metal : per-device tuned (Q, NE) for flash-attn vec (#25750)

* rebase Q-generic FA vec body from 01dc93607 (#23114)

* add 53 f16 (Q,NE) flash-attn vec instantiations (vec 80 -> 133)

* add FA vec (Q,NE) tuning table + dispatch wiring + SMEM cap fallback

* add  FA vec (Q,NE) perf sweep

* fill tuning result

* fold family table into a per-family representative SKU

* refactor tuning result format

* extend FA vec tuning to quantized KV caches

* sync fa vec tuner bucketing with runtime, use pointwise tuning regret

* update tuned table

* format and cleanup

* prefix fa_vec tuning procs with ggml_backend_metal_tuning_, drop unused fa_vec_override_active

CONTRIBUTING.md

@@ -73,6 +73,7 @@ For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 - When merging a PR, make sure you have a good understanding of the changes
 - If a PR does not warrant a new release, add `[no release]` in the squashed commit to spare CI resources
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
+- Add the ["merge ready"](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+is%3Aopen+draft%3Ano+sort%3Aupdated-desc+label%3A%22merge+ready%22+) label to a PR to indicate when a PR can be fast-merged without waiting for 2 independent reviews. [(more info)](https://github.com/ggml-org/llama.cpp/pull/26178)
 
 Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:
 - The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.

common/arg.cpp

@@ -1061,6 +1061,31 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
     return devices;
 }
 
+void common_print_available_devices() {
+    constexpr size_t MiB = 1024 * 1024;
+    std::vector<ggml_backend_dev_t> devices;
+
+    ggml_backend_load_all();
+
+    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+        auto * dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU) {
+            devices.push_back(dev);
+        }
+    }
+    printf("Available devices:\n");
+
+    if (devices.empty()) {
+        printf("  (none)\n");
+        return;
+    }
+    for (auto * dev : devices) {
+        size_t free, total;
+        ggml_backend_dev_memory(dev, &free, &total);
+        printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / MiB, free / MiB);
+    }
+}
+
 static void add_rpc_devices(const std::string & servers) {
     auto rpc_servers = string_split<std::string>(servers, ',');
     if (rpc_servers.empty()) {
@@ -2588,20 +2613,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--list-devices"},
         "print list of available devices and exit",
         [](common_params &) {
-            ggml_backend_load_all();
-            std::vector<ggml_backend_dev_t> devices;
-            for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-                auto * dev = ggml_backend_dev_get(i);
-                if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU) {
-                    devices.push_back(dev);
-                }
-            }
-            printf("Available devices:\n");
-            for (auto * dev : devices) {
-                size_t free, total;
-                ggml_backend_dev_memory(dev, &free, &total);
-                printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
-            }
+            common_print_available_devices();
             exit(0);
         }
     ));

common/arg.h

@@ -123,6 +123,9 @@ struct common_params_context {
 // if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
 
+// load all backends and print the list of available (non-CPU) devices to stdout
+void common_print_available_devices();
+
 // parse input arguments from CLI into a map
 bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map);
 

common/chat-peg-parser.cpp

@@ -1056,3 +1056,141 @@ void common_chat_peg_gemma4_mapper::visit(const common_peg_ast_arena & arena, co
         visit(arena, child_id);
     }
 }
+
+static void minimax_m3_collect(const common_peg_ast_arena &     arena,
+                               const common_peg_ast_node &      node,
+                               const std::string &              tag,
+                               std::vector<common_peg_ast_id> & out) {
+    for (auto child_id : node.children) {
+        const auto & child = arena.get(child_id);
+        if (child.tag == tag) {
+            out.push_back(child_id);
+        } else {
+            minimax_m3_collect(arena, child, tag, out);
+        }
+    }
+}
+
+static common_peg_ast_id minimax_m3_value_of(const common_peg_ast_arena & arena, const common_peg_ast_node & node) {
+    for (auto child_id : node.children) {
+        const auto & tag = arena.get(child_id).tag;
+        if (tag == common_chat_peg_builder::TOOL_ARG_VALUE ||
+            tag == common_chat_peg_builder::TOOL_ARG_STRING_VALUE ||
+            tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_OBJECT ||
+            tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_ARRAY) {
+            return child_id;
+        }
+    }
+    return COMMON_PEG_INVALID_AST_ID;
+}
+
+static std::string minimax_m3_value_to_json(const common_peg_ast_arena & arena, common_peg_ast_id id, bool closed);
+
+static std::string minimax_m3_member_to_json(const common_peg_ast_arena & arena, const common_peg_ast_node & node) {
+    auto name_id = arena.find_by_tag(node, common_chat_peg_builder::TOOL_ARG_NAME);
+    if (name_id == COMMON_PEG_INVALID_AST_ID) {
+        return "";
+    }
+
+    return ordered_json(arena.get(name_id).text).dump() + ":" +
+           minimax_m3_value_to_json(arena, minimax_m3_value_of(arena, node), !node.is_partial);
+}
+
+static std::string minimax_m3_container_to_json(const common_peg_ast_arena & arena,
+                                                const common_peg_ast_node & node,
+                                                bool                        is_object,
+                                                bool                        closed) {
+    const std::string tag = is_object ? common_chat_peg_builder::TOOL_ARG
+                                      : common_chat_peg_minimax_m3_mapper::TOOL_ARG_ITEM;
+
+    std::vector<common_peg_ast_id> entries;
+    minimax_m3_collect(arena, node, tag, entries);
+
+    std::string result = is_object ? "{" : "[";
+
+    bool add_comma = false;
+    for (auto entry_id : entries) {
+        const auto & entry = arena.get(entry_id);
+
+        std::string text;
+        if (is_object) {
+            text = minimax_m3_member_to_json(arena, entry);
+        } else {
+            text = minimax_m3_value_to_json(arena, minimax_m3_value_of(arena, entry), !entry.is_partial);
+        }
+
+        if (text.empty()) {
+            continue;
+        }
+
+        if (add_comma) {
+            result += ",";
+        }
+        add_comma = true;
+        result += text;
+    }
+
+    if (closed) {
+        result += is_object ? "}" : "]";
+    }
+    return result;
+}
+
+static std::string minimax_m3_value_to_json(const common_peg_ast_arena & arena, common_peg_ast_id id, bool closed) {
+    if (id == COMMON_PEG_INVALID_AST_ID) {
+        return "";
+    }
+
+    const auto & node = arena.get(id);
+
+    if (node.tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_OBJECT) {
+        return minimax_m3_container_to_json(arena, node, /* is_object = */ true, closed);
+    }
+
+    if (node.tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_ARRAY) {
+        return minimax_m3_container_to_json(arena, node, /* is_object = */ false, closed);
+    }
+
+    if (node.tag == common_chat_peg_builder::TOOL_ARG_STRING_VALUE) {
+        return "\"" + escape_json_string_inner(std::string(node.text)) + (closed ? "\"" : "");
+    }
+
+    // Numbers and booleans are written verbatim by the template
+    return std::string(node.text);
+}
+
+void common_chat_peg_minimax_m3_mapper::from_ast(const common_peg_ast_arena &    arena,
+                                                 const common_peg_parse_result & result) {
+    for (const auto & node : result.nodes) {
+        visit(arena, node);
+    }
+}
+
+void common_chat_peg_minimax_m3_mapper::visit(const common_peg_ast_arena & arena, common_peg_ast_id id) {
+    const auto & node = arena.get(id);
+
+    if (node.tag == common_chat_peg_builder::REASONING) {
+        result.reasoning_content += std::string(node.text);
+        return;
+    }
+
+    if (node.tag == common_chat_peg_builder::CONTENT) {
+        result.content += std::string(node.text);
+        return;
+    }
+
+    if (node.tag == common_chat_peg_builder::TOOL) {
+        auto name_id = arena.find_by_tag(node, common_chat_peg_builder::TOOL_NAME);
+        if (name_id != COMMON_PEG_INVALID_AST_ID) {
+            common_chat_tool_call call;
+            call.name      = std::string(arena.get(name_id).text);
+            call.arguments = minimax_m3_container_to_json(arena, node, /* is_object = */ true, !node.is_partial);
+            result.tool_calls.push_back(call);
+        }
+        return;
+    }
+
+    for (auto child_id : node.children) {
+        visit(arena, child_id);
+    }
+}

common/chat-peg-parser.h

@@ -40,6 +40,18 @@ class common_chat_peg_gemma4_mapper : public common_chat_peg_mapper {
     void visit(const common_peg_ast_arena & arena, common_peg_ast_id id);
 };
 
+class common_chat_peg_minimax_m3_mapper : public common_chat_peg_mapper {
+  public:
+    static constexpr const char * TOOL_ARG_OBJECT = "tool-arg-object";
+    static constexpr const char * TOOL_ARG_ARRAY  = "tool-arg-array";
+    static constexpr const char * TOOL_ARG_ITEM   = "tool-arg-item";
+
+    common_chat_peg_minimax_m3_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
+    virtual void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
+  private:
+    void visit(const common_peg_ast_arena & arena, common_peg_ast_id id);
+};
+
 struct content_structure;
 struct tool_call_structure;
 

common/chat.cpp

@@ -816,6 +816,8 @@ const char * common_chat_format_name(common_chat_format format) {
             return "peg-native";
         case COMMON_CHAT_FORMAT_PEG_GEMMA4:
             return "peg-gemma4";
+        case COMMON_CHAT_FORMAT_PEG_MINIMAX_M3:
+            return "peg-minimax-m3";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -2270,6 +2272,264 @@ static common_chat_params common_chat_params_init_cohere2moe(const common_chat_t
     return data;
 }
 
+static common_chat_params common_chat_params_init_minimax_m3(const common_chat_template &          tmpl,
+                                                             const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format             = COMMON_CHAT_FORMAT_PEG_MINIMAX_M3;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = "<mm:think>";
+    data.thinking_end_tags  = {"</mm:think>"};
+
+    // M3 prefixes every tool tag with the namespace token "]<]minimax[>[";
+    // params use the parameter name as the tag (<file_path>...</file_path>).
+    const std::string NS          = "]<]minimax[>[";
+    const std::string THINK_START = "<mm:think>";
+    const std::string THINK_END   = "</mm:think>";
+    const std::string FC_START    = NS + "<tool_call>";
+    const std::string FC_END      = NS + "</tool_call>";
+    const std::string INVOKE_END  = NS + "</invoke>";
+
+    data.preserved_tokens = {
+        NS,
+        "<tool_call>",
+        "</tool_call>",
+        THINK_START,
+        THINK_END,
+    };
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "]~b]ai"        },
+        { COMMON_CHAT_ROLE_USER,      "]~b]user"      },
+        { COMMON_CHAT_ROLE_TOOL,      "]~b]tool"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "]~b]developer" },
+        { COMMON_CHAT_ROLE_SYSTEM,    "]~b]system"    },
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    const std::string GEN_PROMPT = data.generation_prompt;
+
+    using mm3 = common_chat_peg_minimax_m3_mapper;
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.prefix(GEN_PROMPT, THINK_START);
+        auto end = p.end();
+
+        auto reasoning = p.eps();
+        if (extract_reasoning) {
+            auto block = inputs.enable_thinking
+                             ? p.literal(THINK_START) + p.space() +
+                                   p.ac(p.reasoning(p.until(THINK_END)) + p.literal(THINK_END), THINK_END)
+                             : p.literal(THINK_START) + p.ac(p.until(THINK_END) + p.literal(THINK_END), THINK_END);
+
+            // A turn without reasoning is prefixed with a bare </mm:think>, written either by the
+            // generation prompt (thinking_mode = "disabled") or by the model itself.
+            reasoning = p.optional(p.choice({ block, p.literal(THINK_END) }));
+        }
+
+        if (has_response_format) {
+            auto response_format = p.rule("response-format",
+                p.literal("```json") + p.space() +
+                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
+                p.space() + p.literal("```"));
+            return generation_prompt + reasoning + response_format + end;
+        }
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
+        }
+
+        auto alternatives_of = [](const json & schema) -> std::optional<json> {
+            for (const auto * keyword : { "oneOf", "anyOf" }) {
+                if (schema.contains(keyword) && schema.at(keyword).is_array() && !schema.at(keyword).empty()) {
+                    return schema.at(keyword);
+                }
+            }
+            return std::nullopt;
+        };
+
+        auto tool_choice = p.choice();
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string  name     = function.at("name");
+            auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+            auto schema_info = common_schema_info();
+            schema_info.resolve_refs(params);
+
+            // The template expands argument values recursively in XML (see the to_xml() macro)
+            std::function<common_peg_parser(const json &, const std::string &, const std::string &)> value_of;
+            std::function<common_peg_parser(const json &, const std::string &)>                      members_of;
+
+            auto element_of = [&](const std::string & tag, const json & schema, const std::string & rule_name) {
+                const std::string close = NS + "</" + tag + ">";
+                return p.rule(rule_name,
+                    p.tool_arg(
+                        p.tool_arg_open(
+                            p.literal(NS + "<") +
+                            p.tool_arg_name(p.literal(tag)) +
+                            p.literal(">")) +
+                        value_of(schema, rule_name, close)));
+            };
+
+            value_of = [&](const json & schema,
+                           const std::string & rule_name,
+                           const std::string & close) -> common_peg_parser {
+                auto close_tag = p.tool_arg_close(p.literal(close));
+
+                // A string accepts anything, so a union with a string alternative is a string
+                if (schema_info.resolves_to_string(schema)) {
+                    return p.ac(p.tool_arg_string_value(p.until(close)) + close_tag, close);
+                }
+
+                if (auto alternatives = alternatives_of(schema)) {
+                    std::vector<common_peg_parser> choices;
+
+                    size_t index = 0;
+                    for (const auto & alternative : *alternatives) {
+                        const std::string alt_name = rule_name + "-" + std::to_string(index++);
+
+                        // There is a risk that this breaks streaming deltas, but that's a risk we
+                        // assume to provide tool arg streaming.
+                        choices.push_back(value_of(alternative, alt_name, close));
+                    }
+
+                    return p.choice(choices);
+                }
+
+                const std::string type = schema.contains("type") && schema.at("type").is_string()
+                                             ? schema.at("type").get<std::string>()
+                                             : "";
+
+                if (type == "object" && schema.contains("properties")) {
+                    return p.tag(mm3::TOOL_ARG_OBJECT, members_of(schema, rule_name)) + p.space() + close_tag;
+                }
+
+                if (type == "array" && schema.contains("items")) {
+                    const std::string item_close = NS + "</item>";
+                    auto item = p.rule(rule_name + "-item",
+                        p.tag(mm3::TOOL_ARG_ITEM,
+                              p.literal(NS + "<item>") +
+                                  value_of(schema.at("items"), rule_name + "-item", item_close)));
+                    return p.tag(mm3::TOOL_ARG_ARRAY, p.repeat(p.space() + item, 0, -1)) + p.space() + close_tag;
+                }
+
+                return p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", schema, false)) + close_tag;
+            };
+
+            // Required properties in schema order, then any number of optional ones in any order.
+            members_of = [&](const json & schema, const std::string & rule_prefix) -> common_peg_parser {
+                const auto & props = schema.at("properties");
+
+                std::set<std::string> required;
+                if (schema.contains("required")) {
+                    schema.at("required").get_to(required);
+                }
+
+                std::vector<common_peg_parser> required_elements;
+                std::vector<common_peg_parser> optional_elements;
+                for (const auto & [key, key_schema] : props.items()) {
+                    auto element = element_of(key, key_schema, rule_prefix + "-" + key);
+                    if (required.find(key) != required.end()) {
+                        required_elements.push_back(element);
+                    } else {
+                        optional_elements.push_back(element);
+                    }
+                }
+
+                common_peg_parser members = p.eps();
+                for (size_t i = 0; i < required_elements.size(); i++) {
+                    if (i > 0) {
+                        members = members + p.space();
+                    }
+                    members = members + required_elements[i];
+                }
+
+                if (!optional_elements.empty()) {
+                    common_peg_parser any_optional = p.choice();
+                    for (const auto & element : optional_elements) {
+                        any_optional |= element;
+                    }
+                    members = members + p.repeat(p.space() + any_optional, 0, -1);
+                }
+
+                return members;
+            };
+
+            common_peg_parser invoke_body =
+                params.contains("properties") ? members_of(params, "tool-" + name + "-arg") : p.eps();
+
+            auto func_parser = p.tool(
+                p.tool_open(p.literal(NS + "<invoke name=\"") +
+                            p.tool_name(p.literal(name)) + p.literal("\">")) +
+                p.space() + invoke_body + p.space() +
+                p.tool_close(p.literal(INVOKE_END)));
+
+            tool_choice |= p.rule("tool-" + name, func_parser);
+        });
+
+        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+        common_peg_parser tool_calls = p.eps();
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice +
+                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
+        } else {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
+        }
+
+        if (!require_tools) {
+            tool_calls = p.optional(tool_calls);
+        }
+
+        auto content_before_tools = p.content(p.until(FC_START));
+        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, FC_START },
+        };
+    }
+
+    return data;
+}
+
 namespace workaround {
 
 static void map_developer_role_to_system(json & messages) {
@@ -2707,6 +2967,15 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_gigachat_v3(tmpl, params);
     }
 
+    // MiniMax-M3: the namespace token "]<]minimax[>[" collides with the autoparser's
+    // markup delimiters, so detect the template and use a dedicated parser.
+    if (src.find("]<]minimax[>[") != std::string::npos &&
+        src.find("<tool_call>") != std::string::npos &&
+        src.find("<invoke name=") != std::string::npos) {
+        LOG_DBG("Using specialized template: MiniMax-M3\n");
+        return common_chat_params_init_minimax_m3(tmpl, params);
+    }
+
     // DeepSeek V3.2/V4 format detection: template defines dsml_token and uses it for tool calls.
     // The template source contains the token as a variable assignment, not as a literal in markup.
     // V3.2 names the tool call block "function_calls", V4 names it "tool_calls".
@@ -2998,6 +3267,8 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
             std::unique_ptr<common_chat_peg_mapper> mapper;
             if (params.format == COMMON_CHAT_FORMAT_PEG_GEMMA4) {
                 mapper = std::make_unique<common_chat_peg_gemma4_mapper>(msg);
+            } else if (params.format == COMMON_CHAT_FORMAT_PEG_MINIMAX_M3) {
+                mapper = std::make_unique<common_chat_peg_minimax_m3_mapper>(msg);
             } else {
                 mapper = std::make_unique<common_chat_peg_mapper>(msg);
             }
@@ -3020,6 +3291,8 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
     std::unique_ptr<common_chat_peg_mapper> mapper;
     if (params.format == COMMON_CHAT_FORMAT_PEG_GEMMA4) {
         mapper = std::make_unique<common_chat_peg_gemma4_mapper>(msg);
+    } else if (params.format == COMMON_CHAT_FORMAT_PEG_MINIMAX_M3) {
+        mapper = std::make_unique<common_chat_peg_minimax_m3_mapper>(msg);
     } else {
         mapper = std::make_unique<common_chat_peg_mapper>(msg);
     }

common/chat.h

@@ -233,6 +233,7 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_PEG_SIMPLE,
     COMMON_CHAT_FORMAT_PEG_NATIVE,
     COMMON_CHAT_FORMAT_PEG_GEMMA4,
+    COMMON_CHAT_FORMAT_PEG_MINIMAX_M3,
 
     COMMON_CHAT_FORMAT_COUNT,  // Not a format, just the # formats
 };

common/common.cpp

@@ -1518,23 +1518,49 @@ done:
     return res;
 }
 
-void common_context_seq_rm(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+static void common_context_seq_rm(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
     auto * mem = llama_get_memory(ctx);
     if (!llama_memory_seq_rm(mem, seq_id, p0, p1)) {
         GGML_ABORT("%s", string_format("failed to remove sequence %d with p0=%d, p1=%d\n", seq_id, p0, p1).c_str());
     }
 }
 
-void common_context_seq_cp(llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+static void common_context_seq_cp(llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     auto * mem = llama_get_memory(ctx);
     llama_memory_seq_cp(mem, seq_id_src, seq_id_dst, p0, p1);
 }
 
-void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
+static void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
     auto * mem = llama_get_memory(ctx);
     llama_memory_seq_add(mem, seq_id, p0, p1, delta);
 }
 
+void common_memory::init(llama_context * ctx_tgt, llama_context * ctx_dft) {
+    this->ctx_tgt = ctx_tgt;
+    this->ctx_dft = ctx_dft;
+}
+
+void common_memory::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const {
+    common_context_seq_rm(ctx_tgt, seq_id, p0, p1);
+    if (ctx_dft) {
+        common_context_seq_rm(ctx_dft, seq_id, p0, p1);
+    }
+}
+
+void common_memory::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) const {
+    common_context_seq_cp(ctx_tgt, seq_id_src, seq_id_dst, p0, p1);
+    if (ctx_dft) {
+        common_context_seq_cp(ctx_dft, seq_id_src, seq_id_dst, p0, p1);
+    }
+}
+
+void common_memory::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) const {
+    common_context_seq_add(ctx_tgt, seq_id, p0, p1, delta);
+    if (ctx_dft) {
+        common_context_seq_add(ctx_dft, seq_id, p0, p1, delta);
+    }
+}
+
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     std::vector<llama_adapter_lora *> loras;
     std::vector<float> scales;

common/common.h

@@ -173,6 +173,7 @@ enum common_speculative_type {
     COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3,  // Eagle3 speculative decoding
     COMMON_SPECULATIVE_TYPE_DRAFT_MTP,     // Multi-token prediction
     COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH,  // DFlash speculative decoding
+    COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK,  // DSpark speculative decoding (DFlash + Markov head)
     COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,  // simple self-speculative decoding based on n-grams
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,   // self-speculative decoding with n-gram keys only
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values
@@ -388,7 +389,7 @@ struct common_params_speculative {
 
     uint32_t need_n_rs_seq() const {
         bool needs_rs_seq = std::any_of(types.begin(), types.end(), [&](auto t) {
-            return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH;
+            return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH || t == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK;
         });
 
         return needs_rs_seq ? draft.n_max : 0u;
@@ -948,10 +949,17 @@ enum common_context_seq_rm_type {
 // note: clears the memory of the context
 common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx);
 
-// aborts execution on failure
-void common_context_seq_rm (llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1);
-void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta);
-void common_context_seq_cp (llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1);
+struct common_memory {
+    llama_context * ctx_tgt = nullptr;
+    llama_context * ctx_dft = nullptr;
+
+    void init(llama_context * ctx_tgt, llama_context * ctx_dft = nullptr);
+
+    // aborts execution on failure
+    void seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) const;
+    void seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) const;
+    void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) const;
+};
 
 //
 // Batch utils

common/fit.cpp

@@ -136,7 +136,7 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
         devs.push_back(llama_model_get_device(model, i));
     }
 
-    hp_ngl         = llama_model_n_layer(model);
+    hp_ngl         = llama_model_n_layer(model) + llama_model_n_layer_nextn(model);
     hp_n_ctx_train = llama_model_n_ctx_train(model);
     hp_n_expert    = llama_model_n_expert(model);
 

common/speculative.cpp

@@ -34,6 +34,7 @@ const std::map<std::string, common_speculative_type> common_speculative_type_fro
     {"draft-eagle3",  COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3},
     {"draft-mtp",     COMMON_SPECULATIVE_TYPE_DRAFT_MTP},
     {"draft-dflash",  COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH},
+    {"draft-dspark",  COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK},
     {"ngram-simple",  COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE},
     {"ngram-map-k",   COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K},
     {"ngram-map-k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V},
@@ -437,6 +438,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
     int32_t n_embd_dec = 0;       // draft hidden size
     int32_t n_embd_enc = 0;       // target_layer_ids_n * target_hidden_size
     int32_t n_embd_tgt = 0;       // target model hidden size
+    int32_t n_layer_tgt = 0;      // target model layer count
 
     const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
     uint32_t        target_layer_ids_n = 0;
@@ -478,6 +480,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
         n_embd_tgt = llama_model_n_embd(model_tgt);
         n_embd_dec = llama_model_n_embd(model_dft);
         n_embd_enc = (int32_t) target_layer_ids_n * n_embd_tgt;
+        n_layer_tgt = llama_model_n_layer(model_tgt);
 
         const int32_t n_b = (int32_t) llama_n_batch(ctx_dft);
         batch = llama_batch_init(/*n_tokens=*/ n_b, /*embd=*/ n_embd_dec, /*n_seq_max=*/ 1);
@@ -510,9 +513,15 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
             }
         }
 
-        // turn on extraction of the target layers' input embeddings
+        // turn on extraction of the target layers' hidden states
         for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
-            llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
+            if (target_layer_ids[k] < n_layer_tgt) {
+                llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
+            } else if (target_layer_ids[k] == n_layer_tgt) {
+                llama_set_embeddings_nextn(ctx_tgt, true, /*masked*/ false);
+            } else {
+                GGML_ABORT("EAGLE3: target layer id %d exceeds target n_layer %d", target_layer_ids[k], n_layer_tgt);
+            }
         }
 
         // turn on extraction of the draft model's pre-norm hidden state
@@ -600,7 +609,9 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
         features_buf.resize((size_t) n_tokens * n_embd_enc, 0.0f);
 
         for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
-            const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]);
+            const float * layer = target_layer_ids[k] < n_layer_tgt
+                ? llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k])
+                : llama_get_embeddings_nextn(ctx_tgt);
             if (!layer) {
                 GGML_ABORT("EAGLE3: target layer %d input not extracted.", target_layer_ids[k]);
             }
@@ -918,15 +929,20 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
     int32_t     block_size    = 0;
     llama_token mask_token_id = 0;
 
+    // draft-dspark: the draft carries a Markov head and uses an anchor-first block layout
+    const bool is_dspark;
+
     const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
     uint32_t        target_layer_ids_n = 0;
 
     // scratch buffer for concatenated target features [n_tokens, n_embd_enc]
     std::vector<float> features_buf;
 
-    common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, n_seq)
+    common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq,
+            common_speculative_type type = COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)
+        : common_speculative_impl(type, n_seq)
         , params(params.draft)
+        , is_dspark(type == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK)
     {
         auto * ctx_tgt = this->params.ctx_tgt;
         auto * ctx_dft = this->params.ctx_dft;
@@ -953,16 +969,18 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
         }
         mask_token_id = llama_vocab_mask(llama_model_get_vocab(model_dft));
 
-        LOG_INF("%s: adding speculative implementation 'draft-dflash'\n", __func__);
+        LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(type).c_str());
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min);
         LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u\n", __func__, block_size, mask_token_id, target_layer_ids_n);
 
-        // DFlash input is [id_last, <mask> * (block_size-1)], so it can draft at most block_size-1 tokens per step
-        if (this->params.n_max > block_size - 1 || this->params.n_min > block_size - 1) {
-            LOG_WRN("%s: requested draft size (n_max=%d, n_min=%d) exceeds the trained DFlash block size %d -- clamping to %d\n",
-                    __func__, this->params.n_max, this->params.n_min, block_size, block_size - 1);
-            this->params.n_max = std::min(this->params.n_max, block_size - 1);
-            this->params.n_min = std::min(this->params.n_min, block_size - 1);
+        // DFlash input is [id_last, <mask> * (block_size-1)]: in-place denoising yields at most
+        // block_size-1 draft tokens, DSpark yield a full block_size draft tokens
+        const int32_t n_draft_max = is_dspark ? block_size : block_size - 1;
+        if (this->params.n_max > n_draft_max || this->params.n_min > n_draft_max) {
+            LOG_WRN("%s: requested draft size (n_max=%d, n_min=%d) exceeds the trained block size %d -- clamping to %d\n",
+                    __func__, this->params.n_max, this->params.n_min, block_size, n_draft_max);
+            this->params.n_max = std::min(this->params.n_max, n_draft_max);
+            this->params.n_min = std::min(this->params.n_min, n_draft_max);
         }
 
         batch        = llama_batch_init(llama_n_batch(ctx_dft), 0,          n_seq);
@@ -1126,12 +1144,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
             const int32_t n = (int32_t) dp.n_past;
 
-            int32_t n_draft = params.n_max;
-            if (dp.n_max > 0) {
-                n_draft = std::min(n_draft, dp.n_max);
-            }
+            const int32_t n_draft = params.n_max;
 
-            const int32_t n_block_tokens = n_draft + 1; // id_last + n_draft * <mask>
+            const int32_t n_block_tokens = n_draft + (is_dspark ? 0 : 1);
             i_block_beg[seq_id] = batch.n_tokens;
             n_block    [seq_id] = n_block_tokens;
             for (int32_t i = 0; i < n_block_tokens; ++i) {
@@ -1163,27 +1178,57 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
             auto & result = *dp.result;
 
-            // greedily read the predicted block at this sequence's noise positions 1..n_block_tokens-1
-            for (int32_t i = 1; i < n_block_tokens; ++i) {
-                common_sampler_sample(smpl, ctx_dft, beg + i, true);
+            if (is_dspark) {
+                // DSpark predicts the next token from position 0 and optionally truncates
+                // at the first position below the confidence threshold.
+                const float * conf = params.p_min > 0.0f ? llama_get_embeddings_nextn(ctx_dft) : nullptr;
 
-                const auto * cur_p = common_sampler_get_candidates(smpl, true);
+                for (int32_t i = 0; i < n_block_tokens; ++i) {
+                    const int32_t idx = beg + i;
 
-                for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
-                    LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
-                            seq_id, k, i - 1, cur_p->data[k].id, cur_p->data[k].p,
-                            common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+                    if (conf && conf[(size_t) idx * n_embd_dec] < params.p_min) {
+                        break;
+                    }
+
+                    common_sampler_sample(smpl, ctx_dft, idx, true);
+
+                    const auto * cur_p = common_sampler_get_candidates(smpl, true);
+
+                    for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
+                        LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                                seq_id, k, i, cur_p->data[k].id, cur_p->data[k].p,
+                                common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+                    }
+
+                    const llama_token id = cur_p->data[0].id;
+
+                    common_sampler_accept(smpl, id, true);
+
+                    result.push_back(id);
                 }
+            } else {
+                // greedily read the predicted block at this sequence's noise positions 1..n_block_tokens-1
+                for (int32_t i = 1; i < n_block_tokens; ++i) {
+                    common_sampler_sample(smpl, ctx_dft, beg + i, true);
 
-                const llama_token id = cur_p->data[0].id;
+                    const auto * cur_p = common_sampler_get_candidates(smpl, true);
 
-                if (cur_p->data[0].p < params.p_min) {
-                    break;
+                    for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
+                        LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                                seq_id, k, i - 1, cur_p->data[k].id, cur_p->data[k].p,
+                                common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+                    }
+
+                    const llama_token id = cur_p->data[0].id;
+
+                    if (cur_p->data[0].p < params.p_min) {
+                        break;
+                    }
+
+                    common_sampler_accept(smpl, id, true);
+
+                    result.push_back(id);
                 }
-
-                common_sampler_accept(smpl, id, true);
-
-                result.push_back(id);
             }
 
             if (result.size() < (size_t) params.n_min) {
@@ -2145,6 +2190,7 @@ std::string common_speculative_type_to_str(common_speculative_type type) {
         case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3:  return "draft-eagle3";
         case COMMON_SPECULATIVE_TYPE_DRAFT_MTP:     return "draft-mtp";
         case COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH:  return "draft-dflash";
+        case COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK:  return "draft-dspark";
         case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:  return "ngram-simple";
         case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:   return "ngram-map-k";
         case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram-map-k4v";
@@ -2198,6 +2244,7 @@ int32_t common_speculative_n_max(const common_params_speculative * spec) {
             case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3:
             case COMMON_SPECULATIVE_TYPE_DRAFT_MTP:
             case COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH:
+            case COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK:
                 n_max = std::max(n_max, std::max(0, spec->draft.n_max));
                 break;
             case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:
@@ -2342,6 +2389,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
         bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr;
         bool has_draft_mtp    = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP))    && params.draft.ctx_dft != nullptr;
         bool has_draft_dflash = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)) && params.draft.ctx_dft != nullptr;
+        bool has_draft_dspark = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK)) && params.draft.ctx_dft != nullptr;
 
 
 
@@ -2352,7 +2400,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
         bool has_ngram_mod     = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MOD));
 
         // when adding a new type - update here the logic above
-        static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 10);
+        static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 11);
 
         // this list here defines the priority of the speculators
         // the one with highest priority are listed first
@@ -2385,6 +2433,9 @@ common_speculative * common_speculative_init(common_params_speculative & params,
         if (has_draft_dflash) {
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, params));
         }
+        if (has_draft_dspark) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK, params));
+        }
     }
 
     std::vector<std::unique_ptr<common_speculative_impl>> impls = {};
@@ -2409,6 +2460,11 @@ common_speculative * common_speculative_init(common_params_speculative & params,
                 impls.push_back(std::make_unique<common_speculative_impl_draft_dflash>(config.params, n_seq));
                 break;
             }
+            case COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK: {
+                impls.push_back(std::make_unique<common_speculative_impl_draft_dflash>(
+                        config.params, n_seq, COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK));
+                break;
+            }
             case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: {
                 common_ngram_map ngram_map = get_common_ngram_map(config.type, config.params.ngram_simple);
 

conversion/__init__.py

@@ -53,6 +53,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "DeepseekV3ForCausalLM": "deepseek",
     "DeepseekV32ForCausalLM": "deepseek",
     "DFlashDraftModel": "qwen",
+    "Qwen3DSparkModel": "qwen",
     "DeepseekV4ForCausalLM": "deepseek",
     "DistilBertForMaskedLM": "bert",
     "DistilBertForSequenceClassification": "bert",
@@ -167,6 +168,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "ModernBertForMaskedLM": "bert",
     "ModernBertForSequenceClassification": "bert",
     "ModernBertModel": "bert",
+    "NanbeigeForCausalLM": "nanbeige",
     "NemotronForCausalLM": "nemotron",
     "NemotronHForCausalLM": "nemotron",
     "NeoBERT": "bert",

conversion/llama.py

@@ -69,9 +69,14 @@ class LlamaModel(TextModel):
                 target_config = {**target_config, **target_config["text_config"]}
             self.target_vocab_size = target_config["vocab_size"]
 
-            # target_layers: derived from target model layer count (low/mid/high)
+            # target_layers: use the eagle3 config's explicit aux hidden-state layer ids
+            # if present, else derive from the target layer count.
             target_num_layers = target_config["num_hidden_layers"]
-            target_layers = [2, target_num_layers // 2, target_num_layers - 3]
+            aux_layer_ids = eagle3_raw_config.get("eagle_aux_hidden_state_layer_ids")
+            if aux_layer_ids:
+                target_layers = aux_layer_ids
+            else:
+                target_layers = [2, target_num_layers // 2, target_num_layers - 3]
             logger.info(f"EAGLE-3: target_layers = {target_layers} (target model has {target_num_layers} layers)")
             self.gguf_writer.add_target_layers(target_layers)
 
@@ -90,6 +95,12 @@ class LlamaModel(TextModel):
             logger.info(f"EAGLE-3: norm_before_residual = {norm_before_residual}")
             self.gguf_writer.add_norm_before_residual(norm_before_residual)
 
+            # norm_before_fc: RMSNorm applied to the fused target features before the
+            # fc projection (e.g. nvidia/gpt-oss-120b-Eagle3-v3)
+            norm_before_fc = eagle3_raw_config.get("norm_before_fc", False)
+            logger.info(f"EAGLE-3: norm_before_fc = {norm_before_fc}")
+            self.gguf_writer.add_norm_before_fc(norm_before_fc)
+
     def set_vocab(self):
         # eagle3: use tokenizer from target model if provided
         original_dir_model = None
@@ -222,6 +233,9 @@ class LlamaModel(TextModel):
             if name == "fc.weight":
                 yield (name, data_torch)
                 return
+            if name == "input_norm.weight":
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ENC_OUTPUT_NORM), data_torch)
+                return
             if name == "d2t":
                 # store for manual int64 handling in prepare_tensors (avoid F32 conversion)
                 if not hasattr(self, '_eagle3_int_tensors'):

conversion/mimo.py

@@ -1,8 +1,9 @@
 from __future__ import annotations
 
+import json
 import re
 
-from typing import Callable, TYPE_CHECKING
+from typing import Any, Callable, Iterable, TYPE_CHECKING
 
 import torch
 
@@ -229,7 +230,13 @@ class MimoV2Model(TextModel):
 
 
 @ModelBase.register("MiMoV2ForCausalLM")
-class MiMoV2VisionModel(MmprojModel):
+class MiMoV2VisionAudioModel(MmprojModel):
+    has_audio_encoder = True
+
+    _audio_tok_hparams: dict[str, Any] | None = None
+    _rvq_codebook_sizes: list[int] | None = None
+    _code_embd: dict[int, Tensor] | None = None
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         assert self.hparams_vision is not None
@@ -253,10 +260,22 @@ class MiMoV2VisionModel(MmprojModel):
         self.visual_token_window_size = int(hp.get("visual_token_window_size", -1))
         self.use_sink = bool(hp.get("use_sink", False))
 
+    def get_audio_config(self) -> dict[str, Any] | None:
+        if self._audio_tok_hparams is None:
+            path = self.dir_model / "audio_tokenizer" / "config.json"
+            with open(path, "r", encoding="utf-8") as f:
+                cfg = json.load(f)
+            # aliases so MmprojModel.find_aparam() / n_block_keys can resolve them
+            cfg["hidden_size"] = cfg["d_model"]
+            cfg["intermediate_size"] = cfg["encoder_ffn_dim"]
+            cfg["num_attention_heads"] = cfg["encoder_attention_heads"]
+            self._audio_tok_hparams = cfg
+        return self._audio_tok_hparams
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MIMOVL)
+        self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.MIMOVL)
         self.gguf_writer.add_vision_use_silu(True)
         self.gguf_writer.add_vision_head_count_kv(self.num_kv_heads)
         self.gguf_writer.add_vision_spatial_merge_size(self.spatial_merge_size)
@@ -266,19 +285,45 @@ class MiMoV2VisionModel(MmprojModel):
         self.gguf_writer.add_vision_min_pixels(int(self.preprocessor_config["min_pixels"]))
         self.gguf_writer.add_vision_max_pixels(int(self.preprocessor_config["max_pixels"]))
 
+        assert self.hparams_audio is not None
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.MIMO_AUDIO)
+        self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["n_mels"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-5))
+
+        assert self._rvq_codebook_sizes is not None
+        self.gguf_writer.add_audio_rvq_num_quantizers(len(self._rvq_codebook_sizes))
+        self.gguf_writer.add_audio_rvq_codebook_size(self._rvq_codebook_sizes)
+
+        n_layer = self.hparams_audio["encoder_layers"]
+        swa_per_block = self.hparams_audio.get("swa_per_block", 1)
+        if self.hparams_audio.get("hybrid_attention") and swa_per_block > 1:
+            wa_pattern = [0 if i % swa_per_block < swa_per_block - 1 else -1 for i in range(n_layer)]
+        else:
+            wa_pattern = [-1] * n_layer
+        self.gguf_writer.add_audio_wa_pattern_mode(wa_pattern)
+        self.gguf_writer.add_audio_window_size(int(self.hparams_audio["encoder_attn_window_size"][0]))
+
+        audio_cfg = self.global_config["audio_config"]
+        self.gguf_writer.add_audio_local_block_count(int(audio_cfg["input_local_layers"]))
+        self.gguf_writer.add_audio_local_group_size(int(audio_cfg["group_size"]))
+
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        # Sinks must be F32: any sink-style softmax/mask add in ggml requires
-        # F32, and we fold sinks into a host-built F32 mask at encode time.
-        if new_name.endswith(".attn_sinks"):
+        # for audio encoder: keep codebook in F32
+        if new_name in (
+            gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_ENC_RVQ_CODEBOOK] + ".weight",
+            gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_MM_CODE_EMBD] + ".weight",
+        ):
+            return gguf.GGMLQuantizationType.F32
+        if ("encoder.conv" in name or "encoder.down_sample_layer" in name) and name.endswith(".weight"):
             return gguf.GGMLQuantizationType.F32
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, _ = item
-        if not name.startswith("visual."):
-            return None
-        return super().filter_tensors(item)
+        if name.startswith("visual.") or name.startswith("speech_embeddings.") or name.startswith("audio_encoder."):
+            return super().filter_tensors(item)
+        return None
 
     def modify_tensors(self, data_torch, name, bid):
         # Conv3D patch embed: split along the temporal axis (kt=2) into two Conv2D
@@ -292,4 +337,64 @@ class MiMoV2VisionModel(MmprojModel):
             yield (embd_name + ".weight.1", data_torch[:, :, 1, ...])
             return
 
+        if m := re.match(r"^speech_embeddings\.(\d+)\.weight$", name):
+            if self._code_embd is None:
+                self._code_embd = {}
+            self._code_embd[int(m.group(1))] = data_torch
+
+            n_channels = int(self.global_config["audio_config"]["audio_channels"])
+            if len(self._code_embd) < n_channels:
+                return
+            merged = torch.stack([self._code_embd.pop(i) for i in range(n_channels)], dim=0)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MM_CODE_EMBD), merged)
+            return
+
+        if "conv1.bias" in name or "conv2.bias" in name:
+            # transpose conv1/conv2 bias so it broadcasts against [n_frames, C_out, 1]
+            data_torch = data_torch.unsqueeze(-1)
+
+        if name == "audio_encoder.projection.mlp.0.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 1), data_torch)
+            return
+        if name == "audio_encoder.projection.mlp.2.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 2), data_torch)
+            return
+
         yield from super().modify_tensors(data_torch, name, bid)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # note: audio encoder is in its own subdir "audio_tokenizer"
+        from safetensors.torch import load_file
+
+        tok_dir = self.dir_model / "audio_tokenizer"
+        state_dict = load_file(tok_dir / "model.safetensors")
+
+        codebook_re = re.compile(r"^encoder\.quantizer\.vq\.layers\.(\d+)\._codebook\.embed$")
+        codebooks: dict[int, Tensor] = {}
+
+        # EMA/training-only RVQ buffers - not needed for inference (nearest-codebook
+        # lookup only reads "_codebook.embed")
+        skip_suffixes = (
+            "_codebook.cluster_size",
+            "_codebook.embed_avg",
+            "_codebook.inited",
+        )
+        for name, tensor in state_dict.items():
+            if name.endswith(skip_suffixes):
+                continue
+            if m := codebook_re.match(name):
+                codebooks[int(m.group(1))] = tensor
+                continue
+            yield name, tensor
+
+        # gather codebooks and merge into 3D tensor, similar to MoE MLP tensors
+        n_q = len(codebooks)
+        ordered = [codebooks[i] for i in range(n_q)]
+        self._rvq_codebook_sizes = [int(cb.shape[0]) for cb in ordered]
+        max_bins = max(self._rvq_codebook_sizes)
+        dim = ordered[0].shape[1]
+        merged = ordered[0].new_zeros(n_q, max_bins, dim)
+        for i, cb in enumerate(ordered):
+            merged[i, : cb.shape[0], :] = cb
+
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_ENC_RVQ_CODEBOOK), merged)

conversion/nanbeige.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from .base import ModelBase, gguf, logger
+from .llama import LlamaModel
+
+
+@ModelBase.register("NanbeigeForCausalLM")
+class NanbeigeModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.NANBEIGE
+    undo_permute = True
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        n_loops = int(hparams.get("num_loops", 1) or 1)
+        if n_loops < 1:
+            n_loops = 1
+        self.gguf_writer.add_num_loops(n_loops)
+        logger.info(f"gguf: num_loops = {n_loops}")
+
+        skip_loop_final_norm = bool(hparams.get("skip_loop_final_norm", False))
+        self.gguf_writer.add_skip_loop_final_norm(skip_loop_final_norm)
+        logger.info(f"gguf: skip_loop_final_norm = {skip_loop_final_norm}")

conversion/nemotron.py

@@ -39,28 +39,48 @@ class NemotronNanoV2VLModel(MmprojModel):
         }
         return vision_config
 
+    def get_audio_config(self) -> dict[str, Any] | None:
+        return self.global_config.get("sound_config")
+
     def set_gguf_parameters(self):
         if "image_mean" not in self.preprocessor_config:
             self.preprocessor_config["image_mean"] = [0.485, 0.456, 0.406]
         if "image_std" not in self.preprocessor_config:
             self.preprocessor_config["image_std"] = [0.229, 0.224, 0.225]
 
+        if self.hparams_audio is not None:
+            self.has_vision_encoder = True
+            self.has_audio_encoder = True
+            self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["num_mel_bins"])
+            self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+            self.gguf_writer.add_audio_subsampling_factor(self.hparams_audio["subsampling_factor"])
+            self.gguf_writer.add_audio_conv_kernel_size(self.hparams_audio["conv_kernel_size"])
+            self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.PARAKEET)
+            self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
+        else:
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
+
         super().set_gguf_parameters()
         hparams = self.global_config
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
         self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
         self.gguf_writer.add_vision_use_gelu(True)
         downsample_ratio = hparams.get("downsample_ratio", 0.5)
         self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        if ".position_embd." in new_name or "pos_embed" in new_name:
-            return gguf.GGMLQuantizationType.F32
+        if "sound_encoder" in name or new_name.startswith("mm.a."):
+            if "bias" in new_name or "norm" in new_name:
+                return gguf.GGMLQuantizationType.F32
+            if "conv" in new_name and "weight" in new_name:
+                return gguf.GGMLQuantizationType.F32
+
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
 
         if "input_conditioner" in name:
             return None
@@ -69,14 +89,18 @@ class NemotronNanoV2VLModel(MmprojModel):
         if "radio_model.model.patch_generator.video_embedder" in name:
             return None
 
-        if not name.startswith("vision_model.radio_model.model.") and not name.startswith("mlp1."):
+        if not name.startswith(("vision_model.radio_model.model.", "mlp1.", "sound_encoder.", "sound_projection.")):
             return None
 
         if "patch_generator.pos_embed" in name:
             if not name.endswith(".weight"):
                 name += ".weight"
 
-        return super().filter_tensors((name, gen))
+        # num_batches is only used for training not inference.
+        if "conv.norm" in name and "num_batches" in name:
+            return None
+
+        return name, gen
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # RADIO's pos_embed doesn't have .weight suffix, but clip.cpp expects it
@@ -104,7 +128,26 @@ class NemotronNanoV2VLModel(MmprojModel):
             n_embd = self.hparams["hidden_size"]
             data_torch = data_torch.reshape(n_embd, 3, patch_size, patch_size)
 
-        yield from super().modify_tensors(data_torch, name, bid)
+        if "depthwise_conv.weight" in name:
+            data_torch = data_torch.unsqueeze(-1)
+            data_torch = data_torch.permute(3, 1, 0, 2).contiguous()
+
+        if "pointwise_conv" in name and name.endswith(".weight"):
+            if len(data_torch.shape) == 3 and data_torch.shape[2] == 1:
+                data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[1])
+
+        if "subsampling.layers" in name and name.endswith(".bias"):
+            if len(data_torch.shape) == 1:
+                data_torch = data_torch.reshape(1, -1, 1, 1)
+
+        if "pointwise_conv" in name and name.endswith(".bias"):
+            if len(data_torch.shape) == 1:
+                data_torch = data_torch.reshape(1, -1, 1, 1)
+
+        for mapped_name, tensor in super().modify_tensors(data_torch, name, bid):
+            if name.startswith("sound_projection.") and mapped_name.startswith("mm.model.mlp."):
+                mapped_name = mapped_name.replace("mm.model.mlp.", "mm.a.mlp.")
+            yield mapped_name, tensor
 
 
 @ModelBase.register("NemotronForCausalLM")

conversion/qwen.py

@@ -688,3 +688,23 @@ class DFlashModel(Qwen3Model):
         if not name.startswith("model."):
             name = "model." + name
         return super().filter_tensors((name, gen))
+
+
+@ModelBase.register("Qwen3DSparkModel")
+class DSparkModel(DFlashModel):
+    # DSpark = DFlash + a semi-autoregressive Markov head
+    model_arch = gguf.MODEL_ARCH.DFLASH
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # normalize the flat DeepSpec schema to DFlash's nested dflash_config
+        self.hparams.setdefault("dflash_config", {
+            k: self.hparams[k] for k in ("target_layer_ids", "mask_token_id") if k in self.hparams
+        })
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if name.endswith(("embed_tokens.weight", "lm_head.weight")):
+            return None
+        return super().filter_tensors((name, gen))

docs/backend/SYCL.md

@@ -794,6 +794,8 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_USE_LEVEL_ZERO_API | 1 (default) or 0 | Use Level Zero API for device memory allocation instead of SYCL. Reduces system RAM usage on Intel dGPUs by avoiding DMA-buf/TTM host memory staging. Requires GGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON at build time. SYCL backend always runs on Level Zero running time even if it's set as OFF (The SYCL api will be usage for memory allocation).|
 | GGML_SYCL_ENABLE_DNN | 0 or 1 (default)| Enable running computations through oneDNN and always use oneMKL. |
+| GGML_SYCL_FA_ONEDNN | 1 (default) or 0 | Enable the oneDNN fused SDPA (flash-attention) path on supported GPUs. Set to 0 to always use the native SYCL flash-attention kernel. |
+| GGML_SYCL_FA_ONEDNN_MAX_KV | 0 (default, disabled) or positive integer | By default (0), all sequences are handled by the oneDNN fused SDPA path, regardless of KV length; a positive value caps that length, past which sequences fall back to the native kernel. If GPU driver watchdog resets (DEVICE_LOST) occur during long-context inference, set this near the context depth where they start, e.g. 24576. |
 | GGML_SYCL_ENABLE_VMM | 0 or 1 (default) | Enable the virtual-memory device pool. |
 | GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute (currently top-k MoE gating). |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |

docs/install.md

@@ -16,22 +16,22 @@ conda-forge provides builds for:
  - Apple Metal (macOS)
 
 ```sh
-conda install -c conda-forge llama-cpp
+conda install -c conda-forge llama.cpp
 ```
 
 ```sh
-mamba install -c conda-forge llama-cpp
+mamba install -c conda-forge llama.cpp
 ```
 
 ```sh
 # Project-local installation
-pixi add llama-cpp
+pixi add llama.cpp
 
 # Global installation
-pixi global install llama-cpp
+pixi global install llama.cpp
 ```
 
-This distribution is managed on [`conda-forge/llama-cpp-feedstock`](https://github.com/conda-forge/llama.cpp-feedstock/).
+This distribution is managed on [`conda-forge/llama.cpp-feedstock`](https://github.com/conda-forge/llama.cpp-feedstock/).
 
 Shall you have any problems, please open an issue on [its issue tracker](https://github.com/conda-forge/llama.cpp-feedstock/issues).
 

docs/speculative.md

@@ -78,6 +78,38 @@ See:
 
 - #22105
 
+### DSpark (`draft-dspark`)
+
+DSpark extends DFlash with a semi-autoregressive _Markov head_: the draft still emits a whole
+block per forward pass, but each block position's logits are biased by a low-rank term keyed on
+the previous token, chained in-graph across the block. This keeps drafting at one decode per
+block while recovering some of the left-to-right signal that pure block diffusion loses.
+
+The draft is a small DeepSpec checkpoint trained for a specific target (for example
+[`deepseek-ai/dspark_qwen3_4b_block7`](https://huggingface.co/deepseek-ai/dspark_qwen3_4b_block7)
+for `Qwen/Qwen3-4B`). Convert it with `--target-model-dir` so it inherits the target's tokenizer
+and token embeddings:
+
+```bash
+python convert_hf_to_gguf.py deepseek-ai/dspark_qwen3_4b_block7 \
+    --target-model-dir Qwen/Qwen3-4B --outtype bf16 --outfile Qwen3-4B-DSpark.gguf
+
+llama-server -m Qwen3-4B.gguf -md Qwen3-4B-DSpark.gguf \
+    --spec-type draft-dspark --spec-draft-n-max 7 -fa on --jinja
+```
+
+`--spec-draft-n-max` is clamped to the draft model's trained block size.
+
+`--spec-draft-conf-min P` truncates each drafted block at the first position whose predicted
+acceptance (from the draft's confidence head, if present) falls below `P` (default 0 = disabled).
+
+Currently only drafts with a Qwen3 backbone are supported; support for other backbones
+(e.g. Gemma4) is planned.
+
+See:
+
+- #25173
+
 ### n-gram Cache (`ngram-cache`)
 
 An n-gram is a sequence of n tokens. The n-gram cache implementation maintains statistics about short n-gram sequences.
@@ -173,7 +205,7 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
 ### General Speculative Parameters
 
 ```
---spec-type [none|draft-simple|draft-eagle3|draft-dflash|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
+--spec-type [none|draft-simple|draft-eagle3|draft-dflash|draft-dspark|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
                                         comma-separated list of types of speculative decoding to use
                                         (default: none)
                                         (env: LLAMA_ARG_SPEC_TYPE)
@@ -314,6 +346,7 @@ Specifies a comma-separated list of speculative decoding types to use.
 | `draft-simple` | Use a simple draft model for speculation |
 | `draft-eagle3` | Use an EAGLE-3 draft model that reads the target's hidden states |
 | `draft-dflash` | Use a DFlash block-diffusion draft model that emits a block per step |
+| `draft-dspark` | Use a DSpark draft model (DFlash backbone + semi-autoregressive Markov head) |
 | `draft-mtp` | Use Multi Token Prediction (MTP) heads from the main model |
 | `ngram-cache` | Use n-gram cache lookup |
 | `ngram-simple` | Use simple n-gram pattern matching |

ggml/CMakeLists.txt

@@ -341,9 +341,6 @@ set(GGML_PUBLIC_HEADERS
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
-#if (GGML_METAL)
-#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
-#endif()
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)
 

ggml/src/ggml-backend.cpp

@@ -906,26 +906,35 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
     }
 
     // operations with weights are preferably run on the same backend as the weights
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        const struct ggml_tensor * src = tensor->src[i];
-        if (src == NULL) {
-            continue;
-        }
-        // skip ROPE since the rope freqs tensor is too small to choose a backend based on it
-        // not an ideal solution
-        if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
-            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
-            // check if a backend with higher prio wants to offload the op
-            if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
-                for (int b = 0; b < src_backend_id; b++) {
-                    if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
-                        SET_CAUSE(tensor, "1.off");
-                        return b;
+    // TODO: there are exceptions (see below) - not an ideal solution
+    bool allow = true;
+
+    // skip ROPE since the rope freqs tensor is too small to choose a backend based on it
+    allow = allow && tensor->op != GGML_OP_ROPE;
+
+    // skip FLASH_ATTN_EXT since the sinks tensor is too small to choose a based based on it
+    allow = allow && tensor->op != GGML_OP_FLASH_ATTN_EXT;
+
+    if (allow) {
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            const struct ggml_tensor * src = tensor->src[i];
+            if (src == NULL) {
+                continue;
+            }
+            if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+                int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
+                // check if a backend with higher prio wants to offload the op
+                if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
+                    for (int b = 0; b < src_backend_id; b++) {
+                        if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
+                            SET_CAUSE(tensor, "1.off");
+                            return b;
+                        }
                     }
                 }
+                SET_CAUSE(tensor, "1.wgt%d", i);
+                return src_backend_id;
             }
-            SET_CAUSE(tensor, "1.wgt%d", i);
-            return src_backend_id;
         }
     }
 

ggml/src/ggml-cuda/ssm-scan.cu

@@ -9,6 +9,21 @@ using namespace cub;
 
 #include "ssm-scan.cuh"
 
+
+// Minimum number of tokens to use SSD (State Space Duality) matmul path instead of scan path.
+// For n_tok <= this threshold, the scan kernel is used (lower overhead for short sequences).
+#define SSM_SSD_MIN_TOKENS 128
+
+// prepare_dt kernel dimensions: one block per (head, seq), each block handles DT_MAX_ITEMS items.
+#define SSM_SSD_DT_BLOCK     256
+#define SSM_SSD_DT_MAX_ITEMS  32
+
+// Maximum tokens the SSD path supports, derived from the prepare_dt kernel block capacity.
+#define SSM_SSD_MAX_TOKENS (SSM_SSD_DT_BLOCK * SSM_SSD_DT_MAX_ITEMS)
+
+// Chunk size for chunked SSD. Caps matmul cost at O(chunk^2) per chunk.
+#define SSM_SSD_CHUNK_SIZE 256
+
 // We would like to keep pragma unroll for cases where L_template is not 0,
 // so we suppress the clang transformation warning.
 #ifdef __clang__
@@ -316,6 +331,429 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
     }
 }
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+// ============================================================================
+// SSD (State Space Duality) kernels for Mamba-2 prefill (n_tok > SSM_SSD_MIN_TOKENS)
+//
+// Instead of a sequential scan, SSD reformulates the output as:
+//   Y = (L (.) (C @ B^T)) @ (X * dt)  +  decay * C @ s_init
+// where L is a causal decay mask derived from A and dt.
+//
+// This converts the O(T*N) sequential scan into parallel matmuls.
+// ============================================================================
+// Softplus(dt) and inclusive prefix sum per head using CUB BlockScan.
+// Grid: (n_head, n_seqs)
+template <int BLOCK_SIZE, int MAX_ITEMS>
+__global__ void ssm_ssd_prepare_dt_kernel(
+        const float * __restrict__ dt_raw,
+        float * __restrict__ dt_sp_out,
+        float * __restrict__ cs_out,
+        const int n_head, const int n_tok,
+        const int dt_stride_tok,   // elements between tokens in dt
+        const int dt_stride_seq) { // elements between sequences in dt
+
+    const int h = blockIdx.x;
+    const int s = blockIdx.y;
+
+    const float * dt_seq = dt_raw + s * dt_stride_seq;
+
+    float * dt_sp_seq = dt_sp_out + s * n_tok * n_head;
+    float * cs_seq    = cs_out    + s * n_tok * n_head;
+
+    const int items_per_thread = (n_tok + BLOCK_SIZE - 1) / BLOCK_SIZE;
+
+    // Phase 1: softplus with interleaved distribution (t = i*BLOCK_SIZE + threadIdx.x).
+    // Each warp reads BLOCK_SIZE consecutive tokens, giving coalesced dt_raw loads
+    // (stride n_head between threads vs. items_per_thread*n_head in blocked layout).
+    float local_vals[MAX_ITEMS];
+    for (int i = 0; i < items_per_thread; i++) {
+        const int t = i * BLOCK_SIZE + threadIdx.x;
+        if (t < n_tok) {
+            float val = dt_seq[h + t * dt_stride_tok];
+            float sp = (val <= 20.0f) ? log1pf(expf(val)) : val;
+            local_vals[i] = sp;
+            dt_sp_seq[t * n_head + h] = sp;
+        } else {
+            local_vals[i] = 0.0f;
+        }
+    }
+
+    // Phase 2+3: per-step inclusive scan to build cs[] in token order.
+    // With interleaved distribution the per-thread total scan would not give token-order
+    // prefix sums, so we scan one BLOCK_SIZE slab at a time and carry a running total.
+#ifdef USE_CUB
+    using BlockScan = cub::BlockScan<float, BLOCK_SIZE>;
+    __shared__ typename BlockScan::TempStorage scan_temp;
+    __shared__ float step_total;
+
+    float running = 0.0f;
+    for (int i = 0; i < items_per_thread; i++) {
+        float inclusive;
+        BlockScan(scan_temp).InclusiveSum(local_vals[i], inclusive);
+        const int t = i * BLOCK_SIZE + threadIdx.x;
+        if (t < n_tok) {
+            cs_seq[t * n_head + h] = running + inclusive;
+        }
+        if (threadIdx.x == BLOCK_SIZE - 1) {
+            step_total = inclusive;
+        }
+        __syncthreads();
+        running += step_total;
+    }
+#else
+    // Fallback: sequential prefix scan in shared memory, one slab at a time.
+    __shared__ float sdata[BLOCK_SIZE];
+    float running = 0.0f;
+    for (int i = 0; i < items_per_thread; i++) {
+        const int t = i * BLOCK_SIZE + threadIdx.x;
+        sdata[threadIdx.x] = local_vals[i];
+        __syncthreads();
+        if (threadIdx.x == 0) {
+            for (int j = 1; j < BLOCK_SIZE; j++) {
+                sdata[j] += sdata[j - 1];
+            }
+        }
+        __syncthreads();
+        if (t < n_tok) {
+            cs_seq[t * n_head + h] = running + sdata[threadIdx.x];
+        }
+        running += sdata[BLOCK_SIZE - 1];
+        __syncthreads();
+    }
+#endif
+}
+
+// Prepare SSD matmul inputs for one chunk: X_dt, B_weighted, C_scaled.
+// T_matmul controls precision for X_dt, B_weighted (float or half).
+// C_scaled is always float (pairs with float s_cur in step 3c).
+// Computation is always FP32; only the final store converts to T_matmul.
+// Also materializes the causal M matrix = exp(A*(cs_out - cs_in)) * CB (fused with prep to save a launch).
+// Grid: (ceil(max(C*head_dim, d_state*C, chunk_len^2) / BLOCK), n_head, n_seqs)
+template <int BLOCK_SIZE, typename T_matmul>
+__global__ void ssm_ssd_pre_matmul_kernel(
+        const float * __restrict__ cs,         // {n_tok, n_head} cumulative dt sums
+        const float * __restrict__ dt_sp,      // {n_tok, n_head} softplus(dt)
+        const float * __restrict__ A,          // {1, n_head}
+        const float * __restrict__ x,          // {head_dim, n_head, n_tok, n_seqs}
+        const float * __restrict__ B,          // {d_state, n_group, n_tok, n_seqs}
+        const float * __restrict__ C_src,      // {d_state, n_group, n_tok, n_seqs}
+        T_matmul * __restrict__ X_dt,          // {head_dim, C, n_head} x * dt, d-fastest
+        T_matmul * __restrict__ B_weighted,    // {d_state, C, n_head} B * decay_from_end
+        float * __restrict__ C_scaled,         // {d_state, C, n_head} C * decay_to_pos (always float)
+        const float * __restrict__ CB,         // {chunk_len, chunk_len, n_group, n_seqs}
+        half * __restrict__ M_out,             // {chunk_len, chunk_len, n_head, n_seqs}
+        const int chunk_len, const int head_dim, const int n_head, const int n_group,
+        const int d_state, const int A_stride,
+        const int x_stride_tok, const int x_stride_seq,
+        const int B_stride_tok, const int B_stride_seq,
+        const int C_stride_tok, const int C_stride_seq,
+        const int chunk_offset,
+        const int n_tok_total) {
+
+    const int h = blockIdx.y;
+    const int s = blockIdx.z;
+    const int g = h / (n_head / n_group);
+
+    const float A_h = A[h * A_stride];
+    const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
+
+    const int cs_seq_off = s * n_tok_total * n_head;
+    const float cs_base = (chunk_offset > 0) ? cs[cs_seq_off + (chunk_offset - 1) * n_head + h] : 0.0f;
+    const float cs_last = cs[cs_seq_off + (chunk_offset + chunk_len - 1) * n_head + h] - cs_base;
+
+    // Prepare X_dt = x * dt, stored d-fastest for coalesced reads and writes.
+    const int n_xdt = chunk_len * head_dim;
+    if (idx < n_xdt) {
+        const int d = idx % head_dim;
+        const int t = idx / head_dim;
+
+        const float x_val = x[s * x_stride_seq + (chunk_offset + t) * x_stride_tok + d + h * head_dim];
+        const float dt_val = dt_sp[cs_seq_off + (chunk_offset + t) * n_head + h];
+
+        X_dt[d + t * head_dim + h * n_xdt + s * n_xdt * n_head] = (T_matmul)(x_val * dt_val);
+    }
+
+    // Prepare B_weighted and C_scaled together: both share the same index space (d_state * chunk_len)
+    // and the same cs_t load, so merging halves the cs[] global memory traffic.
+    const int n_bw = d_state * chunk_len;
+    if (idx < n_bw) {
+        const int n = idx % d_state;
+        const int t = idx / d_state;
+
+        const float cs_t = cs[cs_seq_off + (chunk_offset + t) * n_head + h] - cs_base;
+
+        const float B_val = B[s * B_stride_seq + (chunk_offset + t) * B_stride_tok + g * d_state + n];
+        B_weighted[n + t * d_state + h * n_bw + s * n_bw * n_head] = (T_matmul)(B_val * __expf(A_h * (cs_last - cs_t)));
+
+        const float C_val = C_src[s * C_stride_seq + (chunk_offset + t) * C_stride_tok + g * d_state + n];
+        C_scaled[n + t * d_state + h * n_bw + s * n_bw * n_head] = C_val * __expf(A_h * cs_t);
+    }
+
+    // Materialize M = exp(A*(cs_out - cs_in)) * CB with causal mask.
+    const int n_M = chunk_len * chunk_len;
+    if (idx < n_M) {
+        const int t_out = idx % chunk_len;
+        const int t_in  = idx / chunk_len;
+
+        half val;
+        if (t_in <= t_out) {
+            const float cs_out = cs[cs_seq_off + (chunk_offset + t_out) * n_head + h] - cs_base;
+            const float cs_in  = cs[cs_seq_off + (chunk_offset + t_in)  * n_head + h] - cs_base;
+            const float decay  = __expf(A_h * (cs_out - cs_in));
+            const float * CB_g = CB + (int64_t)s * chunk_len * chunk_len * n_group
+                                   + (int64_t)g * chunk_len * chunk_len;
+            const float cb_val = CB_g[t_out + t_in * chunk_len];
+            val = __float2half(decay * cb_val);
+        } else {
+            val = __float2half(0.0f);
+        }
+
+        M_out[(int64_t)s * n_M * n_head + (int64_t)h * n_M + t_in * chunk_len + t_out] = val;
+    }
+}
+
+// Scale running state in-place: s_cur *= decay_total(chunk).
+// Called BEFORE cuBLAS state update (beta=1) to fuse inter-chunk decay.
+// Eliminates the s_old buffer and D2D memcpy vs the old approach of:
+//   memcpy(s_old, s_cur) -> cuBLAS(beta=0) -> s_cur += decay * s_old
+// Grid: (ceil(d_state * head_dim / BLOCK), n_head, n_seqs)
+template <int BLOCK_SIZE>
+__global__ void ssm_ssd_scale_state_kernel(
+        float * __restrict__ s_cur,            // {d_state, head_dim, n_head, n_seqs}
+        const float * __restrict__ cs,         // {n_tok, n_head} cumulative dt sums
+        const float * __restrict__ A,          // {1, n_head}
+        const int d_state, const int head_dim, const int n_head,
+        const int chunk_offset, const int chunk_len,
+        const int n_tok_total, const int A_stride) {
+
+    const int h = blockIdx.y;
+    const int s = blockIdx.z;
+    const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
+    const int state_per_head = d_state * head_dim;
+    if (idx >= state_per_head) return;
+
+    const float A_h = A[h * A_stride];
+    const int cs_seq_off = s * n_tok_total * n_head;
+    const float cs_base = (chunk_offset > 0) ? cs[cs_seq_off + (chunk_offset - 1) * n_head + h] : 0.0f;
+    const float cs_last = cs[cs_seq_off + (chunk_offset + chunk_len - 1) * n_head + h] - cs_base;
+    const float decay_total = __expf(A_h * cs_last);
+
+    const int off = s * state_per_head * n_head + h * state_per_head + idx;
+    s_cur[off] *= decay_total;
+}
+
+// Copy initial state from src0[ids[s]] into s_cur for each sequence.
+// Grid: (ceil(d_state * head_dim * n_head / BLOCK), n_seqs)
+template <int BLOCK_SIZE>
+__global__ void ssm_ssd_init_state_kernel(
+        const float * __restrict__ src0,       // {d_state, head_dim, n_head, n_rs}
+        const int32_t * __restrict__ ids,      // {n_seqs}
+        float * __restrict__ s_cur,            // {d_state, head_dim, n_head, n_seqs}
+        const int state_size,                  // d_state * head_dim * n_head
+        const int64_t s0_stride_seq) {         // elements between state rows
+    const int s = blockIdx.y;
+    const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
+    if (idx >= state_size) return;
+
+    const float * s_src = src0 + (int64_t)ids[s] * s0_stride_seq;
+    s_cur[s * state_size + idx] = s_src[idx];
+}
+
+// SSD (State Space Duality) dispatch for Mamba-2 prefill.
+// Chunked matmuls: CB, materialize M + cuBLAS Y, S@C, B@X_dt.
+// All strides are in elements (floats), not bytes.
+static void ssm_scan_ssd_f32_cuda(
+        ggml_backend_cuda_context & ctx,
+        const float * src0_d, const float * src1_d, const float * src2_d, const float * src3_d,
+        const float * src4_d, const float * src5_d, const int32_t * src6_d, float * dst_d,
+        const int64_t s0_stride_seq,                                   // state (src0) stride between seqs
+        const int x_stride_tok,  const int x_stride_seq,               // x (src1) strides
+        const int dt_stride_tok, const int dt_stride_seq,              // dt (src2) strides
+        const int A_stride,                                            // A (src3) stride between heads
+        const int B_stride_tok,  const int B_stride_seq,               // B (src4) strides
+        const int C_stride_tok,  const int C_stride_seq,               // C (src5) strides
+        const int64_t s_off, const int64_t d_state, const int64_t head_dim,
+        const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq) {
+
+    cudaStream_t stream = ctx.stream();
+    const int64_t d_inner = head_dim * n_head;
+
+    const int64_t chunk_size = SSM_SSD_CHUNK_SIZE;
+    const int64_t n_chunks = (n_tok + chunk_size - 1) / chunk_size;
+
+    const int64_t state_per_head = d_state * head_dim;
+
+    using matmul_t = half;
+    static constexpr cudaDataType_t matmul_dtype = CUDA_R_16F;
+
+    ggml_cuda_pool_alloc<float>    dt_sp_buf(ctx.pool(), n_tok * n_head * n_seq);
+    ggml_cuda_pool_alloc<float>    cs_buf(ctx.pool(), n_tok * n_head * n_seq);
+    ggml_cuda_pool_alloc<float>    CB_buf(ctx.pool(), chunk_size * chunk_size * n_group * n_seq);
+    ggml_cuda_pool_alloc<matmul_t> X_dt_buf(ctx.pool(), chunk_size * head_dim * n_head * n_seq);
+    ggml_cuda_pool_alloc<matmul_t> B_w_buf(ctx.pool(), d_state * chunk_size * n_head * n_seq);
+    ggml_cuda_pool_alloc<float>    C_s_buf(ctx.pool(), d_state * chunk_size * n_head * n_seq);
+    float    * dt_sp      = dt_sp_buf.get();
+    float    * cs         = cs_buf.get();
+    float    * CB         = CB_buf.get();
+    matmul_t * X_dt       = X_dt_buf.get();
+    matmul_t * B_weighted = B_w_buf.get();
+    float    * C_scaled   = C_s_buf.get();
+    float    * s_cur      = (float *)((char *)dst_d + s_off); // write state directly to dst
+
+    // Step 1: softplus(dt) and parallel prefix sum over full sequence
+    {
+        dim3 grid(n_head, n_seq);
+        ssm_ssd_prepare_dt_kernel<SSM_SSD_DT_BLOCK, SSM_SSD_DT_MAX_ITEMS><<<grid, SSM_SSD_DT_BLOCK, 0, stream>>>(
+            src2_d, dt_sp, cs, n_head, n_tok, dt_stride_tok, dt_stride_seq);
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    // Step 2: initialize running state from src0[ids[s]]
+    {
+        constexpr int BLOCK = 256;
+        const int64_t state_size = d_state * head_dim * n_head;
+        dim3 grid((state_size + BLOCK - 1) / BLOCK, n_seq);
+        ssm_ssd_init_state_kernel<BLOCK><<<grid, BLOCK, 0, stream>>>(
+            src0_d, src6_d, s_cur, state_size, s0_stride_seq);
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    // Step 3: chunked SSD loop
+    // Per chunk: pre_matmul (incl. M) + 4 cuBLAS (CB, Y, S@C, state update) + scale_state
+    cublasHandle_t handle = ctx.cublas_handle();
+    CUBLAS_CHECK(cublasSetStream(handle, stream));
+    const float alpha_one  = 1.0f;
+    const float beta_zero  = 0.0f;
+    const float beta_one   = 1.0f;
+    const int lda_C_src = C_stride_tok;  // leading dim for C in CB = C^T @ B
+    const int ldb_B_src = B_stride_tok;  // leading dim for B in CB = C^T @ B
+
+    // Scratch buffer for causal M matrix, reused across chunks (max size at chunk_size)
+    const int64_t n_M_max = chunk_size * chunk_size;
+    ggml_cuda_pool_alloc<half> M_buf(ctx.pool(), n_M_max * n_head * n_seq);
+    half * M_mat = M_buf.get();
+
+    for (int64_t k = 0; k < n_chunks; k++) {
+        const int64_t chunk_offset = k * chunk_size;
+        const int64_t chunk_len = (chunk_offset + chunk_size <= n_tok) ? chunk_size : (n_tok - chunk_offset);
+
+        // 3a: CB = C^T @ B per group
+        for (int64_t s = 0; s < n_seq; s++) {
+            const float * C_s = src5_d + s * C_stride_seq + chunk_offset * C_stride_tok;
+            const float * B_s = src4_d + s * B_stride_seq + chunk_offset * B_stride_tok;
+            float *      CB_s = CB + s * chunk_len * chunk_len * n_group;
+
+            if (n_group == 1) {
+                CUBLAS_CHECK(cublasSgemm(handle, CUBLAS_OP_T, CUBLAS_OP_N,
+                    chunk_len, chunk_len, d_state,
+                    &alpha_one, C_s, lda_C_src, B_s, ldb_B_src,
+                    &beta_zero, CB_s, (int)chunk_len));
+            } else {
+                CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_T, CUBLAS_OP_N,
+                    chunk_len, chunk_len, d_state,
+                    &alpha_one,
+                    C_s, CUDA_R_32F, lda_C_src, d_state,
+                    B_s, CUDA_R_32F, ldb_B_src, d_state,
+                    &beta_zero,
+                    CB_s, CUDA_R_32F, (int)chunk_len, (long long)(chunk_len * chunk_len),
+                    n_group,
+                    CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+            }
+        }
+
+        // 3b: prepare X_dt, B_weighted, C_scaled + materialize causal M matrix
+        const int64_t n_M = chunk_len * chunk_len;
+        {
+            constexpr int BLOCK = 256;
+            const int64_t n_xdt   = chunk_len * head_dim;
+            const int64_t n_bw    = d_state * chunk_len;
+            int64_t max_work = n_xdt;
+            if (n_bw  > max_work) max_work = n_bw;
+            if (n_M   > max_work) max_work = n_M;
+            dim3 grid((max_work + BLOCK - 1) / BLOCK, n_head, n_seq);
+            ssm_ssd_pre_matmul_kernel<BLOCK, matmul_t><<<grid, BLOCK, 0, stream>>>(
+                cs, dt_sp, src3_d, src1_d, src4_d, src5_d,
+                X_dt, B_weighted, C_scaled,
+                CB, M_mat,
+                chunk_len, head_dim, n_head, n_group, d_state, A_stride,
+                x_stride_tok, x_stride_seq, B_stride_tok, B_stride_seq, C_stride_tok, C_stride_seq,
+                chunk_offset, n_tok);
+            CUDA_CHECK(cudaGetLastError());
+        }
+
+        // 3c: dst = S_cur^T @ C_scaled (state contribution)
+        {
+            const int64_t stride_S  = state_per_head;
+            const int64_t stride_Cs = d_state * chunk_len;
+
+            for (int64_t s = 0; s < n_seq; s++) {
+                float * dst_chunk = dst_d + s * d_inner * n_tok + chunk_offset * d_inner;
+
+                CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_T, CUBLAS_OP_N,
+                    head_dim, chunk_len, d_state,
+                    &alpha_one,
+                    s_cur    + s * stride_S  * n_head, CUDA_R_32F, d_state, stride_S,
+                    C_scaled + s * stride_Cs * n_head, CUDA_R_32F, d_state, stride_Cs,
+                    &beta_zero,
+                    dst_chunk, CUDA_R_32F, d_inner, head_dim,
+                    n_head,
+                    CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+            }
+        }
+
+        // 3d: dst += X_dt @ M^T (intra-chunk contribution, adds to 3c result)
+        // M is stored as M[t_out, t_in] (lower-triangular), transpose needed for Y = X @ M^T.
+        {
+            const int64_t stride_M = n_M;
+            const int64_t stride_X_h = (int64_t)chunk_len * head_dim;
+
+            for (int64_t s = 0; s < n_seq; s++) {
+                float * dst_chunk = dst_d + s * d_inner * n_tok + chunk_offset * d_inner;
+                CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_N, CUBLAS_OP_T,
+                    head_dim, chunk_len, chunk_len,
+                    &alpha_one,
+                    X_dt       + s * stride_X_h * n_head, matmul_dtype, head_dim, stride_X_h,
+                    M_mat      + s * stride_M   * n_head, matmul_dtype, chunk_len, stride_M,
+                    &beta_one,
+                    dst_chunk, CUDA_R_32F, d_inner, head_dim,
+                    n_head,
+                    CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+            }
+        }
+
+        // 3e: s_cur = B_weighted @ X_dt^T + decay_total * s_cur_old (state update)
+        {
+            // Scale s_cur in-place by per-head decay_total BEFORE cuBLAS overwrites it
+            constexpr int BLOCK = 256;
+            dim3 grid((state_per_head + BLOCK - 1) / BLOCK, n_head, n_seq);
+            ssm_ssd_scale_state_kernel<BLOCK><<<grid, BLOCK, 0, stream>>>(
+                s_cur, cs, src3_d,
+                d_state, head_dim, n_head,
+                chunk_offset, chunk_len, n_tok, A_stride);
+            CUDA_CHECK(cudaGetLastError());
+
+            // cuBLAS with beta=1: s_cur = B_weighted @ X_dt^T + 1.0 * s_cur (already scaled)
+            const int64_t stride_Bw = d_state * chunk_len;
+            const int64_t stride_X  = chunk_len * head_dim;
+            const int64_t stride_S  = state_per_head;
+
+            for (int64_t s = 0; s < n_seq; s++) {
+                // X_dt is d-fastest {hd, C}, read as OP_T to get {C, hd}
+                CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_N, CUBLAS_OP_T,
+                    d_state, head_dim, chunk_len,
+                    &alpha_one,
+                    B_weighted + s * stride_Bw * n_head, matmul_dtype, d_state, stride_Bw,
+                    X_dt       + s * stride_X  * n_head, matmul_dtype, head_dim, stride_X,
+                    &beta_one,
+                    s_cur      + s * stride_S  * n_head, CUDA_R_32F, d_state, stride_S,
+                    n_head,
+                    CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+            }
+        }
+    }
+}
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
 void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const struct ggml_tensor * src0 = dst->src[0];  // s
     const struct ggml_tensor * src1 = dst->src[1];  // x
@@ -357,6 +795,49 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(src6->type == GGML_TYPE_I32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
+    // Byte strides are narrowed to int for both scan and SSD paths.
+    GGML_ASSERT(src0->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src0->nb[3] <= (size_t)INT_MAX);
+    GGML_ASSERT(src1->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src1->nb[3] <= (size_t)INT_MAX);
+    GGML_ASSERT(src2->nb[1] <= (size_t)INT_MAX);
+    GGML_ASSERT(src2->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src3->nb[1] <= (size_t)INT_MAX);
+    GGML_ASSERT(src4->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src4->nb[3] <= (size_t)INT_MAX);
+    GGML_ASSERT(src5->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src5->nb[3] <= (size_t)INT_MAX);
+
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    // Mamba-2 with scalar A per head: use SSD matmul path for long sequences.
+    // Requires NVIDIA Turing+ otherwise fallback to scan.
+    const bool is_mamba2 = (src3->nb[1] == sizeof(float));
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const bool use_ssd = is_mamba2 && n_t > SSM_SSD_MIN_TOKENS
+                      && n_t <= SSM_SSD_MAX_TOKENS
+                      && GGML_CUDA_CC_IS_NVIDIA(cc)
+                      && cc >= GGML_CUDA_CC_TURING
+                      && nr % 8 == 0;  // cuBLAS requires 8-element (16-byte) alignment
+
+    if (use_ssd) {
+        // ssm_ssd_init_state_kernel uses flat linear indexing within each sequence,
+        // so src0 must be fully contiguous across all inner dimensions.
+        // The scan path handles non-contiguous nb[2] via src0_nb2 but does not handle nb[1].
+        GGML_ASSERT(src0->nb[1] == nc         * sizeof(float));
+        GGML_ASSERT(src0->nb[2] == nc * nr    * sizeof(float));
+
+        ssm_scan_ssd_f32_cuda(ctx,
+            src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
+            (int64_t)(src0->nb[3] / sizeof(float)),
+            (int)(src1->nb[2] / sizeof(float)), (int)(src1->nb[3] / sizeof(float)),
+            (int)(src2->nb[1] / sizeof(float)), (int)(src2->nb[2] / sizeof(float)),
+            (int)(src3->nb[1] / sizeof(float)),
+            (int)(src4->nb[2] / sizeof(float)), (int)(src4->nb[3] / sizeof(float)),
+            (int)(src5->nb[2] / sizeof(float)), (int)(src5->nb[3] / sizeof(float)),
+            s_off, nc, nr, nh, ng, n_t, n_s);
+        return;
+    }
+#endif
     ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
                       src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
                       src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],

ggml/src/ggml-hip/CMakeLists.txt

@@ -154,5 +154,3 @@ if (GGML_HIP_RCCL)
 endif()
 
 target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)
-
-target_compile_options(ggml-hip PRIVATE "$<$<COMPILE_LANGUAGE:HIP>:-ffast-math;-fno-finite-math-only>")

ggml/src/ggml-metal/CMakeLists.txt

@@ -11,6 +11,7 @@ ggml_add_backend_library(ggml-metal
                          ggml-metal-common.cpp
                          ggml-metal-context.m
                          ggml-metal-ops.cpp
+                         ggml-metal-tuning.cpp
                         )
 
 target_link_libraries(ggml-metal PRIVATE
@@ -24,62 +25,119 @@ if (GGML_METAL_NDEBUG)
 endif()
 
 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
+set(METALLIB_KERNELS_COMMON     "${CMAKE_CURRENT_SOURCE_DIR}/kernels/common.h")
+set(METALLIB_KERNELS_DEQUANTIZE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/dequantize.h")
+set(METALLIB_KERNELS_QUANTIZE   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantize.h")
+
+set(METALLIB_KERNEL_SOURCES
+    kernels/fa.metal
+    kernels/mul_mv.metal
+    kernels/mul_mm.metal
+    kernels/quantize.metal
+    kernels/softmax.metal
+    kernels/norm.metal
+    kernels/unary.metal
+    kernels/binbcast.metal
+    kernels/reduce.metal
+    kernels/tri.metal
+    kernels/ssm.metal
+    kernels/wkv.metal
+    kernels/gated_delta_net.metal
+    kernels/solve_tri.metal
+    kernels/rope.metal
+    kernels/conv.metal
+    kernels/upscale.metal
+    kernels/argsort.metal
+    kernels/pool.metal
+    kernels/misc.metal
+)
+
 if (GGML_METAL_EMBED_LIBRARY)
     enable_language(ASM)
 
     add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
-    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
 
     file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 
-    # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+    set(METALLIB_EMBED_ASM_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(kind ${src} NAME_WE)
+        # symbol names must be valid C identifiers ('-' is not allowed)
+        string(REPLACE "-" "_" kind_sym ${kind})
 
-    add_custom_command(
-        OUTPUT "${METALLIB_EMBED_ASM}"
-        COMMAND echo "Embedding Metal library"
-        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
-        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
-        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
-        COMMENT "Generate assembly for embedded Metal library"
-        VERBATIM
-    )
+        set(SRC   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${kind}.metal")
+        set(EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.metal")
+        set(ASM   "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.s")
 
-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+        # only prepend headers that this source actually includes
+        set(HEADERS_FOR_SRC ${METALLIB_KERNELS_COMMON})
+        file(STRINGS ${SRC} _has_dequantize REGEX "#include \"dequantize\\.h\"")
+        file(STRINGS ${SRC} _has_quantize   REGEX "#include \"quantize\\.h\"")
+        if(_has_dequantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_DEQUANTIZE})
+        endif()
+        if(_has_quantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_QUANTIZE})
+        endif()
+
+        add_custom_command(
+            OUTPUT "${ASM}"
+            # Step 1: concatenate shared headers + this kernel source
+            COMMAND cat ${HEADERS_FOR_SRC} ${SRC} > "${EMBED}.tmp1"
+            # Step 2: remove internal #include and #pragma once
+            COMMAND sed -e "/\#include \"common.h\"/d" -e "/\#include \"dequantize.h\"/d" -e "/\#include \"quantize.h\"/d" -e "/\#pragma once/d" < "${EMBED}.tmp1" > "${EMBED}.tmp2"
+            # Step 3: inline ggml-common.h (replacing __embed_ggml-common.h__ sentinel)
+            COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}" -e "/__embed_ggml-common.h__/d" < "${EMBED}.tmp2" > "${EMBED}.tmp3"
+            # Step 4: inline ggml-metal-impl.h
+            COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${EMBED}.tmp3" > "${EMBED}"
+            # Step 5: emit an asm chunk with kind-specific start/end symbols
+            #   note: '-' is illegal in C symbols, so we use kind_sym; the macOS
+            #   section name is limited to 16 chars so we keep it shared
+            #   across kinds (__ggml_metallib) and only vary the global symbols.
+            COMMAND echo ".section __DATA,__ggml_metallib"                       >  "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_start"               >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_start:"                     >> "${ASM}"
+            COMMAND echo .incbin "\"${EMBED}\""                                  >> "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_end"                 >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_end:"                       >> "${ASM}"
+            DEPENDS ../ggml-common.h ggml-metal-impl.h
+                    kernels/common.h kernels/dequantize.h kernels/quantize.h
+                    kernels/${kind}.metal
+            COMMENT "Generate embedded Metal library for ${kind}"
+            VERBATIM
+        )
+
+        list(APPEND METALLIB_EMBED_ASM_FILES "${ASM}")
+    endforeach()
+
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM_FILES})
 else()
-    # copy metal files to bin directory
+    # copy header files to bin directory
     configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
     configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
 
+    file(MAKE_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels")
+    configure_file(kernels/common.h     ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/common.h     COPYONLY)
+    configure_file(kernels/dequantize.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/dequantize.h COPYONLY)
+    configure_file(kernels/quantize.h   ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/quantize.h   COPYONLY)
+
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
+    endforeach()
+
     if (GGML_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: disabling fast math is needed in order to pass tests/test-backend-ops
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
         #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
         # note: adding -g causes segmentation fault during compile
-        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
         set(XC_FLAGS -fno-fast-math -fno-inline)
     else()
         set(XC_FLAGS -O3)
     endif()
 
-    # Append macOS metal versioning flags
     if (GGML_METAL_MACOSX_VERSION_MIN)
         message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
         list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
@@ -90,35 +148,46 @@ else()
         list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
     endif()
 
+    # Compile each kernel source to .air, then link into default.metallib
+    set(AIR_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(name ${src} NAME_WE)
+        set(AIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${name}.air")
+        list(APPEND AIR_FILES ${AIR})
+        add_custom_command(
+            OUTPUT ${AIR}
+            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
+            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
+            COMMENT "Compiling ${src}"
+            VERBATIM
+        )
+    endforeach()
+
     add_custom_command(
         OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
-        COMMENT "Compiling Metal kernels"
-        )
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
+        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
+        DEPENDS ${AIR_FILES}
+        COMMENT "Linking Metal kernels into default.metallib"
+    )
 
-    # FIXME: only add to the ggml-metal target?
     add_custom_target(
         ggml-metal-lib ALL
         DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
+    )
 endif() # GGML_METAL_EMBED_LIBRARY
 
 if (NOT GGML_METAL_EMBED_LIBRARY)
     install(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
+        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
+        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
+        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
+    )
 
-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
+    install(
+        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        DESTINATION ${CMAKE_INSTALL_BINDIR}
+    )
 endif()

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -1,6 +1,7 @@
 #include "ggml-metal-device.h"
 
 #include "ggml-metal-impl.h"
+#include "ggml-metal-tuning.h"
 
 #include "ggml-impl.h"
 
@@ -1252,6 +1253,21 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort_merge(gg
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_fwht(ggml_metal_library_t lib, int n) {
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_fwht_f32_%d", n);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 // note: reuse the argsort kernel for top_k
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_TOP_K);
@@ -1465,6 +1481,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
         bool    has_bias,
         bool    has_scap,
         bool    has_kvpad,
+        int32_t nqpsg,
+        int32_t ne,
         int32_t nsg,
         int32_t nwg) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
@@ -1478,11 +1496,17 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
     const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
     const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
 
-    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
+    char qne_suffix[16] = {0};
+    if (!(nqpsg == 1 && ne == ggml_metal_tuning::fa_vec_baseline_ne(dk, dv))) {
+        snprintf(qne_suffix, sizeof(qne_suffix), "_q%d_ne%d", nqpsg, ne);
+    }
+
+    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d%s",
             "flash_attn_ext_vec",
             ggml_type_name(op->src[1]->type),
             dk,
-            dv);
+            dv,
+            qne_suffix);
 
     snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
             base,

ggml/src/ggml-metal/ggml-metal-device.h

@@ -139,6 +139,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argmax            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_fwht              (ggml_metal_library_t lib, int n);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse );
@@ -196,6 +197,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
         bool    has_bias,
         bool    has_scap,
         bool    has_kvpad,
+        int32_t nqpsg,
+        int32_t ne,
         int32_t nsg,
         int32_t nwg);
 
@@ -263,6 +266,7 @@ struct ggml_metal_device_props {
     bool supports_gpu_family_apple7;
 
     enum ggml_metal_device_id device_id;
+    int gpu_family;
 
     int op_offload_min_batch_size;
 };

ggml/src/ggml-metal/ggml-metal-device.m

@@ -94,8 +94,63 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
     return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }
 
+//
+// MTLLibrary collection (one library per op-source, compiled separately)
+//
+
+// Single source of truth for the per-kind metal libraries. The order here
+// defines the enum values and every per-kind table below, so adding a library
+// is a one-line change here (plus adding its source to CMakeLists.txt).
+//   X(suffix, name): name is both the kernels/<name>.metal basename and the
+//   ggml_metallib_<name>_{start,end} embed-symbol stem.
+#define GGML_METAL_LIBS \
+    X(FA,              fa)             \
+    X(MUL_MV,          mul_mv)         \
+    X(MUL_MM,          mul_mm)         \
+    X(QUANTIZE,        quantize)       \
+    X(SOFTMAX,         softmax)        \
+    X(NORM,            norm)           \
+    X(UNARY,           unary)          \
+    X(BINBCAST,        binbcast)       \
+    X(REDUCE,          reduce)         \
+    X(TRI,             tri)            \
+    X(SSM,             ssm)            \
+    X(WKV,             wkv)            \
+    X(GATED_DELTA_NET, gated_delta_net)\
+    X(SOLVE_TRI,       solve_tri)      \
+    X(ROPE,            rope)           \
+    X(CONV,            conv)           \
+    X(UPSCALE,         upscale)        \
+    X(ARGSORT,         argsort)        \
+    X(POOL,            pool)           \
+    X(MISC,            misc)
+
+enum ggml_metal_lib_kind {
+#define X(e, s) GGML_METAL_LIB_##e,
+    GGML_METAL_LIBS
+#undef X
+    GGML_METAL_LIB_COUNT,
+};
+
+static const char * const k_lib_names[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = #s,
+    GGML_METAL_LIBS
+#undef X
+};
+
 struct ggml_metal_library {
-    id<MTLLibrary> obj;
+    // Per-kind compiled libraries. When single_library is true, the whole library
+    // (e.g. a pre-compiled default.metallib or a from-source build) lives at
+    // objs[0] and the remaining slots are nil.
+    id<MTLLibrary> objs[GGML_METAL_LIB_COUNT];
+    bool single_library; // true: combined library at objs[0]; false: per-kind libs in objs[*]
+
+    // Routing table: kernel function name -> objs[] index, populated from each
+    // compiled library's -[MTLLibrary functionNames]. The actual compiled
+    // libraries are the single source of truth for which library owns a kernel,
+    // so adding kernels later requires no manual routing maintenance.
+    // nil in single_library mode (everything resolves to objs[0]).
+    NSMutableDictionary<NSString *, NSNumber *> * fn_to_lib;
 
     ggml_metal_device_t dev;
     ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
@@ -103,160 +158,376 @@ struct ggml_metal_library {
     NSLock * lock;
 };
 
-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
-    id<MTLLibrary> library = nil;
-    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+// Build the fn_to_lib routing table by querying each compiled library's public
+// function names. Call once after all per-kind libraries have been compiled.
+static void ggml_metal_library_build_index(ggml_metal_library_t lib) {
+    @autoreleasepool {
+        NSMutableDictionary<NSString *, NSNumber *> * index = [[NSMutableDictionary alloc] init];
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            for (NSString * fname in [lib->objs[kind] functionNames]) {
+                index[fname] = @(kind);
+            }
+        }
+        lib->fn_to_lib = index;
+    }
+}
 
-    // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    //
-    // TODO: move to a function
-    {
-        const int64_t t_start = ggml_time_us();
+// Parse a `#include "name"` line. Returns the quoted name in *include_name on
+// success. Whitespace-tolerant; ignores `#include <...>` (system headers).
+static bool ggml_metal_library_parse_quoted_include(NSString * line, NSString ** include_name) {
+    NSScanner * scanner = [NSScanner scannerWithString:line];
+    scanner.charactersToBeSkipped = [NSCharacterSet whitespaceCharacterSet];
 
-        NSError * error = nil;
-        NSString * src = nil;
+    if (![scanner scanString:@"#" intoString:NULL] ||
+        ![scanner scanString:@"include" intoString:NULL] ||
+        ![scanner scanString:@"\"" intoString:NULL]) {
+        return false;
+    }
 
-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+    NSString * name = nil;
+    if (![scanner scanUpToString:@"\"" intoString:&name]) {
+        return false;
+    }
 
-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
+    if (include_name) {
+        *include_name = name;
+    }
+    return true;
+}
 
-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
+// Recursively inline `#include "name"` directives. System includes (<...>),
+// `#if/#else/#endif`, and other preprocessor lines are passed through to the
+// Metal compiler unchanged. `#pragma once` is dropped since `seen` already
+// guards against double-inclusion.
+static bool ggml_metal_library_flatten_file(NSMutableString * dst, NSString * path,
+                                            NSArray<NSString *> * search_paths,
+                                            NSMutableSet<NSString *> * seen, NSError ** error) {
+    NSString * key = [path stringByStandardizingPath];
+    if ([seen containsObject:key]) {
+        return true;
+    }
+    [seen addObject:key];
 
-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
+    NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:error];
+    if (!src) {
+        return false;
+    }
 
-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+    NSFileManager * fm = [NSFileManager defaultManager];
+    for (NSString * line in [src componentsSeparatedByString:@"\n"]) {
+        NSString * trimmed = [line stringByTrimmingCharactersInSet:[NSCharacterSet whitespaceCharacterSet]];
+        if ([trimmed isEqualToString:@"#pragma once"]) {
+            continue;
+        }
 
-            NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
-
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
-                    if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
-                    }
-                    if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                        // Link to the resource could not be resolved.
-                        path_lib_default = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
-                    }
+        NSString * include_name = nil;
+        if (ggml_metal_library_parse_quoted_include(line, &include_name)) {
+            NSString * resolved = nil;
+            for (NSString * dir in search_paths) {
+                NSString * candidate = [dir stringByAppendingPathComponent:include_name];
+                if ([fm isReadableFileAtPath:candidate]) {
+                    resolved = candidate;
+                    break;
                 }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                path_lib_default = nil;
             }
-
-            path_lib = path_lib_default;
+            if (!resolved) {
+                if (error) {
+                    NSString * msg = [NSString stringWithFormat:@"could not resolve include \"%@\" from '%@'", include_name, path];
+                    *error = [NSError errorWithDomain:@"ggml-metal-source-flatten" code:1
+                                             userInfo:@{NSLocalizedDescriptionKey: msg}];
+                }
+                return false;
+            }
+            if (!ggml_metal_library_flatten_file(dst, resolved, search_paths, seen, error)) {
+                return false;
+            }
+            continue;
         }
 
-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+        [dst appendString:line];
+        [dst appendString:@"\n"];
+    }
 
-            library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+    return true;
+}
 
-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+static NSString * ggml_metal_library_flatten_source(NSString * path_source, NSError ** error) {
+    // Search paths cover both runtime layout (build/bin/kernels + build/bin)
+    // and source-tree layout (ggml/src/ggml-metal/kernels + ggml/src/ggml-metal + ggml/src).
+    NSString * path_kernels = [path_source stringByDeletingLastPathComponent];
+    NSString * path_base    = [path_kernels stringByDeletingLastPathComponent];
+    NSArray<NSString *> * search_paths = @[
+        path_kernels,
+        path_base,
+        [path_base stringByDeletingLastPathComponent],
+    ];
 
-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+    NSMutableString * src = [[NSMutableString alloc] init];
+    NSMutableSet<NSString *> * seen = [NSMutableSet set];
 
-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+    if (!ggml_metal_library_flatten_file(src, path_source, search_paths, seen, error)) {
+        [src release];
+        return nil;
+    }
+    return src;
+}
+
+// Compile all per-kind libraries in parallel. `source_for_kind` returns the MSL
+// source for a kind (the helper takes ownership and releases it), or nil with
+// *err set on failure. On success the objs[] slots are populated and the routing
+// index is built; on any failure every error is logged and false is returned
+// (the caller is responsible for freeing `res`).
+static bool ggml_metal_library_compile_all(
+        ggml_metal_library_t res,
+        id<MTLDevice> device,
+        NSDictionary * prep,
+        NSString * (^source_for_kind)(int kind, NSError ** err),
+        const char * origin) {
+    const int64_t t_start = ggml_time_us();
+
+    int64_t  * t_per_lib   = calloc(GGML_METAL_LIB_COUNT, sizeof(int64_t));
+    NSError ** err_per_lib = calloc(GGML_METAL_LIB_COUNT, sizeof(NSError *));
+    __block atomic_bool any_failure = false;
+
+    dispatch_group_t group = dispatch_group_create();
+    dispatch_queue_t queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        dispatch_group_async(group, queue, ^{
+
+            const int64_t t0 = ggml_time_us();
+
+            NSError * error = nil;
+
+            NSString * src = source_for_kind(kind, &error);
+            if (!src) {
+                err_per_lib[kind] = [error retain];
+                atomic_store(&any_failure, true);
+                return;
             }
 
-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
+            id<MTLLibrary> lib = nil;
 
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        }
-#endif
-
-        if (!library) {
             @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ggml_metal_device_get_props(dev)->has_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
-                }
-
-                if (ggml_metal_device_get_props(dev)->has_tensor) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                 MTLCompileOptions * options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
-                //[options setFastMathEnabled:false];
+                lib = [device newLibraryWithSource:src options:options error:&error];
 
-                library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return nil;
-                }
-
-#if !__has_feature(objc_arc)
                 [options release];
-#endif
+
+                // retain the error before the autorelease pool drains it
+                if (!lib) {
+                    err_per_lib[kind] = [error retain];
+                }
+            }
+
+            [src release];
+
+            t_per_lib[kind] = ggml_time_us() - t0;
+
+            if (!lib) {
+                atomic_store(&any_failure, true);
+                return;
+            }
+
+            res->objs[kind] = lib;
+        });
+    }
+    dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
+    dispatch_release(group);
+
+    const bool ok = !atomic_load(&any_failure);
+
+    if (ok) {
+        const int64_t t_total = ggml_time_us() - t_start;
+        int64_t t_max = 0;
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            GGML_LOG_DEBUG("%s: compiled '%s' library in %.3f sec\n",
+                           __func__, k_lib_names[kind], t_per_lib[kind] / 1e6);
+            if (t_per_lib[kind] > t_max) t_max = t_per_lib[kind];
+        }
+        GGML_LOG_INFO("%s: loaded %d libraries from %s in %.3f sec (max single = %.3f sec)\n",
+                      __func__, GGML_METAL_LIB_COUNT, origin, t_total / 1e6, t_max / 1e6);
+
+        ggml_metal_library_build_index(res);
+    } else {
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            if (err_per_lib[kind]) {
+                GGML_LOG_ERROR("%s: failed to build '%s' library: %s\n", __func__,
+                               k_lib_names[kind], [[err_per_lib[kind] description] UTF8String]);
+                [err_per_lib[kind] release];
             }
         }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
-
-        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
     }
 
-    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    free(err_per_lib);
+    free(t_per_lib);
 
-    res->obj       = library;
+    return ok;
+}
+
+ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
     res->dev       = dev;
     res->pipelines = ggml_metal_pipelines_init();
     res->lock      = [NSLock new];
 
+    // shared MTLCompileOptions preprocessor macros (matches the build-time defines)
+    NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+    if (ggml_metal_device_get_props(dev)->has_bfloat) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
+    }
+    if (ggml_metal_device_get_props(dev)->has_tensor) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+    }
+#if GGML_METAL_EMBED_LIBRARY
+    [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    // start/end symbols emitted by CMake (see CMakeLists.txt), one pair per kind
+#define X(e, s) extern const char ggml_metallib_##s##_start[]; extern const char ggml_metallib_##s##_end[];
+    GGML_METAL_LIBS
+#undef X
+
+    static const char * const lib_start[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_start,
+    GGML_METAL_LIBS
+#undef X
+    };
+    static const char * const lib_end[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_end,
+    GGML_METAL_LIBS
+#undef X
+    };
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            (void) err;
+            return [[NSString alloc] initWithBytes:lib_start[kind]
+                                            length:(lib_end[kind] - lib_start[kind])
+                                          encoding:NSUTF8StringEncoding];
+        }, "embedded data");
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
     return res;
+#else
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    const int64_t t_start = ggml_time_us();
+
+    NSError * error = nil;
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+
+        NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
+
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
+                if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
+                }
+                if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+                    // Link to the resource could not be resolved.
+                    path_lib_default = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            path_lib_default = nil;
+        }
+
+        path_lib = path_lib_default;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found: a single combined default.metallib
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        res->objs[0]        = [device newLibraryWithURL:libURL error:&error];
+        res->single_library = true;
+        if (!res->objs[0]) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            ggml_metal_library_free(res);
+            return NULL;
+        }
+
+        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+        return res;
+    }
+
+    // no pre-compiled metallib: fall back to compiling each kernel source separately
+    GGML_LOG_INFO("%s: default.metallib not found, loading kernel sources\n", __func__);
+
+    NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+    if (path_resource) {
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, [path_resource UTF8String]);
+    }
+
+    // resolve each kind's source path up front (file lookup/logging stays on the calling thread)
+    NSString ** path_per_kind = calloc(GGML_METAL_LIB_COUNT, sizeof(NSString *));
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        NSString * rel = [NSString stringWithFormat:@"kernels/%s.metal", k_lib_names[kind]];
+
+        NSString * path_source = nil;
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:rel];
+        } else {
+            NSString * stem = [NSString stringWithFormat:@"kernels/%s", k_lib_names[kind]];
+            path_source = [bundle pathForResource:stem ofType:@"metal"];
+        }
+
+        if (path_source == nil || ![[NSFileManager defaultManager] isReadableFileAtPath:path_source]) {
+            GGML_LOG_WARN("%s: could not locate %s in bundle, falling back to cwd\n", __func__, [rel UTF8String]);
+            path_source = rel;
+        }
+
+        GGML_LOG_DEBUG("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        path_per_kind[kind] = [path_source retain];
+    }
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            return ggml_metal_library_flatten_source(path_per_kind[kind], err);
+        }, "source");
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        [path_per_kind[kind] release];
+    }
+    free(path_per_kind);
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
+    return res;
+#endif
 }
 
 ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
@@ -318,10 +589,11 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
         return NULL;
     }
 
-    res->obj       = library;
-    res->dev       = dev;
-    res->pipelines = ggml_metal_pipelines_init();
-    res->lock      = [NSLock new];
+    res->objs[0]        = library;
+    res->single_library = true;
+    res->dev            = dev;
+    res->pipelines      = ggml_metal_pipelines_init();
+    res->lock           = [NSLock new];
 
     return res;
 }
@@ -331,8 +603,14 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
         return;
     }
 
-    if (lib->obj) {
-        [lib->obj release];
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        if (lib->objs[kind]) {
+            [lib->objs[kind] release];
+        }
+    }
+
+    if (lib->fn_to_lib) {
+        [lib->fn_to_lib release];
     }
 
     ggml_metal_pipelines_free(lib->pipelines);
@@ -393,11 +671,28 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_
 
         GGML_LOG_DEBUG("%s: compiling pipeline: base = '%s', name = '%s'\n", __func__, base, name);
 
+        // route to the library that actually defines this kernel; fn_to_lib is
+        // built from -[MTLLibrary functionNames] so it's always in sync
+        int lib_idx = 0;
+        if (!lib->single_library) {
+            NSNumber * idx = lib->fn_to_lib[base_func];
+            if (!idx) {
+                [lib->lock unlock];
+
+                GGML_LOG_ERROR("%s: kernel not found in any metal library: base = '%s', name = '%s'\n", __func__, base, name);
+
+                return res;
+            }
+            lib_idx = [idx intValue];
+        }
+
+        id<MTLLibrary> mtl_lib = lib->objs[lib_idx];
+
         id<MTLFunction> mtl_function;
         if (!cv) {
-            mtl_function = [lib->obj newFunctionWithName:base_func];
+            mtl_function = [mtl_lib newFunctionWithName:base_func];
         } else {
-            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+            mtl_function = [mtl_lib newFunctionWithName:base_func constantValues:cv->obj error:&error];
         }
         if (!mtl_function) {
             [lib->lock unlock];
@@ -878,7 +1173,8 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
             {
                 for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
                     if ([dev->mtl_device supportsFamily:i]) {
-                        GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
+                        dev->props.gpu_family = i - (int) MTLGPUFamilyApple1 + 1;
+                        GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, dev->props.gpu_family, i);
                         break;
                     }
                 }

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -1157,6 +1157,10 @@ typedef struct {
     int32_t  len;
 } ggml_metal_kargs_argsort_merge;
 
+typedef struct {
+    int32_t nrows;
+} ggml_metal_kargs_fwht;
+
 typedef struct {
     int64_t  ne0;
     float    start;

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -7,6 +7,7 @@
 #include "ggml-metal-impl.h"
 #include "ggml-metal-common.h"
 #include "ggml-metal-device.h"
+#include "ggml-metal-tuning.h"
 
 #include <cassert>
 #include <algorithm>
@@ -1979,6 +1980,46 @@ int ggml_metal_op_pool_1d(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+// supported FWHT sizes, must stay in sync with the
+// kernel_fwht_f32_<N> templates in ggml-metal.metal
+static bool ggml_metal_fwht_supported_size(int64_t n) {
+    return n == 64 || n == 128 || n == 256 || n == 512;
+}
+
+int ggml_metal_op_fwht(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    ggml_tensor * src1 = op->src[1];
+
+    const int64_t n = src1->ne[0];
+    const int64_t nrows = ggml_nrows(src1);
+
+    ggml_metal_kargs_fwht args = {
+        /*.nrows = */ (int32_t) nrows,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_fwht(lib, n);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(src1), 1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 2);
+
+    const int th_max = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
+    const int simd_size = 32;
+
+    int sg_per_tg = 2;
+    sg_per_tg = std::min(sg_per_tg, th_max/simd_size);
+    sg_per_tg = std::max(sg_per_tg, 1);
+
+    const int64_t n_tg = (nrows + sg_per_tg - 1) / sg_per_tg;
+    ggml_metal_encoder_dispatch_threadgroups(enc, n_tg, 1, 1, 32*sg_per_tg, 1, 1);
+
+    return 1;
+}
 
 int ggml_metal_op_pool_2d(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
@@ -2046,6 +2087,18 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
     ggml_metal_library_t lib = ctx->lib;
     ggml_metal_encoder_t enc = ctx->enc;
 
+    const int32_t hint = ggml_get_op_params_i32(op, 1);
+
+    if (hint == GGML_HINT_SRC0_IS_HADAMARD) {
+        if (op->src[1]->type == GGML_TYPE_F32 &&
+            op->type == GGML_TYPE_F32 &&
+            ggml_is_contiguous(op->src[1]) &&
+            ggml_is_contiguous(op) &&
+            ggml_are_same_shape(op->src[1], op) &&
+            ggml_metal_fwht_supported_size(op->src[1]->ne[0])) {
+            return ggml_metal_op_fwht(ctx, idx);
+        }
+    }
     const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx->dev);
 
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
@@ -2890,12 +2943,18 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
     } else {
         // half4x4 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPSG; // queries per threadgroup
+        auto cfg = ggml_metal_tuning::fa_vec_pick(
+                props_dev->device_id,
+                props_dev->gpu_family,
+                (int) op->src[1]->type,
+                (int) ne00, (int) ne20,   // dk, dv (ne00 == dk for FA)
+                ne11, ne01);
+        int nqptg = cfg.Q;                             // queries per threadgroup
         const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
         const int nhptg = 1;                           // heads per threadgroup
 
         GGML_ASSERT(nqptg <= 32);
-        GGML_ASSERT(nqptg  % 1  == 0);
+        GGML_ASSERT(nqptg == 1 || nqptg == 2 || nqptg == 4);  // only instantiated Q values
         GGML_ASSERT(ncpsg  % 32 == 0);
 
         bool need_sync = false;
@@ -2954,7 +3013,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         // ne20*(nsg)
         // each simdgroup has a full f32 head vector in shared mem to accumulate results
         //
-#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg)*nqptg)*(sizeof(float)/2), 16))
 
         int64_t nsg = 1;
 
@@ -2974,6 +3033,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             }
         }
 
+        // fall back to baseline (Q=1) if the tuned config exceeds threadgroup memory
+        if ((size_t) FATTN_SMEM(nsg) > props_dev->max_theadgroup_memory_size) {
+            cfg   = ggml_metal_tuning::fa_vec_baseline_cfg((int) ne00, (int) ne20);
+            nqptg = cfg.Q;  // = 1
+        }
+
         ggml_metal_kargs_flash_attn_ext_vec args = {
             /*.ne01          =*/ ne01,
             /*.ne02          =*/ ne02,
@@ -3009,7 +3074,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nqptg, cfg.NE, nsg, nwg);
 
         GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -64,6 +64,7 @@ int ggml_metal_op_set               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cpy               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_1d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_2d           (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_fwht              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_mul_mat           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_mul_mat_id        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_add_id            (ggml_metal_op_t ctx, int idx);

ggml/src/ggml-metal/ggml-metal-tuning.cpp

@@ -0,0 +1,347 @@
+#include "ggml-metal-tuning.h"
+
+#include <cstddef>
+#include <cstring>
+#include <iterator>
+
+namespace ggml_metal_tuning {
+
+int fa_vec_ne11_bucket(int64_t ne11) {
+    for (int i = 0; i < (int) std::size(FA_VEC_NE11_BUCKETS); ++i) {
+        if (ne11 < FA_VEC_NE11_BUCKETS[i]) {
+            return i;
+        }
+    }
+    return (int) std::size(FA_VEC_NE11_BUCKETS);
+}
+
+int fa_vec_ne01_bucket(int64_t ne01) {
+    for (int i = 0; i < (int) std::size(FA_VEC_NE01_BUCKETS); ++i) {
+        if (ne01 < FA_VEC_NE01_BUCKETS[i]) {
+            return i;
+        }
+    }
+    return (int) std::size(FA_VEC_NE01_BUCKETS);
+}
+
+int fa_vec_baseline_ne(int dk, int dv) {
+    if (dk == 32 && dv == 32) {
+        return 4;
+    }
+    if (dk == 64 && dv == 64) {
+        return 2;
+    }
+    if (dk == 96 && dv == 96) {
+        return 4;
+    }
+    if (dk == 128 && dv == 128) {
+        return 1;
+    }
+    if (dk == 192 && dv == 192) {
+        return 2;
+    }
+    if (dk == 192 && dv == 128) {
+        return 2;
+    }
+    if (dk == 256 && dv == 256) {
+        return 1;
+    }
+    if (dk == 320 && dv == 256) {
+        return 2;
+    }
+    if (dk == 512 && dv == 512) {
+        return 1;
+    }
+    if (dk == 576 && dv == 512) {
+        return 2;
+    }
+    return 4;  // template default
+}
+
+fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv) {
+    return { 1, (int8_t) fa_vec_baseline_ne(dk, dv) };
+}
+
+// Generated by `test-backend-ops tune --tune-perf`; do not hand-edit.
+// One row per kept bucket, plus per-(dtype,dk,dv) ne11-collapsed domain defaults
+// (ne11_b = FA_VEC_NE11_DEFAULT, ne01_b = domain). To retune or add a device, re-run the
+// sweep and paste its block. See ggml-metal-tuning.h for the row/lookup semantics.
+constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
+    // ---- f16: 13 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 2, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 128, 128, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 192, 192, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 192, 128, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 256, 256, -1, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 320, 256, -1, 1 }, { 2, 2 } },
+    // ---- q4_0: 29 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 4 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 128, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 128, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 4 }, { 1, 2 } },
+    // ---- q4_1: 28 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 4 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 2, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 576, 512, -1, 1 }, { 1, 4 } },
+    // ---- q5_0: 45 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, 1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, 1, 2 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, 1, 3 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, -1, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 1, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 4 }, { 1, 4 } },
+    // ---- q5_1: 49 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, 1, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 2 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 3 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 4 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, 1, 3 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 4 }, { 1, 4 } },
+    // ---- q8_0: 29 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 32, 32, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 32, 32, 3, 4 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 1, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
+};
+
+static enum ggml_metal_device_id fa_vec_family_representative(int gpu_family) {
+    switch (gpu_family) {
+        case 9:  return GGML_METAL_DEVICE_M4_MAX;
+        default: return GGML_METAL_DEVICE_GENERIC;
+    }
+}
+
+static bool         g_override_set = false;
+static fa_vec_cfg_t g_override_cfg = { 1, 4 };
+
+void fa_vec_set_override(fa_vec_cfg_t cfg) {
+    g_override_cfg = cfg;
+    g_override_set = true;
+}
+
+void fa_vec_clear_override() {
+    g_override_set = false;
+}
+
+static const fa_vec_cfg_t * find_cfg(const fa_vec_entry_t * tbl, size_t n, const fa_vec_key_t & k) {
+    for (size_t i = 0; i < n; ++i) {
+        if (memcmp(&tbl[i].key, &k, sizeof(k)) == 0) {
+            return &tbl[i].cfg;
+        }
+    }
+    return nullptr;
+}
+
+fa_vec_cfg_t fa_vec_pick(enum ggml_metal_device_id device_id, int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01) {
+    if (g_override_set) {
+        return g_override_cfg;
+    }
+
+    const fa_vec_cfg_t baseline = fa_vec_baseline_cfg(dk, dv);
+
+    const int ne11_b = fa_vec_ne11_bucket(ne11);
+    if (ne11_b == 0) {
+        return baseline;  // short KV: attention is a small slice of the step, left to baseline
+    }
+    const int ne01_b = fa_vec_ne01_bucket(ne01);
+
+    fa_vec_key_t k{};
+    k.dtype = (int8_t) dtype;
+    k.dk    = (int16_t) dk;
+    k.dv    = (int16_t) dv;
+
+    // exact bucket, then the ne01 domain default (ne11 collapsed); tried under each device tier
+    auto lookup = [&](enum ggml_metal_device_id dev) -> const fa_vec_cfg_t * {
+        k.device_id = (int8_t) dev;
+        k.ne11_b    = (int8_t) ne11_b;
+        k.ne01_b    = (int8_t) ne01_b;
+        if (auto * c = find_cfg(fa_vec_tuned_table, std::size(fa_vec_tuned_table), k)) {
+            return c;
+        }
+        k.ne11_b = FA_VEC_NE11_DEFAULT;
+        k.ne01_b = (ne01_b == 0) ? FA_VEC_DOMAIN_DECODE : FA_VEC_DOMAIN_BATCH;
+        return find_cfg(fa_vec_tuned_table, std::size(fa_vec_tuned_table), k);
+    };
+
+    if (auto * c = lookup(device_id)) {
+        return *c;
+    }
+
+    // family fallback: retry under the family's representative SKU; none -> baseline
+    if (gpu_family > 0) {
+        const enum ggml_metal_device_id rep = fa_vec_family_representative(gpu_family);
+        if (rep != GGML_METAL_DEVICE_GENERIC) {
+            if (auto * c = lookup(rep)) {
+                return *c;
+            }
+        }
+    }
+
+    return baseline;
+}
+
+}  // namespace ggml_metal_tuning

ggml/src/ggml-metal/ggml-metal-tuning.h

@@ -0,0 +1,63 @@
+#pragma once
+
+#include "ggml-metal-device.h"  // enum ggml_metal_device_id
+#include "ggml.h"
+
+#include <cstdint>
+
+namespace ggml_metal_tuning {
+
+// FA vec selection buckets. ne01 (query rows) splits decode (==1) from batch (>=2), the
+// batch side refined into {2,3,4,5}: Q>1 reuses one K/V load across rows, so it only pays
+// off once ne01 aligns with Q. ne11 (KV length) is bucketed too, as the Q>1 crossover is
+// head-size dependent (small dk crosses late, large dk wins even at short KV).
+constexpr int FA_VEC_NE11_BUCKETS[] = { 1024, 4096, 16384 };
+constexpr int FA_VEC_NE01_BUCKETS[] = { 2, 3, 4, 5 };
+
+int fa_vec_ne11_bucket(int64_t ne11);
+int fa_vec_ne01_bucket(int64_t ne01);
+
+// NE baked into each (dk,dv) baseline instantiation in kernels/fa.metal.
+// Hand-maintained mirror; keep in sync with those instantiations (run_fa_vec_tune_check
+// exercises every (Q,NE), so a missing instantiation surfaces there).
+int fa_vec_baseline_ne(int dk, int dv);
+
+// Tuned table has two row kinds. Exact rows key a (ne11_b, ne01_b) bucket. Default rows
+// collapse ne11 over one ne01 domain: ne11_b == FA_VEC_NE11_DEFAULT and ne01_b holds the
+// domain. fa_vec_pick tries exact bucket -> domain default -> baseline; short KV
+// (ne11 < FA_VEC_NE11_BUCKETS[0]) always uses baseline.
+constexpr int8_t FA_VEC_NE11_DEFAULT  = -1;
+constexpr int8_t FA_VEC_DOMAIN_DECODE = 0;  // ne01 == 1
+constexpr int8_t FA_VEC_DOMAIN_BATCH  = 1;  // ne01 >= 2
+
+struct fa_vec_key_t {
+    int8_t  device_id;
+    int8_t  dtype;
+    int16_t dk;
+    int16_t dv;
+    int8_t  ne11_b;
+    int8_t  ne01_b;
+};
+
+static_assert(sizeof(fa_vec_key_t) == 8, "fa_vec_key_t must be tightly packed for memcmp");
+
+struct fa_vec_cfg_t {
+    int8_t Q;
+    int8_t NE;
+};
+
+struct fa_vec_entry_t {
+    fa_vec_key_t key;
+    fa_vec_cfg_t cfg;
+};
+
+// test/tune-only override; when set, fa_vec_pick returns it directly.
+void         fa_vec_set_override(fa_vec_cfg_t cfg);
+void         fa_vec_clear_override();
+fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv);
+
+// device_id selects a per-SKU row; on a miss, gpu_family (0 if unknown) maps to a representative
+// SKU and the table is retried. No match -> baseline.
+fa_vec_cfg_t fa_vec_pick(enum ggml_metal_device_id device_id, int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01);
+
+}  // namespace ggml_metal_tuning

ggml/src/ggml-metal/ggml-metal.cpp

@@ -6,6 +6,7 @@
 #include "ggml-metal-device.h"
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"
+#include "ggml-metal-tuning.h"
 
 #include <mutex>
 #include <string>
@@ -868,10 +869,46 @@ static ggml_backend_feature * ggml_backend_metal_get_features(ggml_backend_reg_t
     GGML_UNUSED(reg);
 }
 
+// test/tune-only override for the FA vec (Q, NE) selection, reached via proc_address.
+static void ggml_backend_metal_tuning_set_fa_vec_override(int Q, int NE) {
+    ggml_metal_tuning::fa_vec_set_override({ (int8_t) Q, (int8_t) NE });
+}
+
+static void ggml_backend_metal_tuning_clear_fa_vec_override(void) {
+    ggml_metal_tuning::fa_vec_clear_override();
+}
+
+static int ggml_backend_metal_tuning_fa_vec_ne11_bucket(int64_t ne11) {
+    return ggml_metal_tuning::fa_vec_ne11_bucket(ne11);
+}
+
+static int ggml_backend_metal_tuning_fa_vec_ne01_bucket(int64_t ne01) {
+    return ggml_metal_tuning::fa_vec_ne01_bucket(ne01);
+}
+
+static int ggml_backend_metal_tuning_fa_vec_baseline_ne(int dk, int dv) {
+    return ggml_metal_tuning::fa_vec_baseline_ne(dk, dv);
+}
+
 static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     if (strcmp(name, "ggml_backend_get_features") == 0) {
         return (void *)ggml_backend_metal_get_features;
     }
+    if (strcmp(name, "ggml_backend_metal_tuning_set_fa_vec_override") == 0) {
+        return (void *)ggml_backend_metal_tuning_set_fa_vec_override;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_clear_fa_vec_override") == 0) {
+        return (void *)ggml_backend_metal_tuning_clear_fa_vec_override;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_ne11_bucket") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_ne11_bucket;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_ne01_bucket") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_ne01_bucket;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_baseline_ne") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_baseline_ne;
+    }
 
     return NULL;
 

ggml/src/ggml-metal/kernels/argsort.metal

@@ -0,0 +1,232 @@
+#include "common.h"
+
+// bitonic sort implementation following the CUDA kernels as reference
+typedef void (argsort_t)(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    // bitonic sort
+    const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
+
+    // initialize indices
+    shmem_i32[col] = i00 + col;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (int k = 2; k <= ntg.x; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                } else {
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                }
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+    }
+
+    const int64_t i0 = ib*args.top_k;
+
+    // copy the result to dst without the padding
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
+
+        dst[col] = shmem_i32[col];
+    }
+}
+
+template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
+
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;

ggml/src/ggml-metal/kernels/binbcast.metal

@@ -0,0 +1,226 @@
+#include "common.h"
+
+// OP: 0 - add, 1 - sub, 2 - mul, 3 - div
+constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
+constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
+constant bool  FC_bin_rb [[function_constant(FC_BIN + 2)]];
+constant bool  FC_bin_cb [[function_constant(FC_BIN + 3)]];
+
+template <typename T0, typename T1, typename T>
+kernel void kernel_bin_fuse_impl(
+        constant ggml_metal_kargs_bin & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP FC_bin_op
+#define FC_F  FC_bin_f
+#define FC_RB FC_bin_rb
+#define FC_CB FC_bin_cb
+
+    if (FC_RB) {
+        // row broadcast
+        const uint i0 = tgpig.y*args.ne00 + tgpig.x;
+        const uint i1 = FC_CB ? tgpig.x%args.ne10 : tgpig.x;
+
+        device const T0 * src0_row = (device const T0 *) (src0);
+        device       T  * dst_row  = (device       T  *) (dst);
+
+        if (FC_F == 1) {
+            device const T1 * src1_row = (device const T1 *) (src1 + args.o1[0]);
+
+            if (FC_OP == 0) {
+                dst_row[i0] = src0_row[i0] + src1_row[i1];
+            }
+
+            if (FC_OP == 1) {
+                dst_row[i0] = src0_row[i0] - src1_row[i1];
+            }
+
+            if (FC_OP == 2) {
+                dst_row[i0] = src0_row[i0] * src1_row[i1];
+            }
+
+            if (FC_OP == 3) {
+                dst_row[i0] = src0_row[i0] / src1_row[i1];
+            }
+        } else {
+            T0 res = src0_row[i0];
+
+            if (FC_OP == 0) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res += ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 1) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res -= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 2) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res *= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 3) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res /= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            dst_row[i0] = res;
+        }
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int i01 = tgpig.x;
+
+        if (i01 >= args.ne01) {
+            return;
+        }
+
+        const int i13 = i03%args.ne13;
+        const int i12 = i02%args.ne12;
+        const int i11 = i01%args.ne11;
+
+        device const T0 * src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+        device       T  * dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+        if (FC_F == 1) {
+            device const T1 * src1_ptr = (device const T1 *) (src1 + args.o1[0] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                if (FC_OP == 0) {
+                    dst_ptr[i0] = src0_ptr[i0] + src1_ptr[i10];
+                }
+
+                if (FC_OP == 1) {
+                    dst_ptr[i0] = src0_ptr[i0] - src1_ptr[i10];
+                }
+
+                if (FC_OP == 2) {
+                    dst_ptr[i0] = src0_ptr[i0] * src1_ptr[i10];
+                }
+
+                if (FC_OP == 3) {
+                    dst_ptr[i0] = src0_ptr[i0] / src1_ptr[i10];
+                }
+            }
+        } else {
+            device const T1 * src1_ptr[8];
+            FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                src1_ptr[j] = (device const T1 *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+            }
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                T res = src0_ptr[i0];
+
+                if (FC_OP == 0) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res += src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 1) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res -= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 2) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res *= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 3) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res /= src1_ptr[j][i10];
+                    }
+                }
+
+                dst_ptr[i0] = res;
+            }
+        }
+    }
+
+#undef FC_OP
+#undef FC_F
+#undef FC_RB
+#undef FC_CB
+}
+
+typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
+
+template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
+template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
+
+kernel void kernel_add_id(
+        constant ggml_metal_kargs_add_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i1 = tgpig.x;
+    const int i2 = tgpig.y;
+
+    const int i11 = *((device const int32_t *) (src2 + i1*sizeof(int32_t) + i2*args.nb21));
+
+    const size_t nb1 = args.ne0 * sizeof(float);
+    const size_t nb2 = args.ne1 * nb1;
+
+    device       float * dst_row  = (device       float *)((device char *)dst  +  i1*nb1       + i2*nb2);
+    device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
+    device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
+template<typename T>
+kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
+
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_repeat_bf16")]] kernel kernel_repeat_t kernel_repeat<bfloat>;
+#endif
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;

ggml/src/ggml-metal/kernels/common.h

@@ -0,0 +1,126 @@
+#pragma once
+
+#include "ggml-metal-impl.h"
+
+#include <metal_stdlib>
+
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
+
+#define PAD2(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
+#define FOR_UNROLL(x) _Pragma("clang loop unroll(full)") for (x)
+
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+//
+// cmd:
+//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal/kernels/<src>.metal
+//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal/kernels/<src>.metal
+//
+#if __METAL_VERSION__ < 310 && defined(GGML_METAL_HAS_BF16)
+#undef GGML_METAL_HAS_BF16
+#endif
+
+#if defined(GGML_METAL_HAS_BF16)
+typedef matrix<bfloat, 4, 4> bfloat4x4;
+typedef matrix<bfloat, 2, 4> bfloat2x4;
+#endif
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+constexpr constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static inline float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    return as_type<float>(bits);
+}
+
+static inline float dot(float x, float y) {
+    return x*y;
+}
+
+static inline float sum(float x) {
+    return x;
+}
+
+static inline float sum(float4 x) {
+    return x[0] + x[1] + x[2] + x[3];
+}
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+inline T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+template<typename T> T elu_approx(T x);
+
+template<> inline float elu_approx<float>(float x) {
+    return (x > 0.f) ? x : (exp(x) - 1);
+}
+
+template<> inline float4 elu_approx<float4>(float4 x) {
+    float4 res;
+
+    res[0] = (x[0] > 0.0f) ? x[0] : (exp(x[0]) - 1.0f);
+    res[1] = (x[1] > 0.0f) ? x[1] : (exp(x[1]) - 1.0f);
+    res[2] = (x[2] > 0.0f) ? x[2] : (exp(x[2]) - 1.0f);
+    res[3] = (x[3] > 0.0f) ? x[3] : (exp(x[3]) - 1.0f);
+
+    return res;
+}

ggml/src/ggml-metal/kernels/conv.metal

@@ -0,0 +1,723 @@
+#include "common.h"
+
+typedef void (im2col_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
+
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
+
+          int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0;
+    const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1;
+
+    int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        while (in < args.N) {
+            pdst[offset_dst] = 0.0f;
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+
+            in += ntg[0];
+        }
+    } else {
+        int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw;
+
+        while (in < args.N) {
+            pdst[offset_dst] = x[offset_src];
+
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+            offset_src += ntg[0]*args.ofs0;
+
+            in += ntg[0];
+        }
+    }
+}
+
+template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
+template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
+
+// TODO: optimize
+typedef void (im2col_ext_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int64_t KHW = (int64_t)args.KHW;
+
+    const int64_t d   = tgpig[0] / args.CHW;
+    const int64_t chw = tgpig[0] % args.CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
+
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= args.N) {
+        return;
+    }
+
+    const int64_t tpitg_1 = HW / args.KW;
+    const int64_t tpitg_2 = HW % args.KW;
+
+    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
+    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
+
+    const int64_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
+        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
+        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+
+template <typename T>
+kernel void kernel_col2im_1d(
+        constant ggml_metal_kargs_col2im_1d & args,
+        device const T * col,
+        device       T * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T_out * args.OC) {
+        return;
+    }
+
+    const int t_out = idx % args.T_out;
+    const int oc    = idx / args.T_out;
+    const int t_abs = t_out + args.p0;  // absolute position in uncropped signal
+
+    int t_in_min = (t_abs - args.K + args.s0) / args.s0;  // ceil((t_abs - K + 1) / s0)
+    if (t_in_min < 0) {
+        t_in_min = 0;
+    }
+    int t_in_max = t_abs / args.s0;
+    if (t_in_max >= args.T_in) {
+        t_in_max = args.T_in - 1;
+    }
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * args.s0;
+        sum += float(col[(oc * args.K + k) + t_in * args.K_OC]);
+    }
+
+    dst[t_out + oc * args.T_out] = T(sum);
+}
+
+template [[host_name("kernel_col2im_1d_f32")]]  kernel void kernel_col2im_1d<float>(constant ggml_metal_kargs_col2im_1d &, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_col2im_1d_f16")]]  kernel void kernel_col2im_1d<half>(constant ggml_metal_kargs_col2im_1d &, device const half *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_col2im_1d_bf16")]] kernel void kernel_col2im_1d<bfloat>(constant ggml_metal_kargs_col2im_1d &, device const bfloat *, device bfloat *, uint, uint, uint);
+#endif
+
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+typedef void (conv_transpose_1d_t)(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_1d(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const     T * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    // For output position j on the time axis, only input positions
+    //   i such that i*s0 <= j < i*s0 + K
+    // contribute -- i.e. i in [ceil((j - K + 1)/s0), floor(j/s0)]
+    // intersected with [0, IL-1]. That's at most ceil(K/s0) values
+    // (typically 2 for stride==K/2 transposed convs).
+    const int32_t j  = tgpig[0];
+    const int32_t s0 = args.s0;
+    const int32_t K  = args.K;
+    const int32_t IL = args.IL;
+
+    int32_t i_min;
+    {
+        int32_t a = j - K + 1;
+        i_min = a <= 0 ? 0 : (a + s0 - 1) / s0; // ceil(a/s0) for a>0
+    }
+    int32_t i_max = j / s0;
+    if (i_max > IL - 1) i_max = IL - 1;
+
+    float v = 0.0f;
+    if (i_min <= i_max) {
+        for (int64_t c = 0; c < args.IC; c++) {
+            const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1];
+            const int32_t input_offset  = c * IL;
+
+            for (int32_t i = i_min; i <= i_max; i++) {
+                v += float(src0[kernel_offset + j - i * s0]) * src1[input_offset + i];
+            }
+        }
+    }
+
+    device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1);
+
+    dst_ptr[0] = v;
+}
+
+template [[host_name("kernel_conv_transpose_1d_f32_f32")]]
+kernel void kernel_conv_transpose_1d<float>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+template [[host_name("kernel_conv_transpose_1d_f16_f32")]]
+kernel void kernel_conv_transpose_1d<half>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+
+typedef void (conv_transpose_2d_t)(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_2d(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const T * src0,
+        device const float * src1,
+        device        char * dst,
+        threadgroup float * shared_sum [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int64_t out_x = tgpig[0];
+    const int64_t out_y = tgpig[1];
+    const int64_t out_c = tgpig[2];
+
+    const int64_t kw = tpitg[0];
+    const int64_t kh = tpitg[1];
+
+    float v = 0.0f;
+
+    for (int64_t in_c = 0; in_c < args.IC; in_c++) {
+        int64_t in_y = out_y - kh;
+
+        if (in_y < 0 || in_y % args.s0) continue;
+
+        in_y /= args.s0;
+
+        if (in_y >= args.IH) continue;
+
+        int64_t in_x = out_x - kw;
+
+        if (in_x < 0 || in_x % args.s0) continue;
+
+        in_x /= args.s0;
+
+        if (in_x >= args.IW) continue;
+
+        const int64_t input_idx = (args.IW * args.IH) * in_c + (args.IW) * in_y + in_x;
+        const int64_t kernel_idx = (args.KH * args.KW * args.OC) * in_c + (args.KH * args.KW) * out_c + (args.KW) * kh + kw;
+
+        v += (float)src0[kernel_idx] * src1[input_idx];
+    }
+
+    const uint tid = tpitg.y * ntg.x + tpitg.x;
+    shared_sum[tid] = v;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tid == 0) {
+        float total = 0.0f;
+        const uint num_threads = ntg.x * ntg.y;
+        for (uint i = 0; i < num_threads; i++) {
+            total += shared_sum[i];
+        }
+
+        device float * dst_ptr = (device float *) (dst + out_x*args.nb0 + out_y * args.nb1 + out_c*args.nb2);
+        dst_ptr[0] = total;
+    }
+}
+
+template [[host_name("kernel_conv_transpose_2d_f32_f32")]]
+kernel void kernel_conv_transpose_2d<float>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_transpose_2d_f16_f32")]]
+kernel void kernel_conv_transpose_2d<half>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+// grid: x = C tile, y = OH, z = OW * N (for channel-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw_tiled(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t c = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (c >= args.C) {
+        return;
+    }
+
+    const int32_t oh = tgpig.y;
+    const int32_t own = tgpig.z;
+    const int32_t ow = own % args.OW;
+    const int32_t n  = own / args.OW;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    float acc = 0.0f;
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        const uint64_t w_base   = (uint64_t) c * args.nb02;
+        const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+// grid: x = OW tile, y = OH, z = C * N (for spatially-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t oh = tgpig.y;
+    const int32_t cn = tgpig.z;
+    const int32_t c  = cn % args.C;
+    const int32_t n  = cn / args.C;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const uint64_t w_base   = (uint64_t) c * args.nb02;
+    const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+    const int32_t ow = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (ow >= args.OW) {
+        return;
+    }
+
+    float acc = 0.0f;
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+template [[host_name("kernel_conv_2d_dw_f32_f32")]]
+kernel void kernel_conv_2d_dw<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_f16_f32")]]
+kernel void kernel_conv_2d_dw<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f32_f32")]]
+kernel void kernel_conv_2d_dw_tiled<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f16_f32")]]
+kernel void kernel_conv_2d_dw_tiled<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_conv_3d(
+        constant ggml_metal_kargs_conv_3d & args,
+        device const  char * src0, // Weights [IC * OC, KD, KH, KW]
+        device const  char * src1, // Inputs  [IC * N,  ID, IH, IW]
+        device       char  * dst,  // Outputs [OC * N,  OD, OH, OW]
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+
+    // 1. Un-flatten the spatial dimension from Grid X
+    int64_t spatial_idx = tgpig.x * 32 + tpitg.x;
+
+    if (spatial_idx >= args.OW * args.OH * args.OD) {
+        return; // Thread falls outside the spatial volume
+    }
+
+    int64_t od = spatial_idx / (args.OW * args.OH);
+    int64_t oh = (spatial_idx / args.OW) % args.OH;
+    int64_t ow = spatial_idx % args.OW;
+
+    // 2. Map Y to Channels, Z to Batch
+    int64_t oc = tgpig.y;
+    int64_t batch_idx = tgpig.z;
+
+    // 3. Calculate anchor coordinates in the Input volume
+    int64_t i_w_base = ow * args.s0 - args.p0;
+    int64_t i_h_base = oh * args.s1 - args.p1;
+    int64_t i_d_base = od * args.s2 - args.p2;
+
+    float sum = 0.0f;
+
+    // 4. Gather Loop (Iterate over Input Channels -> Depth -> Height -> Width)
+    for (int64_t ic = 0; ic < args.IC; ++ic) {
+
+        // ggml packs batch and channel together in the 4th dimension
+        int64_t src_cn_idx = batch_idx * args.IC + ic;
+        int64_t w_cn_idx   = oc * args.IC + ic;
+
+        for (int64_t kz = 0; kz < args.KD; ++kz) {
+            int64_t id = i_d_base + kz * args.d2;
+            if (id < 0 || id >= args.ID) continue; // Boundary check (Padding)
+
+            for (int64_t ky = 0; ky < args.KH; ++ky) {
+                int64_t ih = i_h_base + ky * args.d1;
+                if (ih < 0 || ih >= args.IH) continue;
+
+                for (int64_t kx = 0; kx < args.KW; ++kx) {
+                    int64_t iw = i_w_base + kx * args.d0;
+                    if (iw < 0 || iw >= args.IW) continue;
+
+                    // Convert multi-dimensional coordinates to flat byte offsets
+                    int64_t w_idx = kx*args.nb00 + ky*args.nb01 + kz*args.nb02 + w_cn_idx*args.nb03;
+                    int64_t i_idx = iw*args.nb10 + ih*args.nb11 + id*args.nb12 + src_cn_idx*args.nb13;
+
+                    // Dereference memory and cast weights to f32 if they were f16
+                    float w_val = (float)*(device const T*)((device const char*)src0 + w_idx);
+                    float i_val = *(device const float*)((device const char*)src1 + i_idx);
+
+                    sum += w_val * i_val;
+                }
+            }
+        }
+    }
+
+    // 5. Write the accumulated value out to RAM
+    int64_t dst_cn_idx = batch_idx * args.OC + oc;
+    int64_t d_idx = ow*args.nb0 + oh*args.nb1 + od*args.nb2 + dst_cn_idx*args.nb3;
+
+    *(device float*)(dst + d_idx) = sum;
+}
+
+// Explicit instantiations so the JIT compiler can find them by name
+template [[host_name("kernel_conv_3d_f32_f32")]]
+kernel void kernel_conv_3d<float>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+// Explicit instantiation for f16 weights
+template [[host_name("kernel_conv_3d_f16_f32")]]
+kernel void kernel_conv_3d<half>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char  * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);

ggml/src/ggml-metal/kernels/dequantize.h

@@ -0,0 +1,719 @@
+#pragma once
+
+#include "common.h"
+
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+#include "ggml-common.h"
+#endif
+
+#define QK_NL 16 // shared by mul_mm and get_rows_q instantiations
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+
+#if defined(GGML_METAL_HAS_BF16)
+template <typename type4x4>
+void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+#endif
+
+template <typename type4x4>
+void dequantize_q1_0(device const block_q1_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+    const float neg_d = -d;
+
+    const int byte_offset = il * 2;  // il*16 bits = il*2 bytes
+    const uint8_t b0 = qs[byte_offset];
+    const uint8_t b1 = qs[byte_offset + 1];
+
+    float4x4 reg_f;
+
+    reg_f[0][0] = select(neg_d, d, bool(b0 & 0x01));
+    reg_f[0][1] = select(neg_d, d, bool(b0 & 0x02));
+    reg_f[0][2] = select(neg_d, d, bool(b0 & 0x04));
+    reg_f[0][3] = select(neg_d, d, bool(b0 & 0x08));
+    reg_f[1][0] = select(neg_d, d, bool(b0 & 0x10));
+    reg_f[1][1] = select(neg_d, d, bool(b0 & 0x20));
+    reg_f[1][2] = select(neg_d, d, bool(b0 & 0x40));
+    reg_f[1][3] = select(neg_d, d, bool(b0 & 0x80));
+
+    reg_f[2][0] = select(neg_d, d, bool(b1 & 0x01));
+    reg_f[2][1] = select(neg_d, d, bool(b1 & 0x02));
+    reg_f[2][2] = select(neg_d, d, bool(b1 & 0x04));
+    reg_f[2][3] = select(neg_d, d, bool(b1 & 0x08));
+    reg_f[3][0] = select(neg_d, d, bool(b1 & 0x10));
+    reg_f[3][1] = select(neg_d, d, bool(b1 & 0x20));
+    reg_f[3][2] = select(neg_d, d, bool(b1 & 0x40));
+    reg_f[3][3] = select(neg_d, d, bool(b1 & 0x80));
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const float neg_d = -d;
+    const int base = il * 4;
+    const uint8_t byte = xb->qs[base / 8];
+    const int s = base % 8;
+
+    float4 reg_f;
+    reg_f[0] = select(neg_d, d, bool((byte >> (s    )) & 1));
+    reg_f[1] = select(neg_d, d, bool((byte >> (s + 1)) & 1));
+    reg_f[2] = select(neg_d, d, bool((byte >> (s + 2)) & 1));
+    reg_f[3] = select(neg_d, d, bool((byte >> (s + 3)) & 1));
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q2_0(device const block_q2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    const int byte_offset = il * 4;  // il*16 elements = il*4 bytes (4 elements per byte)
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; i++) {
+        const uint8_t b = qs[byte_offset + i];
+        reg_f[i][0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+        reg_f[i][1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+        reg_f[i][2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+        reg_f[i][3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q2_0_t4(device const block_q2_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const uint8_t b = xb->qs[il];
+
+    float4 reg_f;
+    reg_f[0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+    reg_f[1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+    reg_f[2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+    reg_f[3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
+        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
+
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = ((qs[i] & mask0) * d1) + m;
+        reg_f[i/2][2*(i%2) + 1] = ((qs[i] & mask1) * d2) + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + md;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + m;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 16; i++) {
+        reg_f[i/4][i%4] = (qs[i + 16*il] * d);
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    for (int i = 0; i < 4; i++) {
+        reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d);
+    }
+}
+
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
+    }
+}
+
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
+
+    const uint8_t shr = il >= 4 ? 4 : 0;
+
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint16_t * ql = (device const uint16_t *)xb->ql;
+    device const uint16_t * qh = (device const uint16_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
+    qh = qh + 16*(il/8) + 8*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
+    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
+    const float ml = d_all * sc * 32.f;
+    const float dl0 = d_all * sc;
+    const float dl1 = dl0 / 256.f;
+    const float dl2 = dl0 / (256.f * 256.f);
+    const float dl3 = dl0 / (256.f * 256.f * 256.f);
+    const uint8_t shr_h = il>2 ? 2 : 0;
+    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
+    const uint8_t shr_l = il>1 ? 4 : 0;
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
+        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
+        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
+        reg[i][0] = dl0 *  ((half)(q & 0xFF))       - ml;
+        reg[i][1] = dl1 * ((float)(q & 0xFF00))     - ml;
+        reg[i][2] = dl2 * ((float)(q & 0xFF0000))   - ml;
+        reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+

ggml/src/ggml-metal/kernels/gated_delta_net.metal

@@ -0,0 +1,250 @@
+#include "common.h"
+
+constant short FC_gated_delta_net_ne20 [[function_constant(FC_GATED_DELTA_NET + 0)]];
+constant short FC_gated_delta_net_ne30 [[function_constant(FC_GATED_DELTA_NET + 1)]];
+constant short FC_gated_delta_net_K    [[function_constant(FC_GATED_DELTA_NET + 2)]];
+
+#if 1
+template<short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+#define K   FC_gated_delta_net_K
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B (n_seqs)
+    const uint i21 = tgpig.y; // H (head)
+    const uint i20 = tgpig.x*NSG + ty; // row within S_v
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    // input state layout [S_v, S_v, H, n_seqs] (s0 only): per-seq stride is H*D.
+    // state is stored transposed: M[i20][is] = S[is][i20], so row i20 is contiguous
+    const uint state_in_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+    device const float * s_ptr = (device const float *) (s) + state_in_base;
+
+    float ls[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        ls[j] = s_ptr[is];
+    }
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
+
+    // output state base offset: after attention scores
+    const uint attn_size = args.ne22 * args.ne21 * S_v * args.ne23;
+    // output state per-slot size: S_v * S_v * H * n_seqs
+    const uint state_size_per_snap = S_v * S_v * args.ne21 * args.ne23;
+    // per-(seq,head) offset within a slot
+    const uint state_out_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+
+    for (short t = 0; t < args.ne22; t++) {
+        float s_k = 0.0f;
+
+        if (G == 1) {
+            const float g_exp = exp(g_ptr[0]);
+
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= g_exp;
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        } else {
+            // KDA
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= exp(g_ptr[is]);
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        }
+
+        s_k = simd_sum(s_k);
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        float y = 0.0f;
+
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            ls[j] += k_ptr[is]*d;
+
+            y += ls[j]*q_ptr[is];
+        }
+
+        y = simd_sum(y);
+
+        if (tx == 0) {
+            dst_attn[t*args.ne21*S_v] = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        if (K > 1) {
+            const int target_slot = (int)args.ne22 - 1 - (int)t;
+            if (target_slot >= 0 && target_slot < (int)K) {
+                device float * dst_state = (device float *) (dst) + attn_size + (uint)target_slot * state_size_per_snap + state_out_base;
+                FOR_UNROLL (short j = 0; j < NSG; j++) {
+                    const short is = tx*NSG + j;
+                    dst_state[is] = ls[j];
+                }
+            }
+        }
+    }
+
+    if (K == 1) {
+        device float * dst_state = (device float *) (dst) + attn_size + state_out_base;
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            dst_state[is] = ls[j];
+        }
+    }
+
+#undef S_v
+#undef G
+#undef K
+}
+
+typedef decltype(kernel_gated_delta_net_impl<4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<4>;
+
+#else
+// a simplified version of the above
+// no performance improvement, so keep the above version for now
+
+template<typename T, short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B
+    const uint i21 = tgpig.y; // H
+    const uint i20 = tgpig.x*NSG + ty;
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    float lsf[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        lsf[j] = s_ptr[is*S_v];
+    }
+
+    thread T * ls = (thread T *) (lsf);
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr  = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr  = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    for (short t = 0; t < args.ne22; t++) {
+        device const T * qt_ptr = (device const T *) (q_ptr);
+        device const T * kt_ptr = (device const T *) (k_ptr);
+        device const T * gt_ptr = (device const T *) (g_ptr);
+
+        if (G == 1) {
+            *ls *= exp(g_ptr[0]);
+        } else {
+            // KDA
+            *ls *= exp(gt_ptr[tx]);
+        }
+
+        const float s_k = simd_sum(dot(*ls, kt_ptr[tx]));
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        *ls += kt_ptr[tx]*d;
+
+        const float y = simd_sum(dot(*ls, qt_ptr[tx]));
+
+        if (tx == 0) {
+            *dst_attn = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        dst_attn += args.ne21*S_v;
+    }
+
+    device float * dst_state  = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
+    device T     * dstt_state = (device T     *) (dst_state);
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        dst_state[is*S_v] = lsf[j];
+    }
+
+#undef S_v
+#undef G
+}
+
+typedef decltype(kernel_gated_delta_net_impl<float4, 4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float,  1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float2, 2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float4, 4>;
+#endif

ggml/src/ggml-metal/kernels/misc.metal

@@ -0,0 +1,438 @@
+#include "common.h"
+
+kernel void kernel_argmax_f32(
+        constant ggml_metal_kargs_argmax & args,
+        device   const char * src0,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    device const float * x_row = (device const float *) ((device const char *) src0 + tgpig * args.nb01);
+
+    float   lmax = -INFINITY;
+    int32_t larg = -1;
+
+    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+        if (x_row[i00] > lmax) {
+            lmax = x_row[i00];
+            larg = i00;
+        }
+    }
+
+    // find the argmax value in the block
+    float max_val = simd_max(lmax);
+    int32_t arg_val = simd_max(select(-1, larg, lmax == max_val));
+
+    device int32_t * dst_i32 = (device int32_t *) dst;
+
+    threadgroup   float * shared_maxval = (threadgroup   float *) shmem;
+    threadgroup int32_t * shared_argmax = (threadgroup int32_t *) shmem + N_SIMDWIDTH;
+
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            shared_maxval[tiisg] = -INFINITY;
+            shared_argmax[tiisg] = -1;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            shared_maxval[sgitg] = max_val;
+            shared_argmax[sgitg] = arg_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = shared_maxval[tiisg];
+        arg_val = shared_argmax[tiisg];
+
+        float max_val_reduced   = simd_max(max_val);
+        int32_t arg_val_reduced = simd_max(select(-1, arg_val, max_val == max_val_reduced));
+
+        dst_i32[tgpig] = arg_val_reduced;
+
+        return;
+    }
+
+    dst_i32[tgpig] = arg_val;
+}
+
+kernel void kernel_diag_f32(
+        constant ggml_metal_kargs_diag & args,
+        device   const char * src0,
+        device         char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *)(src0 +                i2*args.nb02 + i3*args.nb03);
+    device       float * dst_ptr  = (device       float *)(dst  + i1*args.nb01 + i2*args.nb2  + i3*args.nb3);
+
+    for (int i0 = tiitg; i0 < args.ne0; i0 += NW) {
+        dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
+    }
+}
+
+kernel void kernel_roll_f32(
+    constant ggml_metal_kargs_roll & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *) src0;
+    device       float * dst_ptr  = (device       float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        // apply shifts and wrap around
+        int64_t i00 = i0 - args.s0;
+        int64_t i01 = i1 - args.s1;
+        int64_t i02 = i2 - args.s2;
+        int64_t i03 = i3 - args.s3;
+
+        if (i00 < 0) { i00 += args.ne00; } else if (i00 >= args.ne00) { i00 -= args.ne00; }
+        if (i01 < 0) { i01 += args.ne01; } else if (i01 >= args.ne01) { i01 -= args.ne01; }
+        if (i02 < 0) { i02 += args.ne02; } else if (i02 >= args.ne02) { i02 -= args.ne02; }
+        if (i03 < 0) { i03 += args.ne03; } else if (i03 >= args.ne03) { i03 -= args.ne03; }
+
+        int64_t src_idx = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00 + i00;
+        int64_t dst_idx = i3 *args.ne2 *args.ne1 *args.ne0  + i2 *args.ne1 *args.ne0  + i1 *args.ne0  + i0;
+
+        dst_ptr[dst_idx] = src0_ptr[src_idx];
+    }
+}
+
+template <typename T>
+kernel void kernel_pad_impl(
+    constant ggml_metal_kargs_pad & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t k0 = tgpig.x/args.ne1;
+    const int32_t i1 = tgpig.x - k0*args.ne1;
+
+    const int32_t i03 = i3;
+    const int32_t i02 = i2;
+    const int32_t i01 = i1;
+
+    device const T * src0_ptr = (device const T *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T * dst_ptr  = (device       T *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    for (int32_t l0 = 0; l0 < 1024; l0 += ntg.x) {
+        const int32_t i0 = k0*1024 + tpitg.x + l0;
+        if (i0 >= args.ne0) {
+            break;
+        }
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            dst_ptr[i0] = src0_ptr[i0];
+        } else {
+            dst_ptr[i0] = 0.0f;
+        }
+    }
+}
+
+typedef decltype(kernel_pad_impl<float>) kernel_pad_t;
+
+template [[host_name("kernel_pad_f32")]]   kernel kernel_pad_t kernel_pad_impl<float>;
+template [[host_name("kernel_pad_f32_4")]] kernel kernel_pad_t kernel_pad_impl<float4>;
+
+// TODO: this is slow - optimize
+kernel void kernel_pad_reflect_1d_f32(
+    constant   ggml_metal_kargs_pad_reflect_1d & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3  tgpg[[threadgroups_per_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            if (i0 < args.p0) {
+                dst_ptr[i0] = src0_ptr[args.p0 - i0];
+            } else if (i0 < args.ne0 - args.p1) {
+                dst_ptr[i0] = src0_ptr[i0 - args.p0];
+            } else {
+                dst_ptr[i0] = src0_ptr[(args.ne0 - args.p1 - args.p0) - (args.p1 + 1 - (args.ne0 - i0)) - 1];
+            }
+        }
+    }
+}
+
+kernel void kernel_arange_f32(
+    constant   ggml_metal_kargs_arange & args,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    device float * dst_ptr = (device float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_ptr[i0] = args.start + args.step * i0;
+    }
+}
+
+kernel void kernel_timestep_embedding_f32(
+    constant  ggml_metal_kargs_timestep_embedding & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    int i = tgpig.x;
+    device float * embed_data = (device float *)(dst + i*args.nb1);
+
+    int half_ = args.dim / 2;
+    for (int j = tpitg.x; j < half_; j += ntg.x) {
+        float timestep = ((device float *)src0)[i];
+        float freq = (float)exp(-log((float)args.max_period) * j / half_);
+        float arg = timestep * freq;
+        embed_data[j        ] = cos(arg);
+        embed_data[j + half_] = sin(arg);
+    }
+
+    if (args.dim % 2 != 0 && tpitg.x == 0) {
+        embed_data[2 * half_] = 0.f;
+    }
+}
+
+kernel void kernel_opt_step_adamw_f32(
+        constant    ggml_metal_kargs_opt_step_adamw & args,
+        device       float * x,
+        device const float * g,
+        device       float * g_m,
+        device       float * g_v,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
+    const float gi = g[gid];
+    const float gmi = g_m[gid] * beta1 +      gi * (1.0f - beta1);
+    const float gvi = g_v[gid] * beta2 + gi * gi * (1.0f - beta2);
+
+    g_m[gid] = gmi;
+    g_v[gid] = gvi;
+
+    const float mh =      gmi * beta1h;
+    const float vh = sqrt(gvi * beta2h) + eps;
+
+    x[gid] = x[gid] * (1.0f - alpha * wd) - alpha * mh / vh;
+}
+
+kernel void kernel_opt_step_sgd_f32(
+        constant    ggml_metal_kargs_opt_step_sgd & args,
+        device       float * x,
+        device const float * g,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    x[gid] = x[gid] * (1.0f - pars[0] * pars[1]) - pars[0] * g[gid];
+}
+
+template<typename T>
+kernel void kernel_memset(
+        constant ggml_metal_kargs_memset & args,
+        device T * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = args.val;
+}
+
+typedef decltype(kernel_memset<int64_t>) kernel_memset_t;
+
+template [[host_name("kernel_memset_i64")]] kernel kernel_memset_t kernel_memset<int64_t>;
+
+constant short FC_count_equal_nsg [[function_constant(FC_COUNT_EQUAL + 0)]];
+
+template<typename T>
+kernel void kernel_count_equal(
+        constant ggml_metal_kargs_count_equal & args,
+        device   const char * src0,
+        device   const char * src1,
+        device   atomic_int * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const short NSG = FC_count_equal_nsg;
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    int sum = 0;
+
+    device const char * base0 = src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03;
+    device const char * base1 = src1 + i1*args.nb11 + i2*args.nb12 + i3*args.nb13;
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        const T v0 = *(device const T *)(base0 + i0*args.nb00);
+        const T v1 = *(device const T *)(base1 + i0*args.nb10);
+        sum += (v0 == v1);
+    }
+
+    sum = simd_sum(sum);
+
+    if (tiisg == 0) {
+        shmem_i32[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        float v = 0.0f;
+        if (tpitg.x < NSG) {
+            v = shmem_i32[tpitg.x];
+        }
+
+        float total = simd_sum(v);
+        if (tpitg.x == 0) {
+            atomic_fetch_add_explicit(dst, (int32_t) total, memory_order_relaxed);
+        }
+    }
+}
+
+typedef decltype(kernel_count_equal<int32_t>) kernel_count_equal_t;
+
+template [[host_name("kernel_count_equal_i32")]] kernel kernel_count_equal_t kernel_count_equal<int32_t>;
+
+template <typename T>
+kernel void kernel_snake(
+        constant ggml_metal_kargs_snake & args,
+        device const T     * x,
+        device const float * a,
+        device const float * inv_b,
+        device       T     * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T * args.C) {
+        return;
+    }
+
+    const int   c  = idx / args.T;  // x is [T, C], a / inv_b collapse to [1, C]
+    const float xi = float(x[idx]);
+    const float si = sin(a[c] * xi);
+    dst[idx] = T(xi + si * si * inv_b[c]);
+}
+
+template [[host_name("kernel_snake_f32")]]  kernel void kernel_snake<float>(constant ggml_metal_kargs_snake &, device const float *, device const float *, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_snake_f16")]]  kernel void kernel_snake<half>(constant ggml_metal_kargs_snake &, device const half *, device const float *, device const float *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_snake_bf16")]] kernel void kernel_snake<bfloat>(constant ggml_metal_kargs_snake &, device const bfloat *, device const float *, device const float *, device bfloat *, uint, uint, uint);
+#endif
+
+template<int N>
+kernel void kernel_fwht_f32(
+        constant ggml_metal_kargs_fwht & args,
+        device const float * src,
+        device float * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort3  ntg[[threads_per_threadgroup]]) {
+
+    constexpr int NW = N_SIMDWIDTH;
+    constexpr int NE = N / NW;
+
+    const float scale = 1.0f / sqrt((float) N);
+
+    const int sg_per_tg = ntg.x / NW;
+    const int64_t r = tgpig.x * sg_per_tg + sgitg;
+    if (r >= args.nrows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    const int lane = tiisg;
+
+    float reg[NE];
+    for (int i = 0; i < NE; i++) {
+        reg[i] = src[i*NW + lane]*scale;
+    }
+    for (int i = 1; i < NW; i *= 2) {
+        for (int j = 0; j < NE; j++) {
+            const float val = reg[j];
+            const float val2 = simd_shuffle_xor(val, i);
+            reg[j] = (lane & i) == 0 ? val2 + val : val2 - val;
+        }
+    }
+
+    for (int i = NW; i < N; i *= 2) {
+        const int step = i / NW;
+        for (int j = 0; j < NE; j += (2 * step)) {
+            for (int k = 0; k < step; k++) {
+                const float x = reg[j + k ];
+                const float y = reg[j + k + step];
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+    for (int i = 0; i < NE; i++) {
+        dst[i*NW + lane] = reg[i];
+    }
+}
+
+typedef decltype(kernel_fwht_f32<64>) kernel_fwht_t;
+
+template [[host_name("kernel_fwht_f32_64")]]  kernel kernel_fwht_t kernel_fwht_f32<64>;
+template [[host_name("kernel_fwht_f32_128")]] kernel kernel_fwht_t kernel_fwht_f32<128>;
+template [[host_name("kernel_fwht_f32_256")]] kernel kernel_fwht_t kernel_fwht_f32<256>;
+template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_t kernel_fwht_f32<512>;
+

ggml/src/ggml-metal/kernels/mul_mm.metal

@@ -0,0 +1,842 @@
+#include "common.h"
+#include "dequantize.h"
+
+constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
+constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
+constant short FC_mul_mm_ne12  [[function_constant(FC_MUL_MM + 2)]];
+constant short FC_mul_mm_ne13  [[function_constant(FC_MUL_MM + 3)]];
+constant short FC_mul_mm_r2    [[function_constant(FC_MUL_MM + 4)]];
+constant short FC_mul_mm_r3    [[function_constant(FC_MUL_MM + 5)]];
+
+// each block_q contains 16*nl weights
+#ifdef GGML_METAL_HAS_TENSOR
+template<
+    typename SA, typename SA_4x4, typename SA_8x8,
+    typename SB, typename SB_2x4, typename SB_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread SA_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    // Matrix dimensions: A(M,K) x B(K,N) -> C(M,N)
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    // Batch dimension handling
+    const int im = tgpig.z;
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    // Batch offsets for srcA and srcB
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+
+    // Tile dimensions
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+
+    // Tile offsets in output matrix
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // Threadgroup memory for dequantized A tile only
+    threadgroup SA * sa = (threadgroup SA *)(shmem);
+
+    // Work-item count for A loading
+    constexpr int A_WORK_ITEMS = NRA * N_MM_NK;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+
+    // tA wraps threadgroup memory
+    auto tA = tensor(sa, dextents<int32_t, 2>(N_MM_NK_TOTAL, NRA));
+
+    // tB wraps device memory directly
+    device T1 * ptrB = (device T1 *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(T1);
+    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
+
+    // Configure matmul operation
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    // Accumulate partial results over K dimension
+    for (int loop_k = 0; loop_k < K; loop_k += N_MM_NK_TOTAL) {
+        // === PHASE 1: Dequantization of A into threadgroup memory ===
+        for (int work = tiitg; work < A_WORK_ITEMS; work += NUM_THREADS) {
+            const int row = work / N_MM_NK;
+            const int k_chunk = work % N_MM_NK;
+            const int k_pos = loop_k + k_chunk * 16;
+            const short k_base = k_chunk * 16;
+
+            // Bounds check: skip device read if row is out of matrix bounds
+            if (ra + row < M) {
+                if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+                    // Element-wise reads when K is not aligned (nb01 not aligned for half4x4/float4x4).
+                    // MSL spec Table 2.5: half4x4 requires 8-byte alignment. When K is odd,
+                    // nb01 = K*2 is not 8-byte aligned, so odd-row pointers are misaligned.
+                    // Mirrors the legacy kernel's existing guard.
+                    device const T0 * row_ptr = (device const T0 *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? (SA) row_ptr[k_pos + i] : (SA)0;
+                    }
+                } else {
+                    const int block_idx = k_pos / (16 * nl);
+                    const short il = (k_pos / 16) % nl;
+
+                    device const block_q * row_ptr = (device const block_q *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    SA_4x4 temp_a;
+                    dequantize_func(row_ptr + block_idx, il, temp_a);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        // Zero-pad A for K positions beyond valid range (handles partial K iterations)
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? temp_a[i/4][i%4] : (SA)0;
+                    }
+                }
+            } else {
+                // Zero-pad rows beyond matrix bounds
+                FOR_UNROLL (short i = 0; i < 16; i++) {
+                    sa[row * N_MM_NK_TOTAL + (k_base + i)] = (SA)0;
+                }
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // === PHASE 2: Tensor matmul ===
+        auto mA = tA.slice(0, 0);
+        auto mB = tB.slice(loop_k, rb);
+
+        mm.run(mB, mA, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // Store result tile to output matrix (with batch offset)
+    // cT.store handles bounds checking via tD's extents (M, N)
+    device float * dstBatch = (device float *)dst + im * N * M;
+
+    auto tD = tensor(dstBatch, dextents<int32_t, 2>(M, N), array<int, 2>({1, M}));
+    cT.store(tD.slice(ra, rb));
+}
+
+#else
+
+template<
+    typename S0, typename S0_4x4, typename S0_8x8,
+    typename S1, typename S1_2x4, typename S1_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+    }
+
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
+        // if no bounds checks on the output are needed, we can directly write to device memory
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (sgitg == 0) {
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*NR0);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < nr0/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < nr0; i++) {
+                    *(D + i) = *(C + i);
+                }
+            }
+        }
+    }
+}
+
+#endif // GGML_METAL_HAS_TENSOR
+
+template<short ne20> // n_expert_used
+kernel void kernel_mul_mm_id_map0(
+        constant ggml_metal_kargs_mul_mm_id_map0 & args,
+        device  const char * src2,
+        device        char * htpe,
+        device        char * hids,
+        threadgroup   char * shmem [[threadgroup(0)]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const short ide = tpitg; // expert id
+
+    uint32_t n_all = 0;
+
+    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;
+
+    for (int i21 = 0; i21 < args.ne21; i21 += ntg) { // n_tokens
+        if (i21 + tpitg < args.ne21) {
+            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);
+
+            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;
+
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sids[i20] = src2_i32[i20];
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (short t = 0; t < ntg; t++) {
+            if (i21 + t >= args.ne21) {
+                break;
+            }
+
+            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;
+
+            short sel = 0;
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sel += (sids[i20] == ide)*(i20 + 1);
+            }
+
+            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;
+
+            n_all += sel > 0;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
+    tpe_u32[ide] = n_all;
+}
+
+typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
+
+template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;
+
+template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm_id(
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * htpe,
+        device const char * hids,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
+    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
+
+    const int32_t neh1 = tpe_u32[im];
+
+    if (r1 >= neh1) {
+        return;
+    }
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
+
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? (S0) *((device T0 *) x + i) : (S0) 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifndef GGML_METAL_HAS_TENSOR
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
+    }
+
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+    for (short i = 0; i < 8; i++) {
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+    }
+#endif
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
+
+        const short ide = id % args.ne20;
+        const short idt = id / args.ne20;
+
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float4 * D4 = (device float4 *) D;
+
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
+        threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+        int i = tiisg;
+        for (; i < nr0/4; i += 32) {
+            *(D4 + i) = *(C4 + i);
+        }
+
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
+            *(D + i) = *(C + i);
+        }
+    }
+}
+
+//
+// matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_t;
+
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_q1_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_f32_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_f16_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_q1_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_id;
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_id_q1_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q1_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;

ggml/src/ggml-metal/kernels/norm.metal

@@ -0,0 +1,308 @@
+#include "common.h"
+
+// F == 1 : norm (no fuse)
+// F == 2 : norm + mul
+// F == 3 : norm + mul + add
+template <typename T, short F>
+kernel void kernel_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    T sumft(0.0f);
+
+    float sumf = 0.0f;
+
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumft += x[i00];
+    }
+    sumf = dot(sumft, T(1.0f));
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean = sumf/args.ne00;
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+
+    sumf = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        y[i00] = x[i00] - mean;
+        sumf += dot(y[i00], y[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float variance = sumf/args.ne00;
+
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (y[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (y[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (y[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_norm_fuse_impl<float4, 1>) kernel_norm_fuse_t;
+
+template [[host_name("kernel_norm_f32")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_norm_mul_f32")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_norm_mul_add_f32")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_norm_f32_4")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_norm_mul_f32_4")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_norm_mul_add_f32_4")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 3>;
+
+// F == 1 : rms_norm (no fuse)
+// F == 2 : rms_norm + mul
+// F == 3 : rms_norm + mul + add
+template <typename T, short F>
+kernel void kernel_rms_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean  = sumf/args.ne00;
+    const float scale = 1.0f/sqrt(mean + args.eps);
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (x[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (x[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (x[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_rms_norm_fuse_impl<float4, 1>) kernel_rms_norm_fuse_t;
+
+template [[host_name("kernel_rms_norm_f32")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_rms_norm_mul_f32")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_rms_norm_f32_4")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_rms_norm_mul_f32_4")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32_4")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 3>;
+
+template <typename T0, typename T>
+kernel void kernel_l2_norm_impl(
+        constant ggml_metal_kargs_l2_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const T0 * x = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T  * y = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float scale = 1.0f/max(sqrt(sumf), args.eps);
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        y[i00] = x[i00] * scale;
+    }
+}
+
+typedef decltype(kernel_l2_norm_impl<float, float>) kernel_l2_norm_t;
+
+template [[host_name("kernel_l2_norm_f32_f32")]]   kernel kernel_l2_norm_t kernel_l2_norm_impl<float,  float>;
+template [[host_name("kernel_l2_norm_f32_f32_4")]] kernel kernel_l2_norm_t kernel_l2_norm_impl<float4, float4>;
+
+kernel void kernel_group_norm_f32(
+        constant ggml_metal_kargs_group_norm & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    const int64_t ne = args.ne00*args.ne01*args.ne02;
+    const int64_t gs = args.ne00*args.ne01*((args.ne02 + args.ngrp - 1) / args.ngrp);
+
+    int start = tgpig * gs;
+    int end   = start + gs;
+
+    start += tpitg;
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += ntg) {
+        tmp += src0[j];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float mean = tmp / gs;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += ntg) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float variance = tmp / gs;
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int j = start; j < end; j += ntg) {
+        dst[j] *= scale;
+    }
+}

ggml/src/ggml-metal/kernels/pool.metal

@@ -0,0 +1,148 @@
+#include "common.h"
+
+kernel void kernel_pool_2d_max_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+
+    float res = -INFINITY;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            res = MAX(res, i_ptr[i * args.IW + j]);
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+kernel void kernel_pool_2d_avg_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+    // const float scale = 1. / ((eh - bh) * (ew - bw));
+    const float scale = 1. / (args.k0 * args.k1);
+
+    float res = 0;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            float cur = i_ptr[i * args.IW + j];
+            res += cur * scale;
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+
+kernel void kernel_pool_1d_max_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = -INFINITY;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        int j = base + ki;
+        if (j < 0 || j >= args.IW){
+            continue;
+        }
+        float v = src[src_off + j];
+        acc = max(acc, v);
+    }
+
+    dst[dst_off + ow] = acc;
+}
+
+kernel void kernel_pool_1d_avg_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = 0.0f;
+    int   cnt = 0;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        const int j = base + ki;
+        if (j < 0 || j >= args.IW) {
+            continue;
+        }
+        acc += src[src_off + j];
+        cnt += 1;
+    }
+
+    dst[dst_off + ow] = (cnt > 0) ? (acc / (float)cnt) : 0.0f;
+}

ggml/src/ggml-metal/kernels/quantize.h

@@ -0,0 +1,234 @@
+#pragma once
+
+#include "common.h"
+
+void quantize_q1_0(device const float * src, device block_q1_0 & dst) {
+    float sum_abs = 0.0f;
+    for (int j = 0; j < QK1_0; j++) {
+        sum_abs += fabs(src[j]);
+    }
+    dst.d = sum_abs / QK1_0;
+
+    for (int j = 0; j < QK1_0 / 8; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK1_0; j++) {
+        if (src[j] >= 0.0f) {
+            dst.qs[j / 8] |= (1 << (j % 8));
+        }
+    }
+}
+
+void quantize_q2_0(device const float * src, device block_q2_0 & dst) {
+    float amax = 0.0f;
+    for (int j = 0; j < QK2_0; j++) {
+        float a = fabs(src[j]);
+        if (a > amax) amax = a;
+    }
+    const float d = amax;
+    dst.d = d;
+
+    const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+    for (int j = 0; j < QK2_0 / 4; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK2_0; j++) {
+        int q = (int)round(src[j] * id) + 1;
+        q = max(0, min(3, q));
+        dst.qs[j / 4] |= (q << (2 * (j % 4)));
+    }
+}
+
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}

ggml/src/ggml-metal/kernels/quantize.metal

@@ -0,0 +1,393 @@
+#include "common.h"
+#include "dequantize.h"
+#include "quantize.h"
+
+template<typename T0, typename T1>
+kernel void kernel_cpy_t_t(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.ne00;) {
+        device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        dst_data[i00] = (T1) src[0];
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_t_t<float, float>) kernel_cpy_t;
+
+template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   float>;
+template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
+template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
+template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
+#endif
+template [[host_name("kernel_cpy_f16_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    float>;
+template [[host_name("kernel_cpy_f16_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  float>;
+template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  bfloat>;
+#endif
+
+template<short QK,
+         typename block_q,
+         void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_cpy_f32_q(
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK;
+
+    device block_q * dst_data = (device block_q *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        device const float * src = (device const float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + (i00*QK)*args.nb00);
+
+        quantize_func(src, dst_data[i00]);
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_f32_q<QK8_0,  block_q8_0,  quantize_q8_0>)  cpy_f_q_t;
+
+template [[host_name("kernel_cpy_f32_q8_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK8_0,  block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_cpy_f32_q1_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK1_0,  block_q1_0,   quantize_q1_0>;
+template [[host_name("kernel_cpy_f32_q2_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK2_0,  block_q2_0,   quantize_q2_0>;
+template [[host_name("kernel_cpy_f32_q4_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_0,  block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_cpy_f32_q4_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_1,  block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_cpy_f32_q5_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_0,  block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_cpy_f32_q5_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_1,  block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_cpy_f32_iq4_nl")]] kernel cpy_f_q_t kernel_cpy_f32_q<QK4_NL, block_iq4_nl, quantize_iq4_nl>;
+
+template<typename T4x4, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_cpy_q_f32(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device const block_q * src_data = (device const block_q *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T4x4    * dst_data = (device       T4x4    *)(dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        T4x4 temp;
+        dequantize_func(src_data + i00/nl, i00%nl, temp);
+        dst_data[i00] = temp;
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>) cpy_q_f_t;
+
+template [[host_name("kernel_cpy_q1_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_q1_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template<typename T>
+kernel void kernel_concat(
+        constant ggml_metal_kargs_concat & args,
+        device  const char * src0,
+        device  const char * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = ntg.y == 1 ? tgpig.x : tgpig.x*ntg.y + tpitg.y;
+
+    if (i1 >= args.ne1) {
+        return;
+    }
+
+    int o[4] = {0, 0, 0, 0};
+    o[args.dim] = args.dim == 0 ? args.ne00 : (args.dim == 1 ? args.ne01 : (args.dim == 2 ? args.ne02 : args.ne03));
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        device const T * x;
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            x = (device const T *)(src0 + (i3       )*args.nb03 + (i2       )*args.nb02 + (i1       )*args.nb01 + (i0       )*args.nb00);
+        } else {
+            x = (device const T *)(src1 + (i3 - o[3])*args.nb13 + (i2 - o[2])*args.nb12 + (i1 - o[1])*args.nb11 + (i0 - o[0])*args.nb10);
+        }
+
+        device T * y = (device T *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+        *y = *x;
+    }
+}
+
+typedef decltype(kernel_concat<float>) kernel_concat_t;
+
+template [[host_name("kernel_concat_f32")]]  kernel kernel_concat_t kernel_concat<float>;
+template [[host_name("kernel_concat_f16")]]  kernel kernel_concat_t kernel_concat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_concat_bf16")]] kernel kernel_concat_t kernel_concat<bfloat>;
+#endif
+template [[host_name("kernel_concat_i8")]]   kernel kernel_concat_t kernel_concat<char>;
+template [[host_name("kernel_concat_i16")]]  kernel kernel_concat_t kernel_concat<short>;
+template [[host_name("kernel_concat_i32")]]  kernel kernel_concat_t kernel_concat<int>;
+template [[host_name("kernel_concat_i64")]]  kernel kernel_concat_t kernel_concat<long>;
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
+kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg  [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (device const block_q *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (device      float4x4 *) ((      device char *) dst  + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        float4x4 temp;
+        dequantize_func(psrc + ind/nl, ind%nl, temp);
+        pdst[ind] = temp;
+
+        break;
+    }
+}
+
+template<typename T0, typename T>
+kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (const device T0 *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (      device T  *) ((      device char *)  dst + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        pdst[ind] = psrc[ind];
+
+        break;
+    }
+}
+
+typedef decltype(kernel_get_rows_f<float, float>) get_rows_f_t;
+
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_f_t kernel_get_rows_f<float, float>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_f_t kernel_get_rows_f<half,  float>;
+template [[host_name("kernel_get_rows_i32")]]  kernel get_rows_f_t kernel_get_rows_f<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_rows_f<bfloat, float>;
+#endif
+
+typedef decltype(kernel_get_rows_q<block_q4_0, 2, dequantize_q4_0>) get_rows_q_t;
+
+template [[host_name("kernel_get_rows_q1_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q1_0,    8, dequantize_q1_0>;
+template [[host_name("kernel_get_rows_q2_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_0,    4, dequantize_q2_0>;
+template [[host_name("kernel_get_rows_q4_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_0,    2, dequantize_q4_0>;
+template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_1,    2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
+template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_mxfp4")]]   kernel get_rows_q_t kernel_get_rows_q<block_mxfp4,   2, dequantize_mxfp4>;
+template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_get_rows_q5_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_get_rows_q6_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+
+template<typename TS, typename TI, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + 32*ind, dst_row[ind]);
+    }
+}
+
+template<typename TS, typename TI, typename TD>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device TD * dst_row = (      device TD *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS * src_row = (const device TS *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (TD) src_row[ind];
+    }
+}
+
+typedef decltype(kernel_set_rows_f<float, int64_t, float>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32_i64_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, float>;
+template [[host_name("kernel_set_rows_f32_i32_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, float>;
+template [[host_name("kernel_set_rows_f32_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, half>;
+template [[host_name("kernel_set_rows_f32_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_f32_i64_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_f32_i32_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int32_t, bfloat>;
+#endif
+
+template [[host_name("kernel_set_rows_f16_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int64_t, half>;
+template [[host_name("kernel_set_rows_f16_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_bf16_i64_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_bf16_i32_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int32_t, bfloat>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<float, int64_t, block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_f32_i64_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i32_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i32_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i32_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i64_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i32_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i64_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i32_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i64_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_set_rows_f32_i32_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_iq4_nl, quantize_iq4_nl>;
+

ggml/src/ggml-metal/kernels/reduce.metal

@@ -0,0 +1,228 @@
+#include "common.h"
+
+kernel void kernel_op_sum_f32(
+        constant ggml_metal_kargs_sum & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    if (args.np == 0) {
+        return;
+    }
+
+    // TODO: become function constant
+    const uint nsg = (ntg.x + 31) / 32;
+
+    float sumf = 0;
+
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+        sumf += src0[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float total = 0;
+
+    if (sgitg == 0) {
+        float v = 0;
+
+        if (tpitg.x < nsg) {
+            v = shmem_f32[tpitg.x];
+        }
+
+        total = simd_sum(v);
+
+        if (tpitg.x == 0) {
+            dst[0] = total;
+        }
+    }
+}
+
+constant short FC_sum_rows_op [[function_constant(FC_SUM_ROWS + 0)]];
+
+template <typename T0, typename T>
+kernel void kernel_sum_rows_impl(
+        constant ggml_metal_kargs_sum_rows & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_sum_rows_op
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    threadgroup T0 * shmem_t = (threadgroup T0 *) shmem;
+
+    if (sgitg == 0) {
+        shmem_t[tiisg] = 0.0f;
+    }
+
+    device const T0 * src_row = (device const T0 *) (src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T  * dst_row = (device       T  *) (dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    T0 sumf = T0(0.0f);
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_t[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_t[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        if (FC_OP == OP_SUM_ROWS_NUM_MEAN) {
+            if (is_same<float4, T0>::value) {
+                dst_row[0] = sum(sumf) / (4*args.ne00);
+            } else {
+                dst_row[0] = sum(sumf) / args.ne00;
+            }
+        } else {
+            dst_row[0] = sum(sumf);
+        }
+    }
+
+#undef FC_OP
+}
+
+typedef decltype(kernel_sum_rows_impl<float, float>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows_f32_f32")]]   kernel kernel_sum_rows_t kernel_sum_rows_impl<float,  float>;
+template [[host_name("kernel_sum_rows_f32_f32_4")]] kernel kernel_sum_rows_t kernel_sum_rows_impl<float4, float>;
+
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;

ggml/src/ggml-metal/kernels/rope.metal

@@ -0,0 +1,318 @@
+#include "common.h"
+
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+constant bool FC_rope_is_back   [[function_constant(FC_ROPE + 1)]];
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+    if (FC_rope_is_back) {
+        *sin_theta *= -1.0f;
+    }
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
+}
+
+template<typename T>
+kernel void kernel_rope_norm(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[1];
+
+            dst_data[0] = x0*cos_theta - x1*sin_theta;
+            dst_data[1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_neox(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_multi(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            // mrope theta calculations
+            // note: the rest is the same as kernel_rope_neox
+            const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3;
+            const int sec_w01   = args.sect_0 + args.sect_1;               // end of section 1
+            const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
+            const int sector    = ic % sect_dims;
+
+            float theta_base;
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            } else {
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            }
+            // end of mrope
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_vision(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < 2*args.n_dims) { // different from kernel_rope_multi
+            const int ic = i0/2;
+
+            // mrope theta calculations (only support 2 dimensions)
+            const int sect_dims = args.sect_0 + args.sect_1;
+            const int sector    = ic % sect_dims;
+
+            float p;
+            float theta_base;
+            if (sector < args.sect_1) {
+                p = (float) sector;
+                theta_base = (float) pos[i2];
+            } else {
+                p = (float) sector - args.sect_0;
+                theta_base = (float) pos[i2 + args.ne02];
+            }
+
+            const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
+            // end of mrope
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims]; // different from kernel_rope_multi
+
+            dst_data[0]           = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
+typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+typedef decltype(kernel_rope_multi<float>) kernel_rope_multi_t;
+typedef decltype(kernel_rope_vision<float>) kernel_rope_vision_t;
+
+template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
+template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
+
+template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
+template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
+
+template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi<float>;
+template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi<half>;
+
+template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision<float>;
+template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision<half>;

ggml/src/ggml-metal/kernels/softmax.metal

@@ -0,0 +1,223 @@
+#include "common.h"
+
+template<typename T>
+kernel void kernel_soft_max(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float * psrc2 = src2 != src0 ? (device const float *) (src2)                                                 : nullptr;
+    device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+
+    // find the max value in the block
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
+        lsum += exp_psrc0;
+        pdst[i00] = exp_psrc0;
+    }
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        pdst[i00] *= inv_sum;
+    }
+}
+
+template<typename T>
+kernel void kernel_soft_max_4(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float *  psrc2 = src2 != src0 ? (device const float * ) (src2)                                                 : nullptr;
+    device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
+    }
+
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        pdst4[i00] *= inv_sum;
+    }
+}
+
+typedef decltype(kernel_soft_max<float>)    kernel_soft_max_t;
+typedef decltype(kernel_soft_max_4<float4>) kernel_soft_max_4_t;
+
+template [[host_name("kernel_soft_max_f16")]]   kernel kernel_soft_max_t   kernel_soft_max<half>;
+template [[host_name("kernel_soft_max_f32")]]   kernel kernel_soft_max_t   kernel_soft_max<float>;
+template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<half4>;
+template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<float4>;

ggml/src/ggml-metal/kernels/solve_tri.metal

@@ -0,0 +1,75 @@
+#include "common.h"
+
+constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
+constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
+constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
+
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device   const char * src0,
+        device   const char * src1,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        ushort3 tgpig[[threadgroup_position_in_grid]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const short NSG = FC_solve_tri_nsg;
+    const short N   = FC_solve_tri_n;
+    const short K   = FC_solve_tri_k;
+    const short NP  = PAD2(N, NW);
+
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x*NSG + sgitg;
+
+    threadgroup float * sh0 = (threadgroup float *) shmem;
+
+    device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
+    device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
+    device       float * dst_ptr  = (device       float *)(dst  + i02 * args.nb2  + i03 * args.nb3)  + i01;
+
+    for (short rr = 0; rr < N; rr += NSG) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        {
+            threadgroup float * sh0_cur = sh0 + sgitg*NP;
+
+            for (short t = 0; t*NW < N; ++t) {
+                const short idx = t*NW + tiisg;
+                sh0_cur[idx] = src0_ptr[idx];
+            }
+
+            src0_ptr += NSG*N;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (i01 >= args.ne10) {
+            continue;
+        }
+
+        for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
+            const short r = rr + ir;
+
+            threadgroup float * sh0_cur = sh0 + ir*NP;
+
+            float sum = 0.0f;
+
+            for (short t = 0; t*NW < r; ++t) {
+                const short idx = t*NW + tiisg;
+                sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
+            }
+
+            sum = simd_sum(sum);
+
+            if (tiisg == 0) {
+                const float diag = sh0_cur[r];
+
+                dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
+            }
+        }
+    }
+}

ggml/src/ggml-metal/kernels/ssm.metal

@@ -0,0 +1,279 @@
+#include "common.h"
+
+// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
+kernel void kernel_ssm_conv_f32_f32(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+    device       float * x = (device       float *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+    device       float  * x = (device       float  *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+constant short FC_ssm_conv_bs   [[function_constant(FC_SSM_CONV + 0)]];
+
+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+kernel void kernel_ssm_conv_f32_f32_batched(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_batched_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// Optimized version: reduces redundant memory loads by having one thread load shared values
+kernel void kernel_ssm_scan_f32(
+        constant ggml_metal_kargs_ssm_scan & args,
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device const void * src6,
+        device      float * dst,
+        threadgroup float * shared [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgptg[[simdgroups_per_threadgroup]],
+        uint3    tgpg[[threadgroups_per_grid]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    // Shared memory layout:
+    // [0..sgptg*NW-1]: partial sums for reduction (existing)
+    // [sgptg*NW..sgptg*NW+sgptg-1]: pre-computed x_dt values for each token in batch
+    // [sgptg*NW+sgptg..sgptg*NW+2*sgptg-1]: pre-computed dA values for each token in batch
+    threadgroup float * shared_sums = shared;
+    threadgroup float * shared_x_dt = shared + sgptg * NW;
+    threadgroup float * shared_dA   = shared + sgptg * NW + sgptg;
+
+    shared_sums[tpitg.x] = 0.0f;
+
+    const int32_t i0 = tpitg.x;
+    const int32_t i1 = tgpig.x;
+    const int32_t ir = tgpig.y; // current head
+    const int32_t i3 = tgpig.z; // current seq
+
+    const int32_t nc  = args.d_state;
+    const int32_t nr  = args.d_inner;
+    const int32_t nh  = args.n_head;
+    const int32_t ng  = args.n_group;
+    const int32_t n_t = args.n_seq_tokens;
+
+    const int32_t s_off = args.s_off;
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0_buff = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s_buff  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+
+    const int32_t i = i0 + i1*nc;
+    const int32_t g = ir / (nh / ng); // repeat_interleave
+
+    float s0 = s0_buff[i];
+    float s  = 0.0f;
+
+    device const float * A = (device const float *) ((device const char *) src3 + ir*args.nb31); // {ne30, nh}
+
+    const float A0 = A[i0%args.ne30];
+
+    device const float * x  = (device const float *)((device const char *) src1 + i1*args.nb10  + ir*args.nb11 + i3*args.nb13); // {dim, nh, nt, ns}
+    device const float * dt = (device const float *)((device const char *) src2 + ir*args.nb20  + i3*args.nb22);                // {nh, nt, ns}
+    device const float * B  = (device const float *)((device const char *) src4 +  g*args.nb41  + i3*args.nb43);                // {d_state, ng, nt, ns}
+    device const float * C  = (device const float *)((device const char *) src5 +  g*args.nb51  + i3*args.nb53);                // {d_state, ng, nt, ns}
+
+    device float * y = dst + (i1 + ir*(nr) + i3*(n_t*nh*nr)); // {dim, nh, nt, ns}
+
+    for (int i2 = 0; i2 < n_t; i2 += sgptg) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // Pre-compute x_dt and dA for this batch of tokens
+        // Only first sgptg threads do the loads and expensive math
+        if (i0 < sgptg && i2 + i0 < n_t) {
+            // ns12 and ns21 are element strides (nb12/nb10, nb21/nb20)
+            device const float * x_t  = x  + i0 * args.ns12;
+            device const float * dt_t = dt + i0 * args.ns21;
+
+            const float dt0  = dt_t[0];
+            const float dtsp = dt0 <= 20.0f ? log(1.0f + exp(dt0)) : dt0;
+            shared_x_dt[i0] = x_t[0] * dtsp;
+            shared_dA[i0]   = dtsp;  // Store dtsp, compute exp(dtsp * A0) per-thread since A0 varies
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
+            const float x_dt = shared_x_dt[t];
+            const float dA   = exp(shared_dA[t] * A0);
+
+            s = (s0 * dA) + (B[i0] * x_dt);
+
+            const float sumf = simd_sum(s * C[i0]);
+
+            if (tiisg == 0) {
+                shared_sums[t*NW + sgitg] = sumf;
+            }
+
+            // recurse
+            s0 = s;
+
+            B  += args.ns42;
+            C  += args.ns52;
+        }
+
+        // Advance pointers for next batch
+        x  += sgptg * args.ns12;
+        dt += sgptg * args.ns21;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float sumf = simd_sum(shared_sums[sgitg*NW + tiisg]);
+
+        if (tiisg == 0 && i2 + sgitg < n_t) {
+            y[sgitg*nh*nr] = sumf;
+        }
+
+        y += sgptg*nh*nr;
+    }
+
+    s_buff[i] = s;
+}

ggml/src/ggml-metal/kernels/tri.metal

@@ -0,0 +1,69 @@
+#include "common.h"
+
+template<uint32_t ttype>
+bool _ggml_vec_tri_cmp(const int i, const int r);
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER */ 3>(const int i, const int r) {
+    return i < r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER_DIAG */ 2>(const int i, const int r) {
+    return i <= r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER */ 1>(const int i, const int r) {
+    return i > r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER_DIAG */ 0>(const int i, const int r) {
+    return i >= r;
+}
+
+template<typename T, int ttype>
+kernel void kernel_tri(
+        constant ggml_metal_kargs_tri & args,
+        device const char * src0,
+        device const char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    device const T * src_row = (device const T *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T * dst_row = (device       T *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    // Each thread is a single element of the row if ne00 < max threads per
+    // threadgroup, so this will loop once for each index that this thread is
+    // responsible for
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        // Use the comparison as a mask for branchless
+        dst_row[i0] = static_cast<T>(_ggml_vec_tri_cmp<ttype>(i0, i1)) * src_row[i0];
+    }
+}
+
+typedef decltype(kernel_tri<float, 0>) kernel_tri_t;
+
+template [[host_name("kernel_tri_f32_0")]] kernel kernel_tri_t kernel_tri<float, 0>;
+template [[host_name("kernel_tri_f32_1")]] kernel kernel_tri_t kernel_tri<float, 1>;
+template [[host_name("kernel_tri_f32_2")]] kernel kernel_tri_t kernel_tri<float, 2>;
+template [[host_name("kernel_tri_f32_3")]] kernel kernel_tri_t kernel_tri<float, 3>;
+template [[host_name("kernel_tri_f16_0")]] kernel kernel_tri_t kernel_tri<half, 0>;
+template [[host_name("kernel_tri_f16_1")]] kernel kernel_tri_t kernel_tri<half, 1>;
+template [[host_name("kernel_tri_f16_2")]] kernel kernel_tri_t kernel_tri<half, 2>;
+template [[host_name("kernel_tri_f16_3")]] kernel kernel_tri_t kernel_tri<half, 3>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_tri_bf16_0")]] kernel kernel_tri_t kernel_tri<bfloat, 0>;
+template [[host_name("kernel_tri_bf16_1")]] kernel kernel_tri_t kernel_tri<bfloat, 1>;
+template [[host_name("kernel_tri_bf16_2")]] kernel kernel_tri_t kernel_tri<bfloat, 2>;
+template [[host_name("kernel_tri_bf16_3")]] kernel kernel_tri_t kernel_tri<bfloat, 3>;
+#endif

ggml/src/ggml-metal/kernels/unary.metal

@@ -0,0 +1,360 @@
+#include "common.h"
+
+constant short FC_unary_op [[function_constant(FC_UNARY + 0)]];
+constant bool  FC_unary_cnt[[function_constant(FC_UNARY + 1)]];
+
+template <typename T0, typename T, typename TC>
+kernel void kernel_unary_impl(
+        constant ggml_metal_kargs_unary & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_unary_op
+#define FC_CNT FC_unary_cnt
+
+    device const T0 * src0_ptr;
+    device       T  * dst_ptr;
+
+    int i0;
+
+    if (FC_CNT) {
+        i0 = tgpig.x;
+
+        src0_ptr = (device const T0 *) (src0);
+        dst_ptr  = (device       T  *) (dst);
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int k0  = tgpig.x/args.ne01;
+        const int i01 = tgpig.x - k0*args.ne01;
+
+        i0 = k0*ntg.x + tpitg.x;
+
+        src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+        dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1 );
+    }
+
+    {
+        //threadgroup_barrier(mem_flags::mem_none);
+
+        if (!FC_CNT) {
+            if (i0 >= args.ne0) {
+                return;
+            }
+        }
+
+        const TC x = (TC) src0_ptr[i0];
+
+        if (FC_OP == OP_UNARY_NUM_SCALE) {
+            dst_ptr[i0] = (T) (args.scale * x + args.bias);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FILL) {
+            dst_ptr[i0] = (T) args.val;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CLAMP) {
+            dst_ptr[i0] = (T) clamp(x, args.min, args.max);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQR) {
+            dst_ptr[i0] = (T) (x * x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQRT) {
+            dst_ptr[i0] = (T) sqrt(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIN) {
+            dst_ptr[i0] = (T) sin(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_COS) {
+            dst_ptr[i0] = (T) cos(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LOG) {
+            dst_ptr[i0] = (T) log(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LEAKY_RELU) {
+            dst_ptr[i0] = (T) (TC(x > 0)*x + TC(x <= 0)*(x * args.slope));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TANH) {
+            dst_ptr[i0] = (T) precise::tanh(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_RELU) {
+            dst_ptr[i0] = (T) fmax(0, x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIGMOID) {
+            dst_ptr[i0] = (T) (1 / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + precise::tanh(SQRT_2_OVER_PI*x*(1 + GELU_COEF_A*x*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_ERF) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + erf_approx(SQRT_2_INV*x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_QUICK) {
+            dst_ptr[i0] = (T) (x * (1/(1 + exp(GELU_QUICK_COEF*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SILU) {
+            dst_ptr[i0] = (T) (x / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ELU) {
+            dst_ptr[i0] = (T) elu_approx(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_NEG) {
+            dst_ptr[i0] = (T) -x;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ABS) {
+            dst_ptr[i0] = (T) fabs(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SGN) {
+            dst_ptr[i0] = T(x > 0) - T(x < 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_STEP) {
+            dst_ptr[i0] = T(x > 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSWISH) {
+            dst_ptr[i0] = (T) (x * fmax(0, fmin(1, x/6 + 0.5)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSIGMOID) {
+            dst_ptr[i0] = (T) fmax(0, fmin(1, x/6 + 0.5));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXP) {
+            dst_ptr[i0] = (T) exp(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SOFTPLUS) {
+            dst_ptr[i0] = (T) select(log(1 + exp(x)), x, x > 20);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXPM1) {
+            // TODO: precise implementation
+            dst_ptr[i0] = (T) (exp(x) - 1);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FLOOR) {
+            dst_ptr[i0] = (T) floor(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CEIL) {
+            dst_ptr[i0] = (T) ceil(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ROUND) {
+            dst_ptr[i0] = (T) round(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TRUNC) {
+            dst_ptr[i0] = (T) trunc(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_XIELU) {
+            const TC xi      = x;
+            const TC gate    = TC(xi > TC(0.0f));
+            const TC clamped = fmin(xi, TC(args.val));
+            const TC y_pos   = TC(args.scale) * xi * xi + TC(args.bias) * xi;
+            const TC y_neg   = (exp(clamped) - TC(1.0f) - xi) * TC(args.slope) + TC(args.bias) * xi;
+            dst_ptr[i0] = (T) (gate * y_pos + (TC(1.0f) - gate) * y_neg);
+        }
+    }
+
+#undef FC_OP
+#undef FC_CNT
+}
+
+typedef decltype(kernel_unary_impl<float, float, float>) kernel_unary_t;
+
+template [[host_name("kernel_unary_f32_f32")]]   kernel kernel_unary_t kernel_unary_impl<float,  float,  float>;
+template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_unary_impl<float4, float4, float4>;
+template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
+template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;
+
+template<typename T>
+kernel void kernel_reglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = (T)(x0*x1*(x0 > 0.0f));
+    }
+}
+
+typedef decltype(kernel_reglu<float>) kernel_reglu_t;
+
+template [[host_name("kernel_reglu_f32")]] kernel kernel_reglu_t kernel_reglu<float>;
+template [[host_name("kernel_reglu_f16")]] kernel kernel_reglu_t kernel_reglu<half>;
+
+template<typename T>
+kernel void kernel_geglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + precise::tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = (T)(gelu*x1);
+    }
+}
+
+typedef decltype(kernel_geglu<float>) kernel_geglu_t;
+
+template [[host_name("kernel_geglu_f32")]] kernel kernel_geglu_t kernel_geglu<float>;
+template [[host_name("kernel_geglu_f16")]] kernel kernel_geglu_t kernel_geglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = (T)(silu*x1);
+    }
+}
+
+typedef decltype(kernel_swiglu<float>) kernel_swiglu_t;
+
+template [[host_name("kernel_swiglu_f32")]] kernel kernel_swiglu_t kernel_swiglu<float>;
+template [[host_name("kernel_swiglu_f16")]] kernel kernel_swiglu_t kernel_swiglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu_oai(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, args.limit);
+        x1 = max(min(x1, args.limit), -args.limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * args.alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = (T)out_glu;
+    }
+}
+
+typedef decltype(kernel_swiglu_oai<float>) kernel_swiglu_oai_t;
+
+template [[host_name("kernel_swiglu_oai_f32")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<float>;
+template [[host_name("kernel_swiglu_oai_f16")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<half>;
+
+template<typename T>
+kernel void kernel_geglu_erf(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
+
+        dst_row[i0] = (T)(gelu_erf*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_erf<float>) kernel_geglu_erf_t;
+
+template [[host_name("kernel_geglu_erf_f32")]] kernel kernel_geglu_erf_t kernel_geglu_erf<float>;
+template [[host_name("kernel_geglu_erf_f16")]] kernel kernel_geglu_erf_t kernel_geglu_erf<half>;
+
+template<typename T>
+kernel void kernel_geglu_quick(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = (T)(gelu_quick*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_quick<float>) kernel_geglu_quick_t;
+
+template [[host_name("kernel_geglu_quick_f32")]] kernel kernel_geglu_quick_t kernel_geglu_quick<float>;
+template [[host_name("kernel_geglu_quick_f16")]] kernel kernel_geglu_quick_t kernel_geglu_quick<half>;

ggml/src/ggml-metal/kernels/upscale.metal

@@ -0,0 +1,179 @@
+#include "common.h"
+
+constant bool FC_upscale_aa [[function_constant(FC_UPSCALE + 0)]];
+
+kernel void kernel_upscale_nearest_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3/args.sf3;
+    const int64_t i02 = i2/args.sf2;
+    const int64_t i01 = i1/args.sf1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int64_t i00 = i0/args.sf0;
+
+        device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1  +  i0*args.nb0);
+
+        dst_ptr[0] = src0_ptr[0];
+    }
+}
+
+static inline float bilinear_tri(float x) {
+    return MAX(0.0f, 1.0f - fabs(x));
+}
+
+kernel void kernel_upscale_bilinear_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01  = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01  = MAX(0, MIN(args.ne01 - 1, (int64_t)floor(f01)));
+    const int64_t i01p = MAX(0, MIN(args.ne01 - 1, i01 + 1));
+    const float   fd1  = MAX(0.0f, MIN(1.0f, f01 - (float)i01));
+
+    src0 += i03*args.nb03 + i02*args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1);
+
+    if (FC_upscale_aa) {
+        const float support0  = MAX(1.0f, 1.0f / args.sf0);
+        const float invscale0 = 1.0f / support0;
+        const float support1  = MAX(1.0f, 1.0f / args.sf1);
+        const float invscale1 = 1.0f / support1;
+
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+
+            int64_t x_min = MAX((int64_t)0, (int64_t)floor(f00 - support0 + args.poffs));
+            int64_t x_max = MIN(args.ne00,  (int64_t)ceil (f00 + support0 + args.poffs));
+
+            int64_t y_min = MAX((int64_t)0, (int64_t)floor(f01 - support1 + args.poffs));
+            int64_t y_max = MIN(args.ne01,  (int64_t)ceil (f01 + support1 + args.poffs));
+
+            float sum = 0.0f;
+            float wsum = 0.0f;
+
+            for (int64_t sy = y_min; sy < y_max; ++sy) {
+                const float wy = MAX(0.0f, 1.0f - fabs((float)sy - f01) * invscale1);
+                for (int64_t sx = x_min; sx < x_max; ++sx) {
+                    const float wx = MAX(0.0f, 1.0f - fabs((float)sx - f00) * invscale0);
+                    const float w  = wx * wy;
+                    device const float * src_ptr = (device const float *)(src0 + sy*args.nb01 + sx*args.nb00);
+                    sum  += (*src_ptr) * w;
+                    wsum += w;
+                }
+            }
+
+            const float v = (wsum > 0.0f) ? (sum / wsum) : 0.0f;
+            dst_ptr[i0] = v;
+        }
+    } else {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float   f00  = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+            const int64_t i00  = MAX(0, MIN(args.ne00 - 1, (int64_t)floor(f00)));
+            const int64_t i00p = MAX(0, MIN(args.ne00 - 1, i00 + 1));
+            const float   fd0  = MAX(0.0f, MIN(1.0f, f00 - (float)i00));
+
+            device const float * src00 = (device const float *)(src0 + i01*args.nb01  + i00*args.nb00);
+            device const float * src10 = (device const float *)(src0 + i01*args.nb01  + i00p*args.nb00);
+            device const float * src01 = (device const float *)(src0 + i01p*args.nb01 + i00*args.nb00);
+            device const float * src11 = (device const float *)(src0 + i01p*args.nb01 + i00p*args.nb00);
+
+            const float v =
+                (*src00) * (1.0f - fd0) * (1.0f - fd1) +
+                (*src10) * fd0          * (1.0f - fd1) +
+                (*src01) * (1.0f - fd0) * fd1 +
+                (*src11) * fd0          * fd1;
+
+            dst_ptr[i0] = v;
+        }
+    }
+}
+
+static inline float bicubic_weight1(float x) {
+    const float a = -0.75f;
+    return ((a + 2) * x - (a + 3)) * x * x + 1;
+}
+
+static inline float bicubic_weight2(float x) {
+    const float a = -0.75f;
+    return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a;
+}
+
+kernel void kernel_upscale_bicubic_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01 = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01 = (int64_t)floor(f01);
+    const float   fd1 = f01 - (float)i01;
+
+    const float w_y0 = bicubic_weight2(fd1 + 1.0f);
+    const float w_y1 = bicubic_weight1(fd1);
+    const float w_y2 = bicubic_weight1(1.0f - fd1);
+    const float w_y3 = bicubic_weight2(2.0f - fd1);
+
+    const device const char * src_slice = src0 + i03 * args.nb03 + i02 * args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3 * args.nb3 + i2 * args.nb2 + i1 * args.nb1);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const float   f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+        const int64_t i00 = (int64_t)floor(f00);
+        const float   fd0 = f00 - (float)i00;
+
+        const float w_x0 = bicubic_weight2(fd0 + 1.0f);
+        const float w_x1 = bicubic_weight1(fd0);
+        const float w_x2 = bicubic_weight1(1.0f - fd0);
+        const float w_x3 = bicubic_weight2(2.0f - fd0);
+
+        float sum = 0.0f;
+
+        for (int dy = -1; dy <= 2; ++dy) {
+            const int64_t iy = MAX(0, MIN(args.ne01 - 1, i01 + dy));
+            const float wy = (dy == -1) ? w_y0 : (dy == 0) ? w_y1 : (dy == 1) ? w_y2 : w_y3;
+
+            for (int dx = -1; dx <= 2; ++dx) {
+                const int64_t ix = MAX(0, MIN(args.ne00 - 1, i00 + dx));
+                const float wx = (dx == -1) ? w_x0 : (dx == 0) ? w_x1 : (dx == 1) ? w_x2 : w_x3;
+
+                device const float * src_ptr = (device const float *)(src_slice + iy * args.nb01 + ix * args.nb00);
+                sum += (*src_ptr) * wx * wy;
+            }
+        }
+
+        dst_ptr[i0] = sum;
+    }
+}

ggml/src/ggml-metal/kernels/wkv.metal

@@ -0,0 +1,179 @@
+#include "common.h"
+
+kernel void kernel_rwkv_wkv6_f32(
+    device const float * k,
+    device const float * v,
+    device const float * r,
+    device const float * tf,
+    device const float * td,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _k[head_size];
+    threadgroup float _r[head_size];
+    threadgroup float _tf[head_size];
+    threadgroup float _td[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    _tf[tid] = tf[head_id * head_size + tid];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0;
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 tf_vec = float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
+            float4 td_vec = float4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            float4 temp = tf_vec * kv + s_vec;
+            y += dot(r_vec, temp);
+
+            s_vec = s_vec * td_vec + kv;
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
+
+kernel void kernel_rwkv_wkv7_f32(
+    device const float * r,
+    device const float * w,
+    device const float * k,
+    device const float * v,
+    device const float * a,
+    device const float * b,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _r[head_size];
+    threadgroup float _w[head_size];
+    threadgroup float _k[head_size];
+    threadgroup float _a[head_size];
+    threadgroup float _b[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + tid * head_size + i];
+    }
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _r[tid] = r[t];
+        _w[tid] = w[t];
+        _k[tid] = k[t];
+        _a[tid] = a[t];
+        _b[tid] = b[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0, sa = 0.0;
+
+        float4 sa_vec(0.0);
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 a_vec = float4(_a[j], _a[j+1], _a[j+2], _a[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+            sa_vec += a_vec * s_vec;
+        }
+        sa = sa_vec[0] + sa_vec[1] + sa_vec[2] + sa_vec[3];
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 w_vec = float4(_w[j], _w[j+1], _w[j+2], _w[j+3]);
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 b_vec = float4(_b[j], _b[j+1], _b[j+2], _b[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            s_vec = s_vec * w_vec + kv + sa * b_vec;
+            y += dot(s_vec, r_vec);
+
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + tid * head_size + i] = state[i];
+    }
+}

ggml/src/ggml-sycl/CMakeLists.txt

@@ -199,9 +199,20 @@ if (GGML_SYCL_DEVICE_ARCH)
         -fsycl-targets=spir64_gen
         "SHELL:-Xsycl-target-backend=spir64_gen \"-device ${GGML_SYCL_DEVICE_ARCH}\""
     )
+
+    # Pass through parallel job (process) count for parallelising the
+    # `llvm-foreach -- ocloc` invocation for compiling AOT device images.
+    include(ProcessorCount)
+    ProcessorCount(_ggml_sycl_nproc)
+    if (_ggml_sycl_nproc LESS 1)
+        set(_ggml_sycl_nproc 1)
+    endif()
+    set(GGML_SYCL_MAX_PARALLEL_LINK_JOBS ${_ggml_sycl_nproc} CACHE STRING
+        "Parallel ocloc jobs for spir64_gen AOT device-image lowering")
     target_link_options(
         ggml-sycl PRIVATE
         -fsycl-targets=spir64_gen
         "SHELL:-Xsycl-target-backend=spir64_gen \"-device ${GGML_SYCL_DEVICE_ARCH}\""
+        -fsycl-max-parallel-link-jobs=${GGML_SYCL_MAX_PARALLEL_LINK_JOBS}
     )
 endif()

ggml/src/ggml-sycl/common.hpp

@@ -65,6 +65,7 @@ extern int g_ggml_sycl_prioritize_dmmv;
 extern int g_ggml_sycl_enable_flash_attention;
 extern int g_ggml_sycl_dev2dev_memcpy;
 extern int g_ggml_sycl_fa_onednn;
+extern int g_ggml_sycl_fa_onednn_max_kv;
 
 
 #if defined(__clang__) && __has_builtin(__builtin_expect)

ggml/src/ggml-sycl/fattn-onednn.cpp

@@ -38,6 +38,12 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
     if (K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16) {
         return false;
     }
+    // Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
+    // very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
+    // some stacks; past the cap we fall back to the native FA kernel instead.
+    if (g_ggml_sycl_fa_onednn_max_kv > 0 && K->ne[1] > g_ggml_sycl_fa_onednn_max_kv) {
+        return false;
+    }
     // gate for the following cases
     // 1. if the oneDNN graph Add node has no input --> skip
     // 2. types other than f16 need different logical_tensor declaration
@@ -208,9 +214,17 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
     cont_to_f16_sycl<sycl::half>((const char *) V->data, Vf.get(), d, seq, Hkv, mb, V->nb[1], V->nb[2], V->nb[3], stream);
 
     // divide-by-(1/scale) reproduces ggml's score *= kq_scale on the proven probe graph.
+    //
+    // The scale must not be uploaded with an async memcpy from a stack local: on the in-order
+    // queue that copy waits behind the K/V staging kernels, and once those take long enough
+    // (n_kv >= ~26k on B70) the host frame is recycled before the copy runs, feeding the SDPA a
+    // garbage scale (output collapses to a repeated token). Write the scalar from a kernel
+    // instead -- the value is captured into the command, so no host memory has to outlive the
+    // call, and the enqueue stays async.
     const sycl::half scale_h = (sycl::half) (1.0f / kq_scale);
     ggml_sycl_pool_alloc<sycl::half> scbuf(ctx.pool(), 1);
-    stream->memcpy(scbuf.get(), &scale_h, sizeof(sycl::half));
+    sycl::half * const scale_dev = scbuf.get();
+    stream->single_task([=]() { *scale_dev = scale_h; });
 
     ggml_sycl_pool_alloc<sycl::half> outf(ctx.pool(), (size_t) H * q * d);   // f16 contiguous SDPA out [mb,H,q,d]
 
@@ -232,7 +246,7 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
         if (r == E.id_q)     return Qf.get();
         if (r == E.id_k)     return Kf.get();
         if (r == E.id_v)     return Vf.get();
-        if (r == E.id_scale) return scbuf.get();
+        if (r == E.id_scale) return scale_dev;
         if (r == E.id_mask)  return (void *) mask->data;
         return nullptr;
     };
@@ -245,14 +259,12 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
     E.cp.execute(strm, ti, {to});
 
     permute_sdpa_out_sycl(outf.get(), (float *) dst->data, mb, H, q, d, stream);
-    // Single device: no sync is required, and actually PP perf is ~6% > wait_and_throw() (tested on llama-3.1-8b & qwen3.6-27b, both Q8_0, with Arc B70).
-    // Any future multi-GPU refactor MUST re-measure this single-device path and keep the best
-    // single-device PP speed. Otherwise (multiple devices/streams can race the reuse):
+    // Single device needs no sync: the dnnl stream wraps this same in-order queue, so the SDPA
+    // serializes with the staging kernels before it and the permute/pool reuse after it. The
+    // garbage output formerly blamed on the missing sync here was the scale use-after-return
+    // fixed above. Keep the conservative wait for multi-GPU, where other devices' streams can
+    // race the pool:
     if (ggml_sycl_info().device_count > 1) {
-        // cont_to_f16 -> oneDNN execute -> permute is async on this stream, but the
-        // pool_alloc*s above free their device buffers at host return. Without this wait the next
-        // scheduler op re-acquires those bytes while the GPU is still computing the SDPA, turning
-        // it into garbage and collapsing multi-turn output to a single repeated token ("GGGGG...").
         stream->wait_and_throw();
     }
 }

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -85,6 +85,7 @@ int g_ggml_sycl_enable_optimize = 1;
 int g_ggml_sycl_enable_graph = 0;
 int g_ggml_sycl_enable_dnn = 1;
 int g_ggml_sycl_fa_onednn = 1;
+int g_ggml_sycl_fa_onednn_max_kv = 0;
 int g_ggml_sycl_enable_vmm = 1;
 int g_ggml_sycl_enable_fusion = 1;
 int g_ggml_sycl_prioritize_dmmv = 0;
@@ -287,6 +288,7 @@ static void ggml_check_sycl() try {
         g_ggml_sycl_enable_graph = ggml_sycl_get_env("GGML_SYCL_ENABLE_GRAPH", 0);
         g_ggml_sycl_enable_dnn = ggml_sycl_get_env("GGML_SYCL_ENABLE_DNN", 1);
         g_ggml_sycl_fa_onednn = ggml_sycl_get_env("GGML_SYCL_FA_ONEDNN", 1);
+        g_ggml_sycl_fa_onednn_max_kv = ggml_sycl_get_env("GGML_SYCL_FA_ONEDNN_MAX_KV", 0);
         g_ggml_sycl_enable_vmm = ggml_sycl_get_env("GGML_SYCL_ENABLE_VMM", 1);
         g_ggml_sycl_enable_fusion = ggml_sycl_get_env("GGML_SYCL_ENABLE_FUSION", 1);
         g_ggml_sycl_prioritize_dmmv = ggml_sycl_get_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
@@ -359,6 +361,7 @@ static void ggml_check_sycl() try {
         GGML_LOG_INFO("  GGML_SYCL_ENABLE_DNN: DNN disabled by compile flag\n");
         GGML_LOG_INFO("  GGML_SYCL_FA_ONEDNN: %d\n", g_ggml_sycl_fa_onednn);
 #endif
+        GGML_LOG_INFO("  GGML_SYCL_FA_ONEDNN_MAX_KV: %d\n", g_ggml_sycl_fa_onednn_max_kv);
 #ifdef SYCL_FLASH_ATTN
         GGML_LOG_INFO("  GGML_SYCL_ENABLE_FLASH_ATTN: %d\n", g_ggml_sycl_enable_flash_attention);
 #else

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

@@ -3490,7 +3490,7 @@ struct vk_fa_tuning_params {
 };
 
 static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type k_type, ggml_type v_type);
-static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type k_type = GGML_TYPE_F16);
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type k_type = GGML_TYPE_F16, ggml_type v_type = GGML_TYPE_F16);
 
 static vk_fa_tuning_params get_fa_tuning_params_scalar(const vk_device& device, uint32_t hsk, uint32_t hsv, uint32_t n_rows, uint32_t n_kv, ggml_type k_type, ggml_type v_type, bool f32acc) {
 
@@ -3646,7 +3646,7 @@ static vk_fa_tuning_params get_fa_tuning_params(const vk_device& device, uint32_
         bool shape_ok = (f32acc && device->coopmat_support_16x16x16_f32acc) ||
                         (!f32acc && device->coopmat_support_16x16x16_f16acc);
         const vk_fa_tuning_params params = get_fa_tuning_params_coopmat1(device, hsk, hsv, n_rows, n_kv, k_type, v_type, f32acc);
-        bool shmem_ok = ggml_vk_flash_attn_coopmat_shmem_support(device, params, hsk, hsv, f32acc, k_type);
+        bool shmem_ok = ggml_vk_flash_attn_coopmat_shmem_support(device, params, hsk, hsv, f32acc, k_type, v_type);
 
         if (!shape_ok || !shmem_ok) {
             path = FA_SCALAR;
@@ -3658,11 +3658,6 @@ static vk_fa_tuning_params get_fa_tuning_params(const vk_device& device, uint32_
         path = FA_SCALAR;
     }
 
-    // Q1_0 K/V is only implemented on coopmat2 (flash_attn_cm2); there is no scalar FA shader for it.
-    if ((k_type == GGML_TYPE_Q1_0 || v_type == GGML_TYPE_Q1_0) && device->coopmat2) {
-        path = FA_COOPMAT2;
-    }
-
     switch (path) {
     case FA_SCALAR:
         return get_fa_tuning_params_scalar(device, hsk, hsv, n_rows, n_kv, k_type, v_type, f32acc);
@@ -3904,16 +3899,27 @@ static uint32_t get_subgroup_size(const std::string &pipeline_name, const vk_dev
     return 0; // If no matching configuration is found
 }
 
-// Whether scalar flash attention will use the MMQ path for the given k_type.
-static bool ggml_vk_fa_scalar_uses_mmq(const vk_device& device, ggml_type k_type) {
+// Whether scalar flash attention will use the MMQ path for the given K/V types.
+static bool ggml_vk_fa_type_needs_shmem(ggml_type type) {
+    switch (type) {
+    case GGML_TYPE_IQ4_NL:
+        return true;
+    default:
+        return false;
+    }
+}
+
+static bool ggml_vk_fa_scalar_uses_mmq(const vk_device& device, ggml_type k_type, ggml_type v_type) {
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
     return device->integer_dot_product && device->subgroup_clustered &&
+           !ggml_vk_fa_type_needs_shmem(v_type) &&
            (k_type == GGML_TYPE_Q4_0 || k_type == GGML_TYPE_Q4_1 ||
             k_type == GGML_TYPE_Q5_0 || k_type == GGML_TYPE_Q5_1 ||
             k_type == GGML_TYPE_Q8_0);
 #else
     GGML_UNUSED(device);
     GGML_UNUSED(k_type);
+    GGML_UNUSED(v_type);
     return false;
 #endif
 }
@@ -4246,7 +4252,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         const bool fa_ds = fa.first.subgroup_size == 0;
 
         const bool bf16_kv = fa.first.k_type == GGML_TYPE_BF16;
-        const bool use_mmq = ggml_vk_fa_scalar_uses_mmq(device, fa.first.k_type);
+        const bool use_mmq = ggml_vk_fa_scalar_uses_mmq(device, fa.first.k_type, fa.first.v_type);
         const void * spv_data = nullptr;
         size_t spv_size = 0;
         const char *name = nullptr;
@@ -10380,7 +10386,6 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
 
 static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type k_type, ggml_type v_type) {
     GGML_UNUSED(f32acc);
-    GGML_UNUSED(v_type);
     // Needs to be kept up to date on shader changes
     const uint32_t wg_size = params.workgroup_size;
     const uint32_t Br = params.block_rows;
@@ -10389,13 +10394,15 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con
     // BF16 uses the fp32 shader (FLOAT_TYPE=float)
     const uint32_t float_type_size = (device->fp16 && k_type != GGML_TYPE_BF16) ? sizeof(ggml_fp16_t) : sizeof(float);
 
-    const bool mmq = ggml_vk_fa_scalar_uses_mmq(device, k_type);
+    const bool mmq = ggml_vk_fa_scalar_uses_mmq(device, k_type, v_type);
 
     // tmpsh is overestimated slightly
     const uint32_t tmpsh = wg_size * sizeof(float);
     const uint32_t tmpshv4 = wg_size * 4 * float_type_size;
 
     const uint32_t masksh = Bc * (Br + 1) * float_type_size;
+    // DATA_A_IQ4_NL is compiled into the FA shaders unconditionally, so its shared table is always allocated.
+    const uint32_t iq_shmem = 16 * float_type_size;
 
     uint32_t Qf, kvsh, kblocksh_size;
     if (mmq) {
@@ -10420,7 +10427,7 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con
         kblocksh_size = 0;
     }
 
-    const uint32_t total_size = tmpsh + tmpshv4 + masksh + Qf + kvsh + kblocksh_size;
+    const uint32_t total_size = tmpsh + tmpshv4 + masksh + iq_shmem + Qf + kvsh + kblocksh_size;
     const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
 
     VK_LOG_DEBUG("ggml_vk_flash_attn_scalar_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", mmq=" << mmq << ", total_size=" << total_size << ", supported=" << supported);
@@ -10428,7 +10435,8 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con
     return supported;
 }
 
-static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type k_type) {
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type k_type, ggml_type v_type) {
+    GGML_UNUSED(v_type);
     // Needs to be kept up to date on shader changes
     const uint32_t Br = params.block_rows;
     const uint32_t Bc = params.block_cols;
@@ -10444,6 +10452,8 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     const uint32_t f16vec4 = 8;
 
     const uint32_t tmpsh = (Bc / MatBc) * sizeof(float);
+    // DATA_A_IQ4_NL is compiled into the FA shaders unconditionally, so its shared table is always allocated.
+    const uint32_t iq_shmem = 16 * sizeof(ggml_fp16_t);
 
     const uint32_t qstride = hsk_pad / 4 + 2;
     const uint32_t Qf = Br * qstride * f16vec4;
@@ -10465,7 +10475,7 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
 
     const uint32_t slope = Br * acctype;
 
-    const uint32_t total_size = tmpsh + Qf + Psh + sfsh + ksh + pvsh + slope;
+    const uint32_t total_size = tmpsh + iq_shmem + Qf + Psh + sfsh + ksh + pvsh + slope;
     const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
 
     VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
@@ -17617,7 +17627,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) {
                     return false;
                 }
-                auto fa_kv_ok = [coopmat2](ggml_type t) {
+                auto fa_kv_ok = [](ggml_type t) {
                     switch (t) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
@@ -17627,9 +17637,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_Q5_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_IQ4_NL:
                         return true;
-                    case GGML_TYPE_Q1_0:
-                        return coopmat2;
                     default:
                         return false;
                     }

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

@@ -80,7 +80,9 @@ shared vec4 occupancy_limiter[LIMIT_OCCUPANCY_SHMEM > 0 ? LIMIT_OCCUPANCY_SHMEM
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
-    init_iq_shmem(gl_WorkGroupSize);
+    if (fa_type_needs_shmem(FaTypeK) || fa_type_needs_shmem(FaTypeV)) {
+        init_iq_shmem(gl_WorkGroupSize);
+    }
 #endif
 
     init_indices();

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

@@ -97,8 +97,8 @@ layout (binding = 6) readonly buffer MO {uint32_t data_mask_opt[];};
 #define FA_TYPE_Q5_0  6u
 #define FA_TYPE_Q5_1  7u
 #define FA_TYPE_Q8_0  8u
+#define FA_TYPE_IQ4_NL 20u
 #define FA_TYPE_BF16 30u
-#define FA_TYPE_Q1_0 41u
 
 #if defined(BFLOAT16)
 #define O_TYPE float
@@ -120,8 +120,8 @@ uint fa_block_elems(uint ty) {
         case FA_TYPE_Q5_0: return uint(QUANT_K_Q5_0);
         case FA_TYPE_Q5_1: return uint(QUANT_K_Q5_1);
         case FA_TYPE_Q8_0: return uint(QUANT_K_Q8_0);
+        case FA_TYPE_IQ4_NL: return uint(QUANT_K_IQ4_NL);
         case FA_TYPE_BF16: return 1u;
-        case FA_TYPE_Q1_0: return uint(QUANT_K_Q1_0); // cm2-only, harmless elsewhere
         default:           return 1u;
     }
 }
@@ -140,6 +140,13 @@ uint fa_quant_r_mmq(uint ty) {
     }
 }
 
+bool fa_type_needs_shmem(uint ty) {
+    switch (ty) {
+        case FA_TYPE_IQ4_NL: return true;
+        default:             return false;
+    }
+}
+
 // These can't be `const` globals because GLSL forbids function calls in global
 // const initializers, even when the spec constants would let the driver fold
 // them. Macros expand at the use site and fold after specialization.

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -64,7 +64,9 @@ shared ACC_TYPE slope[Br];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
-    init_iq_shmem(gl_WorkGroupSize);
+    if (fa_type_needs_shmem(FaTypeK) || fa_type_needs_shmem(FaTypeV)) {
+        init_iq_shmem(gl_WorkGroupSize);
+    }
 #endif
 
     init_indices();

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -46,7 +46,7 @@ float16_t faDecodeK(const decodeBufFA_K bl_in, const uint blockCoords[2], const
         case FA_TYPE_Q5_0: return dequantFuncQ5_0(decodeBufQ5_0(bl_in), blockCoords, coordInBlock);
         case FA_TYPE_Q5_1: return dequantFuncQ5_1(decodeBufQ5_1(bl_in), blockCoords, coordInBlock);
         case FA_TYPE_Q8_0: return dequantFuncQ8_0(decodeBufQ8_0(bl_in), blockCoords, coordInBlock);
-        case FA_TYPE_Q1_0: return dequantFuncQ1_0(decodeBufQ1_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_IQ4_NL: return dequantFuncIQ4_NL(decodeBufIQ4_NL(bl_in), blockCoords, coordInBlock);
         default: return float16_t(0);
     }
 }
@@ -59,7 +59,7 @@ float16_t faDecodeV(const decodeBufFA_V bl_in, const uint blockCoords[2], const
         case FA_TYPE_Q5_0: return dequantFuncQ5_0(decodeBufQ5_0(bl_in), blockCoords, coordInBlock);
         case FA_TYPE_Q5_1: return dequantFuncQ5_1(decodeBufQ5_1(bl_in), blockCoords, coordInBlock);
         case FA_TYPE_Q8_0: return dequantFuncQ8_0(decodeBufQ8_0(bl_in), blockCoords, coordInBlock);
-        case FA_TYPE_Q1_0: return dequantFuncQ1_0(decodeBufQ1_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_IQ4_NL: return dequantFuncIQ4_NL(decodeBufIQ4_NL(bl_in), blockCoords, coordInBlock);
         default: return float16_t(0);
     }
 }
@@ -67,26 +67,26 @@ float16_t faDecodeV(const decodeBufFA_V bl_in, const uint blockCoords[2], const
 // V=4 vector decode for K/V; dispatches to per-format _v decoders.
 f16vec4 faDecodeKVector(const decodeBufFA_K bl_in, const uint blockCoords[2], const uint coordInBlock[2]) {
     switch (FaTypeK) {
-        case 0u: return f16vec4(decodeBufF32(bl_in).block);
-        case 2u: return dequantFuncQ4_0_v(decodeBufQ4_0(bl_in), blockCoords, coordInBlock);
-        case 3u: return dequantFuncQ4_1_v(decodeBufQ4_1(bl_in), blockCoords, coordInBlock);
-        case 6u: return dequantFuncQ5_0_v(decodeBufQ5_0(bl_in), blockCoords, coordInBlock);
-        case 7u: return dequantFuncQ5_1_v(decodeBufQ5_1(bl_in), blockCoords, coordInBlock);
-        case 8u: return dequantFuncQ8_0_v(decodeBufQ8_0(bl_in), blockCoords, coordInBlock);
-        case 41u: return dequantFuncQ1_0_v(decodeBufQ1_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_F32: return f16vec4(decodeBufF32(bl_in).block);
+        case FA_TYPE_Q4_0: return dequantFuncQ4_0_v(decodeBufQ4_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_Q4_1: return dequantFuncQ4_1_v(decodeBufQ4_1(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_Q5_0: return dequantFuncQ5_0_v(decodeBufQ5_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_Q5_1: return dequantFuncQ5_1_v(decodeBufQ5_1(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_Q8_0: return dequantFuncQ8_0_v(decodeBufQ8_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_IQ4_NL: return dequantFuncIQ4_NL_v(decodeBufIQ4_NL(bl_in), blockCoords, coordInBlock);
         default: return f16vec4(0);
     }
 }
 
 f16vec4 faDecodeVVector(const decodeBufFA_V bl_in, const uint blockCoords[2], const uint coordInBlock[2]) {
     switch (FaTypeV) {
-        case 0u: return f16vec4(decodeBufF32(bl_in).block);
-        case 2u: return dequantFuncQ4_0_v(decodeBufQ4_0(bl_in), blockCoords, coordInBlock);
-        case 3u: return dequantFuncQ4_1_v(decodeBufQ4_1(bl_in), blockCoords, coordInBlock);
-        case 6u: return dequantFuncQ5_0_v(decodeBufQ5_0(bl_in), blockCoords, coordInBlock);
-        case 7u: return dequantFuncQ5_1_v(decodeBufQ5_1(bl_in), blockCoords, coordInBlock);
-        case 8u: return dequantFuncQ8_0_v(decodeBufQ8_0(bl_in), blockCoords, coordInBlock);
-        case 41u: return dequantFuncQ1_0_v(decodeBufQ1_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_F32: return f16vec4(decodeBufF32(bl_in).block);
+        case FA_TYPE_Q4_0: return dequantFuncQ4_0_v(decodeBufQ4_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_Q4_1: return dequantFuncQ4_1_v(decodeBufQ4_1(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_Q5_0: return dequantFuncQ5_0_v(decodeBufQ5_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_Q5_1: return dequantFuncQ5_1_v(decodeBufQ5_1(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_Q8_0: return dequantFuncQ8_0_v(decodeBufQ8_0(bl_in), blockCoords, coordInBlock);
+        case FA_TYPE_IQ4_NL: return dequantFuncIQ4_NL_v(decodeBufIQ4_NL(bl_in), blockCoords, coordInBlock);
         default: return f16vec4(0);
     }
 }
@@ -169,6 +169,12 @@ ACC_TYPE perElemOpNonGqaSplitKStoreCol0(const in uint32_t r, const in uint32_t c
 }
 
 void main() {
+#ifdef NEEDS_INIT_IQ_SHMEM
+    if (fa_type_needs_shmem(FaTypeK) || fa_type_needs_shmem(FaTypeV)) {
+        init_iq_shmem(gl_WorkGroupSize);
+    }
+#endif
+
     init_indices();
 
     tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutQ = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
@@ -302,7 +308,7 @@ void main() {
         coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, HSK_pad, Bc, gl_MatrixUseB> K_T;
 
         uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
-        // F16: bs_k==1 (direct load). F32: bs_k==4 (vec4 / dequantFuncF32). Q4/Q8 family: bs_k==32. Q1_0: bs_k==128.
+        // F16: bs_k==1 (direct load). F32: bs_k==4 (vec4 / dequantFuncF32). Quantized types: bs_k==32.
 #if defined(BFLOAT16)
         coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK_pad), tensorViewTranspose);
 #else

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

@@ -27,6 +27,8 @@ layout (binding = 1) readonly buffer K_PACKED_Q5_1 { block_q5_1_packed16 data[];
 layout (binding = 2) readonly buffer V_PACKED_Q5_1 { block_q5_1_packed16 data[]; } v_packed_q5_1;
 layout (binding = 1) readonly buffer K_PACKED_Q8_0 { block_q8_0_packed16 data[]; } k_packed_q8_0;
 layout (binding = 2) readonly buffer V_PACKED_Q8_0 { block_q8_0_packed16 data[]; } v_packed_q8_0;
+layout (binding = 1) readonly buffer K_PACKED_IQ4_NL { block_iq4_nl_packed16 data[]; } k_packed_iq4_nl;
+layout (binding = 2) readonly buffer V_PACKED_IQ4_NL { block_iq4_nl_packed16 data[]; } v_packed_iq4_nl;
 
 layout (binding = 1) readonly buffer K_PACKED_BF16 { u16vec4 data[]; } k_packed_bf16;
 layout (binding = 2) readonly buffer V_PACKED_BF16 { u16vec4 data[]; } v_packed_bf16;
@@ -102,6 +104,17 @@ layout (binding = 1) readonly buffer K_PACKED_Q5_1_P32 { block_q5_1_packed32 dat
     return FLOAT_TYPE(BUF.data[a_offset + ib].d) * FLOAT_TYPEV4(v0.x, v0.y, v1.x, v1.y);          \
 }
 
+#define FA_DEQUANT4_IQ4_NL(BUF) {                                                                 \
+    const uint shift = (iqs & 0x10) >> 2;                                                         \
+    const uint qs_i = (iqs & 0xC) >> 1;                                                           \
+    const uint qsw = uint(BUF.data[a_offset + ib].qs[qs_i])                                       \
+                   | (uint(BUF.data[a_offset + ib].qs[qs_i + 1u]) << 16);                         \
+    const FLOAT_TYPE d = FLOAT_TYPE(BUF.data[a_offset + ib].d);                                   \
+    const u8vec4 q = unpack8((qsw >> shift) & 0x0F0F0F0Fu);                                       \
+    return d * FLOAT_TYPEV4(kvalues_iq4nl[q.x], kvalues_iq4nl[q.y],                               \
+                            kvalues_iq4nl[q.z], kvalues_iq4nl[q.w]);                              \
+}
+
 #define FA_DEQUANT4_BF16(BUF) \
     return FLOAT_TYPEV4(bf16_to_fp32(uvec4(BUF.data[(a_offset + ib) / 4])));
 
@@ -114,6 +127,7 @@ FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
             case FA_TYPE_Q5_0: FA_DEQUANT4_Q5_0(k_packed_q5_0)
             case FA_TYPE_Q5_1: FA_DEQUANT4_Q5_1(k_packed_q5_1)
             case FA_TYPE_Q8_0: FA_DEQUANT4_Q8_0(k_packed_q8_0)
+            case FA_TYPE_IQ4_NL: FA_DEQUANT4_IQ4_NL(k_packed_iq4_nl)
             case FA_TYPE_BF16: FA_DEQUANT4_BF16(k_packed_bf16)
         }
     } else {
@@ -124,6 +138,7 @@ FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
             case FA_TYPE_Q5_0: FA_DEQUANT4_Q5_0(v_packed_q5_0)
             case FA_TYPE_Q5_1: FA_DEQUANT4_Q5_1(v_packed_q5_1)
             case FA_TYPE_Q8_0: FA_DEQUANT4_Q8_0(v_packed_q8_0)
+            case FA_TYPE_IQ4_NL: FA_DEQUANT4_IQ4_NL(v_packed_iq4_nl)
             case FA_TYPE_BF16: FA_DEQUANT4_BF16(v_packed_bf16)
         }
     }

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

@@ -673,6 +673,8 @@ void process_shaders() {
             fa_base_dict["ACC_TYPE"] = fp16 && f16acc ? "float16_t" : "float";
             fa_base_dict["ACC_TYPEV2"] = fp16 && f16acc ? "f16vec2" : "vec2";
             fa_base_dict["ACC_TYPEV4"] = fp16 && f16acc ? "f16vec4" : "vec4";
+            // Compile IQ4_NL support into all FA variants so its shared LUT is available when K or V uses it.
+            fa_base_dict["DATA_A_IQ4_NL"] = "1";
             if (fp16 && f16acc) {
                 fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
             }

ggml/src/ggml.c

@@ -7854,7 +7854,9 @@ void ggml_set_input(struct ggml_tensor * tensor) {
 }
 
 void ggml_set_output(struct ggml_tensor * tensor) {
-    tensor->flags |= GGML_TENSOR_FLAG_OUTPUT;
+    for (struct ggml_tensor * cur = tensor; cur != NULL; cur = cur->view_src) {
+        cur->flags |= GGML_TENSOR_FLAG_OUTPUT;
+    }
 }
 
 void ggml_set_param(struct ggml_tensor * tensor) {

gguf-py/gguf/constants.py

@@ -145,6 +145,8 @@ class Keys:
         TOKEN_SHIFT_COUNT                 = "{arch}.token_shift_count"
         INTERLEAVE_MOE_LAYER_STEP         = "{arch}.interleave_moe_layer_step"
         FULL_ATTENTION_INTERVAL           = "{arch}.full_attention_interval"
+        NUM_LOOPS                         = "{arch}.num_loops"
+        SKIP_LOOP_FINAL_NORM              = "{arch}.skip_loop_final_norm"
         HASH_LAYER_COUNT                  = "{arch}.hash_layer_count"
         ACTIVATION_SPARSITY_SCALE         = "{arch}.activation_sparsity_scale"
         ALTUP_ACTIVE_IDX                  = "{arch}.altup.active_idx"
@@ -159,6 +161,7 @@ class Keys:
         TARGET_HIDDEN_SIZE                = "{arch}.target_hidden_size"
         BLOCK_SIZE                        = "{arch}.block_size"
         NORM_BEFORE_RESIDUAL              = "{arch}.norm_before_residual"
+        NORM_BEFORE_FC                    = "{arch}.norm_before_fc"
 
     class Attention:
         HEAD_COUNT                   = "{arch}.attention.head_count"
@@ -370,10 +373,17 @@ class Keys:
         FEED_FORWARD_LENGTH = "clip.audio.feed_forward_length"
         PROJECTION_DIM      = "clip.audio.projection_dim"
         BLOCK_COUNT         = "clip.audio.block_count"
+        SUBSAMPLING_FACTOR  = "clip.audio.subsampling_factor"
         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
+        RVQ_NUM_QUANTIZERS  = "clip.audio.rvq.num_quantizers"
+        RVQ_CODEBOOK_SIZE   = "clip.audio.rvq.codebook_size"
+        WA_PATTERN_MODE     = "clip.audio.wa_pattern_mode" # per-layer -1 (full) / 0 (windowed)
+        WINDOW_SIZE         = "clip.audio.window_size"
+        LOCAL_BLOCK_COUNT   = "clip.audio.local_block_count" # mimo-v2.5: input_local_transformer layer count
+        LOCAL_GROUP_SIZE    = "clip.audio.local_group_size"  # mimo-v2.5: input_local_transformer grouping size
 
         class Attention:
             HEAD_COUNT      = "clip.audio.attention.head_count"
@@ -545,6 +555,7 @@ class MODEL_ARCH(IntEnum):
     KIMI_LINEAR      = auto()
     TALKIE           = auto()
     MELLUM           = auto()
+    NANBEIGE         = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -942,6 +953,9 @@ class MODEL_TENSOR(IntEnum):
     A_ENC_FFN_SCALE_1     = auto() # gemma3n
     A_ENC_FFN_GATE_1      = auto() # lfm2, gemma3n
     A_ENC_FFN_DOWN_1      = auto() # lfm2, gemma3n
+    A_ENC_DOWNSAMPLE_CONV = auto() # mimo-audio-tokenizer: post-transformer downsample conv
+    A_ENC_DOWNSAMPLE_NORM = auto() # mimo-audio-tokenizer: post-transformer downsample norm
+    A_ENC_RVQ_CODEBOOK    = auto() # mimo-audio-tokenizer: residual vector quantizer codebook, per quantizer index
     A_MMPROJ              = auto()
     A_MMPROJ_FC           = auto()
     A_MM_NORM_PRE         = auto()
@@ -950,6 +964,17 @@ class MODEL_TENSOR(IntEnum):
     A_MM_HARD_EMB_NORM    = auto() # gemma3n
     A_MM_SOFT_EMB_NORM    = auto() # gemma3n
     A_MM_INP_PROJ         = auto() # gemma3n
+    A_MM_CODE_EMBD        = auto() # mimo: text-side RVQ code embedding table ("text codebook"), merged 3D [n_channels, vocab, dim]
+    A_MM_LOCAL_ATTN_Q     = auto() # mimo: input_local_transformer (LLM-side connector)
+    A_MM_LOCAL_ATTN_K     = auto()
+    A_MM_LOCAL_ATTN_V     = auto()
+    A_MM_LOCAL_ATTN_OUT   = auto()
+    A_MM_LOCAL_FFN_GATE   = auto()
+    A_MM_LOCAL_FFN_UP     = auto()
+    A_MM_LOCAL_FFN_DOWN   = auto()
+    A_MM_LOCAL_LN1        = auto()
+    A_MM_LOCAL_LN2        = auto()
+    A_MM_LOCAL_NORM       = auto() # final norm after all input_local_transformer layers
     A_PER_DIM_K_SCALE     = auto() # gemma4
     A_PER_DIM_SCALE       = auto() # gemma4
     # nextn/mtp
@@ -964,6 +989,10 @@ class MODEL_TENSOR(IntEnum):
     # eagle3
     FC                     = auto()  # feature fusion layer
     D2T                    = auto()  # draft to target vocabulary mapping
+    # dspark
+    DSPARK_MARKOV_W1       = auto()  # markov head: prev-token embed
+    DSPARK_MARKOV_W2       = auto()  # markov head: bias projection
+    DSPARK_CONF_PROJ       = auto()  # confidence head
     # lfm2 audio
     A_ENC_NORM_CONV        = auto()
     A_ENC_LINEAR_POS       = auto()
@@ -974,6 +1003,10 @@ class MODEL_TENSOR(IntEnum):
     A_ENC_CONV_NORM        = auto() # SSM conv
     A_ENC_CONV_PW1         = auto()
     A_ENC_CONV_PW2         = auto()
+    A_ENC_CONV_NORM_MEAN   = auto() # parakeet
+    A_ENC_CONV_NORM_VAR    = auto() # parakeet
+    A_ENC_MEL_FILTERS      = auto() # parakeet
+    A_ENC_WINDOW           = auto() # parakeet
     A_CTC_OUT              = auto()
     A_CTC_OUT_MID          = auto()
     A_ENC_ATTN_REL_POS_EMB = auto()
@@ -1134,6 +1167,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.KIMI_LINEAR:      "kimi-linear",
     MODEL_ARCH.TALKIE:           "talkie",
     MODEL_ARCH.MELLUM:           "mellum",
+    MODEL_ARCH.NANBEIGE:         "nanbeige",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1528,6 +1562,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.A_ENC_FFN_UP_1:            "a.blk.{bid}.ffn_up_1",
     MODEL_TENSOR.A_ENC_FFN_GATE_1:          "a.blk.{bid}.ffn_gate_1",
     MODEL_TENSOR.A_ENC_FFN_DOWN_1:          "a.blk.{bid}.ffn_down_1",
+    MODEL_TENSOR.A_ENC_DOWNSAMPLE_CONV:     "a.downsample.conv",
+    MODEL_TENSOR.A_ENC_DOWNSAMPLE_NORM:     "a.downsample.norm",
+    MODEL_TENSOR.A_ENC_RVQ_CODEBOOK:        "a.rvq.codebook",
     MODEL_TENSOR.A_MMPROJ:                  "mm.a.mlp.{bid}",
     MODEL_TENSOR.A_MMPROJ_FC:               "mm.a.fc",
     MODEL_TENSOR.A_MM_NORM_PRE:             "mm.a.norm_pre",
@@ -1536,6 +1573,17 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.A_MM_SOFT_EMB_NORM:        "mm.a.soft_emb_norm",         # gemma3n
     MODEL_TENSOR.A_MM_EMBEDDING:            "mm.a.embedding",             # gemma3n
     MODEL_TENSOR.A_MM_HARD_EMB_NORM:        "mm.a.hard_emb_norm",         # gemma3n
+    MODEL_TENSOR.A_MM_CODE_EMBD:            "mm.a.code_embd",
+    MODEL_TENSOR.A_MM_LOCAL_ATTN_Q:         "mm.a.local_blk.{bid}.attn_q",
+    MODEL_TENSOR.A_MM_LOCAL_ATTN_K:         "mm.a.local_blk.{bid}.attn_k",
+    MODEL_TENSOR.A_MM_LOCAL_ATTN_V:         "mm.a.local_blk.{bid}.attn_v",
+    MODEL_TENSOR.A_MM_LOCAL_ATTN_OUT:       "mm.a.local_blk.{bid}.attn_out",
+    MODEL_TENSOR.A_MM_LOCAL_FFN_GATE:       "mm.a.local_blk.{bid}.ffn_gate",
+    MODEL_TENSOR.A_MM_LOCAL_FFN_UP:         "mm.a.local_blk.{bid}.ffn_up",
+    MODEL_TENSOR.A_MM_LOCAL_FFN_DOWN:       "mm.a.local_blk.{bid}.ffn_down",
+    MODEL_TENSOR.A_MM_LOCAL_LN1:            "mm.a.local_blk.{bid}.ln1",
+    MODEL_TENSOR.A_MM_LOCAL_LN2:            "mm.a.local_blk.{bid}.ln2",
+    MODEL_TENSOR.A_MM_LOCAL_NORM:           "mm.a.local_norm",
     MODEL_TENSOR.A_PER_DIM_K_SCALE:         "a.blk.{bid}.per_dim_k_scale", # gemma4
     MODEL_TENSOR.A_PER_DIM_SCALE:           "a.blk.{bid}.per_dim_scale",   # gemma4
     # lfm2 audio
@@ -1548,6 +1596,10 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.A_ENC_CONV_NORM:           "a.blk.{bid}.conv_norm",
     MODEL_TENSOR.A_ENC_CONV_PW1:            "a.blk.{bid}.conv_pw1",
     MODEL_TENSOR.A_ENC_CONV_PW2:            "a.blk.{bid}.conv_pw2",
+    MODEL_TENSOR.A_ENC_CONV_NORM_MEAN:      "a.blk.{bid}.conv_norm_mean",
+    MODEL_TENSOR.A_ENC_CONV_NORM_VAR:       "a.blk.{bid}.conv_norm_var",
+    MODEL_TENSOR.A_ENC_MEL_FILTERS:         "a.mel_filters",
+    MODEL_TENSOR.A_ENC_WINDOW:              "a.window",
     MODEL_TENSOR.A_CTC_OUT:                 "a.enc_ctc_out",
     MODEL_TENSOR.A_CTC_OUT_MID:             "a.enc_ctc_out_mid",
     MODEL_TENSOR.A_ENC_ATTN_REL_POS_EMB:    "a.blk.{bid}.attn_rel_pos_emb",
@@ -1578,6 +1630,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD:    "blk.{bid}.nextn.shared_head_head",
     MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM:    "blk.{bid}.nextn.shared_head_norm",
     MODEL_TENSOR.FC:                        "fc",
+    MODEL_TENSOR.DSPARK_MARKOV_W1:          "markov_w1",
+    MODEL_TENSOR.DSPARK_MARKOV_W2:          "markov_w2",
+    MODEL_TENSOR.DSPARK_CONF_PROJ:          "conf_proj",
     MODEL_TENSOR.D2T:                       "d2t",
 }
 
@@ -1737,10 +1792,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.A_ENC_FFN_UP_1,
         MODEL_TENSOR.A_ENC_FFN_GATE_1,
         MODEL_TENSOR.A_ENC_FFN_DOWN_1,
+        MODEL_TENSOR.A_ENC_DOWNSAMPLE_CONV,
+        MODEL_TENSOR.A_ENC_DOWNSAMPLE_NORM,
+        MODEL_TENSOR.A_ENC_RVQ_CODEBOOK,
         MODEL_TENSOR.A_MMPROJ,
         MODEL_TENSOR.A_MMPROJ_FC,
         MODEL_TENSOR.A_MM_NORM_PRE,
         MODEL_TENSOR.A_MM_NORM_MID,
+        MODEL_TENSOR.A_MM_CODE_EMBD,
+        MODEL_TENSOR.A_MM_LOCAL_ATTN_Q,
+        MODEL_TENSOR.A_MM_LOCAL_ATTN_K,
+        MODEL_TENSOR.A_MM_LOCAL_ATTN_V,
+        MODEL_TENSOR.A_MM_LOCAL_ATTN_OUT,
+        MODEL_TENSOR.A_MM_LOCAL_FFN_GATE,
+        MODEL_TENSOR.A_MM_LOCAL_FFN_UP,
+        MODEL_TENSOR.A_MM_LOCAL_FFN_DOWN,
+        MODEL_TENSOR.A_MM_LOCAL_LN1,
+        MODEL_TENSOR.A_MM_LOCAL_LN2,
+        MODEL_TENSOR.A_MM_LOCAL_NORM,
         MODEL_TENSOR.A_ENC_NORM_CONV,
         MODEL_TENSOR.A_ENC_LINEAR_POS,
         MODEL_TENSOR.A_ENC_POS_BIAS_U,
@@ -1750,6 +1819,10 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.A_ENC_CONV_NORM,
         MODEL_TENSOR.A_ENC_CONV_PW1,
         MODEL_TENSOR.A_ENC_CONV_PW2,
+        MODEL_TENSOR.A_ENC_CONV_NORM_MEAN,
+        MODEL_TENSOR.A_ENC_CONV_NORM_VAR,
+        MODEL_TENSOR.A_ENC_MEL_FILTERS,
+        MODEL_TENSOR.A_ENC_WINDOW,
         MODEL_TENSOR.A_MM_INP_PROJ,
         MODEL_TENSOR.A_MM_SOFT_EMB_NORM,
         MODEL_TENSOR.A_MM_EMBEDDING,
@@ -4291,6 +4364,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.FC,
+        MODEL_TENSOR.ENC_OUTPUT_NORM,
         MODEL_TENSOR.D2T,
     ],
     MODEL_ARCH.DFLASH: [
@@ -4308,6 +4382,10 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.FC,
         MODEL_TENSOR.ENC_OUTPUT_NORM,
+        # optional DSpark heads
+        MODEL_TENSOR.DSPARK_MARKOV_W1,
+        MODEL_TENSOR.DSPARK_MARKOV_W2,
+        MODEL_TENSOR.DSPARK_CONF_PROJ,
     ],
     MODEL_ARCH.MISTRAL4: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -4505,7 +4583,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
-    # TODO
+    MODEL_ARCH.NANBEIGE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
 }
 
 # tensors that will not be serialized
@@ -4572,6 +4665,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
     ],
+    MODEL_ARCH.NANBEIGE: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }
 
 #
@@ -4777,10 +4874,12 @@ class VisionProjectorType:
     YOUTUVL = "youtuvl"
     NEMOTRON_V2_VL = "nemotron_v2_vl"
     HUNYUANVL      = "hunyuanvl"
+    PARAKEET       = "parakeet"  # audio
     MINIMAXM3      = "minimax_m3"
     MINICPMV4_6    = "minicpmv4_6"
     GRANITE_SPEECH = "granite_speech"  # audio
     MIMOVL         = "mimovl"
+    MIMO_AUDIO     = "mimo_audio"
     GRANITE4_VISION = "granite4_vision"
 
 

gguf-py/gguf/gguf_writer.py

@@ -908,6 +908,12 @@ class GGUFWriter:
     def add_token_shift_count(self, count: int) -> None:
         self.add_uint32(Keys.LLM.TOKEN_SHIFT_COUNT.format(arch=self.arch), count)
 
+    def add_num_loops(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.NUM_LOOPS.format(arch=self.arch), count)
+
+    def add_skip_loop_final_norm(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.SKIP_LOOP_FINAL_NORM.format(arch=self.arch), value)
+
     def add_interleave_moe_layer_step(self, value: int) -> None:
         self.add_uint32(Keys.LLM.INTERLEAVE_MOE_LAYER_STEP.format(arch=self.arch), value)
 
@@ -965,6 +971,9 @@ class GGUFWriter:
     def add_norm_before_residual(self, value: bool) -> None:
         self.add_bool(Keys.LLM.NORM_BEFORE_RESIDUAL.format(arch=self.arch), value)
 
+    def add_norm_before_fc(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.NORM_BEFORE_FC.format(arch=self.arch), value)
+
     def add_attention_output_group_count(self, count: int) -> None:
         self.add_uint32(Keys.Attention.OUTPUT_GROUP_COUNT.format(arch=self.arch), count)
 
@@ -1344,9 +1353,30 @@ class GGUFWriter:
     def add_audio_num_mel_bins(self, value: int) -> None:
         self.add_uint32(Keys.ClipAudio.NUM_MEL_BINS, value)
 
+    def add_audio_rvq_num_quantizers(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.RVQ_NUM_QUANTIZERS, value)
+
+    def add_audio_rvq_codebook_size(self, values: Sequence[int]) -> None:
+        self.add_array(Keys.ClipAudio.RVQ_CODEBOOK_SIZE, values)
+
+    def add_audio_wa_pattern_mode(self, modes: Sequence[int]) -> None:
+        self.add_array(Keys.ClipAudio.WA_PATTERN_MODE, modes)
+
+    def add_audio_window_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.WINDOW_SIZE, value)
+
+    def add_audio_local_block_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.LOCAL_BLOCK_COUNT, value)
+
+    def add_audio_local_group_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.LOCAL_GROUP_SIZE, value)
+
     def add_audio_stack_factor(self, value: int) -> None:
         self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value)
 
+    def add_audio_subsampling_factor(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.SUBSAMPLING_FACTOR, value)
+
     def add_audio_chunk_size(self, value: int) -> None:
         self.add_uint32(Keys.ClipAudio.CHUNK_SIZE, value)
 

gguf-py/gguf/tensor_mapping.py

@@ -1304,6 +1304,18 @@ class TensorNameMap:
             "model.fc", # dflash
         ),
 
+        MODEL_TENSOR.DSPARK_MARKOV_W1: (
+            "model.markov_head.markov_w1", # dspark
+        ),
+
+        MODEL_TENSOR.DSPARK_MARKOV_W2: (
+            "model.markov_head.markov_w2", # dspark
+        ),
+
+        MODEL_TENSOR.DSPARK_CONF_PROJ: (
+            "model.confidence_head.proj", # dspark
+        ),
+
         MODEL_TENSOR.CLS: (
             "classifier",       # jina
             "classifier.dense", # roberta
@@ -2095,6 +2107,8 @@ class TensorNameMap:
             "conformer.pre_encode.conv.{bid}", # lfm2
             "model.audio_tower.subsample_conv_projection.conv_{bid}.conv", # gemma3n
             "conformer.subsample_conv_projection.layer{bid}.conv", # gemma4
+            "sound_encoder.encoder.subsampling.layers.{bid}", # parakeet
+            "encoder.conv{bid}", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_CONV1D_NORM: (
@@ -2119,6 +2133,7 @@ class TensorNameMap:
         MODEL_TENSOR.A_POST_NORM: (
             "audio_tower.layer_norm", # ultravox
             "audio_tower.ln_post", # qwen2omni
+            "encoder.layer_norm", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_ATTN_Q: (
@@ -2126,7 +2141,9 @@ class TensorNameMap:
             "conformer.layers.{bid}.self_attn.linear_q", # lfm2
             "conformer.layers.{bid}.attention.attn.q_proj", # gemma3n
             "conformer.layers.{bid}.self_attn.q_proj", # gemma4
+            "sound_encoder.encoder.layers.{bid}.self_attn.q_proj", # parakeet
             "encoder.layers.{bid}.attn.to_q", # granite_speech
+            "encoder.layers.{bid}.self_attn.q_proj", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_ATTN_K: (
@@ -2134,7 +2151,9 @@ class TensorNameMap:
             "conformer.layers.{bid}.self_attn.linear_k", # lfm2
             "conformer.layers.{bid}.attention.attn.k_proj", # gemma3n
             "conformer.layers.{bid}.self_attn.k_proj", # gemma4
+            "sound_encoder.encoder.layers.{bid}.self_attn.k_proj", # parakeet
             "encoder.layers.{bid}.attn.to_k", # granite_speech (split from to_kv)
+            "encoder.layers.{bid}.self_attn.k_proj", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_ATTN_V: (
@@ -2142,7 +2161,9 @@ class TensorNameMap:
             "conformer.layers.{bid}.self_attn.linear_v", # lfm2
             "conformer.layers.{bid}.attention.attn.v_proj", # gemma3n
             "conformer.layers.{bid}.self_attn.v_proj", # gemma4
+            "sound_encoder.encoder.layers.{bid}.self_attn.v_proj", # parakeet
             "encoder.layers.{bid}.attn.to_v", # granite_speech (split from to_kv)
+            "encoder.layers.{bid}.self_attn.v_proj", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_ATTN_K_REL: (
@@ -2170,7 +2191,9 @@ class TensorNameMap:
             "audio_tower.layers.{bid}.self_attn_layer_norm", # ultravox
             "conformer.layers.{bid}.norm_self_att", # lfm2
             "conformer.layers.{bid}.attention.pre_attn_norm", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.norm_self_att", # parakeet
             "encoder.layers.{bid}.attn.pre_norm", # granite_speech
+            "encoder.layers.{bid}.self_attn_layer_norm", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_OUTPUT: (
@@ -2178,20 +2201,25 @@ class TensorNameMap:
             "conformer.layers.{bid}.self_attn.linear_out", # lfm2
             "conformer.layers.{bid}.attention.post", # gemma3n
             "conformer.layers.{bid}.self_attn.post", # gemma4
+            "sound_encoder.encoder.layers.{bid}.self_attn.o_proj", # parakeet
             "encoder.layers.{bid}.attn.to_out", # granite_speech
+            "encoder.layers.{bid}.self_attn.out_proj", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_OUTPUT_NORM: (
             "audio_tower.layers.{bid}.final_layer_norm", # ultravox
             "conformer.layers.{bid}.norm_out", # lfm2
             "conformer.layers.{bid}.attention.post_norm", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.norm_out", # parakeet
             "encoder.layers.{bid}.post_norm", # granite_speech
+            "encoder.layers.{bid}.final_layer_norm", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_FFN_NORM: (
             "conformer.layers.{bid}.norm_feed_forward1", # lfm2
             "conformer.layers.{bid}.ffw_layer_start.pre_layer_norm", # gemma3n
             "conformer.layers.{bid}.feed_forward1.pre_layer_norm", # gemma4
+            "sound_encoder.encoder.layers.{bid}.norm_feed_forward1", # parakeet
             "encoder.layers.{bid}.ff1.pre_norm", # granite_speech
         ),
 
@@ -2209,7 +2237,9 @@ class TensorNameMap:
             "conformer.layers.{bid}.feed_forward1.linear1", # lfm2
             "conformer.layers.{bid}.ffw_layer_start.ffw_layer_1", # gemma3n
             "conformer.layers.{bid}.feed_forward1.ffw_layer_1", # gemma4
+            "sound_encoder.encoder.layers.{bid}.feed_forward1.linear1", # parakeet
             "encoder.layers.{bid}.ff1.up_proj", # granite_speech
+            "encoder.layers.{bid}.fc1", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_FFN_GATE: (),
@@ -2219,13 +2249,16 @@ class TensorNameMap:
             "conformer.layers.{bid}.feed_forward1.linear2", # lfm2
             "conformer.layers.{bid}.ffw_layer_start.ffw_layer_2", # gemma3n
             "conformer.layers.{bid}.feed_forward1.ffw_layer_2", # gemma4
+            "sound_encoder.encoder.layers.{bid}.feed_forward1.linear2", # parakeet
             "encoder.layers.{bid}.ff1.down_proj", # granite_speech
+            "encoder.layers.{bid}.fc2", # mimo-audio-tokenizer
         ),
 
         MODEL_TENSOR.A_ENC_FFN_UP_1: (
             "conformer.layers.{bid}.feed_forward2.linear1", # lfm2
             "conformer.layers.{bid}.ffw_layer_end.ffw_layer_1", # gemma3n
             "conformer.layers.{bid}.feed_forward2.ffw_layer_1", # gemma4
+            "sound_encoder.encoder.layers.{bid}.feed_forward2.linear1", # parakeet
             "encoder.layers.{bid}.ff2.up_proj", # granite_speech
         ),
 
@@ -2233,6 +2266,7 @@ class TensorNameMap:
             "conformer.layers.{bid}.feed_forward2.linear2", # lfm2
             "conformer.layers.{bid}.ffw_layer_end.ffw_layer_2", # gemma3n
             "conformer.layers.{bid}.feed_forward2.ffw_layer_2", # gemma4
+            "sound_encoder.encoder.layers.{bid}.feed_forward2.linear2", # parakeet
             "encoder.layers.{bid}.ff2.down_proj", # granite_speech
         ),
 
@@ -2240,9 +2274,23 @@ class TensorNameMap:
             "conformer.layers.{bid}.norm_feed_forward2", # lfm2
             "conformer.layers.{bid}.ffw_layer_end.pre_layer_norm", # gemma3n
             "conformer.layers.{bid}.feed_forward2.pre_layer_norm", # gemma4
+            "sound_encoder.encoder.layers.{bid}.norm_feed_forward2", # parakeet
             "encoder.layers.{bid}.ff2.pre_norm", # granite_speech
         ),
 
+        MODEL_TENSOR.A_ENC_DOWNSAMPLE_CONV: (
+            "encoder.down_sample_layer.0", # mimo-audio-tokenizer
+        ),
+
+        MODEL_TENSOR.A_ENC_DOWNSAMPLE_NORM: (
+            "encoder.down_sample_norm", # mimo-audio-tokenizer
+        ),
+
+        # note: the raw per-quantizer "encoder.quantizer.vq.layers.{i}._codebook.embed"
+        # tensors are merged (padded + stacked, like MoE experts) into this single 3D
+        # tensor in conversion code, so no raw-name mapping is registered here.
+        MODEL_TENSOR.A_ENC_RVQ_CODEBOOK: (),
+
         MODEL_TENSOR.A_ENC_FFN_POST_NORM_1: (
             "conformer.layers.{bid}.ffw_layer_end.post_layer_norm", # gemma3n
             "conformer.layers.{bid}.feed_forward2.post_layer_norm", # gemma4
@@ -2255,20 +2303,24 @@ class TensorNameMap:
         MODEL_TENSOR.A_ENC_LINEAR_POS: (
             "conformer.layers.{bid}.self_attn.linear_pos", # lfm2
             "conformer.layers.{bid}.attention.attn.relative_position_embedding.pos_proj", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.self_attn.relative_k_proj", # parakeet
         ),
 
         MODEL_TENSOR.A_ENC_POS_BIAS_U: (
             "conformer.layers.{bid}.self_attn.pos_bias_u", # lfm2
+            "sound_encoder.encoder.layers.{bid}.self_attn.bias_u", # parakeet
         ),
 
         MODEL_TENSOR.A_ENC_POS_BIAS_V: (
             "conformer.layers.{bid}.self_attn.pos_bias_v", # lfm2
+            "sound_encoder.encoder.layers.{bid}.self_attn.bias_v", # parakeet
         ),
 
         MODEL_TENSOR.A_ENC_OUT: (
             "conformer.pre_encode.out", # lfm2
             "model.audio_tower.subsample_conv_projection.input_proj_linear", # gemma3n (note: it should be A_ENC_INP_PROJ, this is a mistake; it should be corrected in C++ code when it's supported)
             "conformer.output_proj", # gemma4
+            "sound_encoder.encoder.subsampling.linear", # parakeet
         ),
 
         # note: some tensors below has "audio." pseudo-prefix, to prevent conflicts with vision tensors
@@ -2278,6 +2330,7 @@ class TensorNameMap:
             "audio.multi_modal_projector.linear_{bid}", # ultravox, meralion
             "audio_adapter.model.{bid}", # lfm2
             "audio_tower.proj{bid}", # qwen3omni
+            "sound_projection.linear{bid}", # parakeet (linear1, linear2)
         ),
 
         MODEL_TENSOR.A_MMPROJ_FC: (
@@ -2288,39 +2341,89 @@ class TensorNameMap:
 
         MODEL_TENSOR.A_MM_NORM_PRE: (
             "audio.multi_modal_projector.ln_pre", # ultravox
+            "sound_projection.norm", # parakeet
         ),
 
         MODEL_TENSOR.A_MM_NORM_MID: (
             "audio.multi_modal_projector.ln_mid", # ultravox
         ),
 
+        # note: the raw per-channel "speech_embeddings.{i}" tensors are merged
+        # (stacked, like MoE experts) into this single 3D tensor in conversion
+        # code, so no raw-name mapping is registered here.
+        MODEL_TENSOR.A_MM_CODE_EMBD: (),
+
+        MODEL_TENSOR.A_MM_LOCAL_ATTN_Q: (
+            "audio_encoder.input_local_transformer.layers.{bid}.self_attn.q_proj", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_ATTN_K: (
+            "audio_encoder.input_local_transformer.layers.{bid}.self_attn.k_proj", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_ATTN_V: (
+            "audio_encoder.input_local_transformer.layers.{bid}.self_attn.v_proj", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_ATTN_OUT: (
+            "audio_encoder.input_local_transformer.layers.{bid}.self_attn.o_proj", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_FFN_GATE: (
+            "audio_encoder.input_local_transformer.layers.{bid}.mlp.gate_proj", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_FFN_UP: (
+            "audio_encoder.input_local_transformer.layers.{bid}.mlp.up_proj", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_FFN_DOWN: (
+            "audio_encoder.input_local_transformer.layers.{bid}.mlp.down_proj", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_LN1: (
+            "audio_encoder.input_local_transformer.layers.{bid}.input_layernorm", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_LN2: (
+            "audio_encoder.input_local_transformer.layers.{bid}.post_attention_layernorm", # mimo-v2.5
+        ),
+        MODEL_TENSOR.A_MM_LOCAL_NORM: (
+            "audio_encoder.input_local_transformer.norm", # mimo-v2.5
+        ),
+
         MODEL_TENSOR.A_ENC_CONV_DW: (
             "conformer.layers.{bid}.conv.depthwise_conv", # lfm2
             "conformer.layers.{bid}.lconv1d.depthwise_conv1d", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.conv.depthwise_conv", # parakeet
             "encoder.layers.{bid}.conv.depth_conv.conv", # granite_speech
         ),
 
         MODEL_TENSOR.A_ENC_CONV_NORM: (
             "conformer.layers.{bid}.conv.batch_norm", # lfm2
             "conformer.layers.{bid}.lconv1d.pre_layer_norm", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.conv.norm", # parakeet
+        ),
+
+        MODEL_TENSOR.A_ENC_CONV_NORM_MEAN: (
+            "sound_encoder.encoder.layers.{bid}.conv.norm.running_mean", # parakeet
+        ),
+
+        MODEL_TENSOR.A_ENC_CONV_NORM_VAR: (
+            "sound_encoder.encoder.layers.{bid}.conv.norm.running_var", # parakeet
             "encoder.layers.{bid}.conv.batch_norm", # granite_speech
         ),
 
         MODEL_TENSOR.A_ENC_CONV_PW1: (
             "conformer.layers.{bid}.conv.pointwise_conv1", # lfm2
             "conformer.layers.{bid}.lconv1d.linear_start", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.conv.pointwise_conv1", # parakeet
             "encoder.layers.{bid}.conv.up_conv", # granite_speech
         ),
 
         MODEL_TENSOR.A_ENC_CONV_PW2: (
             "conformer.layers.{bid}.conv.pointwise_conv2", # lfm2
             "conformer.layers.{bid}.lconv1d.linear_end", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.conv.pointwise_conv2", # parakeet
             "encoder.layers.{bid}.conv.down_conv", # granite_speech
         ),
 
         MODEL_TENSOR.A_ENC_NORM_CONV: (
             "conformer.layers.{bid}.norm_conv", # lfm2
             "conformer.layers.{bid}.lconv1d.conv_norm", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.norm_conv", # parakeet
             "encoder.layers.{bid}.conv.norm", # granite_speech
         ),
 
@@ -2332,6 +2435,14 @@ class TensorNameMap:
             "conformer.layers.{bid}.attention.attn.per_dim_scale", # gemma4
         ),
 
+        MODEL_TENSOR.A_ENC_MEL_FILTERS: (
+            "sound_encoder.encoder.feature_extractor.featurizer.fb", # parakeet
+        ),
+
+        MODEL_TENSOR.A_ENC_WINDOW: (
+            "sound_encoder.encoder.feature_extractor.featurizer.window", # parakeet
+        ),
+
         MODEL_TENSOR.A_MM_EMBEDDING: (
             "model.embed_audio.embedding", # gemma3n
         ),

models/templates/MiniMax-M3.jinja

@@ -0,0 +1,247 @@
+{# ---------- special token variables ---------- #}
+{%- set ns_token               = ']<]minimax[>['                  -%}
+{%- set bod_token              = ']~!b['                          -%}
+{%- set bos_token              = ']~b]'                           -%}
+{%- set eos_token              = '[e~['                           -%}
+{%- set toolcall_begin_token   = ns_token ~ '<tool_call>'         -%}
+{%- set toolcall_end_token     = ns_token ~ '</tool_call>'        -%}
+{%- set think_begin_token      = '<mm:think>'                     -%}
+{%- set think_end_token        = '</mm:think>'                    -%}
+{%- set image_token            = ']<]image[>['                    -%}
+{%- set video_token            = ']<]video[>['                    -%}
+{#- Thinking mode: "enabled" / "disabled" / "adaptive" / not defined -#}
+{#- Recursive XML renderer for tool_call arguments ======================== -#}
+{#- None values are intentionally skipped in mapping iteration so that
+    `<key>null</key>` (which would round-trip to the literal string "null")
+    never appears in the rendered tool_call. The convention is: omit the
+    field entirely. The top-level `_args` loop applies the same rule.
+    The `val is none` branch below is a safety net only — upstream cleaning
+    (drop_none_in_tool_arguments) should ensure no None ever reaches here. -#}
+{%- macro to_xml(val, ns) -%}
+{%- if val is mapping -%}
+{%- for k, v in val.items() if v is not none -%}
+{{ ns }}<{{ k }}>{{ to_xml(v, ns) }}{{ ns }}</{{ k }}>
+{%- endfor -%}
+{%- elif val is iterable and val is not string -%}
+{%- for item in val -%}
+{{ ns }}<item>{{ to_xml(item, ns) }}{{ ns }}</item>
+{%- endfor -%}
+{%- elif val is none -%}
+{#- Should be unreachable when upstream cleaning is applied. -#}
+{%- elif val is boolean -%}
+{{ val | tojson }}
+{%- else -%}
+{{ val }}
+{%- endif -%}
+{%- endmacro -%}
+{#- Tool Rendering Functions ============================================== -#}
+{%- macro render_tool_namespace(namespace_name, tool_list) -%}
+{%- for tool in tool_list -%}
+<tool>{{ tool.function | tojson(ensure_ascii=False) }}</tool>
+{% endfor -%}
+{%- endmacro -%}
+{%- macro visible_text(content) -%}
+    {%- if content is string -%}
+        {{ content }}
+    {%- elif content is iterable and content is not mapping -%}
+        {%- for item in content -%}
+            {%- if item is mapping and item.type == 'text' -%}
+                {{- item.text }}
+            {%- elif item is mapping and item.type == 'image' -%}
+                {{- image_token }}
+            {%- elif item is mapping and item.type == 'video' -%}
+                {{- video_token}}
+            {%- elif item is string -%}
+                {{- item }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- elif content is none -%}
+        {{- '' }}
+    {%- else -%}
+        {{- content }}
+    {%- endif -%}
+{%- endmacro -%}
+{#- System Message Construction ============================================ -#}
+{%- macro build_system_message(system_message) -%}
+    {%- if system_message and system_message.content -%}
+        {{- visible_text(system_message.content) }}
+    {%- else -%}
+        {{- 'Your model version is MiniMax-M3, developed by MiniMax. Knowledge cutoff: January 2026. Founded in early 2022, MiniMax is a global AI foundation model company committed to advancing the frontiers of AI towards AGI.' }}
+    {%- endif -%}
+
+    {#- Thinking mode instructions -#}
+    {{- '\n\n<thinking_instructions>\n' }}
+    {{- 'You have a thinking capability that allows you to reason step by step before responding. When thinking is enabled, wrap your reasoning in ' ~ think_begin_token ~ think_end_token ~ ' tags before your response. When thinking is disabled, begin your response directly after the ' ~ think_end_token ~ ' prefix. When thinking is adaptive, decide on your own whether to think for the current turn.\n' }}
+    {%- if thinking_mode is defined -%}
+        {%- if thinking_mode == "enabled" -%}
+            {{- 'Current thinking mode: enabled. You MUST think step by step before every response, including after receiving function/tool results.\n' }}
+        {%- elif thinking_mode == "disabled" -%}
+            {{- 'Current thinking mode: disabled. Do not output any thinking process.\n' }}
+        {%- elif thinking_mode == "adaptive" -%}
+            {{- 'Current thinking mode: adaptive. You are encouraged to think for complex decision-making, multi-step reasoning, or when analyzing function/tool results.\n' }}
+        {%- endif -%}
+    {%- else -%}
+        {{- 'Current thinking mode: adaptive. You are encouraged to think for complex decision-making, multi-step reasoning, or when analyzing function/tool results.\n' }}
+    {%- endif -%}
+    {{- '</thinking_instructions>' }}
+{%- endmacro -%}
+{%- macro build_developer_message(developer_message) -%}
+    {%- if developer_message and developer_message.content -%}
+        {{- visible_text(developer_message.content) }}
+    {%- else -%}
+        {%- if model_identity is not defined -%}
+            {%- set model_identity = "You are a helpful assistant." -%}
+        {%- endif -%}
+        {{- model_identity }}
+    {%- endif -%}
+{%- endmacro -%}
+{#- Main Template Logic ================================================= -#}
+{#- Role mapping: root -> system sp (high priority), system/developer -> developer sp (low priority) -#}
+{%- set system_message = none -%}
+{%- set developer_message = none -%}
+{%- set conversation_messages = messages -%}
+{%- if messages and messages[0].role == "root" -%}
+    {%- set system_message = messages[0] -%}
+    {%- set conversation_messages = messages[1:] -%}
+    {%- if conversation_messages and conversation_messages[0].role in ["system", "developer"] -%}
+        {%- set developer_message = conversation_messages[0] -%}
+        {%- set conversation_messages = conversation_messages[1:] -%}
+    {%- endif -%}
+{%- elif messages and messages[0].role in ["system", "developer"] -%}
+    {%- set developer_message = messages[0] -%}
+    {%- set conversation_messages = messages[1:] -%}
+{%- endif -%}
+{#- Render system sp (higher priority, root role only) -#}
+{{- bod_token ~ bos_token ~ 'system' ~ '\n' }}
+{{- build_system_message(system_message) }}
+{{- eos_token ~ '\n' }}
+
+{#- Render developer sp (lower priority: system/developer role + tools) -#}
+{{- bos_token ~ 'developer' ~ '\n' }}
+{{- build_developer_message(developer_message) }}
+{%- if tools -%}
+    {{- '\n\n' ~ '# Tools' ~ '\n' ~ 'You may call one or more tools to assist with the user query.\nHere are the tools available in JSONSchema format:' ~ '\n' }}
+    {{- '\n' ~ '<tools>' ~ '\n' }}
+    {{- render_tool_namespace("functions", tools) }}
+    {{- '</tools>' ~ '\n\n' }}
+    {{- 'To call tools, wrap all invocations in a single ' ~ toolcall_begin_token ~ toolcall_end_token ~ ' block. Parameter values containing nested objects or arrays are recursively expanded into XML elements. Example:\n' }}
+    {{- '\n' ~ toolcall_begin_token ~ '\n' }}
+    {{- ns_token + '<invoke name="tool-name-1">' }}
+    {{- ns_token + '<param-1>value-1' + ns_token + '</param-1>' }}
+    {{- ns_token + '<param-2>' }}
+    {{- ns_token + '<item>' }}
+    {{- ns_token + '<key-a>val-a' + ns_token + '</key-a>' }}
+    {{- ns_token + '<key-b>val-b' + ns_token + '</key-b>' }}
+    {{- ns_token + '</item>' }}
+    {{- ns_token + '</param-2>' }}
+    {{- ns_token + '</invoke>\n' }}
+    {{- ns_token + '<invoke name="tool-name-2">' }}
+    {{- ns_token + '<param-1>value-1' + ns_token + '</param-1>' }}
+    {{- ns_token + '</invoke>\n' }}
+    {{- toolcall_end_token }}
+{%- endif -%}
+{{- eos_token ~ '\n' }}
+
+{#- Render messages -#}
+{%- set last_tool_call = namespace(name=none) -%}
+{%- for message in conversation_messages -%}
+    {%- if message.role == 'assistant' -%}
+        {{- bos_token ~ 'ai' ~ '\n' }}
+
+        {%- set reasoning_content = '' %}
+        {%- set content = visible_text(message.content) %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if think_end_token in content %}
+                {%- set reasoning_content = content.split(think_end_token)[0].strip('\n').split(think_begin_token)[-1].strip('\n') %}
+                {%- set content = content.split(think_end_token)[-1].strip('\n') %}
+            {%- endif %}
+        {%- endif %}
+
+        {%- if reasoning_content -%}
+            {#- Render thinking for every assistant turn (all-turn visible) -#}
+            {{- think_begin_token ~ reasoning_content ~ think_end_token }}
+        {%- else -%}
+            {#- No thinking rendered → prefix with think_end_token -#}
+            {{- think_end_token }}
+        {%- endif -%}
+
+        {%- if content -%}
+            {{- content }}
+        {%- endif -%}
+        {%- if message.tool_calls -%}
+            {{- toolcall_begin_token ~ '\n' }}
+
+            {%- for tool_call in message.tool_calls -%}
+                {%- if tool_call.function -%}
+                    {%- set tool_call = tool_call.function -%}
+                {%- endif -%}
+{{- ns_token + '<invoke name="' + tool_call.name + '">' }}
+{%- set _args = tool_call.arguments -%}
+{%- for k, v in _args.items() if v is not none %}
+{{- ns_token + '<' + k + '>' -}}
+{{- to_xml(v, ns_token) -}}
+{{- ns_token + '</' + k + '>' }}
+{%- endfor -%}
+{{- ns_token + '</invoke>' ~ '\n' }}
+            {%- endfor -%}
+
+            {{- toolcall_end_token }}
+            {%- if message.tool_calls[-1].function -%}
+                {%- set last_tool_call.name = message.tool_calls[-1].function.name -%}
+            {%- else -%}
+                {%- set last_tool_call.name = message.tool_calls[-1].name -%}
+            {%- endif -%}
+        {%- else -%}
+            {%- set last_tool_call.name = none -%}
+        {%- endif -%}
+        {{- eos_token ~ '\n' }}
+
+    {%- elif message.role == 'tool' -%}
+        {%- if last_tool_call.name is none -%}
+            {{- raise_exception("Message has tool role, but there was no previous assistant message with a tool call!") }}
+        {%- endif -%}
+        {%- if loop.first or (conversation_messages[loop.index0 - 1].role != 'tool') -%}
+            {{- bos_token ~ 'tool' }}
+        {%- endif -%}
+        {{- '\n<response>' }}
+        {%- if message.content is string -%}
+            {{- message.content }}
+        {%- else -%}
+            {%- for tr in message.content -%}
+                {%- if tr is mapping and tr.type is defined and tr.type == 'image' -%}
+                    {{- image_token }}
+                {%- elif tr is mapping and tr.type is defined and tr.type == 'video' -%}
+                    {{- video_token }}
+                {%- else -%}
+                    {{- tr.output if tr.output is defined else (tr.text if tr.type == 'text' and tr.text is defined else tr) }}
+                {%- endif -%}
+            {%- endfor -%}
+        {%- endif -%}
+        {{- '</response>' }}
+        {%- if loop.last or (conversation_messages[loop.index0 + 1].role != 'tool') -%}
+            {{- eos_token ~ '\n' -}}
+        {%- endif -%}
+
+    {%- elif message.role == 'user' -%}
+        {{- bos_token ~ 'user' ~ '\n' }}
+        {{- visible_text(message.content) }}
+        {{- eos_token ~ '\n' }}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- Generation prompt -#}
+{%- if add_generation_prompt -%}
+{{- bos_token ~ 'ai' ~ '\n' }}
+{%- if thinking_mode is defined and thinking_mode == "disabled" -%}
+    {{- think_end_token }}
+{%- elif thinking_mode is defined and thinking_mode == "adaptive" -%}
+    {#- adaptive: no prefix, let model decide -#}
+{%- elif thinking_mode is defined and thinking_mode == "enabled" -%}
+    {#- enabled or not defined: default to think -#}
+    {{- think_begin_token }}
+{%- else -%}
+    {#- adaptive: no prefix, let model decide -#}
+{%- endif -%}
+{%- endif -%}

src/llama-arch.cpp

@@ -143,6 +143,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_KIMI_LINEAR,      "kimi-linear"      },
     { LLM_ARCH_TALKIE,           "talkie"           },
     { LLM_ARCH_MELLUM,           "mellum"           },
+    { LLM_ARCH_NANBEIGE,         "nanbeige"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -221,6 +222,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKEN_SHIFT_COUNT,                 "%s.token_shift_count"                 },
     { LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         "%s.interleave_moe_layer_step"         },
     { LLM_KV_FULL_ATTENTION_INTERVAL,           "%s.full_attention_interval"           },
+    { LLM_KV_NUM_LOOPS,                         "%s.num_loops"                         },
+    { LLM_KV_SKIP_LOOP_FINAL_NORM,              "%s.skip_loop_final_norm"              },
 
     { LLM_KV_ATTENTION_HEAD_COUNT,                   "%s.attention.head_count"                   },
     { LLM_KV_ATTENTION_HEAD_COUNT_KV,                "%s.attention.head_count_kv"                },
@@ -314,6 +317,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TARGET_LAYERS,         "%s.target_layers"        },
     { LLM_KV_TARGET_HIDDEN_SIZE,    "%s.target_hidden_size"   },
     { LLM_KV_NORM_BEFORE_RESIDUAL,  "%s.norm_before_residual" },
+    { LLM_KV_NORM_BEFORE_FC,        "%s.norm_before_fc"       },
 
     { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
     // sentence-transformers dense modules feature dims
@@ -612,6 +616,9 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_MASKED_EMBD_ORDERING,                   "masked_embd_ordering" },
     { LLM_TENSOR_FC,                                     "fc" },
     { LLM_TENSOR_D2T,                                    "d2t" },
+    { LLM_TENSOR_DSPARK_MARKOV_W1,                       "markov_w1" },
+    { LLM_TENSOR_DSPARK_MARKOV_W2,                       "markov_w2" },
+    { LLM_TENSOR_DSPARK_CONF_PROJ,                       "conf_proj" },
 };
 
 // declare information about the model weight tensors:
@@ -866,6 +873,10 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     // eagle3
     {LLM_TENSOR_FC,                         {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
     {LLM_TENSOR_D2T,                        {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_GET_ROWS}},
+    // dspark
+    {LLM_TENSOR_DSPARK_MARKOV_W1,           {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_DSPARK_MARKOV_W2,           {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_DSPARK_CONF_PROJ,           {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
 };
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}

src/llama-arch.h

@@ -148,6 +148,7 @@ enum llm_arch {
     LLM_ARCH_EAGLE3,
     LLM_ARCH_MINIMAX_M3,
     LLM_ARCH_DFLASH,
+    LLM_ARCH_NANBEIGE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -226,6 +227,8 @@ enum llm_kv {
     LLM_KV_TOKEN_SHIFT_COUNT,
     LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
     LLM_KV_FULL_ATTENTION_INTERVAL,
+    LLM_KV_NUM_LOOPS,
+    LLM_KV_SKIP_LOOP_FINAL_NORM,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -360,6 +363,7 @@ enum llm_kv {
     LLM_KV_TARGET_LAYERS,
     LLM_KV_TARGET_HIDDEN_SIZE,
     LLM_KV_NORM_BEFORE_RESIDUAL,
+    LLM_KV_NORM_BEFORE_FC,
 
     LLM_KV_SHORTCONV_L_CACHE,
 
@@ -620,6 +624,9 @@ enum llm_tensor {
     LLM_TENSOR_MASKED_EMBD_ORDERING,
     LLM_TENSOR_FC,
     LLM_TENSOR_D2T,
+    LLM_TENSOR_DSPARK_MARKOV_W1,
+    LLM_TENSOR_DSPARK_MARKOV_W2,
+    LLM_TENSOR_DSPARK_CONF_PROJ,
 };
 
 

src/llama-context.cpp

@@ -2339,6 +2339,7 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
         model.arch == LLM_ARCH_QWEN35 ||
         model.arch == LLM_ARCH_QWEN35MOE ||
         model.arch == LLM_ARCH_DEEPSEEK4 ||
+        model.arch == LLM_ARCH_NANBEIGE ||
         model.arch == LLM_ARCH_MINIMAX_M3) {
         return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
     }
@@ -2473,11 +2474,12 @@ llm_graph_cb llama_context::graph_get_cb() const {
             ggml_set_name(cur, name);
         }
 
-        // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
+        // - norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
+        // - force the last op of the layer on the specified backend to avoid running it on the backend of the next layer due to scheduling
         // FIXME: fix in ggml_backend_sched
         const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer_all;
         if (ubatch.n_tokens < 32 || full_offload) {
-            if (il != -1 && strcmp(name, "norm") == 0) {
+            if (il != -1 && (strcmp(name, "norm") == 0 || strcmp(name, "l_last") == 0)) {
                 const auto & dev_layer = model.dev_layer(il);
                 for (const auto & backend : backends) {
                     if (ggml_backend_get_device(backend.get()) == dev_layer) {

src/llama-hparams.h

@@ -47,6 +47,7 @@ struct llama_hparams {
     bool use_par_res;
     bool swin_norm;
     bool norm_before_residual = false;
+    bool norm_before_fc       = false;
 
     uint32_t n_ctx_train; // context size the model was trained on
     uint32_t n_embd;

src/llama-model.cpp

@@ -85,6 +85,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_stablelm(params);
         case LLM_ARCH_MELLUM:
             return new llama_model_mellum(params);
+        case LLM_ARCH_NANBEIGE:
+            return new llama_model_nanbeige(params);
         case LLM_ARCH_QWEN:
             return new llama_model_qwen(params);
         case LLM_ARCH_QWEN2:
@@ -2491,6 +2493,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_LLAMA_EMBED:
         case LLM_ARCH_MAINCODER:
         case LLM_ARCH_GLM_DSA:
+        case LLM_ARCH_NANBEIGE:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

src/llama-model.h

@@ -606,6 +606,12 @@ struct llama_model {
     struct ggml_tensor * fc  = nullptr;  // feature fusion layer
     struct ggml_tensor * d2t = nullptr;  // draft to target vocabulary mapping
 
+    // dspark
+    struct ggml_tensor * dspark_markov_w1   = nullptr;
+    struct ggml_tensor * dspark_markov_w2   = nullptr;
+    struct ggml_tensor * dspark_conf_proj   = nullptr;
+    struct ggml_tensor * dspark_conf_proj_b = nullptr;
+
     // unified vector to store target-model extracted layer ids in eagle3, dflash, etc.
     std::vector<int32_t> target_layer_ids;
 

src/llama-quant.cpp

@@ -359,6 +359,10 @@ static bool tensor_allows_quantization(const llama_model_quantize_params * param
     quantize &= name.find(".patch_embd")    == std::string::npos;
     quantize &= name.find(".patch_merger")  == std::string::npos;
 
+    // audio codebook
+    quantize &= name.find("a.rvq.codebook")  == std::string::npos;
+    quantize &= name.find("mm.a.code_embd")  == std::string::npos;
+
     return quantize;
 }
 

src/llama-sampler.cpp

@@ -993,7 +993,9 @@ static void llama_sampler_greedy_backend_apply(
     GGML_UNUSED(gf);
     GGML_UNUSED(smpl);
 
-    struct ggml_tensor * curl = ggml_argmax(ctx, data->logits);
+    struct ggml_tensor * logits = ggml_reshape_1d(ctx, data->logits, ggml_nelements(data->logits));
+
+    struct ggml_tensor * curl = ggml_argmax(ctx, logits);
     ggml_set_name(curl, "greedy_argmax");
 
     data->sampled = curl;
@@ -1158,7 +1160,10 @@ static void llama_sampler_dist_backend_apply(
     ggml_set_name (sctx->inp_uniform, "uniform");
     ggml_set_input(sctx->inp_uniform);
 
-    struct ggml_tensor * probs = ggml_soft_max(ctx, data->logits);
+    // flatten
+    struct ggml_tensor * logits = ggml_reshape_1d(ctx, data->logits, ggml_nelements(data->logits));
+
+    struct ggml_tensor * probs = ggml_soft_max(ctx, logits);
     ggml_set_name(probs, "dist_probs");
 
     struct ggml_tensor * cumsum = ggml_cumsum(ctx, probs);
@@ -1289,22 +1294,22 @@ static void llama_sampler_top_k_backend_apply(
         struct llama_sampler_data * data) {
     auto * sctx = (llama_sampler_top_k *) smpl->ctx;
 
-    struct ggml_tensor * top_k = ggml_top_k(ctx, data->logits, sctx->k);
+    struct ggml_tensor * logits = ggml_reshape_1d(ctx, data->logits, ggml_nelements(data->logits));
+
+    struct ggml_tensor * top_k = ggml_top_k(ctx, logits, sctx->k);
     ggml_set_name(top_k, "top_k");
 
     if (data->candidates) {
         struct ggml_tensor * candidates_rows = ggml_reshape_2d(ctx, data->candidates, 1, data->candidates->ne[0]);
         data->candidates = ggml_get_rows(ctx, candidates_rows, top_k);
-        data->candidates = ggml_reshape_1d(ctx, data->candidates, sctx->k);
         ggml_set_name(data->candidates, "top_k_candidates");
     } else {
         data->candidates = top_k;
     }
 
-    struct ggml_tensor * logits_rows = ggml_reshape_2d(ctx, data->logits, 1, data->logits->ne[0]);
-    struct ggml_tensor * top_k_rows = ggml_get_rows(ctx, logits_rows, top_k);
-    data->logits = ggml_reshape_1d(ctx, top_k_rows, sctx->k);
-    ggml_set_name(top_k_rows, "top_k_rows");
+    struct ggml_tensor * logits_rows = ggml_reshape_2d(ctx, logits, 1, logits->ne[0]);
+    data->logits = ggml_get_rows(ctx, logits_rows, top_k);
+    ggml_set_name(data->logits, "top_k_rows");
 
     GGML_UNUSED(gf);
 }
@@ -1435,21 +1440,25 @@ static void llama_sampler_top_p_backend_apply(
         struct llama_sampler_data * data) {
     auto * sctx = (llama_sampler_top_p *) smpl->ctx;
 
+    // flatten
+    struct ggml_tensor * logits = ggml_reshape_1d(ctx, data->logits, ggml_nelements(data->logits));
+
     auto ggml_sort = [ctx](struct ggml_tensor * a, struct ggml_tensor * b) {
         GGML_ASSERT(ggml_nrows(a) == 1);
         struct ggml_tensor * a_reshaped = ggml_reshape_2d(ctx, a, 1, a->ne[0]);
         struct ggml_tensor * a_sorted   = ggml_get_rows(ctx, a_reshaped, b);
-        return ggml_reshape_1d(ctx, a_sorted, a->ne[0]);
+        return a_sorted;
     };
 
     // Get the sorted logits in descending order.
-    struct ggml_tensor * sorted_idx = ggml_argsort(ctx, data->logits, GGML_SORT_ORDER_DESC);
+    struct ggml_tensor * sorted_idx = ggml_argsort(ctx, logits, GGML_SORT_ORDER_DESC);
     ggml_set_name(sorted_idx, "top_p_sorted_idx");
 
     // Do the sorting via reshape + get_rows
-    struct ggml_tensor * sorted_logits = ggml_sort(data->logits, sorted_idx);
+    struct ggml_tensor * sorted_logits = ggml_sort(logits, sorted_idx);
     ggml_set_name(sorted_logits, "top_p_sorted_logits");
 
+    sorted_logits = ggml_reshape_1d(ctx, sorted_logits, ggml_nelements(sorted_logits));
     struct ggml_tensor * softmax = ggml_soft_max(ctx, sorted_logits);
     ggml_set_name(softmax, "top_p_softmax");
 
@@ -1626,10 +1635,12 @@ static void llama_sampler_min_p_backend_apply(
         struct llama_sampler_data * data) {
     auto * sctx = (llama_sampler_min_p *) smpl->ctx;
 
-    struct ggml_tensor * max_idx = ggml_argmax(ctx, data->logits);
+    struct ggml_tensor * logits = ggml_reshape_1d(ctx, data->logits, ggml_nelements(data->logits));
+
+    struct ggml_tensor * max_idx = ggml_argmax(ctx, logits);
     ggml_set_name(max_idx, "max_idx");
 
-    struct ggml_tensor * logits_rows = ggml_reshape_2d(ctx, data->logits, 1, data->logits->ne[0]);
+    struct ggml_tensor * logits_rows = ggml_reshape_2d(ctx, logits, 1, logits->ne[0]);
     ggml_set_name(logits_rows, "logits_rows");
 
     struct ggml_tensor * max_logit = ggml_get_rows(ctx, logits_rows, max_idx);
@@ -1640,7 +1651,7 @@ static void llama_sampler_min_p_backend_apply(
     ggml_set_name(threshold, "min_p_threshold");
 
     // Subtract the threshold from logits.
-    struct ggml_tensor * sub = ggml_sub(ctx, data->logits, threshold);
+    struct ggml_tensor * sub = ggml_sub(ctx, logits, threshold);
 
     // Create a mask where logits below the threshold are 0 (discard),
     // and others are 1 (keep).
@@ -1652,7 +1663,7 @@ static void llama_sampler_min_p_backend_apply(
     struct ggml_tensor * min_p_bias = ggml_log(ctx, mask);
     ggml_set_name(min_p_bias, "min_p_bias");
 
-    data->logits = ggml_add(ctx, data->logits, min_p_bias);
+    data->logits = ggml_add(ctx, logits, min_p_bias);
     ggml_set_name(data->logits, "min_p_logits");
 
     GGML_UNUSED(gf);
@@ -1829,18 +1840,20 @@ static void llama_sampler_backend_temp_sampling(
         struct llama_sampler_data * data,
         float                       temp) {
     if (temp <= 0.0f) {
+        struct ggml_tensor * logits = ggml_reshape_1d(ctx, data->logits, ggml_nelements(data->logits));
+
         // Find the most probable token index.
-        struct ggml_tensor * max_idx = ggml_argmax(ctx, data->logits);
+        struct ggml_tensor * max_idx = ggml_argmax(ctx, logits);
         ggml_set_name(max_idx, "temp_max_idx");
 
         if (data->candidates) {
-            struct ggml_tensor * candidates_rows = ggml_reshape_2d(ctx, data->candidates, 1, data->candidates->ne[0]);
+            struct ggml_tensor * candidates_rows = ggml_reshape_2d(ctx, data->candidates, 1, ggml_nelements(data->candidates));
             data->candidates = ggml_get_rows(ctx, candidates_rows, max_idx);
         } else {
             data->candidates = max_idx;
         }
 
-        struct ggml_tensor * logits_rows = ggml_reshape_2d(ctx, data->logits, 1, data->logits->ne[0]);
+        struct ggml_tensor * logits_rows = ggml_reshape_2d(ctx, logits, 1, ggml_nelements(logits));
         data->logits = ggml_get_rows(ctx, logits_rows, max_idx);
 
         return;
@@ -2019,13 +2032,15 @@ static void llama_sampler_temp_ext_backend_apply(
         return;
     }
 
+    struct ggml_tensor * logits = ggml_reshape_1d(ctx, data->logits, ggml_nelements(data->logits));
+
     // Calculate min_temp, max_temp, and max_entropy.
     const float min_temp    = std::max(0.0f, sctx->temp - sctx->delta);
     const float max_temp    = sctx->temp + sctx->delta;
-    const float max_entropy = logf(data->logits->ne[0]);
+    const float max_entropy = logf(logits->ne[0]);
 
     // Calculate the probabilities.
-    struct ggml_tensor * probs = ggml_soft_max(ctx, data->logits);
+    struct ggml_tensor * probs = ggml_soft_max(ctx, logits);
     ggml_set_name(probs, "temp_ext_softmax_probs");
 
     // Clamp probabilities to avoid log(0) which would give -inf
@@ -2063,7 +2078,7 @@ static void llama_sampler_temp_ext_backend_apply(
     ggml_set_name(dyn_temp,         "temp_ext_dyn_temp");
 
     // Scale the logits by the dynamic temperature
-    struct ggml_tensor * scaled_logits = ggml_div(ctx, data->logits, dyn_temp);
+    struct ggml_tensor * scaled_logits = ggml_div(ctx, logits, dyn_temp);
     ggml_set_name(scaled_logits, "temp_ext_scaled_logits");
 
     data->logits = scaled_logits;

src/models/deepseek4.cpp

@@ -1133,6 +1133,10 @@ llama_model_deepseek4::graph::graph(const llama_model & model, const llm_graph_p
                 &post, &comb, il);
         cb(cur, "hc_ffn_pre", il);
 
+        ggml_build_forward_expand(gf, residual);
+        ggml_build_forward_expand(gf, post);
+        ggml_build_forward_expand(gf, comb);
+
         cur = build_norm(cur, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "ffn_norm", il);
 
@@ -1175,7 +1179,7 @@ llama_model_deepseek4::graph::graph(const llama_model & model, const llm_graph_p
 
         inpL = build_hc_post(cur, residual, post, comb, il);
         inpL = build_cvec(inpL, il);
-        cb(inpL, "l_out", il);
+        cb(inpL, "l_last", il);
     }
 
     if (inp_out_ids) {

src/models/dflash.cpp

@@ -37,6 +37,23 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
 
     const int64_t n_embd_inp = hparams.n_embd_inp_enc();
 
+    // DSpark = DFlash + a semi-autoregressive Markov head and Confidence head
+    //
+    // TODO: only Qwen3-style backbones are supported for now; other backbones (e.g. Gemma4)
+    //       need their own conversion path and graph tweaks
+    const struct ggml_tensor * markov_meta = ml->get_tensor_meta("markov_w1.weight");
+    if (markov_meta) {
+        const int64_t dspark_markov_rank = markov_meta->ne[0];
+
+        dspark_markov_w1 = create_tensor(tn(LLM_TENSOR_DSPARK_MARKOV_W1, "weight"), { dspark_markov_rank, n_vocab }, 0);
+        dspark_markov_w2 = create_tensor(tn(LLM_TENSOR_DSPARK_MARKOV_W2, "weight"), { dspark_markov_rank, n_vocab }, 0);
+
+        dspark_conf_proj   = create_tensor(tn(LLM_TENSOR_DSPARK_CONF_PROJ, "weight"), { n_embd + dspark_markov_rank, 1 }, 0);
+        dspark_conf_proj_b = create_tensor(tn(LLM_TENSOR_DSPARK_CONF_PROJ, "bias"),   { 1 },             TENSOR_NOT_REQUIRED);
+
+        LLAMA_LOG_INFO("%s: DFlash with DSpark markov head (rank = %lld)\n", __func__, (long long) dspark_markov_rank);
+    }
+
     fc              = create_tensor(tn(LLM_TENSOR_FC,              "weight"), { n_embd_inp, n_embd }, 0);
     output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), { n_embd }, 0); // encoder hidden_norm (after fc)
     output_norm     = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM,    "weight"), { n_embd }, 0); // decoder final norm
@@ -105,6 +122,94 @@ llama_model_dflash::graph<true>::graph(const llama_model & model, const llm_grap
     ggml_build_forward_expand(gf, cur);
 }
 
+// DSpark (DFlash + Markov & Confidence head): Markov bias on the draft logits, chained per block position
+static void build_dspark_markov_head(llm_graph_context & g, const llama_model & model, ggml_tensor * tokens) {
+    ggml_context * ctx0 = g.ctx0;
+    auto         & res  = g.res;
+
+    ggml_tensor * w1 = model.dspark_markov_w1;
+    ggml_tensor * w2 = model.dspark_markov_w2;
+    GGML_ASSERT(w1 && w2 && model.dspark_conf_proj && "DSpark markov/confidence weights not loaded");
+
+    ggml_tensor * base = res->t_logits; // [n_vocab, n_tokens]
+    const int64_t n_vocab = base->ne[0];
+    const int64_t n_tok   = base->ne[1];
+
+    const auto it = model.gguf_kv.find("dflash.block_size");
+    GGML_ASSERT(it != model.gguf_kv.end() && "DSpark draft requires 'dflash.block_size' in GGUF metadata");
+    const int64_t block_size = std::stoi(it->second);
+    GGML_ASSERT(block_size > 0);
+
+    const int64_t n_blocks = g.ubatch.n_seqs_unq;
+    GGML_ASSERT(n_blocks > 0 && n_tok % n_blocks == 0 && "DSpark markov head requires equal-size blocks");
+    // runtime tokens per block in this ubatch (anchor + drafted positions), bounded by training block_size
+    const int64_t block_drafts = n_tok / n_blocks;
+    if (block_drafts > block_size) {
+        return;
+    }
+
+    // anchor (committed last) token of every block: token 0 of each block, i.e. a strided view
+    const size_t token_stride = (size_t) block_drafts * tokens->nb[0];
+    const size_t base_stride = (size_t) block_drafts * base->nb[1];
+
+    ggml_tensor * prev = ggml_view_2d(ctx0, tokens, 1, n_blocks, token_stride, 0);
+    prev = ggml_cont_1d(ctx0, prev, n_blocks);
+
+    // confidence head input: predicts per-position acceptance
+    ggml_tensor * conf_inp = res->t_embd; // [n_embd, n_tok]
+
+    ggml_tensor * cat      = nullptr;
+    ggml_tensor * cat_conf = nullptr;
+
+    // TODO: the in-graph chain is greedy (argmax); sampling params affect only the final
+    //       token pick, not the Markov conditioning path
+    for (int64_t i = 0; i < block_drafts; ++i) {
+        ggml_tensor * w1_prev = ggml_get_rows(ctx0, w1, prev);   // [R, n_blocks]
+        ggml_tensor * bias    = ggml_mul_mat(ctx0, w2, w1_prev); // [n_vocab, n_blocks]
+
+        // position i of every block: strided view [n_vocab, n_blocks]
+        ggml_tensor * base_i = ggml_view_2d(ctx0, base, n_vocab, n_blocks, base_stride, i*base->nb[1]);
+        ggml_tensor * col    = ggml_add(ctx0, base_i, bias);
+
+        cat = cat ? ggml_concat(ctx0, cat, col, 1) : col;
+
+        // conf(i) = sigmoid(conf_proj . [conf_inp(i); markov_w1[prev(i)]] + b)  -- [1, n_blocks]
+        ggml_tensor * conf_inp_i = ggml_view_2d(ctx0, conf_inp, conf_inp->ne[0], n_blocks,
+                                                (size_t) block_drafts * conf_inp->nb[1], i*conf_inp->nb[1]);
+        ggml_tensor * feat = ggml_concat(ctx0, ggml_cont(ctx0, conf_inp_i), w1_prev, 0);
+        ggml_tensor * conf = ggml_mul_mat(ctx0, model.dspark_conf_proj, feat);
+        if (model.dspark_conf_proj_b) {
+            conf = ggml_add(ctx0, conf, model.dspark_conf_proj_b);
+        }
+        conf = ggml_sigmoid(ctx0, conf);
+
+        cat_conf = cat_conf ? ggml_concat(ctx0, cat_conf, conf, 1) : conf;
+
+        if (i + 1 < block_drafts) {
+            prev = ggml_argmax(ctx0, col);
+        }
+    }
+
+    // cat is position-major; restore ubatch block-major order
+    ggml_tensor * out = ggml_reshape_3d(ctx0, cat, n_vocab, n_blocks, block_drafts);
+    out = ggml_cont(ctx0, ggml_permute(ctx0, out, 0, 2, 1, 3)); // [n_vocab, block_drafts, n_blocks]
+    out = ggml_reshape_2d(ctx0, out, n_vocab, n_tok);
+
+    {
+        ggml_tensor * conf = ggml_reshape_3d(ctx0, cat_conf, 1, n_blocks, block_drafts);
+        conf = ggml_cont(ctx0, ggml_permute(ctx0, conf, 0, 2, 1, 3));
+        conf = ggml_reshape_2d(ctx0, conf, 1, n_tok);
+
+        // note: broadcast the [1, n_tok] confidences to n_embd-wide rows to be able to reuse `llama_get_embeddings_nextn`
+        conf = ggml_repeat(ctx0, conf, res->t_embd);
+        res->t_h_nextn = conf;
+        ggml_build_forward_expand(g.gf, conf);
+    }
+
+    res->t_logits = out;
+    ggml_build_forward_expand(g.gf, out);
+}
+
 // DFlash decoder, dual-mode by batch type:
 //   * embd batch  -> fused target features: project + inject K/V into the cache.
 //   * token batch -> noise-block diffusion: attend over [committed, MASK...] to generate draft tokens
@@ -210,6 +315,8 @@ llama_model_dflash::graph<false>::graph(const llama_model & model, const llm_gra
     inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
     ggml_set_input(inp->tokens);
 
+    ggml_tensor * inp_tokens = inp->tokens;
+
     ggml_tensor * inpL = ggml_get_rows(ctx0, tok_embd, inp->tokens);
     cb(inpL, "inp_noise_embd", -1);
 
@@ -290,4 +397,9 @@ llama_model_dflash::graph<false>::graph(const llama_model & model, const llm_gra
     res->t_logits = cur;
 
     ggml_build_forward_expand(gf, cur);
+
+    // DSpark: bias the draft logits with the Markov head
+    if (model.dspark_markov_w1) {
+        build_dspark_markov_head(*this, model, inp_tokens);
+    }
 }

src/models/eagle3.cpp

@@ -28,6 +28,10 @@ void llama_model_eagle3::load_arch_hparams(llama_model_loader & ml) {
         LLAMA_LOG_INFO("%s: EAGLE3gnorm_before_residual = true\n", __func__);
     }
 
+    // eagle3 norm_before_fc (optional, default false)
+    // compatible with eagle3.1 (e.g. nvidia/gpt-oss-120b-Eagle3-v3)
+    ml.get_key(LLM_KV_NORM_BEFORE_FC, hparams.norm_before_fc, false);
+
     type = LLM_TYPE_UNKNOWN;
 }
 
@@ -53,6 +57,11 @@ void llama_model_eagle3::load_arch_tensors(llama_model_loader &) {
     // Feature fusion layer: projects 3 target layers to draft hidden size
     fc = create_tensor(tn(LLM_TENSOR_FC, "weight"), {n_embd_inp, n_embd}, 0);
 
+    // RMSNorm on the fused target features (input to fc), only when norm_before_fc is set.
+    if (hparams.norm_before_fc) {
+        output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd_inp}, 0);
+    }
+
     // Output layer (uses draft vocab size)
     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
     output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_draft_vocab}, TENSOR_NOT_REQUIRED);
@@ -130,6 +139,12 @@ llama_model_eagle3::graph<true>::graph(const llama_model & model, const llm_grap
 
     cur = build_inp_embd_enc();
 
+    // RMSNorm on the fused target features before fc
+    if (hparams.norm_before_fc) {
+        cur = build_norm(cur, model.output_norm_enc, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "enc_input_norm", -1);
+    }
+
     // Feature fusion layer
     cur = build_lora_mm(model.fc, cur);
     cb(cur, "fc_out", -1);

src/models/models.h

@@ -424,6 +424,22 @@ struct llama_model_mellum : public llama_model_base {
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 
+struct llama_model_nanbeige : public llama_model_base {
+    llama_model_nanbeige(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    int  n_loops = 1;
+    int  n_layer_phys = 0;
+    bool skip_loop_final_norm = false;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
 struct llama_model_qwen : public llama_model_base {
     llama_model_qwen(const struct llama_model_params & params) : llama_model_base(params) {}
     void load_arch_hparams(llama_model_loader & ml) override;

src/models/nanbeige.cpp

@@ -0,0 +1,184 @@
+#include "models.h"
+
+void llama_model_nanbeige::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    uint32_t n_loops_u = 1;
+    ml.get_key(LLM_KV_NUM_LOOPS, n_loops_u, false);
+    GGML_ASSERT(n_loops_u >= 1);
+
+    skip_loop_final_norm = false;
+    ml.get_key(LLM_KV_SKIP_LOOP_FINAL_NORM, skip_loop_final_norm, false);
+
+    n_layer_phys = (int) hparams.n_layer();
+
+    // Bound-check before casting: signed int mul can overflow and bypass the guard.
+    GGML_ASSERT((size_t) n_layer_phys * (size_t) n_loops_u <= (size_t) LLAMA_MAX_LAYERS);
+    n_loops = (int) n_loops_u;
+
+    // Expand logical layer count before load_tensors() allocates layers / KV.
+    if (n_loops > 1) {
+        for (int j = 1; j < n_loops; ++j) {
+            for (int i = 0; i < n_layer_phys; ++i) {
+                const int dst = i + j * n_layer_phys;
+                hparams.n_head_arr[dst]    = hparams.n_head_arr[i];
+                hparams.n_head_kv_arr[dst] = hparams.n_head_kv_arr[i];
+                hparams.n_ff_arr[dst]      = hparams.n_ff_arr[i];
+                hparams.is_swa_impl[dst]   = hparams.is_swa_impl[i];
+                hparams.is_recr_impl[dst]  = hparams.is_recr_impl[i];
+            }
+        }
+        hparams.n_layer_all = (uint32_t) ((size_t) n_layer_phys * (size_t) n_loops);
+    }
+
+    type = LLM_TYPE_UNKNOWN;
+}
+
+void llama_model_nanbeige::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    const int n_phys = n_layer_phys > 0 ? n_layer_phys : n_layer;
+    for (int i = 0; i < n_phys; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2},
+                TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+
+    // Share physical weights across loops; each slot still has its own KV index.
+    if (n_loops > 1) {
+        for (int j = 1; j < n_loops; ++j) {
+            for (int i = 0; i < n_phys; ++i) {
+                layers[i + j * n_phys] = layers[i];
+            }
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_nanbeige::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_nanbeige::graph::graph(const llama_model & model, const llm_graph_params & params) :
+        llm_graph_context(params) {
+    const auto & nb = static_cast<const llama_model_nanbeige &>(model);
+
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+    const int n_phys  = nb.n_layer_phys > 0 ? nb.n_layer_phys : (int) n_layer;
+    const int n_loops = nb.n_loops > 0 ? nb.n_loops : 1;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f
+        ? 1.0f / sqrtf(float(n_embd_head))
+        : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+                    n_embd_head, n_head, n_head_kv, il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur,   inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        inpL = cur;
+
+        if (n_loops > 1 &&
+            ((il + 1) % n_phys) == 0 &&
+            (il + 1) < n_layer &&
+            !nb.skip_loop_final_norm) {
+            cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "loop_norm", il);
+            inpL = cur;
+        }
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur, model.output_s);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/openai-moe.cpp

@@ -116,7 +116,7 @@ llama_model_openai_moe::graph::graph(const llama_model & model, const llm_graph_
 
             cb(cur, "attn_out", il);
         }
-        if (il == n_layer - 1) {
+        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
             // skip computing output for unused tokens
             cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
@@ -154,6 +154,12 @@ llama_model_openai_moe::graph::graph(const llama_model & model, const llm_graph_
     }
     cur = inpL;
 
+    res->t_h_nextn = cur;
+
+    if (!cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     cur = build_norm(cur,
             model.output_norm, NULL,
             LLM_NORM_RMS, -1);

tests/test-backend-ops.cpp

@@ -478,6 +478,7 @@ enum test_mode {
     MODE_PERF,
     MODE_GRAD,
     MODE_SUPPORT,
+    MODE_TUNE,
 };
 
 // Output format support similar to llama-bench
@@ -4000,7 +4001,7 @@ struct test_ssm_scan : public test_case {
 
     test_ssm_scan(ggml_type type = GGML_TYPE_F32,
             int64_t d_state = 32,
-            int64_t head_dim = 1, // non-zero for Mamba-2
+            int64_t head_dim = 1, // 1 = Mamba-1; > 1 = Mamba-2 (scalar A per head)
             int64_t n_head  = 32,
             int64_t n_group = 1,
             int64_t n_seq_tokens = 32,
@@ -4008,6 +4009,11 @@ struct test_ssm_scan : public test_case {
             bool xbc_overlap = false)
         : type(type), d_state(d_state), head_dim(head_dim), n_head(n_head), n_group(n_group), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs), xbc_overlap(xbc_overlap) {}
 
+    double max_nmse_err() override {
+        // SSD path (head_dim > 1) uses FP16 intermediates (M matrix, X_dt); Mamba-1 is pure FP32.
+        return (head_dim > 1) ? 2e-7 : 1e-7;
+    }
+
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * s   = ggml_new_tensor_4d(ctx, type, d_state,  head_dim,     n_head,       n_seqs);
         ggml_tensor * dt  = ggml_new_tensor_3d(ctx, type, n_head,   n_seq_tokens, n_seqs);
@@ -4034,14 +4040,14 @@ struct test_ssm_scan : public test_case {
         return out;
     }
 
-    // similar to test_mul_mat_id
+
     void initialize_tensors(ggml_context * ctx) override {
         std::random_device rd;
         std::default_random_engine rng(rd());
         for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
             if (t->type == GGML_TYPE_I32) {
                 if (ggml_is_view_op(t->op)) { continue; }
-                // ids
+                // ids: permutation of [0..n_seqs)
                 for (int64_t r = 0; r < ggml_nrows(t); r++) {
                     std::vector<int32_t> data(t->ne[0]);
                     for (int i = 0; i < t->ne[0]; i++) {
@@ -4050,6 +4056,11 @@ struct test_ssm_scan : public test_case {
                     std::shuffle(data.begin(), data.end(), rng);
                     ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(int32_t));
                 }
+            } else if (ggml_is_view_op(t->op)) {
+                continue;
+            } else if (t->ne[1] == n_head && t->ne[2] == 1) {
+                // A {1 or d_state, n_head}: negative decay (2-D tensor, ne[2]==1 distinguishes from 3-D/4-D tensors)
+                init_tensor_uniform(t, -1.0f, -0.5f);
             } else {
                 init_tensor_uniform(t);
             }
@@ -8770,6 +8781,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 32, 4)); // Mamba-2
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 256, 64,  8, 2, 32, 4)); // Falcon-H1
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 128, 4, 4, 16, 2, true)); // x/B/C overlap
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 80, 128, 1, 256, 1)); // Nemotron-9B SSD path
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 80, 128, 1, 512, 1)); // Nemotron-9B SSD multi-chunk (2 aligned chunks)
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 80, 8, 300, 2)); // Mamba-2 SSD multi-chunk (partial 2nd chunk, 2 seqs)
 
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 1, 1));
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 32, 1));
@@ -8793,6 +8807,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 512, 1, 512));
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 128, 32, 128));
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 128, 4, 128, {2, 3}));
+    test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 256, 512, 256)); // many rows
+    test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 32, 1, 32)); // too small (N<64)
+    test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 1024, 1, 1024)); // too big (N>512)
 
 #if 0
     // > 4GB A matrix. Too slow to be enabled by default.
@@ -9544,6 +9561,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_flash_attn_ext(64, 128, 4, {1, 1}, 128, 2, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q4_0, GGML_TYPE_Q1_0));
     test_cases.emplace_back(new test_flash_attn_ext(128, 64, 4, {1, 1}, 64, 2, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q1_0, GGML_TYPE_F16));
 
+    // large-KV F16 cases (Qwen3.6-27B geometry and a llama-class control): the upstream matrix
+    // stops at kv=1024, blind to long-context FA bugs (e.g. the oneDNN SDPA ordering race on BMG).
+    for (int64_t kv : { 4096, 16384 }) {
+        test_cases.emplace_back(new test_flash_attn_ext(256, 256, 4, {6, 1}, kv, 512, true, false, 0, 0,
+                                                        GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+        test_cases.emplace_back(new test_flash_attn_ext(128, 128, 8, {4, 1}, kv, 512, true, false, 0, 0,
+                                                        GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+    }
+
     test_cases.emplace_back(new test_cross_entropy_loss     (GGML_TYPE_F32, {   10, 5, 4, 3}));
     test_cases.emplace_back(new test_cross_entropy_loss     (GGML_TYPE_F32, {30000, 1, 1, 1}));
     test_cases.emplace_back(new test_cross_entropy_loss_back(GGML_TYPE_F32, {   10, 5, 4, 3}));
@@ -9803,6 +9829,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 64, 1, 64));
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 256, 1, 256));
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 128, 32, 128));
+    test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 64, 2048, 64));
+    test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 128, 2048, 128));
+    test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 256, 2048, 256));
+    test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 512, 2048, 512));
 
     test_cases.emplace_back(new test_solve_tri(GGML_TYPE_F32, { 64, 64, 4, 4 }, { 32, 64, 4, 4 }));
     test_cases.emplace_back(new test_solve_tri(GGML_TYPE_F32, { 128, 128, 4, 2 }, { 32, 128, 4, 2 }));
@@ -9974,6 +10004,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_ssm_conv_bias_silu(GGML_TYPE_F32, {4,   3328, 1, 1}, {4, 3328, 1, 1}, true));  // generate
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 48, 1, 512, 1)); // prefill
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 48, 1, 1,   1)); // generate
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 80, 128, 1, 512, 1)); // Nemotron-9B prefill
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 80, 128, 1, 1,   1)); // Nemotron-9B generate
 
     // acc
     test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 1, 1}, {256, 16, 1, 1}, -1));
@@ -10080,8 +10112,482 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_from_file(const c
     return test_cases;
 }
 
+// ---- FA vec (Q,NE) tuning: numerical correctness check + drift guard ----
+// metal proc_address bridges (resolved by string, not symbol linkage)
+using set_fa_vec_override_t   = void (*)(int, int);
+using clear_fa_vec_override_t = void (*)(void);
+using fa_vec_bucket_t         = int  (*)(int64_t);  // runtime ne11/ne01 bucketers, shared via proc bridge
+using fa_vec_baseline_ne_t    = int  (*)(int, int); // runtime (dk,dv) -> baseline NE
+
+// legal NE for a (dk,dv): NL = 32/NE, require (dk/4)%NL==0 && (dv/4)%NL==0
+static std::vector<int> fa_vec_legal_ne(int dk, int dv) {
+    std::vector<int> r;
+    for (int ne : {1, 2, 4}) {
+        const int nl = 32 / ne;
+        if ((dk/4) % nl == 0 && (dv/4) % nl == 0) {
+            r.push_back(ne);
+        }
+    }
+    return r;
+}
+
+// Baseline-path smoke test: with no override, fa_vec_pick returns (1, baseline_ne) and
+// dispatches the no-suffix baseline kernel. Checks baseline correctness across all 10 shapes
+// (NE is numerically transparent; the (Q,NE) kernels are covered by run_fa_vec_tune_check).
+static bool run_fa_vec_drift_guard(ggml_backend_t backend_metal, ggml_backend_t backend_cpu) {
+    struct shape_t { int dk, dv; };
+    const shape_t shapes[] = { {32,32},{64,64},{96,96},{128,128},{192,192},
+                               {192,128},{256,256},{320,256},{512,512},{576,512} };
+    bool ok = true;
+    for (auto s : shapes) {
+        // no override + short KV (ne11 < FA_VEC_NE11_BUCKETS[0]) -> fa_vec_pick returns baseline;
+        // compare metal vs CPU-ref
+        test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, {1,1}, /*kv=*/512, /*nb=*/8,
+                               /*mask=*/true, /*sinks=*/false, 0.0f, 0.0f, GGML_PREC_F32,
+                               GGML_TYPE_F16, GGML_TYPE_F16);
+        auto st = tc.eval(backend_metal, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+        if (st == test_status_t::FAIL) {
+            printf("DRIFT/baseline FAIL dk=%d dv=%d\n", s.dk, s.dv);
+            ok = false;
+        }
+    }
+    return ok;
+}
+
+// Numerical gate: for each (dk,dv) x legal (Q,NE), force the override and compare metal
+// vs CPU-ref. The ne01/ne11/mask/sinks points exercise the rebased body's padded rows
+// (ne01 not a multiple of Q), per-qq sinks, skip-INF, kvpad, and the nsg-dependent shmem
+// offsets / parallel-reduce stride (ne11 -> nsg 1/2/4).
+static bool run_fa_vec_tune_check(ggml_backend_t backend_metal, ggml_backend_t backend_cpu) {
+    auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_metal));
+    auto set_ov   = (set_fa_vec_override_t)   ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_set_fa_vec_override");
+    auto clear_ov = (clear_fa_vec_override_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_clear_fa_vec_override");
+    if (!set_ov || !clear_ov) { printf("metal fa_vec override proc unavailable\n"); return false; }
+
+    struct shape_t { int dk, dv; };
+    const shape_t shapes[] = { {32,32},{64,64},{96,96},{128,128},{192,192},
+                               {192,128},{256,256},{320,256},{512,512},{576,512} };
+    const int  ne01_pts[]   = { 1, 2, 3, 8, 17 };       // 1, Q+/-1, 2Q-1, prime; vec upper bound < 20
+    const int  ne11_pts[]   = { 512, 4096, 8192 };      // drives adaptive nsg 1/2/4
+    const int  ne11_kvpad[] = { 4097, 8193 };           // has_kvpad (non-32-multiple kv)
+    const bool mask_pts[]   = { true, false };
+    const bool sinks_pts[]  = { false, true };
+
+    bool ok = true;
+    int n_run = 0, n_fail = 0;
+    for (auto s : shapes) {
+        for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+            for (int Q : {1, 2, 4}) {
+                for (bool mask : mask_pts) {
+                    for (bool sinks : sinks_pts) {
+                        for (int ne01 : ne01_pts) {
+                            for (int ne11 : ne11_pts) {
+                                set_ov(Q, ne);
+                                test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, {1,1}, /*kv=*/ne11, /*nb=*/ne01,
+                                                       mask, sinks, 0.0f, 0.0f, GGML_PREC_F32,
+                                                       GGML_TYPE_F16, GGML_TYPE_F16);
+                                auto st = tc.eval(backend_metal, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+                                clear_ov();
+                                if (st == test_status_t::FAIL) {
+                                    printf("FAIL dk=%d dv=%d Q=%d ne=%d ne01=%d ne11=%d mask=%d sinks=%d\n",
+                                           s.dk, s.dv, Q, ne, ne01, ne11, (int) mask, (int) sinks);
+                                    ok = false; n_fail++;
+                                }
+                                n_run++;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+    // kvpad pass (non-32-multiple kv); dk128/256 only to keep the case count down
+    for (auto s : shapes) {
+        if (!((s.dk == 128 && s.dv == 128) || (s.dk == 256 && s.dv == 256))) {
+            continue;
+        }
+        for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+            for (int Q : {1, 2, 4}) {
+                for (int ne11 : ne11_kvpad) {
+                    set_ov(Q, ne);
+                    test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, {1,1}, /*kv=*/ne11, /*nb=*/8,
+                                           /*mask=*/true, /*sinks=*/false, 0.0f, 0.0f, GGML_PREC_F32,
+                                           GGML_TYPE_F16, GGML_TYPE_F16);
+                    auto st = tc.eval(backend_metal, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+                    clear_ov();
+                    if (st == test_status_t::FAIL) {
+                        printf("FAIL(kvpad) dk=%d dv=%d Q=%d ne=%d ne11=%d\n", s.dk, s.dv, Q, ne, ne11);
+                        ok = false; n_fail++;
+                    }
+                    n_run++;
+                }
+            }
+        }
+    }
+    // quantized K/V numerical check: the rebased body's dequant path is Q-generic
+    // (dequant once, reuse across Q rows); confirm it stays correct for every quant precision.
+    const ggml_type qtypes[] = { GGML_TYPE_Q4_0, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0 };
+    const int  q_ne01[] = { 1, 2, 3, 8 };
+    const int  q_ne11[] = { 512, 4096, 8192 };
+    for (ggml_type qt : qtypes) {
+        for (auto s : shapes) {
+            for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+                for (int Q : { 1, 2, 4 }) {
+                    for (bool sinks : { false, true }) {
+                        for (int ne01 : q_ne01) {
+                            for (int ne11 : q_ne11) {
+                                set_ov(Q, ne);
+                                test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, {1, 1}, /*kv=*/ne11, /*nb=*/ne01,
+                                                       /*mask=*/true, sinks, 0.0f, 0.0f, GGML_PREC_F32, qt, qt);
+                                auto st = tc.eval(backend_metal, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+                                clear_ov();
+                                if (st == test_status_t::FAIL) {
+                                    printf("FAIL(quant) type=%s dk=%d dv=%d Q=%d ne=%d ne01=%d ne11=%d sinks=%d\n",
+                                           ggml_type_name(qt), s.dk, s.dv, Q, ne, ne01, ne11, (int) sinks);
+                                    ok = false; n_fail++;
+                                }
+                                n_run++;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+    printf("fa_vec tune-check: %d cases run, %d failed\n", n_run, n_fail);
+    return ok;
+}
+
+// A prebuilt FA op graph for one (dk,dv,ne01,ne11) cell. The op is replicated n_runs
+// times so a single graph_compute amortizes dispatch/sync overhead. The graph is reused
+// across (Q,NE) overrides (the override only changes the pipeline picked at encode time),
+// which avoids re-allocating/re-initializing the (large) K/V tensors per candidate.
+struct fa_perf_cell {
+    ggml_context_ptr        ctx;
+    ggml_backend_buffer_ptr buf;
+    ggml_cgraph *           gf     = nullptr;
+    int                     n_runs = 0;
+    bool                    ok     = false;
+};
+
+static fa_perf_cell fa_build_perf_cell(ggml_backend_t backend, int dk, int dv, int ne01, int ne11,
+                                       ggml_type type_kv = GGML_TYPE_F16) {
+    fa_perf_cell cell;
+
+    // GQA shape (nr23=[8,1]) matching real spec-decode / verify workloads: enough query
+    // heads to keep the GPU busy so the Q>1 K/V-reuse benefit is visible (and comparable
+    // to the documented #23114 numbers). nh here is the number of KV heads.
+    test_flash_attn_ext tc(dk, dv, /*nh=*/4, { 8, 1 }, /*kv=*/ne11, /*nb=*/ne01,
+                           /*mask=*/true, /*sinks=*/false, 0.0f, 0.0f, GGML_PREC_F32,
+                           type_kv, type_kv);
+
+    const size_t graph_nodes = 1024;
+    ggml_init_params params = {
+        /* .mem_size  = */ ggml_tensor_overhead() * 128 + ggml_graph_overhead_custom(graph_nodes, false),
+        /* .mem_base  = */ NULL,
+        /* .no_alloc  = */ true,
+    };
+    cell.ctx.reset(ggml_init(params));
+    GGML_ASSERT(cell.ctx);
+
+    ggml_tensor * out = tc.build_graph(cell.ctx.get());
+    if (!ggml_backend_supports_op(backend, out)) {
+        return cell;
+    }
+
+    cell.buf.reset(ggml_backend_alloc_ctx_tensors(cell.ctx.get(), backend));
+    if (cell.buf == NULL) {
+        return cell;
+    }
+
+    tc.initialize_tensors(cell.ctx.get());
+
+    cell.gf = ggml_new_graph_custom(cell.ctx.get(), graph_nodes, false);
+    ggml_build_forward_expand(cell.gf, out);
+
+    // replicate the op to amortize overhead (target ~50 GFLOP/compute, capped to bound graph size)
+    cell.n_runs = 1;
+    if (tc.op_flops(out) > 0) {
+        const uint64_t target_flops = 50ULL * 1000 * 1000 * 1000;
+        const int      cap          = 512;
+        const int      by_flops     = (int) std::min<int64_t>(cap, (int64_t) (target_flops / tc.op_flops(out)));
+        cell.n_runs = std::max(1, std::min<int>(by_flops, (int) (ggml_graph_size(cell.gf) - ggml_graph_n_nodes(cell.gf))));
+    }
+    for (int i = 1; i < cell.n_runs; ++i) {
+        ggml_graph_add_node(cell.gf, out);
+    }
+    cell.ok = true;
+    return cell;
+}
+
+// Median per-op GPU time (us) for the currently-set override over the prebuilt cell graph.
+static double time_fa_cell_median(ggml_backend_t backend, const fa_perf_cell & cell, int reps) {
+    if (!cell.ok) {
+        return -1.0;
+    }
+    ggml_backend_graph_compute(backend, cell.gf);  // warmup (compiles the pipeline for the override)
+    ggml_backend_synchronize(backend);
+
+    std::vector<double> samples;
+    samples.reserve(reps);
+    for (int r = 0; r < reps; ++r) {
+        const int64_t t0 = ggml_time_us();
+        ggml_backend_graph_compute(backend, cell.gf);
+        ggml_backend_synchronize(backend);
+        samples.push_back((double) (ggml_time_us() - t0));
+    }
+    std::nth_element(samples.begin(), samples.begin() + samples.size() / 2, samples.end());
+    return samples[samples.size() / 2] / cell.n_runs;
+}
+
+// Perf sweep over the (Q,NE) grid, emitting pasteable fa_vec_tuned_table rows.
+// nsg/nwg are left to the ops.cpp adaptive heuristic (not part of the table).
+static bool run_fa_vec_tune_perf(ggml_backend_t backend_metal) {
+    auto * reg    = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_metal));
+    auto   set_ov      = (set_fa_vec_override_t)   ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_set_fa_vec_override");
+    auto   clr_ov      = (clear_fa_vec_override_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_clear_fa_vec_override");
+    // Share the runtime's bucketers + baseline NE via read-only proc bridges, so the emitted
+    // (ne11_b, ne01_b) keys match fa_vec_pick and can't silently desync from FA_VEC_*_BUCKETS.
+    auto   ne11_bucket = (fa_vec_bucket_t)      ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_fa_vec_ne11_bucket");
+    auto   ne01_bucket = (fa_vec_bucket_t)      ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_fa_vec_ne01_bucket");
+    auto   baseline_ne = (fa_vec_baseline_ne_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_fa_vec_baseline_ne");
+    if (!set_ov || !clr_ov || !ne11_bucket || !ne01_bucket || !baseline_ne) {
+        printf("metal fa_vec tuning procs unavailable\n");
+        return false;
+    }
+
+    struct shape_t { int dk, dv; };
+    const shape_t shapes[] = { { 32, 32 }, { 64, 64 }, { 96, 96 }, { 128, 128 }, { 192, 192 },
+                               { 192, 128 }, { 256, 256 }, { 320, 256 }, { 512, 512 }, { 576, 512 } };
+    const int ne11_rep[] = { 512, 2048, 8192, 32768 };  // ne11 bucket representatives
+    const int ne01_rep[] = { 1, 2, 3, 4, 5, 6, 7, 8, 16 };  // PR grid: point-bucket reps (1-4) + tail mod-4 cycle (5-8) + large anchor (16); vec serves ne01<20
+    const int REPS = 7;                                 // odd -> exact median
+
+    printf("# fa_vec perf sweep — replace GGML_METAL_DEVICE_M4_MAX with this machine's device\n");
+    printf("# [1] per-bucket timings (us, winner *): '=> cfg Nx' beats baseline, else '=> baseline'\n");
+    printf("# [2] a pasteable fa_vec_tuned_table block (domain defaults + exceptions) is printed after [1]\n");
+
+    struct cand_t { int Q, NE; double t; };  // one timed (Q,NE) candidate
+    struct pt_t   { int dk, dv, ne11, ne01;   // one swept grid point with its candidate times
+                    std::vector<cand_t> cs; double base_t; };
+
+    // group_split compression knobs (see ggml-metal-tuning.h for the row / lookup semantics)
+    const double TUNE_TAU   = 0.05;  // max POINTWISE regret (worst point in a bucket) to ride a domain
+                                     // default instead of emitting the bucket's own exception row
+    const double TUNE_THETA = 1.05;  // min AGGREGATE bucket speedup vs baseline to tune the bucket at all
+
+    auto type_token = [](ggml_type t) -> const char * {
+        switch (t) {
+            case GGML_TYPE_Q4_0: return "GGML_TYPE_Q4_0";
+            case GGML_TYPE_Q4_1: return "GGML_TYPE_Q4_1";
+            case GGML_TYPE_Q5_0: return "GGML_TYPE_Q5_0";
+            case GGML_TYPE_Q5_1: return "GGML_TYPE_Q5_1";
+            case GGML_TYPE_Q8_0: return "GGML_TYPE_Q8_0";
+            default:             return "GGML_TYPE_F16";
+        }
+    };
+
+    const ggml_type types[] = { GGML_TYPE_F16, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1,
+                                GGML_TYPE_Q5_0, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0 };
+    for (ggml_type type_kv : types) {
+        printf("\n### dtype=%s\n", ggml_type_name(type_kv));
+        std::vector<pt_t> pts;  // one per swept (shape, ne11, ne01); bucketed + compressed below
+
+        for (auto s : shapes) {
+            const int base_ne = baseline_ne(s.dk, s.dv);
+            const std::vector<int> legal = fa_vec_legal_ne(s.dk, s.dv);
+            for (int ne11 : ne11_rep) {
+                for (int ne01 : ne01_rep) {
+                    fa_perf_cell cell = fa_build_perf_cell(backend_metal, s.dk, s.dv, ne01, ne11, type_kv);
+                    if (!cell.ok) {
+                        continue;
+                    }
+
+                    std::vector<cand_t> cs;
+                    for (int ne : legal) {
+                        for (int Q : { 1, 2, 4 }) {
+                            cs.push_back({ Q, ne, 0.0 });
+                        }
+                    }
+                    // randomize config order to decorrelate thermal throttling across the cell
+                    std::shuffle(cs.begin(), cs.end(), std::mt19937(1234));
+
+                    double anchor = 0.0;  // periodic baseline re-measure -> throttling detection
+                    for (size_t i = 0; i < cs.size(); ++i) {
+                        set_ov(cs[i].Q, cs[i].NE);
+                        cs[i].t = time_fa_cell_median(backend_metal, cell, REPS);
+                        clr_ov();
+
+                        if (i % 4 == 0) {
+                            const double a = time_fa_cell_median(backend_metal, cell, REPS);
+                            if (anchor > 0.0) {
+                                double drift = (a - anchor) / anchor;
+                                if (drift < 0) {
+                                    drift = -drift;
+                                }
+                                if (drift > 0.10) {
+                                    printf("# WARN throttling? anchor drift %.1f%% dk=%d ne11=%d\n", 100.0 * drift, s.dk, ne11);
+                                }
+                            }
+                            anchor = (anchor > 0.0) ? std::min(anchor, a) : a;
+                        }
+                    }
+
+                    std::sort(cs.begin(), cs.end(), [](const cand_t & a, const cand_t & b) {
+                        return a.Q != b.Q ? a.Q < b.Q : a.NE < b.NE;
+                    });
+                    cand_t best = cs[0];
+                    for (const auto & c : cs) {
+                        if (c.t > 0.0 && (best.t <= 0.0 || c.t < best.t)) {
+                            best = c;
+                        }
+                    }
+                    double base_t = 0.0;  // the swept (Q=1, base_ne) candidate == baseline kernel
+                    for (const auto & c : cs) {
+                        if (c.Q == 1 && c.NE == base_ne) { base_t = c.t; break; }
+                    }
+                    const bool keep = best.t > 0.0 && base_t > 0.0 && best.t < base_t * 0.98;
+
+                    printf("# dtype=%s dk=%d dv=%d ne11=%d ne01=%d:", ggml_type_name(type_kv), s.dk, s.dv, ne11, ne01);
+                    for (const auto & c : cs) {
+                        printf("  Q%dNE%d=%.1f%s", c.Q, c.NE, c.t, (c.Q == best.Q && c.NE == best.NE) ? "*" : "");
+                    }
+                    if (keep) {
+                        printf("  => Q%d,NE%d  %.2fx\n", best.Q, best.NE, base_t / best.t);
+                    } else {
+                        printf("  => baseline\n");
+                    }
+                    pts.push_back({ s.dk, s.dv, ne11, ne01, cs, base_t });
+                }
+            }
+        }
+
+        // [2] Compress into pasteable rows. Per (dk,dv) and ne01 domain {decode==1, batch>=2}, emit
+        // one ne11-collapsed default cfg (fewest rows, ne11_b=-1) plus a per-bucket exception wherever
+        // the default's pointwise regret vs the bucket's min-max-regret target, or its aggregate
+        // slowdown vs baseline, exceeds TUNE_TAU. The target still carries an intrinsic per-point regret
+        // from lumping ne01 into one tail bucket — TUNE_TAU bounds resolved-vs-target, not vs-oracle.
+        std::vector<std::string> rows_out;
+        char rbuf[192];
+        for (auto s : shapes) {
+            const int base_ne = baseline_ne(s.dk, s.dv);
+            std::vector<cand_t> cfgs;  // candidate list (identical for every grid point of this shape)
+            int base_i = 0;
+            for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+                for (int Q : { 1, 2, 4 }) {
+                    if (Q == 1 && ne == base_ne) { base_i = (int) cfgs.size(); }
+                    cfgs.push_back({ Q, ne, 0.0 });
+                }
+            }
+            auto cfg_time = [&](const pt_t & p, int i) {
+                for (const auto & c : p.cs) { if (c.Q == cfgs[i].Q && c.NE == cfgs[i].NE) { return c.t; } }
+                return 0.0;
+            };
+
+            // Bucket the grid points using the runtime's bucketers (via proc), so keys match fa_vec_pick.
+            // Short-KV points (ne11 bucket 0) are dropped — the runtime serves those from baseline. Each
+            // kept bucket picks a min-max-regret target (gated by the aggregate TUNE_THETA benefit gate).
+            struct bkt_t { int b11, b01, Ti; std::vector<double> agg; double base_agg;
+                           std::vector<const pt_t *> bp; };  // bp kept for the pointwise ride test below
+            std::set<std::pair<int, int>> seen;
+            for (const auto & p : pts) {
+                if (p.dk != s.dk || p.dv != s.dv) { continue; }
+                const int b11 = ne11_bucket(p.ne11);
+                if (b11 == 0) { continue; }
+                seen.insert({ b11, ne01_bucket(p.ne01) });
+            }
+            std::vector<bkt_t> bks;
+            for (const auto & bb : seen) {
+                const int b11 = bb.first, b01 = bb.second;
+                std::vector<const pt_t *> bp;
+                for (const auto & p : pts) {
+                    if (p.dk == s.dk && p.dv == s.dv && ne11_bucket(p.ne11) == b11 && ne01_bucket(p.ne01) == b01) {
+                        bp.push_back(&p);
+                    }
+                }
+                std::vector<double> agg(cfgs.size(), 0.0), worst(cfgs.size(), 0.0);
+                for (const auto * p : bp) {
+                    double bestt = 0.0;
+                    for (size_t i = 0; i < cfgs.size(); ++i) {
+                        double t = cfg_time(*p, (int) i);
+                        if (t > 0.0 && (bestt == 0.0 || t < bestt)) { bestt = t; }
+                    }
+                    for (size_t i = 0; i < cfgs.size(); ++i) {
+                        double t = cfg_time(*p, (int) i);
+                        agg[i] += t;
+                        if (t > 0.0 && bestt > 0.0) { worst[i] = std::max(worst[i], t / bestt); }
+                    }
+                }
+                int robust = 0;
+                for (size_t i = 1; i < cfgs.size(); ++i) {
+                    if (worst[i] < worst[robust] ||
+                        (worst[i] == worst[robust] && (cfgs[i].Q < cfgs[robust].Q ||
+                        (cfgs[i].Q == cfgs[robust].Q && cfgs[i].NE < cfgs[robust].NE)))) {
+                        robust = (int) i;
+                    }
+                }
+                const bool tune = robust != base_i && agg[base_i] / agg[robust] >= TUNE_THETA;
+                bks.push_back({ b11, b01, tune ? robust : base_i, agg, agg[base_i], bp });
+            }
+
+            // Pointwise regret of default cfg d vs the bucket target Ti. Per-point (not aggregate):
+            // a ratio-of-sums lets a default that wins on aligned ne01 (8, 16) hide a large penalty
+            // on a misaligned point (ne01=5), so the ride/exception decision must be per point.
+            auto reg_pointwise = [&](const bkt_t * b, int d) {
+                double r = 0.0;
+                for (const auto * p : b->bp) {
+                    const double td = cfg_time(*p, d), tT = cfg_time(*p, b->Ti);
+                    if (td > 0.0 && tT > 0.0) { r = std::max(r, td / tT - 1.0); }
+                }
+                return r;
+            };
+
+            for (int dom = 0; dom <= 1; ++dom) {  // 0 = decode (ne01==1), 1 = batch (ne01>=2)
+                std::vector<const bkt_t *> db;
+                for (const auto & b : bks) { if ((dom == 0) == (b.b01 == 0)) { db.push_back(&b); } }
+                if (db.empty()) { continue; }
+
+                // default cfg = the one minimizing (#rows, total achieved time, Q, NE)
+                int bestD = -1, bestRows = 1 << 30; double bestTot = 0.0;
+                for (size_t d = 0; d < cfgs.size(); ++d) {
+                    int rows = ((int) d != base_i) ? 1 : 0; double tot = 0.0;
+                    for (const auto * b : db) {
+                        const double reg  = reg_pointwise(b, (int) d);       // pointwise vs bucket target
+                        const double slow = b->agg[d] / b->base_agg - 1.0;   // aggregate vs baseline (safety net)
+                        if (reg > TUNE_TAU || slow > TUNE_TAU) { rows++; tot += b->agg[b->Ti]; }
+                        else                                   {         tot += b->agg[d];      }
+                    }
+                    const bool better = bestD < 0 || rows < bestRows ||
+                        (rows == bestRows && (tot < bestTot ||
+                        (tot == bestTot && (cfgs[d].Q < cfgs[bestD].Q ||
+                        (cfgs[d].Q == cfgs[bestD].Q && cfgs[d].NE < cfgs[bestD].NE)))));
+                    if (better) { bestD = (int) d; bestRows = rows; bestTot = tot; }
+                }
+
+                const int dom_id = (dom == 0) ? 0 : 1;  // FA_VEC_DOMAIN_DECODE / FA_VEC_DOMAIN_BATCH
+                if (bestD != base_i) {
+                    snprintf(rbuf, sizeof(rbuf), "    { { GGML_METAL_DEVICE_M4_MAX, %s, %d, %d, -1, %d }, { %d, %d } },",
+                             type_token(type_kv), s.dk, s.dv, dom_id, cfgs[bestD].Q, cfgs[bestD].NE);
+                    rows_out.emplace_back(rbuf);
+                }
+                for (const auto * b : db) {
+                    const double reg  = reg_pointwise(b, bestD);            // pointwise vs bucket target
+                    const double slow = b->agg[bestD] / b->base_agg - 1.0;  // aggregate vs baseline
+                    if (reg <= TUNE_TAU && slow <= TUNE_TAU) { continue; }  // rides the default / baseline
+                    snprintf(rbuf, sizeof(rbuf), "    { { GGML_METAL_DEVICE_M4_MAX, %s, %d, %d, %d, %d }, { %d, %d } },",
+                             type_token(type_kv), s.dk, s.dv, b->b11, b->b01, cfgs[b->Ti].Q, cfgs[b->Ti].NE);
+                    rows_out.emplace_back(rbuf);
+                }
+            }
+        }
+        printf("\n    // ---- %s: %zu rows ----\n", ggml_type_name(type_kv), rows_out.size());
+        for (const auto & r : rows_out) { printf("%s\n", r.c_str()); }
+    }  // for type_kv
+    return true;
+}
+
 static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mode mode, const char * op_names_filter, const char * params_filter,
-                         printer * output_printer, const char * test_file_path, int parallel_workers) {
+                         printer * output_printer, const char * test_file_path, int parallel_workers, bool tune_perf = false) {
     auto filter_test_cases = [](std::vector<std::unique_ptr<test_case>> & test_cases, const char * params_filter) {
         if (params_filter == nullptr) {
             return;
@@ -10111,6 +10617,9 @@ static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mo
         case MODE_PERF:
             test_cases = make_test_cases_perf();
             break;
+        case MODE_TUNE:
+            // MODE_TUNE routes to its own dispatch below; no generic test_cases needed
+            break;
         }
     } else {
         test_cases = make_test_cases_from_file(test_file_path);
@@ -10246,6 +10755,29 @@ static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mo
         return true;
     }
 
+    if (mode == MODE_TUNE) {
+        // self-create a CPU backend with the reference implementation as golden.
+        // (backend_cpu in the MODE_TEST block above is out of scope here.)
+        ggml_backend_t backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, NULL);
+        GGML_ASSERT(backend_cpu != NULL);
+        {
+            using ggml_backend_cpu_set_use_ref_t = void (*)(ggml_backend_t, bool);
+            auto * cpu_reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu));
+            auto * set_use_ref = (ggml_backend_cpu_set_use_ref_t) ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_cpu_set_use_ref");
+            if (set_use_ref) {
+                set_use_ref(backend_cpu, true);
+            }
+        }
+
+        // backend here is the MODE_TUNE-selected metal backend (-b MTL0)
+        const bool ok = tune_perf
+            ? run_fa_vec_tune_perf(backend)
+            : (run_fa_vec_drift_guard(backend, backend_cpu) && run_fa_vec_tune_check(backend, backend_cpu));
+
+        ggml_backend_free(backend_cpu);
+        return ok;
+    }
+
     if (mode == MODE_SUPPORT) {
         // Filter out fusion cases
         test_cases.erase(
@@ -10371,6 +10903,7 @@ static void usage(char ** argv) {
     printf("      - grad (compare gradients from backpropagation with method of finite differences)\n");
     printf("      - perf (performance evaluation)\n");
     printf("      - support (probe backend operation support)\n");
+    printf("      - tune (FA vec (Q,NE) numerical correctness check vs CPU reference; --tune-perf for the perf sweep)\n");
     printf("    op names for -o are as given by ggml_op_desc() (e.g. ADD, MUL_MAT, etc),\n");
     printf("        optionally including the full test case string (e.g. \"ADD(type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1)\")\n");
     printf("    --output specifies output format (default: console, options: console, sql, csv)\n");
@@ -10388,6 +10921,7 @@ int main(int argc, char ** argv) {
     const char * params_filter = nullptr;
     const char * test_file_path = nullptr;
     int parallel_workers = 1;
+    bool tune_perf = false;
 
     for (int i = 1; i < argc; i++) {
         if (strcmp(argv[i], "test") == 0) {
@@ -10398,6 +10932,10 @@ int main(int argc, char ** argv) {
             mode = MODE_GRAD;
         } else if (strcmp(argv[i], "support") == 0) {
             mode = MODE_SUPPORT;
+        } else if (strcmp(argv[i], "tune") == 0) {
+            mode = MODE_TUNE;
+        } else if (strcmp(argv[i], "--tune-perf") == 0) {
+            tune_perf = true;
         } else if (strcmp(argv[i], "-o") == 0) {
             if (i + 1 < argc) {
                 op_names_filter = argv[++i];
@@ -10505,7 +11043,7 @@ int main(int argc, char ** argv) {
                                                              false, "", ggml_backend_dev_description(dev),
                                                              total / 1024 / 1024, free / 1024 / 1024, true));
 
-        bool ok = test_backend(backend.get(), dev, mode, op_names_filter, params_filter, output_printer.get(), test_file_path, parallel_workers);
+        bool ok = test_backend(backend.get(), dev, mode, op_names_filter, params_filter, output_printer.get(), test_file_path, parallel_workers, tune_perf);
 
         if (ok) {
             n_ok++;

tests/test-chat.cpp

@@ -730,6 +730,71 @@ static common_chat_tool imaginary_number_tool{
     })",
 };
 
+static common_chat_tool nested_args_tool{
+    /* .name = */ "nested_args",
+    /* .description = */ "Tool with nested array arguments",
+    /* .parameters = */ R"({
+        "type": "object",
+        "properties": {
+            "tags": {
+                "type": "array",
+                "items": { "type": "string" }
+            },
+            "entries": {
+                "type": "array",
+                "items": {
+                    "type": "object",
+                    "properties": {
+                        "id": { "type": "integer" },
+                        "label": { "type": "string" }
+                    },
+                    "required": ["id", "label"]
+                }
+            }
+        },
+        "required": ["tags", "entries"]
+    })",
+};
+
+static common_chat_tool union_args_tool{
+    /* .name = */ "union_args",
+    /* .description = */ "Tool with union arguments",
+    /* .parameters = */ R"({
+        "type": "object",
+        "properties": {
+            "filter": {
+                "anyOf": [
+                    { "type": "array", "items": { "type": "string" } },
+                    {
+                        "type": "object",
+                        "properties": {
+                            "field": { "type": "string" },
+                            "op": { "type": "string" }
+                        },
+                        "required": ["field", "op"]
+                    }
+                ]
+            },
+            "label": {
+                "oneOf": [
+                    { "type": "string" },
+                    { "type": "object", "properties": { "text": { "type": "string" } } }
+                ]
+            },
+            "limit": {
+                "oneOf": [
+                    { "type": "integer" },
+                    {
+                        "type": "object",
+                        "properties": { "max": { "type": "integer" } },
+                        "required": ["max"]
+                    }
+                ]
+            }
+        }
+    })",
+};
+
 static common_chat_tool nullable_string_tool{
     /* .name = */ "set_nullable_str",
     /* .description = */ "Set a nullable string value",
@@ -4850,6 +4915,370 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .run();
     }
 
+    // MiniMax-M3 tests - namespaced XML invoke format, the parameter name is the tag
+    // Format:
+    //   ]<]minimax[>[<tool_call>
+    //   ]<]minimax[>[<invoke name="get_time">]<]minimax[>[<city>Tokyo]<]minimax[>[</city>]<]minimax[>[</invoke>
+    //   ]<]minimax[>[</tool_call>
+    // Reasoning uses <mm:think>...</mm:think>. The generation prompt is only "]~b]ai\n", so the model
+    // opens the thinking block itself; a turn without reasoning is prefixed with a bare </mm:think>.
+    {
+        auto tst = peg_tester("models/templates/MiniMax-M3.jinja", detailed_debug);
+
+        // Content only (bare </mm:think> prefix)
+        tst.test("</mm:think>Hello, world!\nWhat's up?")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect(message_assist)
+            .expect_reconstruction()
+            .run();
+
+        // Thinking + content
+        tst.test("<mm:think>I'm\nthinking</mm:think>Hello, world!\nWhat's up?")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect(message_assist_thoughts)
+            .expect_reconstruction()
+            .run();
+
+        // Thinking + tool call (single, string param)
+        tst.test(
+               "<mm:think>Let me check the time</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"get_time\">"
+               "]<]minimax[>[<city>Tokyo]<]minimax[>[</city>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ get_time_tool })
+            .expect(message_with_tool_calls_and_reasoning("get_time", R"({"city": "Tokyo"})", "Let me check the time"))
+            .expect_reconstruction()
+            .run();
+
+        // Tool call without reasoning, integer param
+        tst.test(
+               "</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"special_function\">"
+               "]<]minimax[>[<arg1>1]<]minimax[>[</arg1>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ special_function_tool })
+            .expect(message_assist_call)
+            .expect_reconstruction()
+            .run();
+
+        // Tool call with no parameters
+        tst.test(
+               "<mm:think>Let's call a tool:</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"empty_args\">"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ empty_args_tool })
+            .expect(message_with_reasoning_and_tool_call("Let's call a tool:", "empty_args", "{}"))
+            .expect_reconstruction()
+            .run();
+
+        // Multiple parallel tool calls in one block
+        tst.test(
+               "<mm:think>Calling both</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"get_time\">"
+               "]<]minimax[>[<city>Paris]<]minimax[>[</city>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[<invoke name=\"get_weather\">"
+               "]<]minimax[>[<city>Paris]<]minimax[>[</city>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .parallel_tool_calls(true)
+            .tools({ get_time_tool, get_weather_tool })
+            .expect(message_with_reasoning_content_and_multiple_tool_calls(
+                "Calling both", "",
+                { { "get_time", R"({"city": "Paris"})" }, { "get_weather", R"({"city": "Paris"})" } }))
+            .expect_reconstruction()
+            .run();
+
+        // Content before the tool call block
+        tst.test(
+               "<mm:think>Thinking about it</mm:think>"
+               "Let me call the function."
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"special_function\">"
+               "]<]minimax[>[<arg1>1]<]minimax[>[</arg1>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ special_function_tool })
+            .expect_reasoning("Thinking about it")
+            .expect_content("Let me call the function.")
+            .expect_tool_calls({
+                { "special_function", R"({"arg1": 1})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // Negative number
+        tst.test(
+               "<mm:think>Test negative</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"magic_int\">"
+               "]<]minimax[>[<ref>-14]<]minimax[>[</ref>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ magic_int_tool })
+            .expect_reasoning("Test negative")
+            .expect_tool_calls({
+                { "magic_int", R"({"ref": -14})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // Decimal number
+        tst.test(
+               "<mm:think>Test decimal</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"amount\">"
+               "]<]minimax[>[<orig>3.14]<]minimax[>[</orig>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ amount_tool })
+            .expect_reasoning("Test decimal")
+            .expect_tool_calls({
+                { "amount", R"({"orig": 3.14})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // Boolean
+        tst.test(
+               "<mm:think>Test boolean</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"toggle\">"
+               "]<]minimax[>[<enabled>true]<]minimax[>[</enabled>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ toggle_tool })
+            .expect_reasoning("Test boolean")
+            .expect_tool_calls({
+                { "toggle", R"({"enabled": true})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // Multiple params of mixed types (required int first, then optional string)
+        tst.test(
+               "<mm:think>Multi-arg call</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"magic_int\">"
+               "]<]minimax[>[<ref>42]<]minimax[>[</ref>"
+               "]<]minimax[>[<name>foo bar]<]minimax[>[</name>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ magic_int_tool })
+            .expect_reasoning("Multi-arg call")
+            .expect_tool_calls({
+                { "magic_int", R"({"ref": 42, "name": "foo bar"})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // Nested object param, expanded into one element per key
+        tst.test(
+               "<mm:think>Nested object</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"imaginary_number\">"
+               "]<]minimax[>[<number>"
+               "]<]minimax[>[<real>1.5]<]minimax[>[</real>"
+               "]<]minimax[>[<imaginary>-2.5]<]minimax[>[</imaginary>"
+               "]<]minimax[>[</number>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ imaginary_number_tool })
+            .expect_reasoning("Nested object")
+            .expect_tool_calls({
+                { "imaginary_number", R"({"number": {"real": 1.5, "imaginary": -2.5}})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // Array params, expanded into <item> elements (of scalars and of objects)
+        tst.test(
+               "<mm:think>Nested arrays</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"nested_args\">"
+               "]<]minimax[>[<tags>"
+               "]<]minimax[>[<item>alpha]<]minimax[>[</item>"
+               "]<]minimax[>[<item>beta]<]minimax[>[</item>"
+               "]<]minimax[>[</tags>"
+               "]<]minimax[>[<entries>"
+               "]<]minimax[>[<item>"
+               "]<]minimax[>[<id>1]<]minimax[>[</id>"
+               "]<]minimax[>[<label>one]<]minimax[>[</label>"
+               "]<]minimax[>[</item>"
+               "]<]minimax[>[</entries>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ nested_args_tool })
+            .expect_reasoning("Nested arrays")
+            .expect_tool_calls({
+                { "nested_args", R"({"tags": ["alpha", "beta"], "entries": [{"id": 1, "label": "one"}]})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // Union params (anyOf/oneOf), expanded as a choice of the alternatives
+        tst.test(
+               "<mm:think>Union array</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"union_args\">"
+               "]<]minimax[>[<filter>"
+               "]<]minimax[>[<item>alpha]<]minimax[>[</item>"
+               "]<]minimax[>[<item>beta]<]minimax[>[</item>"
+               "]<]minimax[>[</filter>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ union_args_tool })
+            .expect_reasoning("Union array")
+            .expect_tool_calls({
+                { "union_args", R"({"filter": ["alpha", "beta"]})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // oneOf between a scalar and an object
+        tst.test(
+               "<mm:think>Union scalar</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"union_args\">"
+               "]<]minimax[>[<limit>5]<]minimax[>[</limit>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ union_args_tool })
+            .expect_reasoning("Union scalar")
+            .expect_tool_calls({
+                { "union_args", R"({"limit": 5})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        tst.test(
+               "<mm:think>Union nested</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"union_args\">"
+               "]<]minimax[>[<limit>"
+               "]<]minimax[>[<max>10]<]minimax[>[</max>"
+               "]<]minimax[>[</limit>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ union_args_tool })
+            .expect_reasoning("Union nested")
+            .expect_tool_calls({
+                { "union_args", R"({"limit": {"max": 10}})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // A union with a string alternative is a string
+        tst.test(
+               "<mm:think>Union string</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"union_args\">"
+               "]<]minimax[>[<label>hi]<]minimax[>[</label>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ union_args_tool })
+            .expect_reasoning("Union string")
+            .expect_tool_calls({
+                { "union_args", R"({"label": "hi"})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // ... even when the value looks structured
+        tst.test(
+               "<mm:think>Union string</mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"union_args\">"
+               "]<]minimax[>[<label>"
+               "]<]minimax[>[<text>hi]<]minimax[>[</text>"
+               "]<]minimax[>[</label>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ union_args_tool })
+            .expect_reasoning("Union string")
+            .expect_tool_calls({
+                { "union_args", R"({"label": "]<]minimax[>[<text>hi]<]minimax[>[</text>"})", {} },
+            })
+            .expect_reconstruction()
+            .run();
+
+        // Edge case: empty reasoning followed by a tool call
+        tst.test(
+               "<mm:think></mm:think>"
+               "]<]minimax[>[<tool_call>\n"
+               "]<]minimax[>[<invoke name=\"get_time\">"
+               "]<]minimax[>[<city>XYZCITY]<]minimax[>[</city>"
+               "]<]minimax[>[</invoke>\n"
+               "]<]minimax[>[</tool_call>")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ get_time_tool })
+            .expect(message_with_tool_calls("get_time", R"({"city": "XYZCITY"})"))
+            .run();
+
+        // Continuation tests
+        tst.test("world!\nWhat's up?")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .enable_thinking(true)
+            .messages({ message_user, message_assist_prefill_content })
+            .add_generation_prompt(false)
+            .continue_final_message(COMMON_CHAT_CONTINUATION_CONTENT)
+            .expect_reasoning("I'm thinking")
+            .expect_content("Hello, world!\nWhat's up?")
+            .run();
+
+        tst.test(" thinking</mm:think>Hello, world!\nWhat's up?")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .enable_thinking(true)
+            .messages({ message_user, message_assist_prefill_reasoning })
+            .add_generation_prompt(false)
+            .continue_final_message(COMMON_CHAT_CONTINUATION_REASONING)
+            .expect_reasoning("I'm thinking")
+            .expect_content("Hello, world!\nWhat's up?")
+            .run();
+    }
+
     // NVIDIA-Nemotron-Nano-v2 tests - <TOOLCALL>...</TOOLCALL> format
     // Format: <TOOLCALL>[{"name": "func", "arguments": {...}}]</TOOLCALL>
     {

tests/test-save-load-state.cpp

@@ -44,8 +44,6 @@ static llama_tokens generate_tokens(llama_context * ctx, llama_sampler * smpl, i
         n_past++;
     }
 
-    llama_synchronize(ctx);
-
     return result;
 }
 

tools/cli/README.md

@@ -203,7 +203,7 @@
 | `--spec-draft-device, -devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
 | `--spec-draft-ngl, -ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
 | `--spec-draft-model, -md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_SPEC_DRAFT_MODEL) |
-| `--spec-type none,draft-simple,draft-eagle3,draft-mtp,draft-dflash,ngram-simple,ngram-map-k,ngram-map-k4v,ngram-mod,ngram-cache` | comma-separated list of types of speculative decoding to use (default: none)<br/><br/>(env: LLAMA_ARG_SPEC_TYPE) |
+| `--spec-type none,draft-simple,draft-eagle3,draft-mtp,draft-dflash,draft-dspark,ngram-simple,ngram-map-k,ngram-map-k4v,ngram-mod,ngram-cache` | comma-separated list of types of speculative decoding to use (default: none)<br/><br/>(env: LLAMA_ARG_SPEC_TYPE) |
 | `--spec-ngram-mod-n-min N` | minimum number of ngram tokens to use for ngram-based speculative decoding (default: 48) |
 | `--spec-ngram-mod-n-max N` | maximum number of ngram tokens to use for ngram-based speculative decoding (default: 64) |
 | `--spec-ngram-mod-n-match N` | ngram-mod lookup length (default: 24) |

tools/llama-bench/llama-bench.cpp

@@ -670,22 +670,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                     break;
                 }
             } else if (arg == "--list-devices") {
-                std::vector<ggml_backend_dev_t> devices;
-                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-                    auto * dev = ggml_backend_dev_get(i);
-                    if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU) {
-                        devices.push_back(dev);
-                    }
-                }
-                printf("Available devices:\n");
-                if (devices.empty()) {
-                    printf("  (none)\n");
-                }
-                for (auto * dev : devices) {
-                    size_t free, total;
-                    ggml_backend_dev_memory(dev, &free, &total);
-                    printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
-                }
+                common_print_available_devices();
                 exit(0);
             } else if (arg == "-t" || arg == "--threads") {
                 if (++i >= argc) {

tools/mtmd/CMakeLists.txt

@@ -51,6 +51,7 @@ add_library(mtmd
             models/qwen3vl.cpp
             models/mimovl.cpp
             models/qwen3a.cpp
+            models/mimo-audio.cpp
             models/step3vl.cpp
             models/siglip.cpp
             models/whisper-enc.cpp
@@ -59,6 +60,7 @@ add_library(mtmd
             models/mobilenetv5.cpp
             models/youtuvl.cpp
             models/yasa2.cpp
+            models/parakeet.cpp
             )
 
 set_target_properties(mtmd PROPERTIES

tools/mtmd/clip-graph.h

@@ -13,6 +13,14 @@
 
 struct build_vit_opts {
     ggml_tensor * attn_mask = nullptr;
+    // TODO @ngxson : merge attn_mask and attn_mask_layers into one call
+    std::vector<ggml_tensor *> attn_mask_layers; // one per layer
+
+    // hook at layer output embeddings
+    std::function<void(ggml_tensor * cur, int il)> callback_layer_out = nullptr;
+
+    // whether to skip the automatic post-layernorm (model.post_ln_w) applied at the end
+    bool skip_post_ln = false;
 };
 
 struct clip_graph {

tools/mtmd/clip-impl.h

@@ -82,6 +82,13 @@
 #define KEY_A_PROJ_WINDOW_SIZE     "clip.audio.projector.window_size"
 #define KEY_A_PROJ_DOWNSAMPLE_RATE "clip.audio.projector.downsample_rate"
 #define KEY_A_PROJ_HEAD_COUNT      "clip.audio.projector.head_count"
+#define KEY_A_RVQ_NUM_QUANTIZERS   "clip.audio.rvq.num_quantizers"   // mimo-audio-tokenizer
+#define KEY_A_RVQ_CODEBOOK_SIZE    "clip.audio.rvq.codebook_size"    // mimo-audio-tokenizer: per-quantizer bin count
+#define KEY_A_WA_PATTERN_MODE      "clip.audio.wa_pattern_mode"      // mimo-audio-tokenizer, per-layer -1 (full) / 0 (windowed)
+#define KEY_A_ATTN_WINDOW_SIZE     "clip.audio.window_size"          // mimo-audio-tokenizer: sliding-window radius
+#define KEY_A_LOCAL_BLOCK_COUNT    "clip.audio.local_block_count"    // mimo-v2.5: input_local_transformer layer count
+#define KEY_A_LOCAL_GROUP_SIZE     "clip.audio.local_group_size"     // mimo-v2.5: input_local_transformer grouping size
+#define KEY_AUDIO_SUBSAMPLING_FACTOR "clip.audio.subsampling_factor"
 
 //
 // tensor name constants
@@ -175,6 +182,24 @@
 #define TN_MM_NORM_PRE  "mm.a.norm_pre.%s"
 #define TN_MM_NORM_MID  "mm.a.norm_mid.%s"
 
+// mimo-audio-tokenizer
+#define TN_A_DOWNSAMPLE_CONV "a.downsample.conv.%s"
+#define TN_A_DOWNSAMPLE_NORM "a.downsample.norm.%s"
+#define TN_A_RVQ_CODEBOOK    "a.rvq.codebook.%s"
+// mimo-v2.5: text-side RVQ code embedding ("text codebook")
+#define TN_MM_A_CODE_EMBD    "mm.a.code_embd.%s"
+// mimo-v2.5: LLM-side connector (input_local_transformer)
+#define TN_MM_A_LOCAL_ATTN_Q   "mm.a.local_blk.%d.attn_q.%s"
+#define TN_MM_A_LOCAL_ATTN_K   "mm.a.local_blk.%d.attn_k.%s"
+#define TN_MM_A_LOCAL_ATTN_V   "mm.a.local_blk.%d.attn_v.%s"
+#define TN_MM_A_LOCAL_ATTN_OUT "mm.a.local_blk.%d.attn_out.%s"
+#define TN_MM_A_LOCAL_FFN_GATE "mm.a.local_blk.%d.ffn_gate.%s"
+#define TN_MM_A_LOCAL_FFN_UP   "mm.a.local_blk.%d.ffn_up.%s"
+#define TN_MM_A_LOCAL_FFN_DOWN "mm.a.local_blk.%d.ffn_down.%s"
+#define TN_MM_A_LOCAL_LN1      "mm.a.local_blk.%d.ln1.%s"
+#define TN_MM_A_LOCAL_LN2      "mm.a.local_blk.%d.ln2.%s"
+#define TN_MM_A_LOCAL_NORM     "mm.a.local_norm.%s"
+
 // cogvlm
 #define TN_MM_POST_FC_NORM "mm.post_fc_norm.%s"
 #define TN_MM_H_TO_4H      "mm.up.%s"
@@ -314,6 +339,12 @@
 #define TN_YASA_STAGE_DOWN_CONV  "v.stage.%d.down.conv.%s"
 #define TN_YASA_STAGE_BLK        "v.stage.%d.blk.%d.%s.%s"
 
+// parakeet
+#define TN_MEL_FILTERS           "a.mel_filters"
+#define TN_WINDOW                "a.window"
+#define TN_CONV_NORM_MEAN        "%s.blk.%d.conv_norm_mean"
+#define TN_CONV_NORM_VAR         "%s.blk.%d.conv_norm_var"
+
 // align x to upper multiple of n
 #define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
 
@@ -368,12 +399,14 @@ enum projector_type {
     PROJECTOR_TYPE_KIMIK25,
     PROJECTOR_TYPE_NEMOTRON_V2_VL,
     PROJECTOR_TYPE_HUNYUANVL,
+    PROJECTOR_TYPE_PARAKEET,
     PROJECTOR_TYPE_EXAONE4_5,
     PROJECTOR_TYPE_MINICPMV4_6,
     PROJECTOR_TYPE_GRANITE_SPEECH,
     PROJECTOR_TYPE_MIMOVL,
     PROJECTOR_TYPE_MINIMAX_M3,
     PROJECTOR_TYPE_GRANITE4_VISION,
+    PROJECTOR_TYPE_MIMO_AUDIO,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -429,6 +462,8 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MIMOVL,            "mimovl"},
     { PROJECTOR_TYPE_MINIMAX_M3,        "minimax_m3"},
     { PROJECTOR_TYPE_GRANITE4_VISION,   "granite4_vision"},
+    { PROJECTOR_TYPE_MIMO_AUDIO,        "mimo_audio"},
+    { PROJECTOR_TYPE_PARAKEET,          "parakeet"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

tools/mtmd/clip-model.h

@@ -110,6 +110,8 @@ struct clip_hparams {
     // audio
     int32_t n_mel_bins = 0; // whisper preprocessor
     int32_t proj_stack_factor = 0; // ultravox
+    int32_t subsampling_factor = 0; // parakeet
+
     int32_t audio_chunk_size           = 0;
     int32_t audio_conv_kernel_size     = 0;
     int32_t audio_max_pos_emb          = 0;
@@ -124,6 +126,18 @@ struct clip_hparams {
     int32_t audio_window_len  = -1;
     int32_t audio_hop_len     = -1;
 
+    // parakeet
+    std::vector<float> mel_filters;
+    std::vector<float> window;
+
+    // mimo-audio-tokenizer: residual vector quantizer
+    int32_t rvq_num_quantizers = 0;
+    std::vector<int32_t> rvq_codebook_size; // per-quantizer bin count (ragged, e.g. 1024/1024/256/128x17)
+
+    // mimo-v2.5: LLM-side connector (input_local_transformer)
+    int32_t audio_local_n_layer = 0;
+    int32_t audio_local_group_size = 0;
+
     // legacy
     bool has_llava_projector = false;
     int minicpmv_version = 0;
@@ -237,14 +251,16 @@ struct clip_layer {
     ggml_tensor * norm_conv_b   = nullptr;
     ggml_tensor * linear_pos_w  = nullptr;
 
-    ggml_tensor * conv_norm_w   = nullptr;
-    ggml_tensor * conv_norm_b   = nullptr;
-    ggml_tensor * conv_dw_w     = nullptr;
-    ggml_tensor * conv_dw_b     = nullptr;
-    ggml_tensor * conv_pw1_w    = nullptr;
-    ggml_tensor * conv_pw1_b    = nullptr;
-    ggml_tensor * conv_pw2_w    = nullptr;
-    ggml_tensor * conv_pw2_b    = nullptr;
+    ggml_tensor * conv_norm_w    = nullptr;
+    ggml_tensor * conv_norm_b    = nullptr;
+    ggml_tensor * conv_norm_mean = nullptr;  // parakeet
+    ggml_tensor * conv_norm_var  = nullptr;  // parakeet
+    ggml_tensor * conv_dw_w      = nullptr;
+    ggml_tensor * conv_dw_b      = nullptr;
+    ggml_tensor * conv_pw1_w     = nullptr;
+    ggml_tensor * conv_pw1_b     = nullptr;
+    ggml_tensor * conv_pw2_w     = nullptr;
+    ggml_tensor * conv_pw2_b     = nullptr;
 
     // gemma4 audio conformer per-layer
     ggml_tensor * attn_pre_norm_w   = nullptr;
@@ -537,6 +553,20 @@ struct clip_model {
     ggml_tensor * mm_norm_pre_b = nullptr;
     ggml_tensor * mm_norm_mid_w = nullptr;
 
+    // mimo-audio-tokenizer: post-transformer downsample + RVQ codebook
+    ggml_tensor * downsample_conv_w = nullptr; // no bias
+    ggml_tensor * downsample_norm_w = nullptr;
+    ggml_tensor * downsample_norm_b = nullptr;
+    ggml_tensor * rvq_codebook = nullptr; // merged 3D [n_q, max_bins, dim]
+
+    // mimo-v2.5: text-side RVQ code embedding ("text codebook")
+    ggml_tensor * mm_a_code_embd = nullptr; // merged 3D [n_channels, vocab, dim]
+
+    // mimo-v2.5: LLM-side connector (input_local_transformer, separate from the
+    // audio_tokenizer's own encoder `layers`)
+    std::vector<clip_layer> mm_a_local_layers;
+    ggml_tensor * mm_a_local_norm_w = nullptr;
+
     // qwen3a
     ggml_tensor * conv2d_1_w = nullptr;
     ggml_tensor * conv2d_1_b = nullptr;

tools/mtmd/clip.cpp

@@ -340,6 +340,11 @@ ggml_tensor * clip_graph::build_vit(
         auto & layer = model.layers[il];
         ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
 
+        ggml_tensor * attn_mask = opts.attn_mask;
+        if (opts.attn_mask_layers.size() > (size_t) il) {
+            attn_mask = opts.attn_mask_layers[il];
+        }
+
         // layernorm1
         cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
         cb(cur, "layer_inp_normed", il);
@@ -452,7 +457,7 @@ ggml_tensor * clip_graph::build_vit(
 
             // build_attn returns a flat 2D [n_embd, n_pos*B]
             cur = build_attn(layer.o_w, layer.o_b,
-                Qcur, Kcur, Vcur, opts.attn_mask, kq_scale, il);
+                Qcur, Kcur, Vcur, attn_mask, kq_scale, il);
             cb(cur, "attn_out", il);
         }
 
@@ -471,6 +476,10 @@ ggml_tensor * clip_graph::build_vit(
 
         inpL = cur; // inpL = residual, cur = hidden_states
 
+        if (opts.callback_layer_out) {
+            opts.callback_layer_out(cur, il);
+        }
+
         cb(cur, "ffn_inp", il);
 
         // layernorm2 (pre-ffn norm)
@@ -519,7 +528,7 @@ ggml_tensor * clip_graph::build_vit(
     }
 
     // post-layernorm
-    if (model.post_ln_w) {
+    if (model.post_ln_w && !opts.skip_post_ln) {
         inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, -1);
     }
 
@@ -1012,6 +1021,10 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
             {
                 builder = std::make_unique<clip_graph_qwen3a>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_MIMO_AUDIO:
+            {
+                builder = std::make_unique<clip_graph_mimo_audio>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_YOUTUVL:
             {
                 builder = std::make_unique<clip_graph_youtuvl>(ctx, img);
@@ -1020,6 +1033,10 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
             {
                 builder = std::make_unique<clip_graph_yasa2>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_PARAKEET:
+            {
+                builder = std::make_unique<clip_graph_parakeet>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_GRANITE4_VISION:
             {
                 builder = std::make_unique<clip_graph_granite4_vision>(ctx, img);
@@ -1343,6 +1360,20 @@ struct clip_model_loader {
                     {
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                     } break;
+                case PROJECTOR_TYPE_PARAKEET:
+                    {
+                        get_u32(KEY_AUDIO_SUBSAMPLING_FACTOR, hparams.subsampling_factor);
+                        GGML_ASSERT(hparams.subsampling_factor == 8 &&
+                            "subsampling_factor must match the conv strides in clip_graph_parakeet::build()");
+                        get_u32(KEY_A_CONV_KERNEL_SIZE,       hparams.audio_conv_kernel_size);
+                        GGML_ASSERT(hparams.audio_conv_kernel_size > 0 && hparams.audio_conv_kernel_size % 2 == 1 &&
+                            "audio_conv_kernel_size must be a positive odd integer");
+                        hparams.audio_chunk_len    = 0;
+                        hparams.audio_sample_rate  = 16000;
+                        hparams.audio_n_fft        = 512;
+                        hparams.audio_window_len   = 400;
+                        hparams.audio_hop_len      = 160;
+                    } break;
                 case PROJECTOR_TYPE_IDEFICS3:
                     {
                         // use default llava-uhd preprocessing params
@@ -1575,6 +1606,45 @@ struct clip_model_loader {
                         hparams.audio_window_len   = 400;
                         hparams.audio_hop_len      = 160;
                     } break;
+                case PROJECTOR_TYPE_MIMO_AUDIO:
+                    {
+                        get_u32(KEY_A_RVQ_NUM_QUANTIZERS, hparams.rvq_num_quantizers, false);
+                        get_arr_int(KEY_A_RVQ_CODEBOOK_SIZE, hparams.rvq_codebook_size, false);
+                        if (hparams.rvq_num_quantizers <= 0) {
+                            throw std::runtime_error(string_format("%s: mimo_audio: missing %s\n", __func__, KEY_A_RVQ_NUM_QUANTIZERS));
+                        }
+                        if ((int) hparams.rvq_codebook_size.size() != hparams.rvq_num_quantizers) {
+                            throw std::runtime_error(string_format(
+                                "%s: mimo_audio: %s length (%zu) must equal %s (%d)\n", __func__,
+                                KEY_A_RVQ_CODEBOOK_SIZE, hparams.rvq_codebook_size.size(),
+                                KEY_A_RVQ_NUM_QUANTIZERS, hparams.rvq_num_quantizers));
+                        }
+                        hparams.ffn_op = FFN_GELU_ERF; // PyTorch F.gelu default (approximate="none")
+                        hparams.rope_theta = 10000.0f;
+
+                        // audio preprocessing params (mel spectrogram)
+                        hparams.audio_sample_rate = 24000;
+                        hparams.audio_n_fft       = 960;
+                        hparams.audio_window_len  = 960;
+                        hparams.audio_hop_len     = 240;
+
+                        get_u32(KEY_A_ATTN_WINDOW_SIZE, hparams.attn_window_size);
+                        std::vector<int> wa_pattern;
+                        get_arr_int(KEY_A_WA_PATTERN_MODE, wa_pattern, true);
+                        if ((int) wa_pattern.size() != hparams.n_layer) {
+                            throw std::runtime_error(string_format(
+                                "%s: mimo_audio: %s length (%zu) must equal n_layer (%d)\n", __func__,
+                                KEY_A_WA_PATTERN_MODE, wa_pattern.size(), hparams.n_layer));
+                        }
+                        hparams.wa_pattern_mode.assign(wa_pattern.begin(), wa_pattern.end());
+
+                        get_u32(KEY_A_LOCAL_BLOCK_COUNT, hparams.audio_local_n_layer);
+                        get_u32(KEY_A_LOCAL_GROUP_SIZE, hparams.audio_local_group_size);
+                        if (hparams.audio_local_group_size <= 0) {
+                            throw std::runtime_error(string_format(
+                                "%s: mimo_audio: %s must be > 0\n", __func__, KEY_A_LOCAL_GROUP_SIZE));
+                        }
+                    } break;
                 case PROJECTOR_TYPE_PADDLEOCR:
                     {
                         hparams.n_merge = 2;
@@ -1841,16 +1911,46 @@ struct clip_model_loader {
             return cur;
         };
 
-        auto get_scalar = [&](const std::string & name, float default_val) {
+        auto get_vector = [&](const std::string & name) {
+            std::vector<float> result;
             auto it = tensor_offset.find(name);
             if (it == tensor_offset.end()) {
+                return result;
+            }
+
+            const int64_t idx = gguf_find_tensor(ctx_gguf.get(), name.c_str());
+            if (idx < 0) {
+                throw std::runtime_error(string_format("%s: failed to find tensor %s\n", __func__, name.c_str()));
+            }
+
+            if (const auto type = gguf_get_tensor_type(ctx_gguf.get(), idx); type != GGML_TYPE_F32) {
+                throw std::runtime_error(string_format("%s: %s must be %s, was %s\n", __func__,
+                            name.c_str(), ggml_type_name(GGML_TYPE_F32), ggml_type_name(type)));
+            }
+
+            const size_t n_bytes = gguf_get_tensor_size(ctx_gguf.get(), idx);
+            if (n_bytes == 0) {
+                throw std::runtime_error(string_format("%s: tensor %s is empty\n", __func__, name.c_str()));
+            }
+
+            const size_t n_elems = n_bytes / sizeof(float);
+            result.resize(n_elems);
+            fin.seekg(it->second, std::ios::beg);
+            fin.read(reinterpret_cast<char*>(result.data()), n_bytes);
+            return result;
+        };
+
+        auto get_scalar = [&](const std::string & name, float default_val) {
+            auto v = get_vector(name);
+            if (v.empty()) {
                 return default_val;
             }
-            size_t offset = it->second;
-            fin.seekg(offset, std::ios::beg);
-            float value;
-            fin.read(reinterpret_cast<char*>(&value), sizeof(float));
-            return value;
+            if (v.size() != 1) {
+                throw std::runtime_error(string_format("%s: expected scalar tensor '%s' but got %d elements\n",
+                            __func__, name.c_str(), (int) v.size()));
+            }
+
+            return v[0];
         };
 
         model.class_embedding = get_tensor(TN_CLASS_EMBD, false);
@@ -2444,6 +2544,54 @@ struct clip_model_loader {
                     model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
                     model.mm_2_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "bias"));
                 } break;
+            case PROJECTOR_TYPE_MIMO_AUDIO:
+                {
+                    model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
+                    model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 1, "bias"));
+                    model.conv1d_2_w = get_tensor(string_format(TN_CONV1D, 2, "weight"));
+                    model.conv1d_2_b = get_tensor(string_format(TN_CONV1D, 2, "bias"));
+                    model.downsample_conv_w = get_tensor(string_format(TN_A_DOWNSAMPLE_CONV, "weight"));
+                    model.downsample_norm_w = get_tensor(string_format(TN_A_DOWNSAMPLE_NORM, "weight"));
+                    model.downsample_norm_b = get_tensor(string_format(TN_A_DOWNSAMPLE_NORM, "bias"));
+                    model.rvq_codebook   = get_tensor(string_format(TN_A_RVQ_CODEBOOK, "weight"), false);
+                    model.mm_a_code_embd = get_tensor(string_format(TN_MM_A_CODE_EMBD, "weight"), false);
+                    if (!model.rvq_codebook || !model.mm_a_code_embd) {
+                        throw std::runtime_error(string_format("%s: mimo_audio: missing %s or %s\n", __func__,
+                            TN_A_RVQ_CODEBOOK, TN_MM_A_CODE_EMBD));
+                    }
+                    // hparams.rvq_codebook_size comes from GGUF metadata and is independent of the
+                    // tensors' actual shapes - bound it so codebook/code_embd views built from it
+                    // (mimo-audio.cpp) can never read past either tensor's allocated bins/vocab.
+                    for (int32_t bins : hparams.rvq_codebook_size) {
+                        if (bins <= 0 || bins > model.rvq_codebook->ne[1] || bins > model.mm_a_code_embd->ne[1]) {
+                            throw std::runtime_error(string_format(
+                                "%s: mimo_audio: %s entry (%d) out of range for codebook/code_embd tensors\n",
+                                __func__, KEY_A_RVQ_CODEBOOK_SIZE, bins));
+                        }
+                    }
+
+                    // LLM-side connector: input_local_transformer + projection
+                    model.mm_a_local_layers.resize(hparams.audio_local_n_layer);
+                    for (int il = 0; il < hparams.audio_local_n_layer; il++) {
+                        auto & layer = model.mm_a_local_layers[il];
+                        layer.q_w    = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_Q,   il, "weight"));
+                        layer.q_b    = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_Q,   il, "bias"));
+                        layer.k_w    = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_K,   il, "weight"));
+                        layer.k_b    = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_K,   il, "bias"));
+                        layer.v_w    = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_V,   il, "weight"));
+                        layer.v_b    = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_V,   il, "bias"));
+                        layer.o_w    = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_OUT, il, "weight"));
+                        layer.ff_gate_w = get_tensor(string_format(TN_MM_A_LOCAL_FFN_GATE, il, "weight"));
+                        layer.ff_up_w   = get_tensor(string_format(TN_MM_A_LOCAL_FFN_UP,   il, "weight"));
+                        layer.ff_down_w = get_tensor(string_format(TN_MM_A_LOCAL_FFN_DOWN, il, "weight"));
+                        layer.ln_1_w = get_tensor(string_format(TN_MM_A_LOCAL_LN1, il, "weight"));
+                        layer.ln_2_w = get_tensor(string_format(TN_MM_A_LOCAL_LN2, il, "weight"));
+                    }
+                    model.mm_a_local_norm_w = get_tensor(string_format(TN_MM_A_LOCAL_NORM, "weight"));
+
+                    model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
+                    model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
+                } break;
             case PROJECTOR_TYPE_VOXTRAL:
                 {
                     model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
@@ -2700,6 +2848,68 @@ struct clip_model_loader {
                         layer.conv_pw2_b   = get_tensor(string_format(TN_CONV_PW2,  prefix, il, "bias"));
                     }
                 } break;
+            case PROJECTOR_TYPE_PARAKEET:
+                {
+
+                    hparams.mel_filters = get_vector(TN_MEL_FILTERS);
+                    hparams.window      = get_vector(TN_WINDOW);
+
+                    // Subsampling layers (conv1d)
+                    for (int i : {0, 2, 3, 5, 6}) {
+                        model.pre_encode_conv_X_w[i] = get_tensor(string_format(TN_CONV1D, i, "weight"));
+                        model.pre_encode_conv_X_b[i] = get_tensor(string_format(TN_CONV1D, i, "bias"));
+                    }
+                    model.pre_encode_out_w = get_tensor(string_format(TN_PRE_ENCODE_OUT, "weight"));
+                    model.pre_encode_out_b = get_tensor(string_format(TN_PRE_ENCODE_OUT, "bias"));
+
+                    // Projection layers
+                    model.mm_norm_pre_w = get_tensor(string_format(TN_MM_NORM_PRE, "weight"), false);
+                    model.mm_0_w        = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"), false);
+                    model.mm_1_w        = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"), false);
+
+                    // Encoder layers
+                    for (int il = 0; il < hparams.n_layer; ++il) {
+                        auto & layer = model.layers[il];
+
+                        // Attention (from shared above)
+
+                        // Relative position encoding
+                        layer.linear_pos_w = get_tensor(string_format(TN_LINEAR_POS, prefix, il, "weight"));
+                        layer.pos_bias_u   = get_tensor(string_format(TN_POS_BIAS_U, prefix, il));
+                        layer.pos_bias_v   = get_tensor(string_format(TN_POS_BIAS_V, prefix, il));
+
+                        // Convolution module
+                        layer.conv_pw1_w     = get_tensor(string_format(TN_CONV_PW1,       prefix, il, "weight"));
+                        layer.conv_pw1_b     = get_tensor(string_format(TN_CONV_PW1,       prefix, il, "bias"), false);
+                        layer.conv_dw_w      = get_tensor(string_format(TN_CONV_DW,        prefix, il, "weight"));
+                        layer.conv_dw_b      = get_tensor(string_format(TN_CONV_DW,        prefix, il, "bias"), false);
+                        layer.conv_norm_w    = get_tensor(string_format(TN_CONV_NORM,      prefix, il, "weight"));
+                        layer.conv_norm_b    = get_tensor(string_format(TN_CONV_NORM,      prefix, il, "bias"));
+                        layer.conv_norm_mean = get_tensor(string_format(TN_CONV_NORM_MEAN, prefix, il));
+                        layer.conv_norm_var  = get_tensor(string_format(TN_CONV_NORM_VAR,  prefix, il));
+                        layer.conv_pw2_w     = get_tensor(string_format(TN_CONV_PW2,       prefix, il, "weight"));
+                        layer.conv_pw2_b     = get_tensor(string_format(TN_CONV_PW2,       prefix, il, "bias"), false);
+
+                        // Feed-forward networks
+                        layer.ff_norm_w   = get_tensor(string_format(TN_FFN_NORM,   prefix, il, "weight"));
+                        layer.ff_norm_b   = get_tensor(string_format(TN_FFN_NORM,   prefix, il, "bias"));
+
+                        layer.ff_norm_1_w = get_tensor(string_format(TN_FFN_NORM_1, prefix, il, "weight"));
+                        layer.ff_norm_1_b = get_tensor(string_format(TN_FFN_NORM_1, prefix, il, "bias"));
+                        layer.ff_up_1_w   = get_tensor(string_format(TN_FFN_UP_1,   prefix, il, "weight"));
+                        layer.ff_up_1_b   = get_tensor(string_format(TN_FFN_UP_1,   prefix, il, "bias"), false);
+                        layer.ff_down_1_w = get_tensor(string_format(TN_FFN_DOWN_1, prefix, il, "weight"));
+                        layer.ff_down_1_b = get_tensor(string_format(TN_FFN_DOWN_1, prefix, il, "bias"), false);
+
+                        // Layer norms
+                        layer.norm_conv_w = get_tensor(string_format(TN_NORM_CONV, prefix, il, "weight"));
+                        layer.norm_conv_b = get_tensor(string_format(TN_NORM_CONV, prefix, il, "bias"));
+                    }
+
+                    model.mm_model_mlp_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
+                    model.mm_model_mlp_2_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
+                    model.mm_model_mlp_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
+                } break;
             case PROJECTOR_TYPE_GRANITE_SPEECH:
                 {
                     model.inp_proj_w     = get_tensor(string_format(TN_INP_PROJ,    "weight"));
@@ -3545,10 +3755,23 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
                 }
                 n_patches = n;
             } break;
+        case PROJECTOR_TYPE_PARAKEET:
+            {
+                n_patches = (img->nx() + (params.subsampling_factor - 1)) / params.subsampling_factor;
+            } break;
         case PROJECTOR_TYPE_GEMMA4UA:
             {
                 n_patches = img->nx();  // no downsampling: one token per raw waveform frame
             } break;
+        case PROJECTOR_TYPE_MIMO_AUDIO:
+            {
+                // conv1(s=1) + conv2(s=2) -> RVQ-encoder downsample conv(k=2,s=2)
+                int n = img->nx();
+                n = (n - 1) / 2 + 1;         // conv1 + conv2
+                n = (n - 2) / 2 + 1;         // downsample conv
+                const int group_size = params.audio_local_group_size;
+                n_patches = (n + group_size - 1) / group_size;
+            } break;
         case PROJECTOR_TYPE_GRANITE_SPEECH:
             {
                 const int ws = ctx->model.hparams.audio_proj_window_size;
@@ -4376,6 +4599,58 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
                     set_input_f32("pos_emb", pos_emb);
                 }
             } break;
+        case PROJECTOR_TYPE_MIMO_AUDIO:
+            {
+                GGML_ASSERT(imgs.entries.size() == 1);
+                const int n_frames = imgs.entries.front().nx();
+                const int n_pos    = (n_frames - 1) / 2 + 1; // matches conv1(s=1)+conv2(s=2) output length
+
+                std::vector<int32_t> positions(n_pos);
+                for (int i = 0; i < n_pos; i++) {
+                    positions[i] = i;
+                }
+                set_input_i32("mimo_audio_positions", positions);
+
+                const int window = hparams.attn_window_size;
+                GGML_ASSERT(window > 0);
+
+                const float neg_inf = std::numeric_limits<float>::lowest();
+                std::vector<float> full_mask((size_t) n_pos * n_pos);
+                std::vector<float> window_mask((size_t) n_pos * n_pos);
+                for (int q = 0; q < n_pos; q++) {
+                    for (int k = 0; k < n_pos; k++) {
+                        const bool causal_ok = k <= q;
+                        full_mask[(size_t) q * n_pos + k]   = causal_ok ? 0.0f : neg_inf;
+                        window_mask[(size_t) q * n_pos + k] = (causal_ok && (q - k) <= window) ? 0.0f : neg_inf;
+                    }
+                }
+                set_input_f32("mimo_audio_full_mask", full_mask);
+                set_input_f32("mimo_audio_window_mask", window_mask);
+
+                // input_local_transformer: block-diagonal mask + in-group positions
+                {
+                    const int n_pos_ds   = (n_pos - 2) / 2 + 1; // matches downsample conv (k=2,s=2,p=0)
+                    const int group_size = hparams.audio_local_group_size;
+                    GGML_ASSERT(group_size > 0);
+                    const int n_groups = (n_pos_ds + group_size - 1) / group_size;
+                    const int n_padded = n_groups * group_size;
+
+                    std::vector<int32_t> local_positions(n_padded);
+                    for (int i = 0; i < n_padded; i++) {
+                        local_positions[i] = i % group_size;
+                    }
+                    set_input_i32("mimo_audio_local_positions", local_positions);
+
+                    std::vector<float> local_mask((size_t) n_padded * n_padded);
+                    for (int q = 0; q < n_padded; q++) {
+                        for (int k = 0; k < n_padded; k++) {
+                            const bool same_group = (q / group_size) == (k / group_size);
+                            local_mask[(size_t) q * n_padded + k] = same_group ? 0.0f : neg_inf;
+                        }
+                    }
+                    set_input_f32("mimo_audio_local_mask", local_mask);
+                }
+            } break;
         case PROJECTOR_TYPE_LFM2A:
             {
                 GGML_ASSERT(imgs.entries.size() == 1);
@@ -4397,6 +4672,88 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
                 }
                 set_input_f32("pos_emb", pos_emb);
             } break;
+        case PROJECTOR_TYPE_PARAKEET:
+            {
+                GGML_ASSERT(imgs.entries.size() == 1);
+                struct ggml_tensor * attn_mask = ggml_graph_get_tensor(gf, "attn_mask");
+                const int n_q = attn_mask->ne[1];
+                const int n_k = attn_mask->ne[0];
+                const int n_frames = imgs.entries.front().nx();
+                const int n_tokens_real = (n_frames + hparams.subsampling_factor-1) / hparams.subsampling_factor;
+                const float mask_value  = -1e30f;
+
+                std::vector<float> mask_data(n_q * n_k);
+                if (n_k == n_q) {
+                    // full attention: mask keys that are padding
+                    for (int q = 0; q < n_q; ++q) {
+                        for (int k = 0; k < n_k; ++k) {
+                            mask_data[q * n_k + k] = (k >= n_tokens_real) ? mask_value : 0.0f;
+                        }
+                    }
+                } else {
+                    // local attention: mask keys outside the valid window
+                    const int att_left = n_k / 2;
+                    for (int q = 0; q < n_q; ++q) {
+                        for (int k = 0; k < n_k; ++k) {
+                            const int key = q - att_left + k;
+                            mask_data[q * n_k + k] = (key >= 0 && key < n_tokens_real) ? 0.0f : mask_value;
+                        }
+                    }
+                }
+                set_input_f32(attn_mask->name, mask_data);
+
+                // local attention skew mask: zeroes out the probs that were
+                // computed for keys outside the valid sliding window.
+                if (struct ggml_tensor * local_mask = ggml_graph_get_tensor(gf, "local_mask")) {
+                    const int lm_k = local_mask->ne[0];
+                    const int lm_q = local_mask->ne[1];
+                    const int window_size = lm_k - lm_q + 1;
+                    std::vector<float> lm_data(lm_q * lm_k);
+                    for (int q = 0; q < lm_q; ++q) {
+                        for (int k = 0; k < lm_k; ++k) {
+                            const int rel = k - q;
+                            lm_data[q * lm_k + k] = (rel >= 0 && rel < window_size) ? 1.0f : 0.0f;
+                        }
+                    }
+                    set_input_f32(local_mask->name, lm_data);
+                }
+
+                // Generate rotation frequencies for relative positional encoding.
+                {
+                    const int n_state     = hparams.n_embd;
+                    const int d_half      = n_state / 2;
+                    const float log_10000 = logf(10000.0f);
+                    std::vector<float> freqs(d_half);
+                    for (int k = 0; k < d_half; ++k) {
+                        freqs[k] = expf(-(float(k * 2) * log_10000 / float(n_state)));
+                    }
+                    set_input_f32("pos_freqs", freqs);
+                }
+
+                // Generate relative positional distance values which scaled by
+                // the frequency to produce the angles for sin/cos.
+                {
+                    // window_size is only known after graph construction since it depends on
+                    // n_time from the conv output, so we read it back from the graph tensor.
+                    struct ggml_tensor * rel_pos = ggml_graph_get_tensor(gf, "rel_positions");
+                    const int window_size = rel_pos->ne[1];
+                    std::vector<float> pos(window_size);
+                    // local attention: window is fixed at [att_left, att_right]
+                    // full attention: window covers the full sequence, centered
+                    if (ggml_graph_get_tensor(gf, "local_mask")) {
+                        const int att_left = window_size / 2;
+                        for (int t = 0; t < window_size; ++t) {
+                            pos[t] = float(att_left - t);
+                        }
+                    } else {
+                        const int n_time = (window_size + 1) / 2;
+                        for (int t = 0; t < window_size; ++t) {
+                            pos[t] = float(n_time - 1 - t);
+                        }
+                    }
+                    set_input_f32(rel_pos->name, pos);
+                }
+            } break;
         case PROJECTOR_TYPE_GRANITE_SPEECH:
             {
                 const int context_size = ctx->model.hparams.audio_chunk_size;
@@ -4678,6 +5035,10 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.qf_proj_blocks.size() * ctx->model.hparams.projection_dim;
         case PROJECTOR_TYPE_GLM4V:
             return ctx->model.mm_ffn_down_w->ne[1];
+        case PROJECTOR_TYPE_MIMO_AUDIO:
+            return ctx->model.mm_2_w->ne[1];
+        case PROJECTOR_TYPE_PARAKEET:
+            return ctx->model.mm_1_w->ne[1];
         default:
             GGML_ABORT("Unknown projector type");
     }

tools/mtmd/models/mimo-audio.cpp

@@ -0,0 +1,218 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_mimo_audio::build() {
+    ggml_tensor * inp = build_inp_raw(1); // [n_frames, n_mel, 1]
+
+    ggml_tensor * cur = ggml_conv_1d_ph(ctx0, model.conv1d_1_w, inp, 1, 1);
+    cur = ggml_add(ctx0, cur, model.conv1d_1_b);
+    cur = ggml_gelu_erf(ctx0, cur);
+
+    cur = ggml_conv_1d_ph(ctx0, model.conv1d_2_w, cur, 2, 1);
+    cur = ggml_add(ctx0, cur, model.conv1d_2_b);
+    cur = ggml_gelu_erf(ctx0, cur);
+
+    ggml_tensor * inpL = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); // [n_embd, n_pos]
+    const int64_t n_pos = inpL->ne[1];
+    cb(inpL, "after_conv1d", -1);
+
+    GGML_ASSERT((int) hparams.wa_pattern_mode.size() == n_layer);
+
+    ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(inp_pos, "mimo_audio_positions");
+    ggml_set_input(inp_pos);
+
+    ggml_tensor * full_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
+    ggml_set_name(full_mask, "mimo_audio_full_mask");
+    ggml_set_input(full_mask);
+
+    ggml_tensor * window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
+    ggml_set_name(window_mask, "mimo_audio_window_mask");
+    ggml_set_input(window_mask);
+
+    build_vit_opts opts;
+    opts.attn_mask_layers.resize(n_layer);
+    for (int il = 0; il < n_layer; il++) {
+        opts.attn_mask_layers[il] = hparams.wa_pattern_mode[il] == -1 ? full_mask : window_mask;
+    }
+    // the skip connection below must be added before the post-transformer  norm,
+    // so build_vit must not apply that norm itself
+    opts.skip_post_ln = true;
+
+    // encoder_skip_layer_id=3 (1-indexed) -> capture output of layer index 2
+    const int skip_capture_il = 2;
+    GGML_ASSERT(n_layer > skip_capture_il);
+    ggml_tensor * skip_hidden = nullptr;
+    opts.callback_layer_out = [&](ggml_tensor * layer_cur, int il) {
+        if (il == skip_capture_il) {
+            skip_hidden = layer_cur;
+        }
+    };
+
+    auto add_pos = [&](ggml_tensor * x, const clip_layer &) {
+        return ggml_rope_ext(ctx0, x, inp_pos, nullptr, d_head,
+                             GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+    };
+
+    inpL = build_vit(inpL, n_pos, NORM_TYPE_NORMAL, hparams.ffn_op, nullptr, add_pos, opts);
+    inpL = ggml_reshape_2d(ctx0, inpL, n_embd, n_pos); // build_vit restores a (size-1) batch dim
+
+    GGML_ASSERT(skip_hidden != nullptr);
+    inpL = ggml_add(ctx0, inpL, skip_hidden);
+
+    inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, NORM_TYPE_NORMAL, eps, -1);
+    cb(inpL, "after_transformer", -1);
+
+    // downsample: strided conv (no bias) + gelu + layernorm
+    {
+        ggml_tensor * ds = ggml_cont(ctx0, ggml_transpose(ctx0, inpL)); // [n_pos, n_embd]
+        ds = ggml_conv_1d(ctx0, model.downsample_conv_w, ds, 2, 0, 1);
+        ds = ggml_gelu_erf(ctx0, ds);
+        ds = ggml_cont(ctx0, ggml_transpose(ctx0, ds)); // [n_embd, n_pos/2]
+        ds = build_norm(ds, model.downsample_norm_w, model.downsample_norm_b, NORM_TYPE_NORMAL, eps, -1);
+        inpL = ds;
+    }
+    cb(inpL, "after_downsample", -1);
+
+    // RVQ quantize: codebook ne=[dim, max_bins, n_q]
+    // quantize input vector to codes (type=I32)
+    std::vector<ggml_tensor *> codes;
+    {
+        GGML_ASSERT(model.rvq_codebook != nullptr);
+        const int64_t dim = model.rvq_codebook->ne[0];
+        GGML_ASSERT(dim == inpL->ne[0]);
+        GGML_ASSERT((int64_t) hparams.rvq_codebook_size.size() == model.rvq_codebook->ne[2]);
+
+        ggml_tensor * residual = inpL; // [dim, n_pos_ds]
+
+        for (size_t q = 0; q < hparams.rvq_codebook_size.size(); q++) {
+            const int64_t bins = hparams.rvq_codebook_size[q];
+            ggml_tensor * codebook_q = ggml_view_2d(ctx0, model.rvq_codebook, dim, bins,
+                model.rvq_codebook->nb[1], q * model.rvq_codebook->nb[2]);
+            codebook_q = ggml_cont(ctx0, codebook_q);
+
+            ggml_tensor * codebook_norm = ggml_sum_rows(ctx0, ggml_sqr(ctx0, codebook_q)); // [1, bins]
+            codebook_norm = ggml_cont(ctx0, ggml_transpose(ctx0, codebook_norm));          // [bins, 1]
+
+            ggml_tensor * dot    = ggml_mul_mat(ctx0, codebook_q, residual); // [bins, n_pos_ds]
+            ggml_tensor * scores = ggml_sub(ctx0, ggml_scale(ctx0, dot, 2.0f), codebook_norm);
+
+            ggml_tensor * idx = ggml_argmax(ctx0, scores); // [n_pos_ds]
+            codes.push_back(idx);
+
+            ggml_tensor * quant = ggml_get_rows(ctx0, codebook_q, idx); // [dim, n_pos_ds]
+            residual = ggml_sub(ctx0, residual, quant);
+            cb(idx, "rvq_code", (int) q);
+        }
+    }
+
+    // convert codes to LLM embeddings
+    ggml_tensor * code_embd_sum = nullptr;
+    {
+        GGML_ASSERT(model.mm_a_code_embd != nullptr);
+        const int64_t dim   = model.mm_a_code_embd->ne[0];
+        const int64_t vocab = model.mm_a_code_embd->ne[1];
+        GGML_ASSERT((int64_t) codes.size() == model.mm_a_code_embd->ne[2]);
+        GGML_ASSERT(dim == inpL->ne[0]);
+
+        for (size_t i = 0; i < codes.size(); i++) {
+            ggml_tensor * table_i = ggml_view_2d(ctx0, model.mm_a_code_embd, dim, vocab,
+                model.mm_a_code_embd->nb[1], i * model.mm_a_code_embd->nb[2]);
+            table_i = ggml_cont(ctx0, table_i);
+
+            ggml_tensor * embd_i = ggml_get_rows(ctx0, table_i, codes[i]); // [dim, n_pos_ds]
+            code_embd_sum = code_embd_sum ? ggml_add(ctx0, code_embd_sum, embd_i) : embd_i;
+        }
+        cb(code_embd_sum, "code_embd_sum", -1);
+    }
+
+    // input_local_transformer
+    // groups of `group_size` consecutive downsampled frames are processed together, attending only within their own group.
+    // Implemented as a block-diagonal mask + in-group-repeating positions
+    // (rather than a real batch dim) - same technique as the encoder's masks above, and as gemma4a's / deepseekocr2's chunked attention.
+
+    // note: hand-rolled here instead of build_vit() because this is a second, independent layer stack
+    // (own layer array/count, RMSNorm instead of LN, SiLU FFN, own RoPE theta)
+
+    ggml_tensor * projected;
+    {
+        const int group_size = hparams.audio_local_group_size;
+        GGML_ASSERT(group_size > 0);
+        const int64_t n_pos_ds = code_embd_sum->ne[1];
+        const int64_t n_groups = (n_pos_ds + group_size - 1) / group_size;
+        const int64_t n_padded = n_groups * group_size;
+
+        ggml_tensor * cur_local = code_embd_sum;
+        if (n_padded != n_pos_ds) {
+            cur_local = ggml_pad(ctx0, cur_local, 0, (int) (n_padded - n_pos_ds), 0, 0);
+        }
+
+        ggml_tensor * local_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_padded);
+        ggml_set_name(local_pos, "mimo_audio_local_positions");
+        ggml_set_input(local_pos);
+
+        ggml_tensor * local_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_padded, n_padded);
+        ggml_set_name(local_mask, "mimo_audio_local_mask");
+        ggml_set_input(local_mask);
+
+        const float local_rope_theta = 640000.0f; // audio_config.rope_theta (differs from the encoder's)
+        auto apply_local_rope = [&](ggml_tensor * x) {
+            return ggml_rope_ext(ctx0, x, local_pos, nullptr, d_head,
+                                 GGML_ROPE_TYPE_NEOX, 0, local_rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+        };
+
+        for (int il = 0; il < hparams.audio_local_n_layer; il++) {
+            auto & layer = model.mm_a_local_layers[il];
+
+            ggml_tensor * attn_in = build_norm(cur_local, layer.ln_1_w, nullptr, NORM_TYPE_RMS, eps, il);
+
+            ggml_tensor * Qcur = build_mm(layer.q_w, attn_in);
+            if (layer.q_b) {
+                Qcur = ggml_add(ctx0, Qcur, layer.q_b);
+            }
+            ggml_tensor * Kcur = build_mm(layer.k_w, attn_in);
+            if (layer.k_b) {
+                Kcur = ggml_add(ctx0, Kcur, layer.k_b);
+            }
+            ggml_tensor * Vcur = build_mm(layer.v_w, attn_in);
+            if (layer.v_b) {
+                Vcur = ggml_add(ctx0, Vcur, layer.v_b);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_padded);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_padded);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_padded);
+
+            Qcur = apply_local_rope(Qcur);
+            Kcur = apply_local_rope(Kcur);
+
+            ggml_tensor * attn_out = build_attn(layer.o_w, nullptr, Qcur, Kcur, Vcur, local_mask, kq_scale, il);
+            cur_local = ggml_add(ctx0, cur_local, attn_out);
+
+            ggml_tensor * ffn_in = build_norm(cur_local, layer.ln_2_w, nullptr, NORM_TYPE_RMS, eps, il);
+            ggml_tensor * ffn_out = build_ffn(ffn_in,
+                layer.ff_up_w, nullptr,
+                layer.ff_gate_w, nullptr,
+                layer.ff_down_w, nullptr,
+                FFN_SILU, il);
+            cur_local = ggml_add(ctx0, cur_local, ffn_out);
+        }
+
+        cur_local = build_norm(cur_local, model.mm_a_local_norm_w, nullptr, NORM_TYPE_RMS, eps, -1);
+        cb(cur_local, "after_local_transformer", -1);
+
+        // flatten each group of `group_size` frames into one (group_size*n_embd)-dim vector
+        // (matching AudioProjection's flattened input)
+        ggml_tensor * grouped = ggml_reshape_2d(ctx0, cur_local, n_embd * group_size, n_groups);
+
+        // AudioProjection: Linear (no bias) -> GELU -> Linear (no bias)
+        projected = build_ffn(grouped,
+            model.mm_1_w, nullptr,
+            nullptr, nullptr,
+            model.mm_2_w, nullptr,
+            FFN_GELU_ERF, -1);
+        cb(projected, "after_projection", -1);
+    }
+
+    ggml_build_forward_expand(gf, projected);
+    return gf;
+}

tools/mtmd/models/models.h

@@ -210,6 +210,11 @@ struct clip_graph_qwen3a : clip_graph {
     ggml_cgraph * build() override;
 };
 
+struct clip_graph_mimo_audio : clip_graph {
+    clip_graph_mimo_audio(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_kimik25 : clip_graph {
     clip_graph_kimik25(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
@@ -217,6 +222,11 @@ struct clip_graph_kimik25 : clip_graph {
     ggml_tensor * resize_position_embeddings_3d(uint32_t interpolation_mode);
 };
 
+struct clip_graph_parakeet : clip_graph {
+    clip_graph_parakeet(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_exaone4_5 : clip_graph {
     clip_graph_exaone4_5(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;