kleidiai : support for concurrent sme and neon kernel execution (#20070)

* ggml-cpu: add rvv ggml_quantize_mat_4x8 for q8_0

Co-authored-by: Rehan Qasim <rehan.qasim@10xengineers.ai>

* ggml-cpu: add rvv repacking for iq4_nl

* ggml-cpu: add generic impl for iq4_nl gemm/gemv

* ggml-cpu: add rvv repacking for q8_0

* ggml-cpu: refactor; add rvv repacking for q4_0, q4_K

* ggml-cpu: refactor; add rvv repacking for q2_K

Co-authored-by: Rehan Qasim <rehan.qasim@10xengineers.ai>

* ggml-cpu: refactor rvv repack

---------

Co-authored-by: Rehan Qasim <rehan.qasim@10xengineers.ai>

common/chat-auto-parser-generator.cpp

@@ -90,7 +90,7 @@ common_peg_arena autoparser::build_parser(const templates_params & inputs) const
         // pre-register a json-string rule that accepts both quote styles. This must happen
         // before any call to p.json() so that all JSON parsing inherits the flexible rule.
         if (tools.format.uses_python_dicts) {
-            p.rule("json-string", [&]() { return p.choice({ p.double_quoted_string(), p.single_quoted_string() }); });
+            p.rule("json-string", p.quoted_string());
         }
 
         parser_build_context ctx(p, inputs);

common/chat-peg-parser.cpp

@@ -507,8 +507,8 @@ common_peg_parser common_chat_peg_builder::python_style_tool_calls(
 
                 common_peg_parser arg_value_parser = eps();
                 auto string_value_parser = choice({
-                    literal("\"") + tool_arg_string_value(json_string_content()) + literal("\""),
-                    literal("'") + tool_arg_string_value(json_string_content()) + literal("'")
+                    literal("\"") + tool_arg_string_value(string_content('"')) + literal("\""),
+                    literal("'") + tool_arg_string_value(string_content('\'')) + literal("'")
                 });
 
                 if (is_string_type) {
@@ -577,7 +577,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_function_is_key(
         if (!call_id_key.empty()) {
             auto id_parser = atomic(
                 literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
-                literal("\"") + tool_id(json_string_content()) + literal("\"")
+                literal("\"") + tool_id(string_content('"')) + literal("\"")
             );
             inner_fields.push_back(optional(id_parser + space() + optional(literal(",") + space())));
         }
@@ -586,7 +586,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_function_is_key(
             auto gen_id_parser = atomic(
                 literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
                 choice({
-                    literal("\"") + tool_id(json_string_content()) + literal("\""),
+                    literal("\"") + tool_id(string_content('"')) + literal("\""),
                     tool_id(json_number())
                 })
             );
@@ -675,7 +675,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
             if (id_spec.first.empty()) {
                 auto id_parser = atomic(
                     literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
-                    literal("\"") + tool_id(json_string_content()) + literal("\"")
+                    literal("\"") + tool_id(string_content('"')) + literal("\"")
                 );
                 tool_parser_body = tool_parser_body + optional(id_parser + space() + literal(",") + space());
             }
@@ -687,7 +687,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
                 auto gen_id_parser = atomic(
                     literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
                     choice({
-                        literal("\"") + tool_id(json_string_content()) + literal("\""),
+                        literal("\"") + tool_id(string_content('"')) + literal("\""),
                         tool_id(json_number())
                     })
                 );
@@ -736,7 +736,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
             id_parser = atomic(
                 literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
                 choice({
-                    literal("\"") + tool_id(json_string_content()) + literal("\""),
+                    literal("\"") + tool_id(string_content('"')) + literal("\""),
                     tool_id(json_number())
                 })
             );
@@ -747,7 +747,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
             gen_id_parser = atomic(
                 literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
                 choice({
-                    literal("\"") + tool_id(json_string_content()) + literal("\""),
+                    literal("\"") + tool_id(string_content('"')) + literal("\""),
                     tool_id(json_number())
                 })
             );

common/http.h

@@ -7,6 +7,7 @@ struct common_http_url {
     std::string user;
     std::string password;
     std::string host;
+    int port;
     std::string path;
 };
 
@@ -47,6 +48,20 @@ static common_http_url common_http_parse_url(const std::string & url) {
         parts.host = rest;
         parts.path = "/";
     }
+
+    auto colon_pos = parts.host.find(':');
+
+    if (colon_pos != std::string::npos) {
+        parts.port = std::stoi(parts.host.substr(colon_pos + 1));
+        parts.host = parts.host.substr(0, colon_pos);
+    } else if (parts.scheme == "http") {
+        parts.port = 80;
+    } else if (parts.scheme == "https") {
+        parts.port = 443;
+    } else {
+        throw std::runtime_error("unsupported URL scheme: " + parts.scheme);
+    }
+
     return parts;
 }
 
@@ -68,7 +83,7 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
     }
 #endif
 
-    httplib::Client cli(parts.scheme + "://" + parts.host);
+    httplib::Client cli(parts.scheme + "://" + parts.host + ":" + std::to_string(parts.port));
 
     if (!parts.user.empty()) {
         cli.set_basic_auth(parts.user, parts.password);

common/peg-parser.cpp

@@ -658,7 +658,7 @@ struct parser_executor {
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
     }
 
-    static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos) {
+    static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos, const char delimiter) {
         ++pos; // consume '\'
         if (pos >= ctx.input.size()) {
             if (!ctx.is_lenient()) {
@@ -667,23 +667,14 @@ struct parser_executor {
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
         }
 
-        switch (ctx.input[pos]) {
-            case '"':
-            case '\'':
-            case '\\':
-            case '/':
-            case 'b':
-            case 'f':
-            case 'n':
-            case 'r':
-            case 't':
-                ++pos;
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
-            case 'u':
-                return handle_unicode_escape(ctx, start, pos);
-            default:
-                // Invalid escape sequence
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
+        char c = ctx.input[pos];
+        if (c == delimiter || c == '\\' || c == '/' || c == 'b' || c == 'f' || c == 'n' || c == 'r' || c == 't') {
+            ++pos;
+            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
+        } else if (c == 'u') {
+            return handle_unicode_escape(ctx, start, pos);
+        } else {
+            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
         }
     }
 
@@ -704,62 +695,20 @@ struct parser_executor {
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
     }
 
-    common_peg_parse_result operator()(const common_peg_json_string_parser & /* p */) {
+    common_peg_parse_result operator()(const common_peg_string_parser & p) {
         auto pos = start_pos;
 
         // Parse string content (without quotes)
         while (pos < ctx.input.size()) {
             char c = ctx.input[pos];
 
-            if (c == '"') {
-                // Found closing quote - success (don't consume it)
+            if (c == p.delimiter) {
+                // Found closing delimiter - success (don't consume it)
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
             }
 
             if (c == '\\') {
-                auto result = handle_escape_sequence(ctx, start_pos, pos);
-                if (!result.success()) {
-                    return result;
-                }
-            } else {
-                auto utf8_result = common_parse_utf8_codepoint(ctx.input, pos);
-
-                if (utf8_result.status == utf8_parse_result::INCOMPLETE) {
-                    if (!ctx.is_lenient()) {
-                        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
-                    }
-                    return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos);
-                }
-
-                if (utf8_result.status == utf8_parse_result::INVALID) {
-                    return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
-                }
-
-                pos += utf8_result.bytes_consumed;
-            }
-        }
-
-        // Reached end without finding closing quote
-        if (!ctx.is_lenient()) {
-            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos, pos);
-        }
-        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos);
-    }
-
-    common_peg_parse_result operator()(const common_peg_python_dict_string_parser & /* p */) {
-        auto pos = start_pos;
-
-        // Parse string content (without quotes)
-        while (pos < ctx.input.size()) {
-            char c = ctx.input[pos];
-
-            if (c == '\'') {
-                // Found closing quote - success (don't consume it)
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
-            }
-
-            if (c == '\\') {
-                auto result = handle_escape_sequence(ctx, start_pos, pos);
+                auto result = handle_escape_sequence(ctx, start_pos, pos, p.delimiter);
                 if (!result.success()) {
                     return result;
                 }
@@ -988,8 +937,7 @@ void common_peg_arena::resolve_refs() {
                                  std::is_same_v<T, common_peg_ref_parser> ||
                                  std::is_same_v<T, common_peg_until_parser> ||
                                  std::is_same_v<T, common_peg_literal_parser> ||
-                                 std::is_same_v<T, common_peg_json_string_parser> ||
-                                 std::is_same_v<T, common_peg_python_dict_string_parser> ||
+                                 std::is_same_v<T, common_peg_string_parser> ||
                                  std::is_same_v<T, common_peg_chars_parser> ||
                                  std::is_same_v<T, common_peg_any_parser> ||
                                  std::is_same_v<T, common_peg_space_parser>) {
@@ -1065,10 +1013,8 @@ std::string common_peg_arena::dump_impl(common_peg_parser_id
                 return "CharRepeat(" + p.pattern + ", " + std::to_string(p.min_count) + ", unbounded)";
             }
             return "CharRepeat(" + p.pattern + ", " + std::to_string(p.min_count) + ", " + std::to_string(p.max_count) + ")";
-        } else if constexpr (std::is_same_v<T, common_peg_json_string_parser>) {
-            return "JsonString()";
-        } else if constexpr (std::is_same_v<T, common_peg_python_dict_string_parser>) {
-            return "PythonDictString()";
+        } else if constexpr (std::is_same_v<T, common_peg_string_parser>) {
+            return "String(" + std::string(1, p.delimiter) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_until_parser>) {
             return "Until(" + string_join(p.delimiters, " | ") + ")";
         } else if constexpr (std::is_same_v<T, common_peg_schema_parser>) {
@@ -1281,47 +1227,25 @@ common_peg_arena common_peg_parser_builder::build() {
 
 // String primitives
 
-common_peg_parser common_peg_parser_builder::json_string_content() {
-    return wrap(arena_.add_parser(common_peg_json_string_parser{}));
-}
-
-common_peg_parser common_peg_parser_builder::single_quoted_string_content() {
-    return wrap(arena_.add_parser(common_peg_python_dict_string_parser{}));
+common_peg_parser common_peg_parser_builder::string_content(char delimiter) {
+    return wrap(arena_.add_parser(common_peg_string_parser{delimiter}));
 }
 
 common_peg_parser common_peg_parser_builder::double_quoted_string() {
-    return rule("dq-string",
-                [this]() { return sequence({ literal("\""), json_string_content(), literal("\""), space() }); });
-}
-
-common_peg_parser common_peg_parser_builder::single_quoted_string() {
-    return rule("sq-string",
-                [this]() { return sequence({ literal("'"), single_quoted_string_content(), literal("'"), space() }); });
-}
-
-common_peg_parser common_peg_parser_builder::flexible_string() {
-    return rule("flexible-string", [this]() { return choice({ double_quoted_string(), single_quoted_string() }); });
-}
-
-// Generic helpers for object/array structure
-
-common_peg_parser common_peg_parser_builder::generic_object(const std::string &            name,
-                                                             const common_peg_parser & string_parser,
-                                                             const common_peg_parser & value_parser) {
-    return rule(name, [this, string_parser, value_parser]() {
-        auto ws      = space();
-        auto member  = sequence({ string_parser, ws, literal(":"), ws, value_parser });
-        auto members = sequence({ member, zero_or_more(sequence({ ws, literal(","), ws, member })) });
-        return sequence({ literal("{"), ws, choice({ literal("}"), sequence({ members, ws, literal("}") }) }) });
+    return rule("double-quoted-string", [this]() {
+        return sequence({literal("\""), string_content('"'), literal("\""), space()});
     });
 }
 
-common_peg_parser common_peg_parser_builder::generic_array(const std::string &            name,
-                                                            const common_peg_parser & value_parser) {
-    return rule(name, [this, value_parser]() {
-        auto ws       = space();
-        auto elements = sequence({ value_parser, zero_or_more(sequence({ literal(","), ws, value_parser })) });
-        return sequence({ literal("["), ws, choice({ literal("]"), sequence({ elements, ws, literal("]") }) }) });
+common_peg_parser common_peg_parser_builder::single_quoted_string() {
+    return rule("single-quoted-string", [this]() {
+        return sequence({literal("'"), string_content('\''), literal("'"), space()});
+    });
+}
+
+common_peg_parser common_peg_parser_builder::quoted_string() {
+    return rule("quoted-string", [this]() {
+        return choice({double_quoted_string(), single_quoted_string()});
     });
 }
 
@@ -1344,7 +1268,7 @@ common_peg_parser common_peg_parser_builder::json_number() {
 
 common_peg_parser common_peg_parser_builder::json_string() {
     return rule("json-string", [this]() {
-        return sequence({literal("\""), json_string_content(), literal("\""), space()});
+        return sequence({literal("\""), string_content('"'), literal("\""), space()});
     });
 }
 
@@ -1361,11 +1285,36 @@ common_peg_parser common_peg_parser_builder::json_null() {
 }
 
 common_peg_parser common_peg_parser_builder::json_object() {
-    return generic_object("json-object", json_string(), json());
+    return rule("json-object", [this]() {
+        auto ws = space();
+        auto member = sequence({json_string(), ws, literal(":"), ws, json()});
+        auto members = sequence({member, zero_or_more(sequence({ws, literal(","), ws, member}))});
+        return sequence({
+            literal("{"),
+            ws,
+            choice({
+                literal("}"),
+                sequence({members, ws, literal("}")})
+            }),
+            ws
+        });
+    });
 }
 
 common_peg_parser common_peg_parser_builder::json_array() {
-    return generic_array("json-array", json());
+    return rule("json-array", [this]() {
+        auto ws = space();
+        auto elements = sequence({json(), zero_or_more(sequence({literal(","), ws, json()}))});
+        return sequence({
+            literal("["),
+            ws,
+            choice({
+                literal("]"),
+                sequence({elements, ws, literal("]")})
+            }),
+            ws
+        });
+    });
 }
 
 common_peg_parser common_peg_parser_builder::json() {
@@ -1382,7 +1331,9 @@ common_peg_parser common_peg_parser_builder::json() {
 }
 
 common_peg_parser common_peg_parser_builder::python_string() {
-    return rule("python-string", [this]() { return choice({ double_quoted_string(), single_quoted_string() }); });
+    return rule("python-string", [this]() {
+        return choice({double_quoted_string(), single_quoted_string()});
+    });
 }
 
 common_peg_parser common_peg_parser_builder::python_number() {
@@ -1390,24 +1341,63 @@ common_peg_parser common_peg_parser_builder::python_number() {
 }
 
 common_peg_parser common_peg_parser_builder::python_bool() {
-    return rule("python-bool", [this]() { return sequence({ choice({ literal("True"), literal("False") }), space() }); });
+    return rule("python-bool", [this]() {
+        return sequence({
+            choice({literal("True"), literal("False")}),
+            space()
+        });
+    });
 }
 
 common_peg_parser common_peg_parser_builder::python_null() {
-    return rule("python-none", [this]() { return sequence({ literal("None"), space() }); });
+    return rule("python-none", [this]() {
+        return sequence({literal("None"), space()});
+    });
 }
 
 common_peg_parser common_peg_parser_builder::python_dict() {
-    return generic_object("python-dict", python_string(), python_value());
+    return rule("python-dict", [this]() {
+        auto ws = space();
+        auto member = sequence({python_string(), ws, literal(":"), ws, python_value()});
+        auto members = sequence({member, zero_or_more(sequence({ws, literal(","), ws, member}))});
+        return sequence({
+            literal("{"),
+            ws,
+            choice({
+                literal("}"),
+                sequence({members, ws, literal("}")})
+            }),
+            ws
+        });
+    });
 }
 
 common_peg_parser common_peg_parser_builder::python_array() {
-    return generic_array("python-array", python_value());
+    return rule("python-array", [this]() {
+        auto ws = space();
+        auto elements = sequence({python_value(), zero_or_more(sequence({literal(","), ws, python_value()}))});
+        return sequence({
+            literal("["),
+            ws,
+            choice({
+                literal("]"),
+                sequence({elements, ws, literal("]")})
+            }),
+            ws
+        });
+    });
 }
 
 common_peg_parser common_peg_parser_builder::python_value() {
     return rule("python-value", [this]() {
-        return choice({ python_dict(), python_array(), python_string(), python_number(), python_bool(), python_null() });
+        return choice({
+            python_dict(),
+            python_array(),
+            python_string(),
+            python_number(),
+            python_bool(),
+            python_null()
+        });
     });
 }
 
@@ -1528,8 +1518,7 @@ static std::unordered_set<std::string> collect_reachable_rules(
                           std::is_same_v<T, common_peg_chars_parser> ||
                           std::is_same_v<T, common_peg_space_parser> ||
                           std::is_same_v<T, common_peg_any_parser> ||
-                          std::is_same_v<T, common_peg_json_string_parser> ||
-                          std::is_same_v<T, common_peg_python_dict_string_parser>) {
+                          std::is_same_v<T, common_peg_string_parser>) {
                 // These parsers do not have any children
             } else if constexpr (std::is_same_v<T, common_peg_sequence_parser>) {
                 for (auto child : p.children) {
@@ -1665,10 +1654,9 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
                     return result + "{" + std::to_string(p.min_count) + "}";
                 }
                 return result + "{" + std::to_string(p.min_count) + "," + std::to_string(p.max_count) + "}";
-            } else if constexpr (std::is_same_v<T, common_peg_json_string_parser>) {
-                return R"(( [^"\\] | "\\" ( ["\\/ bfnrt] | "u" [0-9a-fA-F]{4} ) )*)";
-            } else if constexpr (std::is_same_v<T, common_peg_python_dict_string_parser>) {
-                return R"(( [^"\\] | "\\" ( ["\\/ bfnrt] | "u" [0-9a-fA-F]{4} ) )*)";
+            } else if constexpr (std::is_same_v<T, common_peg_string_parser>) {
+                const std::string delim(1, p.delimiter);
+                return R"(( [^)" + delim + R"(\\] | "\\" ( [)" + delim + R"(\\/ bfnrt] | "u" [0-9a-fA-F]{4} ) )*)";
             } else if constexpr (std::is_same_v<T, common_peg_until_parser>) {
                 if (p.delimiters.empty()) {
                     return ".*";
@@ -1798,10 +1786,8 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
                 {"min_count", p.min_count},
                 {"max_count", p.max_count}
             };
-        } else if constexpr (std::is_same_v<T, common_peg_json_string_parser>) {
-            return json{{"type", "json_string"}};
-        } else if constexpr (std::is_same_v<T, common_peg_python_dict_string_parser>) {
-            return json{{ "type", "python_dict_string" }};
+        } else if constexpr (std::is_same_v<T, common_peg_string_parser>) {
+            return json{{"type", "string"}, {"delimiter", std::string(1, p.delimiter)}};
         } else if constexpr (std::is_same_v<T, common_peg_until_parser>) {
             return json{{"type", "until"}, {"delimiters", p.delimiters}};
         } else if constexpr (std::is_same_v<T, common_peg_schema_parser>) {
@@ -1928,11 +1914,15 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         }
         return parser;
     }
-    if (type == "json_string") {
-        return common_peg_json_string_parser{};
-    }
-    if (type == "python_dict_string") {
-        return common_peg_python_dict_string_parser{};
+    if (type == "string") {
+        if (!j.contains("delimiter")) {
+            throw std::runtime_error("string parser missing delimiter field.");
+        }
+        std::string delimiter = j["delimiter"];
+        if (delimiter.empty()) {
+            throw std::runtime_error("string parser delimiter is empty.");
+        }
+        return common_peg_string_parser{delimiter[0]};
     }
     if (type == "until") {
         if (!j.contains("delimiters") || !j["delimiters"].is_array()) {

common/peg-parser.h

@@ -231,8 +231,9 @@ struct common_peg_chars_parser {
     int max_count;  // -1 for unbounded
 };
 
-struct common_peg_json_string_parser {};
-struct common_peg_python_dict_string_parser {};
+struct common_peg_string_parser {
+    char delimiter;
+};
 
 struct common_peg_until_parser {
     std::vector<std::string> delimiters;
@@ -280,8 +281,7 @@ using common_peg_parser_variant = std::variant<
     common_peg_any_parser,
     common_peg_space_parser,
     common_peg_chars_parser,
-    common_peg_json_string_parser,
-    common_peg_python_dict_string_parser,
+    common_peg_string_parser,
     common_peg_until_parser,
     common_peg_schema_parser,
     common_peg_rule_parser,
@@ -340,10 +340,6 @@ class common_peg_parser_builder {
     common_peg_parser wrap(common_peg_parser_id id) { return common_peg_parser(id, *this); }
     common_peg_parser add(const common_peg_parser_variant & p) { return wrap(arena_.add_parser(p)); }
 
-    // Generic helpers for building object/array structures with configurable string/value parsers.
-    common_peg_parser generic_object(const std::string & name, const common_peg_parser & string_parser, const common_peg_parser & value_parser);
-    common_peg_parser generic_array(const std::string & name, const common_peg_parser & value_parser);
-
   public:
     common_peg_parser_builder();
 
@@ -444,13 +440,10 @@ class common_peg_parser_builder {
     common_peg_parser single_quoted_string();
 
     // Matches a string that accepts both double-quoted and single-quoted styles.
-    common_peg_parser flexible_string();
+    common_peg_parser quoted_string();
 
-    // Matches double-quoted string content without the surrounding quotes.
-    common_peg_parser json_string_content();
-
-    // Matches single-quoted string content without the surrounding quotes.
-    common_peg_parser single_quoted_string_content();
+    // Matches string content without the surrounding delimiter.
+    common_peg_parser string_content(char delimiter);
 
     // Creates a complete JSON parser supporting objects, arrays, strings, numbers, booleans, and null.
     //   value -> object | array | string | number | true | false | null

docs/build.md

@@ -599,7 +599,13 @@ If KleidiAI is enabled, the output will contain a line similar to:
 ```
 load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
 ```
-KleidiAI's microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm and SME. llama.cpp selects the most efficient kernel based on runtime CPU feature detection. However, on platforms that support SME, you must manually enable SME microkernels by setting the environment variable `GGML_KLEIDIAI_SME=1`.
+KleidiAI’s microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm, SVE, and SME. Llama.cpp selects the most efficient kernels at runtime based on detected CPU capabilities.
+On CPUs that support SME, SME microkernels are enabled automatically using runtime detection.
+The environment variable GGML_KLEIDIAI_SME can be used to control SME behavior:
+- Not set: enable SME automatically if supported and detected.
+- 0: disable SME.
+- <n> > 0: enable SME and assume <n> available SME units (override auto detection).
+If SME is not supported by the CPU, SME microkernels are always disabled.
 
 Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`.
 

ggml/src/ggml-cpu/arch-fallback.h

@@ -202,8 +202,9 @@
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
+#define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x1_generic ggml_quantize_mat_q8_K_4x1
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -520,7 +520,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
             /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
         },
-        /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
         /* .rhs_type           = */ GGML_TYPE_Q4_0,
         /* .op_type            = */ GGML_TYPE_F32,
@@ -631,7 +631,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
             /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
         },
-        /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
         /* .rhs_type           = */ GGML_TYPE_Q4_0,
         /* .op_type            = */ GGML_TYPE_F32,
@@ -801,7 +801,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
             /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
             /* .pack_func_ex          = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
         },
-        /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
         /* .rhs_type           = */ GGML_TYPE_Q8_0,
         /* .op_type            = */ GGML_TYPE_F32,

ggml/src/ggml-cpu/repack.h

@@ -28,13 +28,17 @@ template <int K, int N> struct block {
 // control size
 static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
 static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<4, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<4,16> size/padding");
 static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
 static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+static_assert(sizeof(block<8, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 16, "wrong block<8,16> size/padding");
 
 using block_q4_0x4 = block<4, 4>;
 using block_q4_0x8 = block<4, 8>;
+using block_q4_0x16 = block<4, 16>;
 using block_q8_0x4 = block<8, 4>;
 using block_q8_0x8 = block<8, 8>;
+using block_q8_0x16 = block<8, 16>;
 
 struct block_q4_Kx8 {
     ggml_half d[8];      // super-block scale for quantized scales
@@ -44,7 +48,14 @@ struct block_q4_Kx8 {
 };
 
 static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+struct block_q4_Kx16 {
+    ggml_half d[16];      // super-block scale for quantized scales
+    ggml_half dmin[16];   // super-block scale for quantized mins
+    uint8_t scales[192];  // scales and mins, quantized with 6 bits
+    uint8_t qs[2048];    // 4--bit quants
+};
 
+static_assert(sizeof(block_q4_Kx16) == sizeof(ggml_half) * 32 + K_SCALE_SIZE * 16 + QK_K * 8, "wrong q4_K block size/padding");
 struct block_q2_Kx8 {
     ggml_half d[8];      // super-block scale for quantized scales
     ggml_half dmin[8];   // super-block scale for quantized mins
@@ -53,6 +64,13 @@ struct block_q2_Kx8 {
 };
 
 static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
+struct block_q2_Kx16 {
+    ggml_half d[16];       // Super-block scale for quantized scales
+    ggml_half dmin[16];    // Super-block scale for quantized mins
+    uint8_t   scales[256]; // Sub-block scales (16 cols * 16 sub-blocks)
+    uint8_t   qs[1024];    // Data (16 cols * 64 bytes per block)
+};
+static_assert(sizeof(block_q2_Kx16) == sizeof(ggml_half) * 32 + QK_K + QK_K * 4, "wrong q2_K block size/padding");
 
 struct block_q5_Kx8 {
     ggml_half d[8];              // super-block scale for quantized scales
@@ -97,6 +115,12 @@ struct block_iq4_nlx8 {
 
 static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
 
+struct block_iq4_nlx16 {
+    ggml_half d[16];            // deltas for 16 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 8];  // nibbles / quants for 16 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx16) == 16 * sizeof(ggml_half) + QK4_NL * 8, "wrong iq4_nlx16 block size/padding");
 struct block_mxfp4x4 {
     uint8_t e[4];
     uint8_t qs[QK_MXFP4 * 2];
@@ -109,7 +133,6 @@ struct block_mxfp4x8 {
 };
 static_assert(sizeof(block_mxfp4x8) == 8 + QK_MXFP4 * 4, "wrong mxfp4x8 block size/padding");
 
-
 #if defined(__cplusplus)
 extern "C" {
 #endif
@@ -132,6 +155,8 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -146,10 +171,22 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#if defined __riscv_zvfh
+void ggml_quantize_mat_q8_0_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif
 
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -170,6 +207,8 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
 void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -184,10 +223,22 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
 void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#if defined __riscv_zvfh
+void ggml_quantize_mat_q8_0_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif
 
 #if defined(__cplusplus)
 } // extern "C"

ggml/src/ggml-metal/ggml-metal-context.m

@@ -75,6 +75,10 @@ struct ggml_metal {
     // abort ggml_metal_graph_compute if callback returns true
     ggml_abort_callback abort_callback;
     void *              abort_callback_data;
+
+    // error state - set when a command buffer fails during synchronize
+    // once set, graph_compute will return GGML_STATUS_FAILED until the backend is recreated
+    bool has_error;
 };
 
 ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
@@ -158,6 +162,8 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
     res->capture_started = false;
     res->capture_scope = nil;
 
+    res->has_error = false;
+
     res->gf = nil;
     res->encode_async = nil;
     for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
@@ -246,7 +252,8 @@ void ggml_metal_synchronize(ggml_metal_t ctx) {
                 if (status == MTLCommandBufferStatusError) {
                     GGML_LOG_ERROR("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);
                 }
-                GGML_ABORT("fatal error");
+                ctx->has_error = true;
+                return;
             }
         }
     }
@@ -262,7 +269,15 @@ void ggml_metal_synchronize(ggml_metal_t ctx) {
                 if (status == MTLCommandBufferStatusError) {
                     GGML_LOG_ERROR("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);
                 }
-                GGML_ABORT("fatal error");
+
+                // release this and all remaining command buffers before returning
+                for (size_t j = i; j < ctx->cmd_bufs_ext.count; ++j) {
+                    [ctx->cmd_bufs_ext[j] release];
+                }
+                [ctx->cmd_bufs_ext removeAllObjects];
+
+                ctx->has_error = true;
+                return;
             }
 
             [cmd_buf release];
@@ -414,6 +429,11 @@ bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, con
 }
 
 enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph * gf) {
+    if (ctx->has_error) {
+        GGML_LOG_ERROR("%s: backend is in error state from a previous command buffer failure - recreate the backend to recover\n", __func__);
+        return GGML_STATUS_FAILED;
+    }
+
     // number of nodes encoded by the main thread (empirically determined)
     const int n_main = MAX(64, 0.1*gf->n_nodes);
 

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

@@ -1963,6 +1963,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
           (
            op->src[0]->type == GGML_TYPE_F32  || // TODO: helper function
            op->src[0]->type == GGML_TYPE_F16  ||
+           op->src[0]->type == GGML_TYPE_BF16 ||
            op->src[0]->type == GGML_TYPE_Q4_0 ||
            op->src[0]->type == GGML_TYPE_Q4_1 ||
            op->src[0]->type == GGML_TYPE_Q5_0 ||
@@ -1977,6 +1978,8 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
            op->src[0]->type == GGML_TYPE_Q4_K ||
            op->src[0]->type == GGML_TYPE_Q5_K ||
            op->src[0]->type == GGML_TYPE_Q6_K ||
+           op->src[0]->type == GGML_TYPE_Q2_K ||
+           op->src[0]->type == GGML_TYPE_Q3_K ||
            false) && (ne11 >= 4 && ne11 <= 8)
          )
         )

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

@@ -3481,6 +3481,13 @@ template [[host_name("kernel_mul_mv_ext_f16_f32_r1_3")]]    kernel mul_mv_ext_q4
 template [[host_name("kernel_mul_mv_ext_f16_f32_r1_4")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, half4,        4,  dequantize_f16_t4>;
 template [[host_name("kernel_mul_mv_ext_f16_f32_r1_5")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, half4,        4,  dequantize_f16_t4>;
 
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mv_ext_bf16_f32_r1_2")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, bfloat4,      4,  dequantize_bf16_t4>;
+template [[host_name("kernel_mul_mv_ext_bf16_f32_r1_3")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, bfloat4,      4,  dequantize_bf16_t4>;
+template [[host_name("kernel_mul_mv_ext_bf16_f32_r1_4")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, bfloat4,      4,  dequantize_bf16_t4>;
+template [[host_name("kernel_mul_mv_ext_bf16_f32_r1_5")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, bfloat4,      4,  dequantize_bf16_t4>;
+#endif
+
 template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_2")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q4_0,   32, dequantize_q4_0_t4>;
 template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_3")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q4_0,   32, dequantize_q4_0_t4>;
 template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_4")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q4_0,   32, dequantize_q4_0_t4>;
@@ -3531,6 +3538,16 @@ template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_3")]] kernel mul_mv_ext_q4x4
 template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q6_K, 256, dequantize_q6_K>;
 template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q6_K, 256, dequantize_q6_K>;
 
+template [[host_name("kernel_mul_mv_ext_q2_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q2_K, 256, dequantize_q2_K>;
+template [[host_name("kernel_mul_mv_ext_q2_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q2_K, 256, dequantize_q2_K>;
+template [[host_name("kernel_mul_mv_ext_q2_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q2_K, 256, dequantize_q2_K>;
+template [[host_name("kernel_mul_mv_ext_q2_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q2_K, 256, dequantize_q2_K>;
+
+template [[host_name("kernel_mul_mv_ext_q3_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q3_K, 256, dequantize_q3_K>;
+template [[host_name("kernel_mul_mv_ext_q3_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q3_K, 256, dequantize_q3_K>;
+template [[host_name("kernel_mul_mv_ext_q3_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q3_K, 256, dequantize_q3_K>;
+template [[host_name("kernel_mul_mv_ext_q3_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q3_K, 256, dequantize_q3_K>;
+
 template<typename T0, typename T1, short NR0, typename args_t>
 void kernel_mul_mv_t_t_impl(
         args_t args,

gguf-py/gguf/constants.py

@@ -177,6 +177,8 @@ class Keys:
         TEMPERATURE_LENGTH           = "{arch}.attention.temperature_length"
         KEY_LENGTH_MLA               = "{arch}.attention.key_length_mla"
         VALUE_LENGTH_MLA             = "{arch}.attention.value_length_mla"
+        KEY_LENGTH_SWA               = "{arch}.attention.key_length_swa"
+        VALUE_LENGTH_SWA             = "{arch}.attention.value_length_swa"
         SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
         SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
         TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
@@ -188,6 +190,7 @@ class Keys:
 
     class Rope:
         DIMENSION_COUNT           = "{arch}.rope.dimension_count"
+        DIMENSION_COUNT_SWA       = "{arch}.rope.dimension_count_swa"
         DIMENSION_SECTIONS        = "{arch}.rope.dimension_sections"
         FREQ_BASE                 = "{arch}.rope.freq_base"
         FREQ_BASE_SWA             = "{arch}.rope.freq_base_swa"

gguf-py/gguf/gguf_writer.py

@@ -773,6 +773,12 @@ class GGUFWriter:
     def add_value_length_mla(self, length: int) -> None:
         self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA.format(arch=self.arch), length)
 
+    def add_key_length_swa(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.KEY_LENGTH_SWA.format(arch=self.arch), length)
+
+    def add_value_length_swa(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.VALUE_LENGTH_SWA.format(arch=self.arch), length)
+
     def add_indexer_head_count(self, count: int) -> None:
         self.add_uint32(Keys.Attention.Indexer.HEAD_COUNT.format(arch=self.arch), count)
 
@@ -946,6 +952,9 @@ class GGUFWriter:
     def add_rope_dimension_count(self, count: int) -> None:
         self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
 
+    def add_rope_dimension_count_swa(self, count: int) -> None:
+        self.add_uint32(Keys.Rope.DIMENSION_COUNT_SWA.format(arch=self.arch), count)
+
     def add_rope_dimension_sections(self, dims: Sequence[int]) -> None:
         self.add_array(Keys.Rope.DIMENSION_SECTIONS.format(arch=self.arch), dims)
 

src/llama-arch.cpp

@@ -230,11 +230,14 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
+    { LLM_KV_ATTENTION_KEY_LENGTH_SWA,               "%s.attention.key_length_swa"               },
+    { LLM_KV_ATTENTION_VALUE_LENGTH_SWA,             "%s.attention.value_length_swa"             },
     { LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,           "%s.attention.indexer.head_count"           },
     { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,           "%s.attention.indexer.key_length"           },
     { LLM_KV_ATTENTION_INDEXER_TOP_K,                "%s.attention.indexer.top_k"                },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,           "%s.rope.dimension_count"                 },
+    { LLM_KV_ROPE_DIMENSION_COUNT_SWA,       "%s.rope.dimension_count_swa"             },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,        "%s.rope.dimension_sections"              },
     { LLM_KV_ROPE_FREQ_BASE,                 "%s.rope.freq_base"                       },
     { LLM_KV_ROPE_FREQ_BASE_SWA,             "%s.rope.freq_base_swa"                   },

src/llama-arch.h

@@ -234,11 +234,14 @@ enum llm_kv {
     LLM_KV_ATTENTION_TEMPERATURE_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
+    LLM_KV_ATTENTION_KEY_LENGTH_SWA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_SWA,
     LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
     LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
     LLM_KV_ATTENTION_INDEXER_TOP_K,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
+    LLM_KV_ROPE_DIMENSION_COUNT_SWA,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
     LLM_KV_ROPE_FREQ_BASE,
     LLM_KV_ROPE_FREQ_BASE_SWA,

src/llama-context.cpp

@@ -2876,19 +2876,23 @@ llama_context * llama_init_from_model(
 
     if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_k)) {
         const uint32_t blck_size = ggml_blck_size(params.type_k);
-        if (model->hparams.n_embd_head_k % blck_size != 0) {
-            LLAMA_LOG_ERROR("%s: K cache type %s with block size %u does not divide n_embd_head_k=%u\n",
-                __func__, ggml_type_name(params.type_k), blck_size, model->hparams.n_embd_head_k);
-            return nullptr;
+        for (uint32_t il = 0; il < model->hparams.n_layer; ++il) {
+            if (model->hparams.n_embd_head_k(il) % blck_size != 0) {
+                LLAMA_LOG_ERROR("%s: K cache type %s with block size %u does not divide n_embd_head_k=%u\n",
+                    __func__, ggml_type_name(params.type_k), blck_size, model->hparams.n_embd_head_k(il));
+                return nullptr;
+            }
         }
     }
 
     if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_v)) {
         const uint32_t blck_size = ggml_blck_size(params.type_v);
-        if (model->hparams.n_embd_head_v % blck_size != 0) {
-            LLAMA_LOG_ERROR("%s: V cache type %s with block size %u does not divide n_embd_head_k=%u\n",
-                __func__, ggml_type_name(params.type_v), blck_size, model->hparams.n_embd_head_v);
-            return nullptr;
+        for (uint32_t il = 0; il < model->hparams.n_layer; ++il) {
+            if (model->hparams.n_embd_head_v(il) % blck_size != 0) {
+                LLAMA_LOG_ERROR("%s: V cache type %s with block size %u does not divide n_embd_head_v=%u\n",
+                    __func__, ggml_type_name(params.type_v), blck_size, model->hparams.n_embd_head_v(il));
+                return nullptr;
+            }
         }
     }
 

src/llama-graph.cpp

@@ -849,13 +849,13 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     ubatch           (params.ubatch),
     n_embd           (hparams.n_embd),
     n_layer          (hparams.n_layer),
-    n_rot            (hparams.n_rot),
+    n_rot            (hparams.n_rot()),
     n_ctx            (cparams.n_ctx),
     n_head           (hparams.n_head()),
     n_head_kv        (hparams.n_head_kv()),
-    n_embd_head_k    (hparams.n_embd_head_k),
+    n_embd_head_k    (hparams.n_embd_head_k()),
     n_embd_k_gqa     (hparams.n_embd_k_gqa()),
-    n_embd_head_v    (hparams.n_embd_head_v),
+    n_embd_head_v    (hparams.n_embd_head_v()),
     n_embd_v_gqa     (hparams.n_embd_v_gqa()),
     n_expert         (hparams.n_expert),
     n_expert_used    (cparams.warmup ? hparams.n_expert : hparams.n_expert_used),

src/llama-hparams.cpp

@@ -62,6 +62,14 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
     return n_head/n_head_kv;
 }
 
+uint32_t llama_hparams::n_rot(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_rot_swa : n_rot_full;
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 uint32_t llama_hparams::n_embd_inp() const {
     uint32_t n_embd_inp = n_embd;
 
@@ -76,16 +84,32 @@ uint32_t llama_hparams::n_embd_out() const {
     return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
 }
 
+uint32_t llama_hparams::n_embd_head_k(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_embd_head_k_swa : n_embd_head_k_full;
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_embd_head_v(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_embd_head_v_swa : n_embd_head_v_full;
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
-    return n_embd_head_k * n_head_kv;
+    return n_embd_head_k(il) * n_head_kv;
 }
 
 uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
-    return n_embd_head_v * n_head_kv;
+    return n_embd_head_v(il) * n_head_kv;
 }
 
 bool llama_hparams::is_n_embd_k_gqa_variable() const {
@@ -197,11 +221,11 @@ bool llama_hparams::is_mla() const {
 }
 
 uint32_t llama_hparams::n_embd_head_k_mla() const {
-    return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k;
+    return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k();
 }
 
 uint32_t llama_hparams::n_embd_head_v_mla() const {
-    return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v;
+    return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v();
 }
 
 bool llama_hparams::has_kv(uint32_t il) const {

src/llama-hparams.h

@@ -44,13 +44,20 @@ struct llama_hparams {
     uint32_t n_embd;
     uint32_t n_layer;
     int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
-    uint32_t n_rot;
-    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
-    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
     uint32_t n_expert = 0;
     uint32_t n_expert_used = 0;
     uint32_t n_rel_attn_bkts = 0;
 
+    // different head size for full_attention and SWA layers
+    uint32_t n_embd_head_k_full; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
+    uint32_t n_embd_head_v_full; // dimension of values (d_v) aka n_embd_head
+    uint32_t n_embd_head_k_swa;
+    uint32_t n_embd_head_v_swa;
+
+    // different RoPE dimensions for full_attention and SWA layers
+    uint32_t n_rot_full;
+    uint32_t n_rot_swa;
+
     // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
     uint32_t n_embd_head_k_mla_impl = 0;
     uint32_t n_embd_head_v_mla_impl = 0;
@@ -247,12 +254,18 @@ struct llama_hparams {
 
     uint32_t n_gqa(uint32_t il = 0) const;
 
+    uint32_t n_rot(uint32_t il = 0) const;
+
     // dimension of main + auxiliary input embeddings
     uint32_t n_embd_inp() const;
 
     // dimension of output embeddings
     uint32_t n_embd_out() const;
 
+    // dimension of key/value embeddings for each head (per layer)
+    uint32_t n_embd_head_k(uint32_t il = 0) const;
+    uint32_t n_embd_head_v(uint32_t il = 0) const;
+
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
 

src/llama-kv-cache.cpp

@@ -1033,8 +1033,8 @@ ggml_tensor * llama_kv_cache::get_k(ggml_context * ctx, int32_t il, uint32_t n_k
     const uint32_t ns = sinfo.s1 - sinfo.s0 + 1;
 
     return ggml_view_4d(ctx, k,
-            hparams.n_embd_head_k, hparams.n_head_kv(il), n_kv, ns,
-            ggml_row_size(k->type, hparams.n_embd_head_k),
+            hparams.n_embd_head_k(il), hparams.n_head_kv(il), n_kv, ns,
+            ggml_row_size(k->type, hparams.n_embd_head_k(il)),
             ggml_row_size(k->type, n_embd_k_gqa),
             ggml_row_size(k->type, n_embd_k_gqa*kv_size),
             ggml_row_size(k->type, n_embd_k_gqa*kv_size)*sinfo.s0);
@@ -1056,8 +1056,8 @@ ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_k
     if (!v_trans) {
         // note: v->nb[1] <= v->nb[2]
         return ggml_view_4d(ctx, v,
-                hparams.n_embd_head_v, hparams.n_head_kv(il), n_kv, ns,
-                ggml_row_size(v->type, hparams.n_embd_head_v),          // v->nb[1]
+                hparams.n_embd_head_v(il), hparams.n_head_kv(il), n_kv, ns,
+                ggml_row_size(v->type, hparams.n_embd_head_v(il)),          // v->nb[1]
                 ggml_row_size(v->type, n_embd_v_gqa),                   // v->nb[2]
                 ggml_row_size(v->type, n_embd_v_gqa*kv_size),           // v->nb[3]
                 ggml_row_size(v->type, n_embd_v_gqa*kv_size)*sinfo.s0);
@@ -1065,8 +1065,8 @@ ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_k
 
     // note: v->nb[1] > v->nb[2]
     return ggml_view_4d(ctx, v,
-            n_kv, hparams.n_head_kv(il), hparams.n_embd_head_v, ns,
-            ggml_row_size(v->type, kv_size*hparams.n_embd_head_v),  // v->nb[1]
+            n_kv, hparams.n_head_kv(il), hparams.n_embd_head_v(il), ns,
+            ggml_row_size(v->type, kv_size*hparams.n_embd_head_v(il)),  // v->nb[1]
             ggml_row_size(v->type, kv_size),                        // v->nb[2]
             ggml_row_size(v->type, kv_size*n_embd_v_gqa),           // v->nb[3]
             ggml_row_size(v->type, kv_size*n_embd_v_gqa)*sinfo.s0);
@@ -1544,7 +1544,8 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
                 ggml_tensor * shift,
                 ggml_tensor * factors,
                       float   freq_base,
-                      float   freq_scale) const {
+                      float   freq_scale,
+                   uint32_t   il) const {
     const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
 
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
@@ -1552,7 +1553,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_beta_slow   = cparams.yarn_beta_slow;
     const auto & yarn_attn_factor = cparams.yarn_attn_factor;
 
-    const auto & n_rot     = hparams.n_rot;
+    const auto & n_rot     = hparams.n_rot(il);
     const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
                                 // @ngxson : this is a workaround
                                 // for M-RoPE, we want to rotate the whole vector when doing KV shift
@@ -1606,13 +1607,6 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
     auto * ctx = res->get_ctx();
     auto * gf  = res->get_gf();
 
-    const auto & n_embd_head_k = hparams.n_embd_head_k;
-  //const auto & n_embd_head_v = hparams.n_embd_head_v;
-
-    const auto & n_rot = hparams.n_rot;
-
-    const auto n_embd_nope = hparams.n_lora_kv > 0 ? n_embd_head_k - n_rot : 0;
-
     auto inp = std::make_unique<llm_graph_input_k_shift>(this);
 
     inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, (int64_t) get_size()*n_stream);
@@ -1626,6 +1620,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
         const int64_t n_head_kv    = hparams.n_head_kv(il);
         const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
 
+        const auto n_rot         = hparams.n_rot(il);
+        const auto n_embd_head_k = hparams.n_embd_head_k(il);
+        const auto n_embd_nope   = hparams.n_lora_kv > 0 ? n_embd_head_k - n_rot : 0;
+
         const float freq_base_l  = model.get_rope_freq_base (cparams, il);
         const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
 
@@ -1638,7 +1636,7 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
                 ggml_row_size(layer.k->type, n_embd_k_gqa),
                 ggml_row_size(layer.k->type, n_embd_nope));
 
-        ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);
+        ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, il);
 
         ggml_build_forward_expand(gf, cur);
     }

src/llama-kv-cache.h

@@ -264,7 +264,8 @@ private:
                     ggml_tensor * shift,
                     ggml_tensor * factors,
                           float   freq_base,
-                          float   freq_scale) const;
+                          float   freq_scale,
+                       uint32_t   il) const;
 
     ggml_cgraph * build_graph_shift(
                llm_graph_result * res,

src/llama-model-loader.cpp

@@ -918,7 +918,7 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_ROPE:
             {
-                const int n_embd_head = hparams.n_embd_head_v;
+                const int n_embd_head = hparams.n_embd_head_v();
                 const int n_head = hparams.n_head();
                 ggml_tensor * a = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd_head, n_head, 512);
                 ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 512);

src/llama-model-saver.cpp

@@ -186,8 +186,10 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);
     add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS,          hparams.f_max_alibi_bias);
     add_kv(LLM_KV_ATTENTION_CLAMP_KQV,               hparams.f_clamp_kqv);
-    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k);
-    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k_full);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v_full);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA,          hparams.n_embd_head_k_swa);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,        hparams.n_embd_head_v_swa);
     add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,           hparams.f_norm_eps);
     add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
     add_kv(LLM_KV_ATTENTION_CAUSAL,                  hparams.causal_attn);
@@ -199,7 +201,8 @@ void llama_model_saver::add_kv_from_model() {
 
     const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
 
-    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot);
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot_full);
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT_SWA,          hparams.n_rot_swa);
     add_kv(LLM_KV_ROPE_FREQ_BASE,                    hparams.rope_freq_base_train);
     // add_kv(LLM_KV_ROPE_SCALE_LINEAR,                 rope_scaling_factor); // old name
     add_kv(LLM_KV_ROPE_SCALING_TYPE,                 llama_rope_scaling_type_name(hparams.rope_scaling_type_train));

src/llama-model.cpp

@@ -459,26 +459,37 @@ void llama_model::load_hparams(llama_model_loader & ml) {
         // gpt-neox n_rot = rotary_pct * (n_embd / n_head)
         // gpt-j n_rot = rotary_dim
 
-        hparams.n_embd_head_k = hparams.n_embd / hparams.n_head();
-        ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false);
+        hparams.n_embd_head_k_full = hparams.n_embd / hparams.n_head();
+        ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k_full, false);
 
-        hparams.n_embd_head_v = hparams.n_embd / hparams.n_head();
-        ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
+        hparams.n_embd_head_v_full = hparams.n_embd / hparams.n_head();
+        ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v_full, false);
 
         // sanity check for n_rot (optional)
-        hparams.n_rot = hparams.n_embd_head_k;
+        hparams.n_rot_full = hparams.n_embd_head_k_full;
 
-        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
+        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot_full, false);
 
         if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON || arch == LLM_ARCH_LLAMA_EMBED) {
-            if (hparams.n_rot != hparams.n_embd_head_k) {
-                throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
+            if (hparams.n_rot_full != hparams.n_embd_head_k_full) {
+                throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot_full, hparams.n_embd_head_k_full));
             }
         }
     } else {
-        hparams.n_rot = 0;
-        hparams.n_embd_head_k = 0;
-        hparams.n_embd_head_v = 0;
+        hparams.n_rot_full = 0;
+        hparams.n_embd_head_k_full = 0;
+        hparams.n_embd_head_v_full = 0;
+    }
+
+    // head size and n_rot for SWA layers
+    {
+        hparams.n_embd_head_k_swa = hparams.n_embd_head_k_full;
+        hparams.n_embd_head_v_swa = hparams.n_embd_head_v_full;
+        ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_SWA, hparams.n_embd_head_k_swa, false);
+        ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_SWA, hparams.n_embd_head_v_swa, false);
+
+        hparams.n_rot_swa = hparams.n_rot_full;
+        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT_SWA, hparams.n_rot_swa, false);
     }
 
     // for differentiating model types
@@ -1114,10 +1125,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                         break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
-
-                // Load attention parameters
-                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH,   hparams.n_embd_head_k, false);
-                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
             } break;
         case LLM_ARCH_PLAMO3:
             {
@@ -1212,7 +1219,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L173
                 hparams.f_attention_scale = type == LLM_TYPE_27B
                     ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
-                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
             } break;
         case LLM_ARCH_GEMMA3:
             {
@@ -1245,7 +1252,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L289
                 hparams.f_attention_scale = type == LLM_TYPE_27B
                     ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
-                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
             } break;
         case LLM_ARCH_GEMMA3N:
             {
@@ -1294,7 +1301,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     case 24: type = LLM_TYPE_0_3B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
-                hparams.f_attention_scale = 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+                hparams.f_attention_scale = 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
 
             } break;
         case LLM_ARCH_STARCODER2:
@@ -2487,7 +2494,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,    hparams.n_embd_head_k_mla_impl);
                 ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA,  hparams.n_embd_head_v_mla_impl);
                 ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,      hparams.n_lora_kv);
-                ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT,        hparams.n_rot);
                 ml.get_key(LLM_KV_SSM_CONV_KERNEL,             hparams.ssm_d_conv);
                 ml.get_key(LLM_KV_KDA_HEAD_DIM,                hparams.n_embd_head_kda);
 
@@ -2518,6 +2524,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
 
+                // full_attention layer only use half of the RoPE dimensions
+                hparams.n_rot_full = hparams.n_rot_full / 2;
+
                 // MoE + SWA parameters
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
                 ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
@@ -2661,13 +2670,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         const int64_t n_embd        = hparams.n_embd;
         const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
         const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
-        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-        const int64_t n_embd_head_v = hparams.n_embd_head_v;
+        const int64_t n_embd_head_k = hparams.n_embd_head_k();
+        const int64_t n_embd_head_v = hparams.n_embd_head_v();
         const int64_t n_ff          = hparams.n_ff();
         const int64_t n_embd_gqa    = n_embd_v_gqa;
         const int64_t n_vocab       = vocab.n_tokens();
         const int64_t n_token_types = vocab.n_token_types();
-        const int64_t n_rot         = hparams.n_rot;
+        const int64_t n_rot         = hparams.n_rot();
         const int64_t n_expert      = hparams.n_expert;
         const int64_t n_expert_used = hparams.n_expert_used;
         const int64_t n_ctx_train   = hparams.n_ctx_train;
@@ -2967,8 +2976,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_MINICPM3:
                 {
-                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
-                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
 
                     const int64_t q_lora_rank  = hparams.n_lora_q;
                     const int64_t kv_lora_rank = hparams.n_lora_kv;
@@ -3840,8 +3849,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t dt_dim              = std::max(64, int(hparams.n_embd / 16));
 
                     // attention parameters
-                    const uint32_t qk_dim = hparams.n_embd_head_k;
-                    const uint32_t v_dim  = hparams.n_embd_head_v;
+                    const uint32_t qk_dim = hparams.n_embd_head_k();
+                    const uint32_t v_dim  = hparams.n_embd_head_v();
 
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -3901,8 +3910,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_PLAMO3:
                 {
-                    const int64_t head_dim_q = hparams.n_embd_head_k;
-                    const int64_t head_dim_v = hparams.n_embd_head_v;
+                    const int64_t head_dim_q = hparams.n_embd_head_k();
+                    const int64_t head_dim_v = hparams.n_embd_head_v();
 
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -4649,7 +4658,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_SEED_OSS:
                 {
-                    const uint32_t head_dim             = hparams.n_embd_head_k;
+                    const uint32_t head_dim             = hparams.n_embd_head_k();
                     const int64_t n_qo_dim              = n_head * head_dim;
                     const int64_t n_kv_dim              = n_head_kv * head_dim;
 
@@ -4878,7 +4887,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
                     const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
 
-                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot();
                     const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
                     GGML_ASSERT(n_embd_head_qk_nope >= 1);
 
@@ -4957,8 +4966,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_PLM:
                 {
-                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
-                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
                     const int64_t kv_lora_rank = hparams.n_lora_kv;
 
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5396,7 +5405,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
                     const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
 
-                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot();
                     const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
 
                     const int64_t q_lora_rank  = hparams.n_lora_q;
@@ -5680,7 +5689,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_expert       = hparams.n_expert;
                     const int64_t n_expert_used  = hparams.n_expert_used;
                     const int64_t n_ff_shexp     = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : n_ff_exp;
-                    const int64_t head_dim       = hparams.n_embd_head_k;
+                    const int64_t head_dim       = hparams.n_embd_head_k();
                     const int64_t n_qo_dim       = n_head * head_dim;
                     const int64_t n_kv_dim       = n_head_kv * head_dim;
 
@@ -6968,7 +6977,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                              // Kimi: qk_rope_head_dim = 64 (actual RoPE dimension for MLA)
                              // Note: hparams.n_rot may be 72 (from conversion) but actual is 64
-                             const int64_t qk_rope_head_dim = hparams.n_rot;  // From config: qk_rope_head_dim
+                             const int64_t qk_rope_head_dim = hparams.n_rot();  // From config: qk_rope_head_dim
                              layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + qk_rope_head_dim}, 0);
                              // Support Legacy GGUFs that don't split wkv_b (MLA KV cache disabled)
                              layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i),
@@ -7339,7 +7348,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     // ("rope_freqs.weight") and ggml uses only the first (n_rot_l/2) entries per layer.
                     uint32_t n_rot_max = 0;
                     for (int i = 0; i < n_layer; ++i) {
-                        n_rot_max = std::max(n_rot_max, hparams.n_rot);
+                        n_rot_max = std::max(n_rot_max, hparams.n_rot(i));
                     }
                     if (n_rot_max == 0) {
                         n_rot_max = n_rot;
@@ -7674,11 +7683,11 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_layer               = %u\n",     __func__, hparams.n_layer);
         LLAMA_LOG_INFO("%s: n_head                = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head(il);    }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_head_kv             = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
-        LLAMA_LOG_INFO("%s: n_rot                 = %u\n",     __func__, hparams.n_rot);
+        LLAMA_LOG_INFO("%s: n_rot                 = %u\n",     __func__, hparams.n_rot_full);
         LLAMA_LOG_INFO("%s: n_swa                 = %u\n",     __func__, hparams.n_swa);
         LLAMA_LOG_INFO("%s: is_swa_any            = %u\n",     __func__, hparams.is_swa_any());
-        LLAMA_LOG_INFO("%s: n_embd_head_k         = %u\n",     __func__, hparams.n_embd_head_k);
-        LLAMA_LOG_INFO("%s: n_embd_head_v         = %u\n",     __func__, hparams.n_embd_head_v);
+        LLAMA_LOG_INFO("%s: n_embd_head_k         = %u\n",     __func__, hparams.n_embd_head_k_full);
+        LLAMA_LOG_INFO("%s: n_embd_head_v         = %u\n",     __func__, hparams.n_embd_head_v_full);
         LLAMA_LOG_INFO("%s: n_gqa                 = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il);        }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_embd_k_gqa          = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_embd_v_gqa          = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer).c_str());
@@ -7702,6 +7711,9 @@ void llama_model::print_info() const {
         if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
             LLAMA_LOG_INFO("%s: freq_base_swa         = %.1f\n",   __func__, hparams.rope_freq_base_train_swa);
             LLAMA_LOG_INFO("%s: freq_scale_swa        = %g\n",     __func__, hparams.rope_freq_scale_train_swa);
+            LLAMA_LOG_INFO("%s: n_embd_head_k_swa     = %u\n",     __func__, hparams.n_embd_head_k_swa);
+            LLAMA_LOG_INFO("%s: n_embd_head_v_swa     = %u\n",     __func__, hparams.n_embd_head_v_swa);
+            LLAMA_LOG_INFO("%s: n_rot_swa             = %u\n",     __func__, hparams.n_rot_swa);
         }
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn       = %u\n",     __func__, hparams.n_ctx_orig_yarn);
         LLAMA_LOG_INFO("%s: rope_yarn_log_mul     = %.4f\n",   __func__, hparams.rope_yarn_log_mul);

src/models/afmoe.cpp

@@ -1,8 +1,8 @@
 #include "models.h"
 
 llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/apertus.cpp

@@ -3,10 +3,10 @@
 
 
 llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/arcee.cpp

@@ -2,10 +2,10 @@
 
 
 llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/arctic.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/baichuan.cpp

@@ -2,10 +2,10 @@
 
 
 llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/bailingmoe2.cpp

@@ -2,10 +2,10 @@
 
 llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/bert.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/bitnet.cpp

@@ -2,9 +2,9 @@
 
 
 llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/bloom.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/chameleon.cpp

@@ -3,10 +3,10 @@
 #include <float.h>
 
 llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/chatglm.cpp

@@ -2,10 +2,10 @@
 
 
 llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/codeshell.cpp

@@ -1,11 +1,11 @@
 #include "models.h"
 
 llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/cogvlm.cpp

@@ -2,11 +2,11 @@
 
 llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const float   kq_scale    = 1.0f / sqrtf(float(n_embd_head));
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * inpL;
     ggml_tensor * cur;

src/models/cohere2-iswa.cpp

@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     const float f_logit_scale = hparams.f_logit_scale;
 

src/models/command-r.cpp

@@ -4,9 +4,9 @@
 
 llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     const float f_logit_scale = hparams.f_logit_scale;
 

src/models/dbrx.cpp

@@ -1,11 +1,11 @@
 #include "models.h"
 
 llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/deci.cpp

@@ -3,10 +3,10 @@
 
 
 llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/deepseek.cpp

@@ -2,10 +2,10 @@
 
 llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/deepseek2.cpp

@@ -8,7 +8,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
     const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
     const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
 
-    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
     const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
 
     const uint32_t kv_lora_rank = hparams.n_lora_kv;

src/models/dots1.cpp

@@ -2,10 +2,10 @@
 
 llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/dream.cpp

@@ -5,10 +5,10 @@
 llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     //copied from qwen2
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/ernie4-5-moe.cpp

@@ -2,10 +2,10 @@
 
 llm_build_ernie4_5_moe::llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/ernie4-5.cpp

@@ -2,10 +2,10 @@
 
 llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/eurobert.cpp

@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_eurobert::llm_build_eurobert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/exaone-moe.cpp

@@ -2,10 +2,10 @@
 
 llm_build_exaone_moe::llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/exaone.cpp

@@ -4,10 +4,10 @@
 
 llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/exaone4.cpp

@@ -4,10 +4,10 @@
 template <bool iswa>
 llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/falcon-h1.cpp

@@ -2,7 +2,7 @@
 
 llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/falcon.cpp

@@ -2,11 +2,11 @@
 
 
 llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/gemma-embedding.cpp

@@ -2,7 +2,7 @@
 
 llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/gemma.cpp

@@ -2,7 +2,7 @@
 
 
 llm_build_gemma::llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/gemma2-iswa.cpp

@@ -1,7 +1,7 @@
 #include "models.h"
 
 llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/gemma3.cpp

@@ -2,7 +2,7 @@
 
 template <bool iswa>
 llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/gemma3n-iswa.cpp

@@ -3,7 +3,7 @@
 llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model),
-    n_embd_head(model.hparams.n_embd_head_k),
+    n_embd_head(model.hparams.n_embd_head_k()),
     n_embd_altup(model.hparams.n_embd_altup),
     n_altup(model.hparams.n_altup),
     i_altup_act(model.hparams.i_altup_act) {

src/models/glm4-moe.cpp

@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);

src/models/glm4.cpp

@@ -3,10 +3,10 @@
 
 
 llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);

src/models/gpt2.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_gpt2::llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * pos;

src/models/gptneox.cpp

@@ -2,10 +2,10 @@
 
 
 llm_build_gptneox::llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/granite-hybrid.cpp

@@ -2,8 +2,8 @@
 
 llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/granite.cpp

@@ -5,10 +5,10 @@ llm_build_granite::llm_build_granite(
     const llm_graph_params & params)
     : llm_graph_context(params) {
 
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/grok.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_grok::llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/grovemoe.cpp

@@ -2,11 +2,11 @@
 
 llm_build_grovemoe::llm_build_grovemoe(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head    = hparams.n_embd_head_v;
+    const int64_t n_embd_head    = hparams.n_embd_head_v();
     const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/hunyuan-dense.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_hunyuan_dense::llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/hunyuan-moe.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_hunyuan_moe::llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/internlm2.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_internlm2::llm_build_internlm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/jais.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_jais::llm_build_jais(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/jais2.cpp

@@ -3,10 +3,10 @@
 // JAIS-2 model graph builder
 // Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings
 llm_build_jais2::llm_build_jais2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/jamba.cpp

@@ -1,7 +1,7 @@
 #include "models.h"
 
 llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/kimi-linear.cpp

@@ -102,7 +102,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
     const int64_t kv_lora_rank = hparams.n_lora_kv;
     // qk_rope_head_dim = 64 (from Kimi config) which is hparams.n_rot
     // Confirmed from tensor shape: wkv_a_mqa [2304, 576] = [n_embd, kv_lora_rank + qk_rope_head_dim]
-    const int64_t n_embd_head_qk_rope = hparams.n_rot;  // config.qk_rope_head_dim
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();  // config.qk_rope_head_dim
     const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;  // 192 - 64 = 128
     // Attention scale for MLA
     const float kq_scale_mla = 1.0f / sqrtf((float)n_embd_head_k_mla);

src/models/lfm2.cpp

@@ -39,7 +39,7 @@ llm_build_lfm2<iswa>::llm_build_lfm2(const llama_model & model, const llm_graph_
                                            inp_attn_type * inp_attn,
                                            int             il) -> ggml_tensor * {
         GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
-        const auto n_embd_head = hparams.n_embd_head_v;
+        const auto n_embd_head = hparams.n_embd_head_v();
         const auto n_head_kv   = hparams.n_head_kv(il);
 
         auto * q = build_lora_mm(model.layers[il].wq, cur);

src/models/llada-moe.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_llada_moe::llm_build_llada_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/llada.cpp

@@ -2,10 +2,10 @@
 
 llm_build_llada::llm_build_llada(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     // LLaDA is similar to LLaMA but uses non-causal attention for diffusion
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/llama-iswa.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/llama.cpp

@@ -2,10 +2,10 @@
 
 template <bool embed>
 llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/maincoder.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_maincoder::llm_build_maincoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/minicpm3.cpp

@@ -5,10 +5,10 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
     const int64_t n_embd_base = 256;
     const float scale_embd  = 12.0f;
     const float scale_depth = 1.4f;
-    const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k));
+    const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k()));
 
-    const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+    const uint32_t n_embd_head_qk_rope = hparams.n_rot();
+    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
 
     const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
@@ -51,21 +51,21 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
                     LLM_NORM_RMS, il);
             cb(q, "q", il);
 
-            // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
+            // {q_lora_rank, n_head * hparams.n_embd_head_k()} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k(), n_tokens}
             q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
             cb(q, "q", il);
 
             // split into {n_head * n_embd_head_qk_nope, n_tokens}
             ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k),
-                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, hparams.n_embd_head_k()),
+                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
                     0);
             cb(q_nope, "q_nope", il);
 
             // and {n_head * n_embd_head_qk_rope, n_tokens}
             ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k),
-                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, hparams.n_embd_head_k()),
+                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
                     ggml_row_size(q->type, n_embd_head_qk_nope));
             cb(q_pe, "q_pe", il);
 
@@ -97,15 +97,15 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
 
             // split into {n_head * n_embd_head_qk_nope, n_tokens}
             ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v()),
+                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v())),
                     0);
             cb(k_nope, "k_nope", il);
 
             // and {n_head * n_embd_head_v, n_tokens}
-            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v(), n_head, n_tokens,
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())*n_head),
                     ggml_row_size(kv->type, (n_embd_head_qk_nope)));
             cb(v_states, "v_states", il);
 

src/models/minimax-m2.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_minimax_m2::llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    // GGML_ASSERT(n_embd_head == hparams.n_rot); this is wrong in case of minimax, head_dim = 128, n_rot = 64
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    // GGML_ASSERT(n_embd_head == n_rot); this is wrong in case of minimax, head_dim = 128, n_rot = 64
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/mistral3.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_mistral3::llm_build_mistral3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/modern-bert.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_modern_bert::llm_build_modern_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/mpt.cpp

@@ -3,10 +3,10 @@
 
 
 llm_build_mpt::llm_build_mpt(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * pos;

src/models/nemotron-h.cpp

@@ -2,8 +2,8 @@
 
 llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/nemotron.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_nemotron::llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    //GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    //GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/neo-bert.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_neo_bert::llm_build_neo_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/olmo.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_olmo::llm_build_olmo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/olmo2.cpp

@@ -2,10 +2,10 @@
 
 template <bool iswa>
 llm_build_olmo2<iswa>::llm_build_olmo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/olmoe.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_olmoe::llm_build_olmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/openelm.cpp

@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_openelm::llm_build_openelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/orion.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_orion::llm_build_orion(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/paddleocr.cpp

@@ -5,10 +5,10 @@ llm_build_paddleocr::llm_build_paddleocr(const llama_model & model, const llm_gr
 
     // NOTE: same with qwen2vl.cpp, but bias tensors are optional
 
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/pangu-embedded.cpp

@@ -2,10 +2,10 @@
 
 
 llm_build_pangu_embedded::llm_build_pangu_embedded(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

src/models/phi2.cpp

@@ -2,10 +2,10 @@
 
 
 llm_build_phi2::llm_build_phi2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * attn_norm_output;