void main() {
+    if (B_idx_NPQ * BS_NPQ >= NPQ) {
+        return;
+    }
+
+#ifdef COOPMAT2
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator> matC;
+    matC = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator>(0.0);
+#elif defined(COOPMAT)
+    coopmat<float16_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
+    [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
+        sums[i] = coopmat<float16_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0);
+    }
+    const uint warp_r = gl_SubgroupID / warps_N;
+    const uint warp_c = gl_SubgroupID % warps_N;
+#else
+    float regC[TS_K][TS_NPQ];
+    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+            regC[T_ly][T_lx] = 0.0;
+        }
+    }
+#endif
+    /* Advance block in CRS dim */
+    [[dont_unroll]] for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
+        uint32_t CRS_idx_a = B_idx_CRS * BS_CRS + Ac;
+        uint32_t IC_idx_a;
+        uint32_t KD_idx_a;
+        uint32_t KH_idx_a;
+        uint32_t KW_idx_a;
+        split_crs(CRS_idx_a, IC_idx_a, KD_idx_a, KH_idx_a, KW_idx_a);
+
+        /* Load kernel to A_block: (BS_K x BS_CRS)*/
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
+            uint32_t B_ly    = r_offset + Ar;
+            uint32_t B_lx    = Ac;
+            uint32_t K_idx   = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
+            uint32_t knl_idx = KW_idx_a + KH_idx_a * p.nb01 + KD_idx_a * p.nb02 + (K_idx * p.IC + IC_idx_a) * p.nb03;
+            if (aligned == 0) {
+                knl_idx = min(knl_idx, K * CRS - 1);
+            }
+            float    val     = knl_data[knl_idx];
+            if (aligned == 0 && (K_idx >= K || CRS_idx_a >= CRS)) {
+                val = 0.0;
+            }
+            Ash[B_ly * Ash_stride + B_lx] = SHMEM_TYPE(val);
+        }
+        /* Load input to B_block: (BS_CRS x BS_NPQ) */
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
+            uint32_t B_ly          = r_offset + Br;             /* Row index of B block */
+            uint32_t B_lx          = Bc;
+            uint32_t NPQ_idx       = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
+            uint32_t N_idx;
+            uint32_t OD_idx;
+            uint32_t OH_idx;
+            uint32_t OW_idx;
+            split_npq(NPQ_idx, N_idx, OD_idx, OH_idx, OW_idx);
+
+            uint32_t CRS_idx_b = B_idx_CRS * BS_CRS + B_ly;
+            uint32_t IC_idx_b;
+            uint32_t KD_idx_b;
+            uint32_t KH_idx_b;
+            uint32_t KW_idx_b;
+            split_crs(CRS_idx_b, IC_idx_b, KD_idx_b, KH_idx_b, KW_idx_b);
+
+            uint32_t ID_idx = OD_idx * s2 + KD_idx_b * d2 - p2;
+            uint32_t IH_idx = OH_idx * s1 + KH_idx_b * d1 - p1;
+            uint32_t IW_idx = OW_idx * s0 + KW_idx_b * d0 - p0;
+
+            uint32_t src_idx = IW_idx + IH_idx * p.nb11 + ID_idx * p.nb12 + (N_idx * p.IC + IC_idx_b) * p.nb13;
+            // skip clamp when address can't go OOB
+            if (aligned == 0 || !dhw_in_bounds) {
+                src_idx = min(src_idx, p.IC * p.N * p.IW * p.IH * p.ID - 1);
+            }
+            float val = src_data[src_idx];
+            bool oob = false;
+            if (aligned == 0 && (CRS_idx_b >= CRS || NPQ_idx >= NPQ)) {
+                oob = true;
+            }
+            // also catches lower-bound underflow (idx wraps to 0x80000000+)
+            if (!dhw_in_bounds && (ID_idx >= p.ID || IH_idx >= p.IH || IW_idx >= p.IW)) {
+                oob = true;
+            }
+            if (oob) {
+                val = 0.0;
+            }
+            Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);
+        }
+        barrier();
+#ifdef COOPMAT2
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_K, BS_CRS, gl_MatrixUseA> matA;
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_CRS, BS_NPQ, gl_MatrixUseB> matB;
+
+        coopMatLoad(matA, Ash, 0, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
+        coopMatLoad(matB, Bsh, 0, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+        matC = coopMatMulAdd(matA, matB, matC);
+#elif defined(COOPMAT)
+        // each subgroup multiplies its grid of fragments per TK-sized CRS chunk
+        [[unroll]] for (uint k_step = 0; k_step < BS_CRS / TK; k_step++) {
+            coopmat<float16_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a[cms_per_row];
+            [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+                const uint a_off = (warp_r * WM + cm_row * TM) * Ash_stride + k_step * TK;
+                coopMatLoad(cache_a[cm_row], Ash, a_off, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
+            }
+            [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+                coopmat<float16_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
+                const uint b_off = k_step * TK * Bsh_stride + warp_c * WN + cm_col * TN;
+                coopMatLoad(cache_b, Bsh, b_off, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+                    sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a[cm_row], cache_b, sums[cm_col * cms_per_row + cm_row]);
+                }
+            }
+        }
+#else
+        if (T_y * TS_K < K) {
+            UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+                float regA[TS_K];
+                float regB[TS_NPQ];
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
+                }
+                for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                    regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
+                }
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                        regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                    }
+                }
+            }
+        }
+#endif
+        barrier();
+    }
+    /* Save C* */
+#if defined(COOPMAT2) || defined(COOPMAT)
+    // stage matC into Csh, then write to dst with coalesced NPQ-contiguous stores
+#ifdef COOPMAT
+    const bool use_staged_store = true;
+#else
+    const bool use_staged_store = (csh_store != 0);
+#endif
+    if (use_staged_store) {
+#ifdef COOPMAT
+        // cm1: each subgroup stores its fragment grid into its Csh slot
+        [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+            [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+                const uint csh_off = (warp_r * WM + cm_row * TM) * Csh_stride + warp_c * WN + cm_col * TN;
+                coopMatStore(sums[cm_col * cms_per_row + cm_row], Csh, csh_off, Csh_stride, gl_CooperativeMatrixLayoutRowMajor);
+            }
+        }
+#else
+        coopMatStore(matC, Csh, 0, Csh_stride, gl_CooperativeMatrixLayoutRowMajor);
+#endif
+        barrier();
+
+        // cooperative shmem->global: WG threads spread across BS_NPQ (the
+        // contiguous direction of dst), each iter covers store_rows_per_iter K-rows
+        const uint32_t store_rows_per_iter = WG_SIZE / BS_NPQ;
+        const uint32_t store_iters         = BS_K / store_rows_per_iter;
+        const uint32_t k_thread_offset     = tid / BS_NPQ;
+        const uint32_t npq_thread          = tid % BS_NPQ;
+        [[unroll]] for (uint32_t i = 0; i < store_iters; i++) {
+            uint32_t k_local = i * store_rows_per_iter + k_thread_offset;
+            uint32_t K_idx   = B_idx_K * BS_K + k_local;
+            uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + npq_thread;
+            uint32_t N_idx;
+            uint32_t OD_idx;
+            uint32_t OH_idx;
+            uint32_t OW_idx;
+            split_npq(NPQ_idx, N_idx, OD_idx, OH_idx, OW_idx);
+            uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + OD_idx * p.nb2 + (N_idx * p.OC + K_idx) * p.nb3;
+            if (aligned != 0 || (K_idx < K && NPQ_idx < NPQ)) {
+                dst_data[dst_idx] = D_TYPE(Csh[k_local * Csh_stride + npq_thread]);
+            }
+        }
+    }
+#ifdef COOPMAT2
+    else {
+        coopMatPerElementNV(matC, matC, perElemOpStore);
+    }
+#endif
+#else
+    if (T_y * TS_K < K) {
+        for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
+                uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
+                uint32_t N_idx;
+                uint32_t OD_idx;
+                uint32_t OH_idx;
+                uint32_t OW_idx;
+                split_npq(NPQ_idx, N_idx, OD_idx, OH_idx, OW_idx);
+                uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + OD_idx * p.nb2 + (N_idx * p.OC + K_idx) * p.nb3;
+                if (aligned != 0 || (K_idx < K && NPQ_idx < NPQ)) {
+                    dst_data[dst_idx] = D_TYPE(regC[T_ly][T_lx]);
+                }
+            }
+        }
+    }
+#endif
+}

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(cos(val));
-}

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

@@ -0,0 +1,25 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_binary_head.glsl"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint col = gl_GlobalInvocationID.x;
+
+    if (col >= p.ne20) {
+        return;
+    }
+
+    for (uint row = gl_GlobalInvocationID.y; row < p.ne21; row += gl_WorkGroupSize.y * gl_NumWorkGroups.y) {
+        float sum = 0.0f;
+        for (uint i = 0; i < p.ne10; ++i) {
+            if (data_b[get_boffset() + i*p.nb10] == int(row)) {
+                sum += data_a[get_aoffset() + i*p.nb01 + col*p.nb00];
+            }
+        }
+
+        data_d[get_doffset() + row*p.nb21 + col*p.nb20] = sum;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp

@@ -14,16 +14,13 @@ void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
-    const uint i3 = row / (p.ne11 * p.ne12);
-    const uint i3_offset = i3 * p.ne12 * p.ne11;
-    const uint i2 = (row - i3_offset) / p.ne11;
-    const uint i2_offset = i2 * p.ne11;
-    const uint i1 = row - i3_offset - i2_offset;
+    const uint a_base = get_aoffset() + src0_idx(row * p.ne00);
+    const uint d_base = get_doffset() + dst_idx(row * p.ne10);
 
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
     [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]);
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_base + i0*p.nb00]);
         sum[tid] += xi * xi;
     }
 
@@ -39,6 +36,6 @@ void main() {
     const FLOAT_TYPE scale = 1.0f / max(sqrt(sum[0]), FLOAT_TYPE(p.param1));
 
     [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
-        data_d[i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0] = D_TYPE(scale * FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]));
+        data_d[d_base + i0*p.nb10] = D_TYPE(scale * FLOAT_TYPE(data_a[a_base + i0*p.nb00]));
     }
 }

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

@@ -1,22 +0,0 @@
-#version 450
-
-#include "generic_head.glsl"
-#include "types.glsl"
-
-#extension GL_EXT_control_flow_attributes : enable
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
-
-void main() {
-    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
-
-    if (i >= p.KX) {
-        return;
-    }
-
-    const float val = float(data_a[i]);
-    data_d[i] = D_TYPE(max(val, 0.0f) + min(val, 0.0f) * p.param1);
-}

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

@@ -38,17 +38,7 @@
 #define LOAD_VEC_B 1
 #endif
 
-// Load 2 values at once without affecting index calculations through LOAD_VEC
-#if (defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)) && !defined(ALIGNED)
-#define LOAD_VEC_BATCH_A 2
-#else
-#define LOAD_VEC_BATCH_A 1
-#endif
-#if !defined(ALIGNED)
-#define LOAD_VEC_BATCH_B 2
-#else
-#define LOAD_VEC_BATCH_B 1
-#endif
+layout (constant_id = 11) const uint ALIGNED = 0;
 
 #if !defined(TO_FLOAT_TYPE)
 #define TO_FLOAT_TYPE FLOAT_TYPE
@@ -57,6 +47,13 @@
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(DATA_A_F32)
+layout (binding = 0) readonly buffer A_SCALAR {float data_a_scalar[];};
+#elif defined(DATA_A_F16)
+layout (binding = 0) readonly buffer A_SCALAR {float16_t data_a_scalar[];};
+#elif defined(DATA_A_BF16)
+layout (binding = 0) readonly buffer A_SCALAR {uint16_t data_a_scalar[];};
+#endif
 #if defined(A_TYPE_PACKED16)
 layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
 #endif
@@ -65,6 +62,7 @@ layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32
 #endif
 
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 1) readonly buffer B_SCALAR {B_TYPE_SCALAR data_b_scalar[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 #ifdef MUL_MAT_ID
@@ -194,13 +192,23 @@ void main() {
     const uint warp_r = warp_i % (BM / WM);
     const uint warp_c = warp_i / (BM / WM);
 
-    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
-    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
-    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
-    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
+#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
+    const uint LOAD_VEC_A_EFF = (ALIGNED != 0) ? LOAD_VEC_A : 1;
+    const uint LOAD_VEC_BATCH_A = (ALIGNED != 0) ? 1 : 2;
+#else
+    const uint LOAD_VEC_A_EFF = LOAD_VEC_A;
+    const uint LOAD_VEC_BATCH_A = 1;
+#endif
+    const uint LOAD_VEC_B_EFF = (ALIGNED != 0) ? LOAD_VEC_B : 1;
+    const uint LOAD_VEC_BATCH_B = (ALIGNED != 0) ? 1 : 2;
 
-    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A * LOAD_VEC_BATCH_A / BK;
-    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B * LOAD_VEC_BATCH_B / BK;
+    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A_EFF / LOAD_VEC_BATCH_A);
+    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A_EFF / LOAD_VEC_BATCH_A);
+    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B_EFF / LOAD_VEC_BATCH_B);
+    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B_EFF / LOAD_VEC_BATCH_B);
+
+    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A_EFF * LOAD_VEC_BATCH_A / BK;
+    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B_EFF * LOAD_VEC_BATCH_B / BK;
 
 #ifdef MUL_MAT_ID
 #ifdef MUL_MAT_ID_USE_SUBGROUPS
@@ -239,15 +247,15 @@ void main() {
 
     uint pos_a =
 #ifdef MUL_MAT_ID
-        expert_idx * (p.batch_stride_a / LOAD_VEC_A) +
+        expert_idx * (p.batch_stride_a / LOAD_VEC_A_EFF) +
 #else
-        batch_idx_a * (p.batch_stride_a / LOAD_VEC_A) +
+        batch_idx_a * (p.batch_stride_a / LOAD_VEC_A_EFF) +
 #endif
-        (ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
+        (ir * BM * p.stride_a + start_k) / LOAD_VEC_A_EFF;
 #ifdef MUL_MAT_ID
     uint pos_b = 0;
 #else
-    uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B;
+    uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B_EFF;
 #endif
 
 #ifdef COOPMAT
@@ -287,8 +295,8 @@ void main() {
 
         barrier();
 
-        pos_a += BK / LOAD_VEC_A;
-        pos_b += BK / LOAD_VEC_B;
+        pos_a += BK / LOAD_VEC_A_EFF;
+        pos_b += BK / LOAD_VEC_B_EFF;
 
 #ifdef COOPMAT
         [[unroll]] for (uint i = 0; i < BK; i += TK) {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -36,6 +36,7 @@ layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working wit
 layout (constant_id = 4) const bool enable_smaller_matrices = false;
 const uint BNover2 = enable_smaller_matrices ? (BN / 2) : BN;
 const uint BNover4 = enable_smaller_matrices ? (BN / 4) : BN;
+layout (constant_id = 5) const uint ALIGNED = 0;
 
 layout (push_constant) uniform parameter
 {
@@ -111,7 +112,7 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB {
 };
 
 uint _ne1;
-layout (constant_id = 5) const uint subgroup_size = 32;
+layout (constant_id = 6) const uint subgroup_size = 32;
 shared uvec4 ballots_sh[BLOCK_SIZE / subgroup_size];
 
 B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const in uint coordInBlock[2])
@@ -297,12 +298,12 @@ void main() {
 
     // Hint to the compiler that values are aligned (want 16B alignment).
     // Quants are always block-aligned, no alignment needed.
-#if ALIGNED
+    if (ALIGNED != 0) {
 #if QUANT_K == 1
-    stride_a &= ~7;
-#endif
-    stride_b &= ~7;
+        stride_a &= ~7;
 #endif
+        stride_b &= ~7;
+    }
 
     // Create layouts for both clamped and unclamped accesses
     tensorLayoutNV<2> tensorLayoutA = createTensorLayoutNV(2);

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl

@@ -1,50 +1,57 @@
 void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uint idx_m, const uint block, const uint end_k) {
 #if defined(DATA_A_F32) || defined(DATA_A_F16)
 #if LOAD_VEC_A == 8
-            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
-            FLOAT_TYPEV8 aa = FLOAT_TYPEV8(data_a[idx]);
-            buf_a[buf_idx    ] = aa[0].xy;
-            buf_a[buf_idx + 1] = aa[0].zw;
-            buf_a[buf_idx + 2] = aa[1].xy;
-            buf_a[buf_idx + 3] = aa[1].zw;
+            if (ALIGNED != 0) {
+                const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
+                FLOAT_TYPEV8 aa = FLOAT_TYPEV8(data_a[idx]);
+                buf_a[buf_idx    ] = aa[0].xy;
+                buf_a[buf_idx + 1] = aa[0].zw;
+                buf_a[buf_idx + 2] = aa[1].xy;
+                buf_a[buf_idx + 3] = aa[1].zw;
+                return;
+            }
 #elif LOAD_VEC_A == 4
-            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
-            FLOAT_TYPEV4 aa = FLOAT_TYPEV4(data_a[idx]);
-            buf_a[buf_idx    ] = aa.xy;
-            buf_a[buf_idx + 1] = aa.zw;
-#else // LOAD_VEC_BATCH_A == 2
+            if (ALIGNED != 0) {
+                const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
+                FLOAT_TYPEV4 aa = FLOAT_TYPEV4(data_a[idx]);
+                buf_a[buf_idx    ] = aa.xy;
+                buf_a[buf_idx + 1] = aa.zw;
+                return;
+            }
+#endif
             const uint idx = pos_a + col * p.stride_a + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_m < p.M && block + row * 2 + 1 < end_k) {
-                buf_a[buf_idx] = FLOAT_TYPEV2(data_a[idx],
-                                              data_a[idx + 1]);
+                buf_a[buf_idx] = FLOAT_TYPEV2(data_a_scalar[idx],
+                                              data_a_scalar[idx + 1]);
             } else if (idx_m < p.M && block + row * 2 < end_k) {
-                buf_a[buf_idx] = FLOAT_TYPEV2(data_a[idx], 0.0f);
+                buf_a[buf_idx] = FLOAT_TYPEV2(data_a_scalar[idx], 0.0f);
             } else {
                 buf_a[buf_idx] = FLOAT_TYPEV2(0.0f);
             }
-#endif
 #elif defined(DATA_A_BF16)
 #if LOAD_VEC_A == 4
-            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
-            FLOAT_TYPEV4 aa = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_a[idx]));
-            buf_a[buf_idx    ] = aa.xy;
-            buf_a[buf_idx + 1] = aa.zw;
-#else // LOAD_VEC_BATCH_A == 2
+            if (ALIGNED != 0) {
+                const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
+                FLOAT_TYPEV4 aa = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_a[idx]));
+                buf_a[buf_idx    ] = aa.xy;
+                buf_a[buf_idx + 1] = aa.zw;
+                return;
+            }
+#endif
             const uint idx = pos_a + col * p.stride_a + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_m < p.M && block + row * 2 + 1 < end_k) {
-                buf_a[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_a[idx]),
-                                              TO_FLOAT_TYPE(data_a[idx + 1]));
+                buf_a[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_a_scalar[idx]),
+                                              TO_FLOAT_TYPE(data_a_scalar[idx + 1]));
             } else if (idx_m < p.M && block + row * 2 < end_k) {
-                buf_a[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_a[idx]), 0.0f);
+                buf_a[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_a_scalar[idx]), 0.0f);
             } else {
                 buf_a[buf_idx] = FLOAT_TYPEV2(0.0f);
             }
-#endif
 #elif defined(DATA_A_Q4_0)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 4;
@@ -526,75 +533,85 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
 #if !defined(MUL_MAT_ID)
 void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uint idx_n, const uint block, const uint end_k) {
 #if LOAD_VEC_B == 8
-            // Not supported for b_type bf16 because bf16mat2x4 does not exist
-            const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
-            FLOAT_TYPEV8 bb = FLOAT_TYPEV8(data_b[idx]);
-            buf_b[buf_idx + 0] = bb[0].xy;
-            buf_b[buf_idx + 1] = bb[0].zw;
-            buf_b[buf_idx + 2] = bb[1].xy;
-            buf_b[buf_idx + 3] = bb[1].zw;
+            if (ALIGNED != 0) {
+                // Not supported for b_type bf16 because bf16mat2x4 does not exist
+                const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
+                FLOAT_TYPEV8 bb = FLOAT_TYPEV8(data_b[idx]);
+                buf_b[buf_idx + 0] = bb[0].xy;
+                buf_b[buf_idx + 1] = bb[0].zw;
+                buf_b[buf_idx + 2] = bb[1].xy;
+                buf_b[buf_idx + 3] = bb[1].zw;
+                return;
+            }
 #elif LOAD_VEC_B == 4
-            const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
+            if (ALIGNED != 0) {
+                const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
 #if defined(DATA_B_BF16)
-            FLOAT_TYPEV4 bb = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_b[idx]));
+                FLOAT_TYPEV4 bb = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_b[idx]));
 #else
-            FLOAT_TYPEV4 bb = FLOAT_TYPEV4(data_b[idx]);
+                FLOAT_TYPEV4 bb = FLOAT_TYPEV4(data_b[idx]);
+#endif
+                buf_b[buf_idx + 0] = bb.xy;
+                buf_b[buf_idx + 1] = bb.zw;
+                return;
+            }
 #endif
-            buf_b[buf_idx + 0] = bb.xy;
-            buf_b[buf_idx + 1] = bb.zw;
-#else // LOAD_VEC_BATCH_B == 2
             const uint idx = pos_b + col * p.stride_b + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_n < p.N && block + row * 2 + 1 < end_k) {
-                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b[idx]),
-                                              TO_FLOAT_TYPE(data_b[idx + 1]));
+                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b_scalar[idx]),
+                                              TO_FLOAT_TYPE(data_b_scalar[idx + 1]));
             } else if (idx_n < p.N && block + row * 2 < end_k) {
-                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b[idx]), 0.0f);
+                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b_scalar[idx]), 0.0f);
             } else {
                 buf_b[buf_idx] = FLOAT_TYPEV2(0.0f);
             }
-#endif
 }
 #else
 void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uint ic, const uint _ne1, const uint block, const uint end_k) {
 #if LOAD_VEC_B == 8
-            // Not supported for b_type bf16 because bf16mat2x4 does not exist
-            const u16vec2 row_idx = row_ids[col];
-            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
-            FLOAT_TYPEV8 bb = FLOAT_TYPEV8(data_b[idx]);
-            buf_b[buf_idx + 0] = bb[0].xy;
-            buf_b[buf_idx + 1] = bb[0].zw;
-            buf_b[buf_idx + 2] = bb[1].xy;
-            buf_b[buf_idx + 3] = bb[1].zw;
+            if (ALIGNED != 0) {
+                // Not supported for b_type bf16 because bf16mat2x4 does not exist
+                const u16vec2 row_idx = row_ids[col];
+                const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
+                FLOAT_TYPEV8 bb = FLOAT_TYPEV8(data_b[idx]);
+                buf_b[buf_idx + 0] = bb[0].xy;
+                buf_b[buf_idx + 1] = bb[0].zw;
+                buf_b[buf_idx + 2] = bb[1].xy;
+                buf_b[buf_idx + 3] = bb[1].zw;
+                return;
+            }
 #elif LOAD_VEC_B == 4
-            const u16vec2 row_idx = row_ids[col];
-            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
+            if (ALIGNED != 0) {
+                const u16vec2 row_idx = row_ids[col];
+                const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
 #if defined(DATA_B_BF16)
-            FLOAT_TYPEV4 bb = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_b[idx]));
+                FLOAT_TYPEV4 bb = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_b[idx]));
 #else
-            FLOAT_TYPEV4 bb = FLOAT_TYPEV4(data_b[idx]);
+                FLOAT_TYPEV4 bb = FLOAT_TYPEV4(data_b[idx]);
+#endif
+                buf_b[buf_idx + 0] = bb.xy;
+                buf_b[buf_idx + 1] = bb.zw;
+                return;
+            }
 #endif
-            buf_b[buf_idx + 0] = bb.xy;
-            buf_b[buf_idx + 1] = bb.zw;
-#else // LOAD_VEC_BATCH_B == 2
             const uint row_i = ic * BN + col;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (row_i < _ne1 && block + row * 2 + 1 < end_k) {
                 const u16vec2 row_idx = row_ids[col];
                 const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
-                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b[idx]),
-                                              TO_FLOAT_TYPE(data_b[idx + 1]));
+                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b_scalar[idx]),
+                                              TO_FLOAT_TYPE(data_b_scalar[idx + 1]));
             } else if (row_i < _ne1 && block + row * 2 < end_k) {
                 const u16vec2 row_idx = row_ids[col];
                 const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
-                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b[idx]), 0.0f);
+                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b_scalar[idx]), 0.0f);
             } else {
                 buf_b[buf_idx] = FLOAT_TYPEV2(0.0f);
             }
-#endif
 }
 #endif

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

@@ -1,26 +1,26 @@
 #version 450
 
-#include "generic_head.glsl"
 #include "types.glsl"
+#include "generic_unary_head.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512
 
 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
-
 shared vec2 sum[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
+    const uint a_base = get_aoffset() + src0_idx(row * p.ne00);
+    const uint d_base = get_doffset() + dst_idx(row * p.ne10);
+
     sum[tid] = vec2(0.0f, 0.0f);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        const float xi = float(data_a[row*p.KX + col]);
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        const float xi = float(data_a[a_base + i0*p.nb00]);
         sum[tid].x += xi;
         sum[tid].y += xi * xi;
     }
@@ -34,11 +34,11 @@ void main() {
         barrier();
     }
 
-    const float mean = sum[0].x / p.KX;
-    const float var = sum[0].y / p.KX - mean * mean;
+    const float mean = sum[0].x / p.ne00;
+    const float var = sum[0].y / p.ne00 - mean * mean;
     const float inv_std = inversesqrt(var + p.param1);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        data_d[row*p.KX + col] = D_TYPE((float(data_a[row*p.KX + col]) - mean) * inv_std);
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        data_d[d_base + i0*p.nb10] = D_TYPE((float(data_a[a_base + i0*p.nb00]) - mean) * inv_std);
     }
 }

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sin(val));
-}

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sqrt(val));
-}

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val * val);
-}

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

@@ -17,6 +17,30 @@ float op_neg(float x) {
     return -x;
 }
 
+float op_sqr(float x) {
+    return x * x;
+}
+
+float op_sqrt(float x) {
+    return sqrt(x);
+}
+
+float op_sin(float x) {
+    return sin(x);
+}
+
+float op_cos(float x) {
+    return cos(x);
+}
+
+float op_clamp(float x) {
+    return clamp(x, p.param1, p.param2);
+}
+
+float op_leaky_relu(float x) {
+    return max(x, 0.0f) + min(x, 0.0f) * p.param1;
+}
+
 float op_step(float x) {
     return x >= 0.0f ? 1.0f : 0.0f;
 }

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -539,11 +539,9 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
     };
 
     // Shaders with f16 B_TYPE
-    string_to_spv(shader_name + "_f32_f16" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"},                                                     {"B_TYPE", "float16_t"},        {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f32_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f32_f16" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
 
-    string_to_spv(shader_name + "_f16" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"},                                                     {"B_TYPE", "float16_t"},            {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16},     {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f16" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
 
     // bf16
     {
@@ -565,8 +563,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 #endif
         {
             if (!dot2) {
-                string_to_spv(shader_name + "_bf16",         source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"},                             {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
-                string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},  {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+                string_to_spv(shader_name + "_bf16", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"B_TYPE_SCALAR", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}), fp16, coopmat, coopmat2, f16acc);
             }
         }
     }
@@ -583,8 +580,6 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
         }
 
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-        // For unaligned, load one at a time for f32/f16, or two at a time for quants
-        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant;
         // For aligned matmul loads
         std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
 
@@ -597,13 +592,11 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 
         // don't generate f32 variants for coopmat2
         if (!coopmat2) {
-            string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"B_TYPE_SCALAR", "float"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
         if (tname != "f16" && tname != "f32") {
-            string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx,              source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx + "_aligned", source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx, source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
@@ -850,21 +843,12 @@ void process_shaders() {
 
     string_to_spv("repeat_i32", "repeat.comp", {{"A_TYPE", "int32_t"}, {"D_TYPE", "int32_t"}});
     string_to_spv("repeat_back_f32", "repeat_back.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    string_to_spv("get_rows_back_f32", "get_rows_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}});
 
     string_to_spv("repeat_i16", "repeat.comp", {{"A_TYPE", "int16_t"}, {"D_TYPE", "int16_t"}});
 
     string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
-    string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
-    string_to_spv("sqrt_f32", "sqrt.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
-    string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
-    string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
-    string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
     string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("concat_i8", "concat.comp", {{"A_TYPE", "uint8_t"}, {"B_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}});
@@ -891,6 +875,18 @@ void process_shaders() {
     string_to_spv("silu_f32",       "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_silu"}});
     string_to_spv("relu_f16",       "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_relu"}});
     string_to_spv("relu_f32",       "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_relu"}});
+    string_to_spv("sqr_f16",        "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_sqr"}});
+    string_to_spv("sqr_f32",        "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_sqr"}});
+    string_to_spv("sqrt_f16",       "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_sqrt"}});
+    string_to_spv("sqrt_f32",       "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_sqrt"}});
+    string_to_spv("sin_f16",        "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_sin"}});
+    string_to_spv("sin_f32",        "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_sin"}});
+    string_to_spv("cos_f16",        "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_cos"}});
+    string_to_spv("cos_f32",        "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_cos"}});
+    string_to_spv("clamp_f16",      "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_clamp"}});
+    string_to_spv("clamp_f32",      "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_clamp"}});
+    string_to_spv("leaky_relu_f16", "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_leaky_relu"}});
+    string_to_spv("leaky_relu_f32", "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_leaky_relu"}});
     string_to_spv("neg_f16",        "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_neg"}});
     string_to_spv("neg_f32",        "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_neg"}});
     string_to_spv("tanh_f16",       "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_tanh"}});
@@ -948,7 +944,6 @@ void process_shaders() {
     string_to_spv("geglu_quick_f16","geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("geglu_quick_f32","geglu_quick.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
-    string_to_spv("leaky_relu_f32", "leaky_relu.comp",  {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("silu_back_f32",  "silu_back.comp",   {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -1060,6 +1055,31 @@ void process_shaders() {
         }
     }
 
+    for (auto unroll : {false, true}) {
+        for (auto a_f16 : {false, true}) {
+            std::map<std::string, std::string> defines = {
+                {"A_TYPE", a_f16 ? "float16_t" : "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"},
+                {"UNROLL", unroll ? "[[unroll]]" : ""},
+            };
+            std::string name = std::string("conv3d") + (a_f16 ? "_f16" : "") + "_f32";
+            string_to_spv(name + (unroll ? "_unroll" : ""), "conv3d_mm.comp", defines);
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+            if (unroll) {
+                auto cm2_defines = defines;
+                cm2_defines["COOPMAT2"] = "1";
+                string_to_spv(name, "conv3d_mm.comp", cm2_defines, true, false, true);
+            }
+#endif
+#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+            if (unroll) {
+                auto cm1_defines = defines;
+                cm1_defines["COOPMAT"] = "1";
+                string_to_spv(name, "conv3d_mm.comp", cm1_defines, true, true, false);
+            }
+#endif
+        }
+    }
+
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
     string_to_spv("conv2d_dw_whcn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));

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);
@@ -4268,7 +4270,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_RMS_NORM:
         case GGML_OP_NORM:
         case GGML_OP_L2_NORM:
-            supports_op = op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32;
+            supports_op = (op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32) && ggml_is_contiguous_rows(src0);
             break;
         case GGML_OP_ROPE:
             supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;

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)
 

include/llama.h

@@ -558,14 +558,15 @@ extern "C" {
     LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
     LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
 
-    LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_embd     (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_embd_out (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_swa      (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_ctx_train  (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd       (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_inp   (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_out   (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_layer      (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_layer_nextn(const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_head       (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_head_kv    (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_swa        (const struct llama_model * model);
 
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);

src/llama-context.cpp

@@ -1156,6 +1156,10 @@ void llama_context::set_embeddings_layer_inp(uint32_t lid, bool enable) {
     sched_need_reserve = true;
 }
 
+void llama_context::set_nextn_layer_offset(int32_t offset) {
+    cparams.nextn_layer_offset = offset;
+}
+
 void llama_context::set_causal_attn(bool value) {
     LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value);
 
@@ -3699,6 +3703,10 @@ void llama_set_embeddings_layer_inp(llama_context * ctx, uint32_t lid, bool valu
     ctx->set_embeddings_layer_inp(lid, value);
 }
 
+void llama_set_nextn_layer_offset(llama_context * ctx, int32_t offset) {
+    ctx->set_nextn_layer_offset(offset);
+}
+
 llama_memory_t llama_get_memory(const struct llama_context * ctx) {
     if (!ctx) {
         return nullptr;

src/llama-context.h

@@ -115,6 +115,7 @@ struct llama_context {
     void set_embeddings (bool value);
     void set_embeddings_nextn(bool value, bool masked);
     void set_embeddings_layer_inp(uint32_t lid, bool enable);
+    void set_nextn_layer_offset(int32_t offset);
     void set_causal_attn(bool value);
     void set_warmup(bool value);
 

src/llama-cparams.h

@@ -18,6 +18,8 @@ struct llama_cparams {
     int32_t  n_threads;       // number of threads to use for generation
     int32_t  n_threads_batch; // number of threads to use for batch processing
 
+    int32_t  nextn_layer_offset = 0;
+
     float rope_freq_base;
     float rope_freq_scale;
 

src/llama-ext.h

@@ -95,6 +95,11 @@ LLAMA_API llama_memory_breakdown llama_get_memory_breakdown(const struct llama_c
 // If masked == false, output the embeddings for all tokens in the batch regardless of batch.logits
 LLAMA_API void llama_set_embeddings_nextn(struct llama_context * ctx, bool value, bool masked);
 
+// Select which appended NextN block the DECODER_MTP graph runs (offset past
+// the trunk: il = n_layer() + offset). Used by the speculative NextN driver to
+// chain multiple trained NextN heads. Default 0 (first head).
+LLAMA_API void llama_set_nextn_layer_offset(struct llama_context * ctx, int32_t offset);
+
 // mirrors:
 // LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 LLAMA_API float * llama_get_embeddings_nextn(struct llama_context * ctx);

src/llama-graph.h

@@ -682,9 +682,16 @@ struct llm_graph_params {
             }
         }
 
+        // TODO: https://github.com/ggml-org/llama.cpp/pull/24340#discussion_r3448035248
+        if (cparams.nextn_layer_offset != other.cparams.nextn_layer_offset) {
+            return false;
+        }
+
         return
-            cparams.embeddings  == other.cparams.embeddings  &&
-            cparams.causal_attn == other.cparams.causal_attn &&
+            cparams.embeddings              == other.cparams.embeddings              &&
+            cparams.embeddings_nextn        == other.cparams.embeddings_nextn        &&
+            cparams.embeddings_nextn_masked == other.cparams.embeddings_nextn_masked &&
+            cparams.causal_attn             == other.cparams.causal_attn             &&
             arch  == other.arch  &&
             gtype == other.gtype &&
             cvec  == other.cvec  &&

src/llama-model.cpp

@@ -2312,6 +2312,10 @@ int32_t llama_model_n_layer(const llama_model * model) {
     return model->hparams.n_layer();
 }
 
+int32_t llama_model_n_layer_nextn(const llama_model * model) {
+    return model->hparams.n_layer_nextn;
+}
+
 int32_t llama_model_n_head(const llama_model * model) {
     return model->hparams.n_head();
 }

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) {

src/models/step35.cpp

@@ -112,7 +112,7 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
         layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED);
     };
 
-    auto load_block_mtp = [&](int i, bool is_first_mtp) {
+    auto load_block_mtp = [&](int i) {
         auto & layer = layers[i];
 
         const uint32_t n_head_l      = hparams.n_head(i);
@@ -121,15 +121,12 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
 
         // The MTP block is a full Step3p5 decoder layer (mtp_block) plus the
         // NextN-specific wiring (enorm/hnorm/eh_proj + optional shared head).
-        // `mtp_flags` becomes NOT_REQUIRED when the GGUF is trunk-only.
-        //
-        // Only the FIRST MTP block (i == n_main) is required for the
-        // single-block MTP runtime; trailing MTP blocks are always tolerated
-        // as missing so pruned GGUFs (block 0 only) load cleanly. Override
-        // mtp_flags to NOT_REQUIRED for those.
-        const int eff_mtp_flags = is_first_mtp ? mtp_flags : (mtp_flags | TENSOR_NOT_REQUIRED);
+        // Multi-block MTP: every declared MTP block is required (the draft chain
+        // runs all n_layer_nextn heads), so each block uses the captured
+        // `mtp_flags` directly — already NOT_REQUIRED for a trunk-only GGUF,
+        // which keeps that path correct.
 
-        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, eff_mtp_flags);
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, mtp_flags);
         layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED);
         layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED);
 
@@ -140,12 +137,12 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
             layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | TENSOR_DUPLICATED);
         }
 
-        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_l, n_embd_k_gqa, n_embd_v_gqa, eff_mtp_flags);
-        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_v * n_head_l, n_embd}, eff_mtp_flags);
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_l, n_embd_k_gqa, n_embd_v_gqa, mtp_flags);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_v * n_head_l, n_embd}, mtp_flags);
 
         layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, TENSOR_NOT_REQUIRED);
 
-        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, eff_mtp_flags);
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, mtp_flags);
 
         // dense MLP (leading dense blocks) — present if the MTP block isn't MoE
         layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED);
@@ -165,9 +162,9 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
         layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED);
 
         // NextN-specific tensors that define the MTP block.
-        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), { 2 * n_embd, n_embd }, eff_mtp_flags);
-        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), { n_embd },              eff_mtp_flags);
-        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), { n_embd },              eff_mtp_flags);
+        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), { 2 * n_embd, n_embd }, mtp_flags);
+        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), { n_embd },              mtp_flags);
+        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), { n_embd },              mtp_flags);
         layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), { n_embd, n_vocab },     TENSOR_NOT_REQUIRED);
         layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab },     TENSOR_NOT_REQUIRED);
         layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd },              TENSOR_NOT_REQUIRED);
@@ -176,13 +173,11 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
     for (int i = 0; i < n_layer; ++i) {
         load_block_trunk(i, trunk_flags);
     }
-    // Only the first MTP block (i == n_main) is required at runtime — the
-    // single-block-MTP graph in build_arch_graph always uses that one.
-    // Trailing MTP blocks are loaded if present (so an un-pruned GGUF with
-    // all MTP layers still works) but tolerated when absent via the pruning
-    // path. See scripts/prune_step35_extra_mtp.py for the pruner.
+    // All n_layer_nextn MTP blocks are required — the multi-block draft chain
+    // runs every head (head k at offset k). The GGUF declares the count via
+    // step35.nextn_predict_layers.
     for (int i = n_layer; i < n_layer_all; ++i) {
-        load_block_mtp(i, /*is_first_mtp=*/ i == n_layer);
+        load_block_mtp(i);
     }
 }
 
@@ -372,13 +367,14 @@ llama_model_step35::graph_mtp::graph_mtp(const llama_model & model, const llm_gr
     : llm_graph_context(params) {
     GGML_ASSERT(hparams.n_layer_nextn > 0 && "STEP35 MTP requires n_layer_nextn > 0");
 
-    // Single-block MTP only: always run the first trained MTP block (Qwen
-    // MTP / vLLM single-MTP-layer style). Multi-block round-robin proved to
-    // be a much deeper refactor than this PR justifies; the trailing MTP
-    // blocks are loaded with TENSOR_NOT_REQUIRED so pruned GGUFs (with just
-    // block 0) also work — see load_arch_tensors below and
-    // scripts/prune_step35_extra_mtp.py.
-    const int il = hparams.n_layer();
+    // Multi-block MTP: the DECODER_MTP graph runs the MTP head selected by
+    // cparams.nextn_layer_offset (0 = first trained head). The speculative driver
+    // bumps the offset per draft step to chain heads 45->46->47. offset 0 keeps
+    // single-block behavior identical to before.
+    const int il = hparams.n_layer() + cparams.nextn_layer_offset;
+    GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&
+                cparams.nextn_layer_offset < (int) hparams.n_layer_nextn &&
+                "nextn_layer_offset out of range [0, n_layer_nextn)");
     const auto & layer = model.layers[il];
 
     GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");
@@ -536,6 +532,9 @@ llama_model_step35::graph_mtp::graph_mtp(const llama_model & model, const llm_gr
     cur = ggml_add(ctx0, cur, ffn_inp);
     cb(cur, "mtp_post_ffn", il);
 
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+
     // Pre-norm hidden state: used by the AR draft loop to seed the next MTP step.
     cb(cur, "h_nextn", -1);
     res->t_h_nextn = cur;

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

@@ -129,7 +129,154 @@ void test_gbnf_generation(testing &t) {
         });
 
         assert_gbnf_equal(t, R"""(
-            root ::= ([^<] | "<" [^/] | "</" [^t] | "</t" [^a] | "</ta" [^g] | "</tag" [^>])* ("<" | "</" | "</t" | "</ta" | "</tag")?
+            root ::= until-0
+            space ::= | " " | "\n"{1,2} [ \t]{0,20}
+            until-0 ::= | [<] until-0-01 | [^<] until-0
+            until-0-01 ::= | [<] until-0-01 | [/] until-0-02 | [^/<] until-0
+            until-0-02 ::= | [<] until-0-01 | [t] until-0-03 | [^<t] until-0
+            until-0-03 ::= | [<] until-0-01 | [a] until-0-04 | [^<a] until-0
+            until-0-04 ::= | [<] until-0-01 | [g] until-0-05 | [^<g] until-0
+            until-0-05 ::= | [<] until-0-01 | [^<>] until-0
+        )""", gbnf);
+    });
+
+    t.test("until grammar overlapping delimiter", [](testing &t) {
+        auto parser = build_peg_parser([](common_peg_parser_builder & p)  {
+            return p.until("\n</parameter>\n");
+        });
+
+        auto gbnf = build_grammar([&](const common_grammar_builder & builder) {
+            parser.build_grammar(builder);
+        });
+
+        assert_gbnf_equal(t, R"""(
+            root ::= until-0
+            space ::= | " " | "\n"{1,2} [ \t]{0,20}
+            until-0 ::= | [\n] until-0-01 | [^\n] until-0
+            until-0-01 ::= | [\n] until-0-01 | [<] until-0-02 | [^\n<] until-0
+            until-0-02 ::= | [\n] until-0-01 | [/] until-0-03 | [^\n/] until-0
+            until-0-03 ::= | [\n] until-0-01 | [p] until-0-04 | [^\np] until-0
+            until-0-04 ::= | [\n] until-0-01 | [a] until-0-05 | [^\na] until-0
+            until-0-05 ::= | [\n] until-0-01 | [r] until-0-06 | [^\nr] until-0
+            until-0-06 ::= | [\n] until-0-01 | [a] until-0-07 | [^\na] until-0
+            until-0-07 ::= | [\n] until-0-01 | [m] until-0-08 | [^\nm] until-0
+            until-0-08 ::= | [\n] until-0-01 | [e] until-0-09 | [^\ne] until-0
+            until-0-09 ::= | [\n] until-0-01 | [t] until-0-10 | [^\nt] until-0
+            until-0-10 ::= | [\n] until-0-01 | [e] until-0-11 | [^\ne] until-0
+            until-0-11 ::= | [\n] until-0-01 | [r] until-0-12 | [^\nr] until-0
+            until-0-12 ::= | [\n] until-0-01 | [>] until-0-13 | [^\n>] until-0
+            until-0-13 ::= | [^\n] until-0
+        )""", gbnf);
+    });
+
+    // DeepSeek-V3.2 tag prefix. The DSML token (｜DSML｜) embeds U+FF5C,
+    // so the delimiter mixes ASCII and multi-byte codepoints.
+    t.test("until grammar unicode delimiter", [](testing &t) {
+        auto parser = build_peg_parser([](common_peg_parser_builder & p)  {
+            return p.until("<｜DSML｜");
+        });
+
+        auto gbnf = build_grammar([&](const common_grammar_builder & builder) {
+            parser.build_grammar(builder);
+        });
+
+        assert_gbnf_equal(t, R"""(
+            root ::= until-0
+            space ::= | " " | "\n"{1,2} [ \t]{0,20}
+            until-0 ::= | [<] until-0-01 | [^<] until-0
+            until-0-01 ::= | [<] until-0-01 | [\uFF5C] until-0-02 | [^<\uFF5C] until-0
+            until-0-02 ::= | [<] until-0-01 | [D] until-0-03 | [^<D] until-0
+            until-0-03 ::= | [<] until-0-01 | [S] until-0-04 | [^<S] until-0
+            until-0-04 ::= | [<] until-0-01 | [M] until-0-05 | [^<M] until-0
+            until-0-05 ::= | [<] until-0-01 | [L] until-0-06 | [^<L] until-0
+            until-0-06 ::= | [<] until-0-01 | [^<\uFF5C] until-0
+        )""", gbnf);
+    });
+
+    t.test("until grammar multiple delimiters", [](testing &t) {
+        auto parser = build_peg_parser([](common_peg_parser_builder & p)  {
+            return p.until_one_of({"ab", "cd", "ef"});
+        });
+
+        auto gbnf = build_grammar([&](const common_grammar_builder & builder) {
+            parser.build_grammar(builder);
+        });
+
+        assert_gbnf_equal(t, R"""(
+            root ::= until-0
+            space ::= | " " | "\n"{1,2} [ \t]{0,20}
+            until-0 ::= | [a] until-0-01 | [c] until-0-03 | [e] until-0-05 | [^ace] until-0
+            until-0-01 ::= | [a] until-0-01 | [c] until-0-03 | [e] until-0-05 | [^abce] until-0
+            until-0-03 ::= | [a] until-0-01 | [c] until-0-03 | [e] until-0-05 | [^acde] until-0
+            until-0-05 ::= | [a] until-0-01 | [c] until-0-03 | [e] until-0-05 | [^acef] until-0
+        )""", 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);
     });

tests/test-backend-ops.cpp

@@ -3298,21 +3298,29 @@ struct test_norm : public test_case {
     const std::array<int64_t, 4> ne;
     const bool v; // whether a is a non-contiguous view
     const float eps;
+    const bool noncontig_rows;
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, v, eps);
+        return VARS_TO_STR5(type, ne, v, eps, noncontig_rows);
     }
 
     test_norm(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 5, 4, 3},
             bool v = false,
-            float eps = 1e-6f)
-        : type(type), ne(ne), v(v), eps(eps) {}
+            float eps = 1e-6f,
+            bool noncontig_rows = false)
+        : type(type), ne(ne), v(v), eps(eps), noncontig_rows(noncontig_rows) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        const std::array<int64_t, 4> ne_a = noncontig_rows ?
+            std::array<int64_t, 4>{ ne[1], ne[0], ne[2], ne[3] } : ne;
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
         ggml_set_name(a, "a");
 
+        if (noncontig_rows) {
+            a = ggml_permute(ctx, a, 1, 0, 2, 3);
+            ggml_set_name(a, "permuted a");
+        }
         if (v) {
             a = ggml_view_4d(ctx, a, a->ne[0]/2, a->ne[1]/2, a->ne[2]/2, a->ne[3]/2, a->nb[1], a->nb[2], a->nb[3], 0);
             ggml_set_name(a, "view of a");
@@ -6193,21 +6201,29 @@ struct test_l2_norm : public test_case {
     const std::array<int64_t, 4> ne;
     const float eps;
     bool v;
+    bool noncontig_rows;
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, eps, v);
+        return VARS_TO_STR5(type, ne, eps, v, noncontig_rows);
     }
 
     test_l2_norm(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 64, 320, 1},
             float eps = 1e-12f,
-            bool v = false)
-        : type(type), ne(ne), eps(eps), v(v) {}
+            bool v = false,
+            bool noncontig_rows = false)
+        : type(type), ne(ne), eps(eps), v(v), noncontig_rows(noncontig_rows) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        const std::array<int64_t, 4> ne_a = noncontig_rows ?
+            std::array<int64_t, 4>{ ne[1], ne[0], ne[2], ne[3] } : ne;
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
         ggml_set_name(a, "a");
 
+        if (noncontig_rows) {
+            a = ggml_permute(ctx, a, 1, 0, 2, 3);
+            ggml_set_name(a, "permuted a");
+        }
         if (v) {
             a = ggml_view_4d(ctx, a, a->ne[0]/2, a->ne[1]/2, a->ne[2]/2, a->ne[3]/2, a->nb[1], a->nb[2], a->nb[3], 0);
             ggml_set_name(a, "view of a");
@@ -8282,9 +8298,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                 test_cases.emplace_back(new test_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, v, eps));
                 test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, v, eps));
             }
+            test_cases.emplace_back(new test_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, false, eps, true));
             test_cases.emplace_back(new test_rms_norm_back(GGML_TYPE_F32, { n, 5, 4, 3 }, eps));
             test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false));
             test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, true));
+            test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false, true));
         }
     }
 
@@ -8433,6 +8451,7 @@ 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 +8468,7 @@ 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}));
@@ -9270,6 +9290,34 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
 
+    struct conv3d_perf_case {
+        int N, IC, ID, IH, IW, OC, KD, KH, KW, s0, s1, s2, p0, p1, p2, d0, d1, d2;
+    };
+
+    const std::vector<conv3d_perf_case> conv3d_cases = {
+        {1,  320, 8,  38,  26, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1, 1280, 8,  38,  26, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1,  320, 8,  76,  52, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1, 1280, 8,  76,  52, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1,  320, 8, 152, 104, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+#if 0
+        // too slow on some devices
+        {1, 1280, 8, 152, 104, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1,  320, 4, 304, 208, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1,  640, 4, 304, 208, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+#endif
+    };
+
+    for (ggml_type kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+        for (const conv3d_perf_case & c : conv3d_cases) {
+            test_cases.emplace_back(new test_conv_3d(
+                c.N, c.IC, c.ID, c.IH, c.IW,
+                c.OC, c.KD, c.KH, c.KW,
+                c.s0, c.s1, c.s2, c.p0, c.p1, c.p2, c.d0, c.d1, c.d2,
+                kernel_type));
+        }
+    }
+
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1,   1, 1, 1}));
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 512, 1, 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-jinja.cpp

@@ -995,6 +995,32 @@ static void test_macros(testing & t) {
         json::object(),
         "Hello, John Smith,Hi, Jane Doe"
     );
+
+    test_template(t, "macro with caller",
+        "\
+{%- macro nest_dict(o, i, ff='') %}\n\
+  {{- caller(ff) }}\n\
+  {%- for k, v in o|items %}\n\
+    {{- i + k + ': ' }}\n\
+    {%- if v is mapping %}\n\
+      {{- '{' }}\n\
+      {% call(f) nest_dict(v, i + '    ') %}\n\
+        {{- 'fail' if ff is undefined }}\n\
+      {%- endcall %}\n\
+      {{- i + '}' }}\n\
+    {% else %}\n\
+      {{- v|string }}\n\
+    {% endif %}\n\
+  {%- endfor %}\n\
+{%- endmacro %}\n\
+{%- call(f) nest_dict({'root1': 1, 'root2': {'nest1': 1, 'nest2': {'nest3': 2}}}, '    ', 'Dict') %}\n\
+  {{- 'fail' if ff is defined }}\n\
+  {{- f + ' {' }}\n\
+{% endcall %}\n\
+{{- '}' }}",
+        json::object(),
+        "Dict {\n    root1: 1\n    root2: {\n        nest1: 1\n        nest2: {\n            nest3: 2\n        }\n    }\n}"
+    );
 }
 
 static void test_namespace(testing & t) {

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/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

@@ -1045,8 +1045,17 @@ struct clip_model_loader {
     bool has_vision = false;
     bool has_audio  = false;
 
+    mtmd_progress_callback progress_callback = nullptr;
+    void * progress_callback_user_data = nullptr;
+
     // TODO @ngxson : we should not pass clip_ctx here, it should be clip_model
-    clip_model_loader(const char * fname, bool skip_tensors = false) : fname(fname) {
+    clip_model_loader(const char * fname,
+            bool skip_tensors = false,
+            mtmd_progress_callback progress_cb = nullptr,
+            void * progress_user_data = nullptr)
+        : fname(fname),
+          progress_callback(progress_cb),
+          progress_callback_user_data(progress_user_data) {
         struct ggml_context * meta = nullptr;
 
         struct gguf_init_params params = {
@@ -1255,12 +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) {
@@ -1642,6 +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:
                     {
@@ -1654,11 +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);
@@ -2731,7 +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];
@@ -2787,37 +2795,60 @@ struct clip_model_loader {
         }
 
         // load data
-        if (!ctx_clip.no_alloc) {
+        {
             std::vector<uint8_t> read_buf;
 
+            // start loading event
+            if (progress_callback){
+                progress_callback(0.0, progress_callback_user_data);
+            }
+
+            // compute total tensor data size for progress reporting
+            size_t total_data_size = 0;
+            for (auto & t : tensors_to_load) {
+                total_data_size += ggml_nbytes(t);
+            }
+
             // alloc memory and offload data
             ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(ctx_clip.backend);
             ctx_clip.buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(ctx_clip.ctx_data.get(), buft));
             ggml_backend_buffer_set_usage(ctx_clip.buf.get(), GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
-            for (auto & t : tensors_to_load) {
-                ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data.get(), t->name);
-                GGML_ASSERT(cur && "tensor not found in ctx_data");
-                auto it_off = tensor_offset.find(t->name);
-                GGML_ASSERT(it_off != tensor_offset.end() && "no offset for tensor");
-                const size_t offset = it_off->second;
-                fin.seekg(offset, std::ios::beg);
-                if (!fin) {
-                    throw std::runtime_error(string_format("%s: failed to seek for tensor %s\n", __func__, t->name));
-                }
-                size_t num_bytes = ggml_nbytes(cur);
-                if (ggml_backend_buft_is_host(buft)) {
-                    // for the CPU and Metal backend, we can read directly into the tensor
-                    fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
-                } else {
-                    // read into a temporary buffer first, then copy to device memory
-                    read_buf.resize(num_bytes);
-                    fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
-                    ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+            // read the weight from file
+            if (!ctx_clip.no_alloc) {
+                size_t data_loaded = 0;
+                for (auto & t : tensors_to_load) {
+                    ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data.get(), t->name);
+                    GGML_ASSERT(cur && "tensor not found in ctx_data");
+                    auto it_off = tensor_offset.find(t->name);
+                    GGML_ASSERT(it_off != tensor_offset.end() && "no offset for tensor");
+                    const size_t offset = it_off->second;
+                    fin.seekg(offset, std::ios::beg);
+                    if (!fin) {
+                        throw std::runtime_error(string_format("%s: failed to seek for tensor %s\n", __func__, t->name));
+                    }
+                    size_t num_bytes = ggml_nbytes(cur);
+                    if (ggml_backend_buft_is_host(buft)) {
+                        // for the CPU and Metal backend, we can read directly into the tensor
+                        fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+                    } else {
+                        // read into a temporary buffer first, then copy to device memory
+                        read_buf.resize(num_bytes);
+                        fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                        ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+                    }
+                    data_loaded += num_bytes;
+                    if (progress_callback && total_data_size > 0) {
+                        const float progress = (float)data_loaded / (float)total_data_size;
+                        if (!progress_callback(progress, progress_callback_user_data)) {
+                            throw std::runtime_error(string_format("%s: model loading cancelled by progress_callback\n", __func__));
+                        }
+                    }
                 }
+                LOG_DBG("%s: loaded %zu tensors from %s\n", __func__, tensors_to_load.size(), fname.c_str());
+            } else {
+                LOG_DBG("%s: no_alloc is set, skipping tensor data loading (%zu tensors)\n", __func__, tensors_to_load.size());
             }
             fin.close();
-
-            LOG_DBG("%s: loaded %zu tensors from %s\n", __func__, tensors_to_load.size(), fname.c_str());
         }
 
     }
@@ -3105,7 +3136,10 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
     clip_ctx * ctx_audio = nullptr;
 
     try {
-        clip_model_loader loader(fname);
+        clip_model_loader loader(fname,
+            /* skip_tensors */ false,
+            ctx_params.progress_callback,
+            ctx_params.progress_callback_user_data);
         bool skip_audio = false;
 
         if (loader.has_vision) {
@@ -4353,7 +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/clip.h

@@ -54,6 +54,8 @@ struct clip_context_params {
     ggml_backend_sched_eval_callback cb_eval;
     void * cb_eval_user_data;
     bool no_alloc;
+    mtmd_progress_callback progress_callback;
+    void * progress_callback_user_data;
 };
 
 struct clip_init_result {

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/mtmd/mtmd.cpp

@@ -251,6 +251,8 @@ mtmd_context_params mtmd_context_params_default() {
         /* cb_eval           */ nullptr,
         /* cb_eval_user_data */ nullptr,
         /* batch_max_tokens  */ 1024,
+        /* progress_callback */ nullptr,
+        /* progress_callback_user_data */ nullptr,
     };
     return params;
 }
@@ -345,6 +347,8 @@ struct mtmd_context {
             /* cb_eval           */ ctx_params.cb_eval,
             /* cb_eval_user_data */ ctx_params.cb_eval_user_data,
             /* no_alloc          */ no_alloc,
+            /* progress_callback */ ctx_params.progress_callback,
+            /* progress_callback_user_data */ ctx_params.progress_callback_user_data,
         };
 
         auto res = clip_init(mmproj_fname, ctx_clip_params);
@@ -2133,9 +2137,12 @@ std::map<ggml_backend_dev_t, size_t> mtmd_get_memory_usage(const char * mmproj_f
     mtmd::context_ptr ctx;
     auto saved_log_callback = g_logger_state.log_callback;
     auto saved_log_user_data = g_logger_state.log_callback_user_data;
+
+    ctx_params.progress_callback = nullptr;
+
     try {
         mtmd_log_set(stub_log_callback, nullptr); // suppress logging
-        ctx.reset(new mtmd_context(mmproj_fname, nullptr, ctx_params));
+        ctx.reset(new mtmd_context(mmproj_fname, nullptr, ctx_params, true));
         mtmd_log_set(saved_log_callback, saved_log_user_data); // restore log callback
         std::map<ggml_backend_dev_t, size_t> total_mem;
         auto merge = [&](const struct clip_ctx * c) {

tools/mtmd/mtmd.h

@@ -83,6 +83,8 @@ typedef struct mtmd_input_chunks mtmd_input_chunks;
 typedef struct mtmd_input_text   mtmd_input_text;
 typedef struct mtmd_batch        mtmd_batch;
 
+typedef bool (*mtmd_progress_callback)(float progress, void * user_data);
+
 struct mtmd_context_params {
     bool use_gpu;
     bool print_timings;
@@ -104,6 +106,12 @@ struct mtmd_context_params {
     int32_t batch_max_tokens; // maximum number of output tokens in a batch
                               // (note: this is not a hard-limit, the first image will always be added even if it exceeds this limit)
                               // (default: 1024)
+
+    // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
+    // If the provided progress_callback returns true, model loading continues.
+    // If it returns false, model loading is immediately aborted.
+    mtmd_progress_callback progress_callback;
+    void * progress_callback_user_data;
 };
 
 MTMD_API const char * mtmd_default_marker(void);

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:*
 
@@ -1859,9 +1866,37 @@ Example events:
 
 {
   "model": "...",
-  "event": "download_finished",
+  "event": "model_status",
   "data": {
-    "status": "loading"
+    "status": "loading",
+    "progress": {
+      "stages": ["text_model", "spec_model", "mmproj_model"],
+      "current": "text_model",
+      "value": 0.5
+    }
+  }
+}
+// note for "loading" status:
+// - subsequent events will follow the same order of "stages" list
+// - mmap is may report incorrect progress on some platforms; if you need exact progress, use --no-mmap
+
+{
+  "model": "...",
+  "event": "model_status",
+  "data": {
+    "status": "loaded",
+    "info": {
+      // note: only include info on first load
+      // waking up from sleep doesn't have this
+    }
+  }
+}
+
+{
+  "model": "...",
+  "event": "model_status",
+  "data": {
+    "status": "sleeping"
   }
 }
 

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.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

@@ -64,6 +64,17 @@ 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 +224,7 @@ 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 +334,7 @@ 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 +387,7 @@ 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) {
@@ -442,6 +453,7 @@ void server_models::load_models() {
                 /* last_used     */ 0,
                 /* args          */ std::vector<std::string>(),
                 /* loaded_info   */ {},
+                /* progress      */ {},
                 /* exit_code     */ 0,
                 /* stop_timeout  */ DEFAULT_STOP_TIMEOUT,
                 /* multimodal    */ mtmd_caps{false, false},
@@ -522,7 +534,7 @@ 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();
@@ -608,6 +620,7 @@ void server_models::load_models() {
                     /* last_used     */ 0,
                     /* args          */ std::vector<std::string>(),
                     /* loaded_info   */ {},
+                    /* progress      */ {},
                     /* exit_code     */ 0,
                     /* stop_timeout  */ DEFAULT_STOP_TIMEOUT,
                     /* multimodal    */ mtmd_caps{false, false},
@@ -620,7 +633,7 @@ 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();
@@ -813,17 +826,23 @@ 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;
@@ -867,6 +886,12 @@ 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());
@@ -884,13 +909,17 @@ 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
 
@@ -923,7 +952,7 @@ 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;
             }
@@ -971,10 +1000,14 @@ 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);
     });
 
@@ -982,7 +1015,7 @@ 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
         }
@@ -999,92 +1032,13 @@ 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;
@@ -1140,6 +1094,9 @@ void server_models::update_status(const std::string & name, const update_status_
         if (!args.loaded_info.is_null()) {
             meta.loaded_info = args.loaded_info;
         }
+        if (!args.progress.is_null()) {
+            meta.progress = args.progress;
+        }
     }
     // broadcast status change to SSE
     {
@@ -1152,6 +1109,9 @@ void server_models::update_status(const std::string & name, const update_status_
         if (!args.loaded_info.is_null()) {
             data["info"] = args.loaded_info;
         }
+        if (!args.progress.is_null()) {
+            data["progress"] = args.progress;
+        }
         // note: notify_sse doesn't acquire the lock, so no deadlock here
         notify_sse("status_change", name, data);
     }
@@ -1190,37 +1150,65 @@ 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) {
@@ -1235,7 +1223,9 @@ 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;
     });
 }
 
@@ -1320,10 +1310,39 @@ 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:
             {
-                // do nothing for now
-                // TODO: report loading progress for first load and wakeup from sleep
+                update_status(name, {
+                    SERVER_MODEL_STATUS_LOADING,
+                    0,
+                    nullptr, // no loaded_info yet
+                    payload,
+                });
             } break;
         case SERVER_STATE_READY:
             {
@@ -1331,7 +1350,8 @@ void server_models::handle_child_state(const std::string & name, const std::stri
                     SERVER_MODEL_STATUS_LOADED,
                     0,
                     // note: payload can be empty if this is a wakeup from sleep
-                    payload.size() > 0 ? payload : nullptr
+                    payload.size() > 0 ? payload : nullptr,
+                    {}, // reset progress info
                 });
             } break;
         case SERVER_STATE_SLEEPING:
@@ -1353,6 +1373,92 @@ 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]() {
@@ -1384,6 +1490,7 @@ void server_child::notify_to_router(const std::string & state, const json & payl
         {"state", state},
         {"payload", payload},
     };
+    std::lock_guard<std::mutex> lk(mtx_stdout);
     common_log_pause(common_log_main());
     fflush(stdout);
     fprintf(stdout, "%s%s\n", CMD_CHILD_TO_ROUTER_STATE, safe_json_to_str(data).c_str());
@@ -1625,7 +1732,7 @@ 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;
         }
@@ -1666,8 +1773,9 @@ 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;
@@ -1688,10 +1796,21 @@ 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;
@@ -1705,10 +1824,7 @@ 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";
@@ -72,6 +77,7 @@ struct server_model_meta {
     int64_t last_used = 0; // for LRU unloading
     std::vector<std::string> args; // args passed to the model instance, will be populated by render_args()
     json loaded_info; // info to be reflected via /v1/models endpoint ; if in DOWNLOADING state, it should contain download progress info
+    json progress; // reflect load or download progress info, if any
     int exit_code = 0; // exit code of the model instance process (only valid if status == FAILED)
     int stop_timeout = 0; // seconds to wait before force-killing the model instance during shutdown
     mtmd_caps multimodal; // multimodal capabilities
@@ -104,7 +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;
@@ -160,22 +165,27 @@ 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);
-
-    // update the status of a model instance (thread-safe)
     struct update_status_args {
         server_model_status status;
         int exit_code = 0; // only valid if status == UNLOADED
         json loaded_info = nullptr;
+        json progress = nullptr;
     };
+    // update the status of a model instance (thread-safe)
+    // also send SSE notification to /models/sse endpoint
     void update_status(const std::string & name, const update_status_args & args);
     void update_download_progress(const std::string & name, const common_download_progress & progress, bool done, bool ok = true);
 
@@ -208,8 +218,14 @@ 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-schema.cpp

@@ -14,6 +14,9 @@ std::vector<std::unique_ptr<field>> make_llama_cmpl_schema(const common_params &
         fields.emplace_back(f);
     };
 
+    add((new field_bool("verbose", params.verbose))
+        ->set_desc("Include __verbose field in the response with additional debug information"));
+
     add((new field_bool("timings_per_token", params.timings_per_token))
         ->set_desc("Include prompt processing and text generation speed information in each response"));
 

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

@@ -11,6 +11,7 @@
 #include <cstring>
 #include <climits>
 #include <algorithm>
+#include <unordered_set>
 
 namespace fs = std::filesystem;
 
@@ -568,9 +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)}};
@@ -611,8 +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

@@ -89,6 +89,17 @@ int llama_server(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
+    // note: router mode also accepts -hf remote-preset, so we need to check that first
+    if (!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 +145,7 @@ 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 +266,21 @@ 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) {
+        // single-model mode (NOT spawned by router)
+        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) {

tools/server/tests/unit/test_chat_completion.py

@@ -603,3 +603,23 @@ def test_chat_completions_token_count():
         })
         assert res.status_code == 200
         assert res.body["input_tokens"] > 5
+
+
+def test_verbose_debug():
+    global server
+    server.start()
+    for verbose in [True, False]:
+        res = server.make_request("POST", "/chat/completions", data={
+            "max_tokens": 2,
+            "messages": [
+                {"role": "system", "content": "Book"},
+                {"role": "user", "content": "What is the best book"},
+            ],
+            "verbose": verbose,
+        })
+        assert res.status_code == 200
+        if verbose:
+            assert "__verbose" in res.body
+            assert "Book" in res.body["__verbose"]["prompt"]
+        else:
+            assert "__verbose" not in res.body

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;