vulkan : fix FA mask load with bounds check (coopmat2) (#17606)

* git mv

* add server-context.h

* add server-context.h

* clean up headers

* cont : cleanup

* also expose server_response_reader (to be used by CLI)

* fix windows build

* decouple server_routes and server_http

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.devops/vulkan.Dockerfile

@@ -50,6 +50,7 @@ WORKDIR /app
 
 RUN apt-get update \
     && apt-get install -y \
+    build-essential \
     git \
     python3 \
     python3-pip \

CODEOWNERS

@@ -2,10 +2,8 @@
 # multiplie collaborators per item can be specified
 
 /.devops/*.Dockerfile                   @ngxson
-/.github/actions/                       @slaren @CISC
+/.github/actions/                       @CISC
 /.github/workflows/                     @CISC
-/.github/workflows/release.yml          @slaren
-/.github/workflows/winget.yml           @slaren
 /ci/                                    @ggerganov
 /cmake/                                 @ggerganov
 /common/CMakeLists.txt                  @ggerganov
@@ -40,21 +38,14 @@
 /examples/passkey/                      @ggerganov
 /examples/retrieval/                    @ggerganov
 /examples/save-load-state/              @ggerganov
-/examples/simple-chat/                  @slaren
-/examples/simple/                       @slaren
 /examples/speculative-simple/           @ggerganov
 /examples/speculative/                  @ggerganov
 /ggml/cmake/                            @ggerganov
-/ggml/include/                          @ggerganov @slaren
-/ggml/src/ggml-alloc.c                  @slaren
-/ggml/src/ggml-backend*                 @slaren
-/ggml/src/ggml-blas/                    @slaren
-/ggml/src/ggml-common.h                 @ggerganov @slaren
-/ggml/src/ggml-cpu/                     @ggerganov @slaren
+/ggml/include/                          @ggerganov
+/ggml/src/ggml-common.h                 @ggerganov
+/ggml/src/ggml-cpu/                     @ggerganov
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
-/ggml/src/ggml-cuda/common.cuh          @slaren
 /ggml/src/ggml-cuda/fattn*              @JohannesGaessler
-/ggml/src/ggml-cuda/ggml-cuda.cu        @slaren
 /ggml/src/ggml-cuda/mmf.*               @JohannesGaessler @am17an
 /ggml/src/ggml-cuda/mmq.*               @JohannesGaessler
 /ggml/src/ggml-cuda/mmvf.*              @JohannesGaessler
@@ -62,19 +53,19 @@
 /ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
 /ggml/src/ggml-hip/                     @IMbackK
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
-/ggml/src/ggml-impl.h                   @ggerganov @slaren
+/ggml/src/ggml-impl.h                   @ggerganov
 /ggml/src/ggml-metal/                   @ggerganov
 /ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
 /ggml/src/ggml-hexagon/                 @max-krasnyansky @lhez
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @rgerganov
-/ggml/src/ggml-threading.*              @ggerganov @slaren
+/ggml/src/ggml-threading.*              @ggerganov
 /ggml/src/ggml-vulkan/                  @0cc4m
 /ggml/src/ggml-webgpu/                  @reeselevine
 /ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
-/ggml/src/ggml.c                        @ggerganov @slaren
-/ggml/src/ggml.cpp                      @ggerganov @slaren
+/ggml/src/ggml.c                        @ggerganov
+/ggml/src/ggml.cpp                      @ggerganov
 /ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky
 /gguf-py/                               @CISC
 /media/                                 @ggerganov
@@ -86,15 +77,11 @@
 /src/llama-arch.*                       @CISC
 /src/llama-chat.*                       @ngxson
 /src/llama-graph.*                      @CISC
-/src/llama-model-loader.*               @slaren
 /src/llama-model.*                      @CISC
 /src/llama-vocab.*                      @CISC
 /src/models/                            @CISC
 /tests/                                 @ggerganov
-/tests/test-backend-ops.cpp             @slaren
-/tests/test-thread-safety.cpp           @slaren
 /tools/batched-bench/                   @ggerganov
-/tools/llama-bench/                     @slaren
 /tools/main/                            @ggerganov
 /tools/mtmd/                            @ngxson
 /tools/perplexity/                      @ggerganov
@@ -106,8 +93,6 @@
 /tools/tokenize/                        @ggerganov
 /tools/tts/                             @ggerganov
 /vendor/                                @ggerganov
-/.clang-format                          @slaren
-/.clang-tidy                            @slaren
 /AUTHORS                                @ggerganov
 /CMakeLists.txt                         @ggerganov
 /CONTRIBUTING.md                        @ggerganov

ci/run.sh

@@ -45,7 +45,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_SCHED_NO_REALLOC=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
@@ -428,10 +428,10 @@ function gg_run_qwen3_0_6b {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -523,8 +523,8 @@ function gg_run_embd_bge_small {
 
     ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
-    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
     set +e
 }
@@ -564,7 +564,7 @@ function gg_run_rerank_tiny {
     model_f16="${path_models}/ggml-model-f16.gguf"
 
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --no-op-offload --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

common/arg.cpp

@@ -694,6 +694,12 @@ static bool is_autoy(const std::string & value) {
 }
 
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
+    // default values specific to example
+    // note: we place it here instead of inside server.cpp to allow llama-gen-docs to pick it up
+    if (ex == LLAMA_EXAMPLE_SERVER) {
+        params.use_jinja = true;
+    }
+
     // load dynamic backends
     ggml_backend_load_all();
 
@@ -1232,6 +1238,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             const auto sampler_names = string_split<std::string>(value, ';');
             params.sampling.samplers = common_sampler_types_from_names(sampler_names, true);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1261,6 +1268,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.sampling.temp = std::stof(value);
             params.sampling.temp = std::max(params.sampling.temp, 0.0f);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1268,6 +1276,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("top-k sampling (default: %d, 0 = disabled)", params.sampling.top_k),
         [](common_params & params, int value) {
             params.sampling.top_k = value;
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1275,6 +1284,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p),
         [](common_params & params, const std::string & value) {
             params.sampling.top_p = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1282,6 +1292,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p),
         [](common_params & params, const std::string & value) {
             params.sampling.min_p = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1296,6 +1307,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
         [](common_params & params, const std::string & value) {
             params.sampling.xtc_probability = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1303,6 +1315,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
         [](common_params & params, const std::string & value) {
             params.sampling.xtc_threshold = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1321,6 +1334,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
             params.sampling.penalty_last_n = value;
             params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1328,6 +1342,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
             params.sampling.penalty_repeat = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1425,6 +1440,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sampling.mirostat),
         [](common_params & params, int value) {
             params.sampling.mirostat = value;
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1432,6 +1448,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta),
         [](common_params & params, const std::string & value) {
             params.sampling.mirostat_eta = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1439,6 +1456,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau),
         [](common_params & params, const std::string & value) {
             params.sampling.mirostat_tau = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -2476,11 +2494,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--jinja"},
-        "use jinja template for chat (default: disabled)",
+        string_format("use jinja template for chat (default: %s)\n", params.use_jinja ? "enabled" : "disabled"),
         [](common_params & params) {
             params.use_jinja = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_JINJA"));
+    add_opt(common_arg(
+        {"--no-jinja"},
+        string_format("disable jinja template for chat (default: %s)\n", params.use_jinja ? "enabled" : "disabled"),
+        [](common_params & params) {
+            params.use_jinja = false;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_NO_JINJA"));
     add_opt(common_arg(
         {"--reasoning-format"}, "FORMAT",
         "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"

common/chat-parser.cpp

@@ -13,6 +13,120 @@
 
 using json = nlohmann::ordered_json;
 
+static void parse_prefixed_json_tool_call_array(common_chat_msg_parser & builder,
+                                                const common_regex &     prefix,
+                                                size_t                   rstrip_prefix = 0) {
+    static const std::vector<std::vector<std::string>> args_paths = { { "arguments" } };
+    if (auto res = builder.try_find_regex(prefix)) {
+        builder.move_back(rstrip_prefix);
+        auto tool_calls = builder.consume_json_with_dumped_args(args_paths);
+        if (!builder.add_tool_calls(tool_calls.value) || tool_calls.is_partial) {
+            throw common_chat_msg_partial_exception("incomplete tool call array");
+        }
+    } else {
+        builder.add_content(builder.consume_rest());
+    }
+}
+
+static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) {
+    std::string arguments;
+    if (builder.is_partial()) {
+        arguments = (json{
+                         { "code", code + builder.healing_marker() }
+        })
+                        .dump();
+        auto idx = arguments.find(builder.healing_marker());
+        if (idx != std::string::npos) {
+            arguments.resize(idx);
+        }
+    } else {
+        arguments = (json{
+                         { "code", code }
+        })
+                        .dump();
+    }
+    return arguments;
+}
+
+/**
+ * Takes a prefix regex that must have 1 group to capture the function name, a closing suffix, and expects json parameters in between.
+ * Aggregates the prefix, suffix and in-between text into the content.
+ */
+static void parse_json_tool_calls(
+    common_chat_msg_parser &            builder,
+    const std::optional<common_regex> & block_open,
+    const std::optional<common_regex> & function_regex_start_only,
+    const std::optional<common_regex> & function_regex,
+    const common_regex &                close_regex,
+    const std::optional<common_regex> & block_close,
+    bool                                allow_raw_python = false,
+    const std::function<std::string(const common_chat_msg_parser::find_regex_result & fres)> & get_function_name =
+        nullptr) {
+    auto parse_tool_calls = [&]() {
+        size_t from  = std::string::npos;
+        auto   first = true;
+        while (true) {
+            auto start_pos = builder.pos();
+            auto res = function_regex_start_only && first ? builder.try_consume_regex(*function_regex_start_only) :
+                       function_regex                     ? builder.try_find_regex(*function_regex, from) :
+                                                            std::nullopt;
+
+            if (res) {
+                std::string name;
+                if (get_function_name) {
+                    name = get_function_name(*res);
+                } else {
+                    GGML_ASSERT(res->groups.size() == 2);
+                    name = builder.str(res->groups[1]);
+                }
+                first = false;
+                if (name.empty()) {
+                    // get_function_name signalled us that we should skip this match and treat it as content.
+                    from = res->groups[0].begin + 1;
+                    continue;
+                }
+                from = std::string::npos;
+
+                auto maybe_raw_python = name == "python" && allow_raw_python;
+                if (builder.input()[builder.pos()] == '{' || !maybe_raw_python) {
+                    if (auto arguments = builder.try_consume_json_with_dumped_args({ {} })) {
+                        if (!builder.add_tool_call(name, "", arguments->value) || arguments->is_partial) {
+                            throw common_chat_msg_partial_exception("incomplete tool call");
+                        }
+                        builder.consume_regex(close_regex);
+                    }
+                    continue;
+                }
+                if (maybe_raw_python) {
+                    auto arguments = wrap_code_as_arguments(builder, builder.consume_rest());
+                    if (!builder.add_tool_call(name, "", arguments)) {
+                        throw common_chat_msg_partial_exception("incomplete tool call");
+                    }
+                    return;
+                }
+                throw common_chat_msg_partial_exception("incomplete tool call");
+            } else {
+                builder.move_to(start_pos);
+            }
+            break;
+        }
+        if (block_close) {
+            builder.consume_regex(*block_close);
+        }
+        builder.consume_spaces();
+        builder.add_content(builder.consume_rest());
+    };
+    if (block_open) {
+        if (auto res = builder.try_find_regex(*block_open)) {
+            parse_tool_calls();
+        } else {
+            builder.add_content(builder.consume_rest());
+        }
+    } else {
+        parse_tool_calls();
+    }
+}
+
 common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax)
     : input_(input), is_partial_(is_partial), syntax_(syntax)
 {
@@ -532,3 +646,857 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
 void common_chat_msg_parser::clear_tools() {
     result_.tool_calls.clear();
 }
+
+/**
+ * All common_chat_parse_* moved from chat.cpp to chat-parser.cpp below
+ * to reduce incremental compile time for parser changes.
+ */
+static void common_chat_parse_generic(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+    static const std::vector<std::vector<std::string>> content_paths = {
+        {"response"},
+    };
+    static const std::vector<std::vector<std::string>> args_paths = {
+        {"tool_call", "arguments"},
+        {"tool_calls", "arguments"},
+    };
+    auto data = builder.consume_json_with_dumped_args(args_paths, content_paths);
+    if (data.value.contains("tool_calls")) {
+        if (!builder.add_tool_calls(data.value.at("tool_calls")) || data.is_partial) {
+            throw common_chat_msg_partial_exception("incomplete tool calls");
+        }
+    } else if (data.value.contains("tool_call")) {
+        if (!builder.add_tool_call(data.value.at("tool_call")) || data.is_partial) {
+            throw common_chat_msg_partial_exception("incomplete tool call");
+        }
+    } else if (data.value.contains("response")) {
+        const auto & response = data.value.at("response");
+        builder.add_content(response.is_string() ? response.template get<std::string>() : response.dump(2));
+        if (data.is_partial) {
+            throw common_chat_msg_partial_exception("incomplete response");
+        }
+    } else {
+        throw common_chat_msg_partial_exception("Expected 'tool_call', 'tool_calls' or 'response' in JSON");
+    }
+}
+
+static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    static const common_regex prefix(regex_escape("[TOOL_CALLS]"));
+    parse_prefixed_json_tool_call_array(builder, prefix);
+}
+
+static void common_chat_parse_magistral(common_chat_msg_parser & builder) {
+    builder.try_parse_reasoning("[THINK]", "[/THINK]");
+
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    static const common_regex prefix(regex_escape("[TOOL_CALLS]"));
+    parse_prefixed_json_tool_call_array(builder, prefix);
+}
+
+static void common_chat_parse_command_r7b(common_chat_msg_parser & builder) {
+    builder.try_parse_reasoning("<|START_THINKING|>", "<|END_THINKING|>");
+
+    static const common_regex start_action_regex("<\\|START_ACTION\\|>");
+    static const common_regex end_action_regex("<\\|END_ACTION\\|>");
+    static const common_regex start_response_regex("<\\|START_RESPONSE\\|>");
+    static const common_regex end_response_regex("<\\|END_RESPONSE\\|>");
+
+    if (auto res = builder.try_find_regex(start_action_regex)) {
+        // If we didn't extract thoughts, prelude includes them.
+        auto tool_calls = builder.consume_json_with_dumped_args({{"parameters"}});
+        for (const auto & tool_call : tool_calls.value) {
+            std::string name = tool_call.contains("tool_name") ? tool_call.at("tool_name") : "";
+            std::string id = tool_call.contains("tool_call_id") ? tool_call.at("tool_call_id") : "";
+            std::string arguments = tool_call.contains("parameters") ? tool_call.at("parameters") : "";
+            if (!builder.add_tool_call(name, id, arguments) || tool_calls.is_partial) {
+                throw common_chat_msg_partial_exception("incomplete tool call");
+            }
+        }
+        if (tool_calls.is_partial) {
+            throw common_chat_msg_partial_exception("incomplete tool call");
+        }
+        builder.consume_regex(end_action_regex);
+    } else if (auto res = builder.try_find_regex(start_response_regex)) {
+        if (!builder.try_find_regex(end_response_regex)) {
+            builder.add_content(builder.consume_rest());
+            throw common_chat_msg_partial_exception(end_response_regex.str());
+        }
+    } else {
+        builder.add_content(builder.consume_rest());
+    }
+}
+
+static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) {
+    builder.try_parse_reasoning("<think>", "</think>");
+
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    static const common_regex function_regex(
+        "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: ");
+    static const common_regex close_regex("\\}\\s*");
+
+    static const common_regex function_name_regex("\\s*(\\w+)\\s*\\.\\s*call\\(");
+    static const common_regex arg_name_regex("\\s*(\\w+)\\s*=\\s*");
+
+    if (with_builtin_tools) {
+        static const common_regex builtin_call_regex("<\\|python_tag\\|>");
+        if (auto res = builder.try_find_regex(builtin_call_regex)) {
+            auto fun_res = builder.consume_regex(function_name_regex);
+            auto function_name = builder.str(fun_res.groups[1]);
+
+            common_healing_marker healing_marker;
+            json args = json::object();
+            while (true) {
+                if (auto arg_res = builder.try_consume_regex(arg_name_regex)) {
+                    auto arg_name = builder.str(arg_res->groups[1]);
+                    auto partial = builder.consume_json();
+                    args[arg_name] = partial.json;
+                    healing_marker.marker = partial.healing_marker.marker;
+                    healing_marker.json_dump_marker = partial.healing_marker.json_dump_marker;
+                    builder.consume_spaces();
+                    if (!builder.try_consume_literal(",")) {
+                        break;
+                    }
+                } else {
+                    break;
+                }
+            }
+            builder.consume_literal(")");
+            builder.consume_spaces();
+
+            auto arguments = args.dump();
+            if (!builder.add_tool_call(function_name, "", arguments)) {
+                throw common_chat_msg_partial_exception("Incomplete tool call");
+            }
+            return;
+        }
+    }
+    parse_json_tool_calls(
+        builder,
+        /* block_open= */ std::nullopt,
+        /* function_regex_start_only= */ function_regex,
+        /* function_regex= */ std::nullopt,
+        close_regex,
+        std::nullopt);
+
+}
+
+static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
+    builder.try_parse_reasoning("<think>", "</think>");
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
+    static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
+    static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?function<｜tool▁sep｜>([^\n]+)\n```json\n");
+    static const common_regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>");
+
+    parse_json_tool_calls(
+        builder,
+        /* block_open= */ tool_calls_begin,
+        /* function_regex_start_only= */ std::nullopt,
+        function_regex,
+        close_regex,
+        tool_calls_end);
+}
+
+static void common_chat_parse_deepseek_v3_1_content(common_chat_msg_parser & builder) {
+    static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?([^\\n<]+)(?:<｜tool▁sep｜>)");
+
+    static const common_regex close_regex("(?:[\\s]*)?<｜tool▁call▁end｜>");
+    static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
+    static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
+
+    if (!builder.syntax().parse_tool_calls) {
+        LOG_DBG("%s: not parse_tool_calls\n", __func__);
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    LOG_DBG("%s: parse_tool_calls\n", __func__);
+
+    parse_json_tool_calls(
+        builder,
+        /* block_open= */ tool_calls_begin,
+        /* function_regex_start_only= */ std::nullopt,
+        function_regex,
+        close_regex,
+        tool_calls_end);
+}
+
+static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) {
+    // DeepSeek V3.1 outputs reasoning content between "<think>" and "</think>" tags, followed by regular content
+    // First try to parse using the standard reasoning parsing method
+    LOG_DBG("%s: thinking_forced_open: %s\n", __func__, std::to_string(builder.syntax().thinking_forced_open).c_str());
+
+    auto start_pos = builder.pos();
+    auto found_end_think = builder.try_find_literal("</think>");
+    builder.move_to(start_pos);
+
+    if (builder.syntax().thinking_forced_open && !builder.is_partial() && !found_end_think) {
+        LOG_DBG("%s: no end_think, not partial, adding content\n", __func__);
+        common_chat_parse_deepseek_v3_1_content(builder);
+    } else if (builder.try_parse_reasoning("<think>", "</think>")) {
+        // If reasoning was parsed successfully, the remaining content is regular content
+        LOG_DBG("%s: parsed reasoning, adding content\n", __func__);
+        // </think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>NAME\n```json\nJSON\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>
+        common_chat_parse_deepseek_v3_1_content(builder);
+    } else {
+        if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE) {
+          LOG_DBG("%s: reasoning_format none, adding content\n", __func__);
+          common_chat_parse_deepseek_v3_1_content(builder);
+          return;
+        }
+        // If no reasoning tags found, check if we should treat everything as reasoning
+        if (builder.syntax().thinking_forced_open) {
+            // If thinking is forced open but no tags found, treat everything as reasoning
+            LOG_DBG("%s: thinking_forced_open, adding reasoning content\n", __func__);
+            builder.add_reasoning_content(builder.consume_rest());
+        } else {
+            LOG_DBG("%s: no thinking_forced_open, adding content\n", __func__);
+            // <｜tool▁call▁begin｜>NAME<｜tool▁sep｜>JSON<｜tool▁call▁end｜>
+            common_chat_parse_deepseek_v3_1_content(builder);
+        }
+    }
+}
+
+static void common_chat_parse_minimax_m2(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form {
+        /* form.scope_start = */ "<minimax:tool_call>",
+        /* form.tool_start  = */ "<invoke name=\"",
+        /* form.tool_sep    = */ "\">",
+        /* form.key_start   = */ "<parameter name=\"",
+        /* form.key_val_sep = */ "\">",
+        /* form.val_end     = */ "</parameter>",
+        /* form.tool_end    = */ "</invoke>",
+        /* form.scope_end   = */ "</minimax:tool_call>",
+    };
+    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
+}
+
+static void common_chat_parse_qwen3_coder_xml(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "<tool_call>";
+        form.tool_start  = "<function=";
+        form.tool_sep    = ">";
+        form.key_start   = "<parameter=";
+        form.key_val_sep = ">";
+        form.val_end     = "</parameter>";
+        form.tool_end    = "</function>";
+        form.scope_end   = "</tool_call>";
+        form.trim_raw_argval = true;
+        return form;
+    })();
+    builder.consume_reasoning_with_xml_tool_calls(form);
+}
+
+static void common_chat_parse_kimi_k2(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "<|tool_calls_section_begin|>";
+        form.tool_start  = "<|tool_call_begin|>";
+        form.tool_sep    = "<|tool_call_argument_begin|>{";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}<|tool_call_end|>";
+        form.scope_end   = "<|tool_calls_section_end|>";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        return form;
+    })();
+    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
+}
+
+static void common_chat_parse_apriel_1_5(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "<tool_calls>[";
+        form.tool_start  = "{\"name\": \"";
+        form.tool_sep    = "\", \"arguments\": {";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}, ";
+        form.scope_end   = "]</tool_calls>";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        form.last_tool_end = "}";
+        return form;
+    })();
+    builder.consume_reasoning_with_xml_tool_calls(form, "<thinking>", "</thinking>");
+}
+
+static void common_chat_parse_xiaomi_mimo(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "";
+        form.tool_start  = "<tool_call>\n{\"name\": \"";
+        form.tool_sep    = "\", \"arguments\": {";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}\n</tool_call>";
+        form.scope_end   = "";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        return form;
+    })();
+    builder.consume_reasoning_with_xml_tool_calls(form);
+}
+
+static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
+    static const std::string constraint = "(?: (<\\|constrain\\|>)?([a-zA-Z0-9_-]+))";
+    static const std::string recipient("(?: to=functions\\.([^<\\s]+))");
+
+    static const common_regex start_regex("<\\|start\\|>assistant");
+    static const common_regex analysis_regex("<\\|channel\\|>analysis");
+    static const common_regex final_regex("<\\|channel\\|>final" + constraint + "?");
+    static const common_regex preamble_regex("<\\|channel\\|>commentary");
+    static const common_regex tool_call1_regex(recipient + "<\\|channel\\|>(analysis|commentary)" + constraint + "?");
+    static const common_regex tool_call2_regex("<\\|channel\\|>(analysis|commentary)" + recipient + constraint + "?");
+
+    auto consume_end = [&](bool include_end = false) {
+        if (auto res = builder.try_find_literal("<|end|>")) {
+            return res->prelude + (include_end ? builder.str(res->groups[0]) : "");
+        }
+        return builder.consume_rest();
+    };
+
+    auto handle_tool_call = [&](const std::string & name) {
+        if (auto args = builder.try_consume_json_with_dumped_args({{}})) {
+            if (builder.syntax().parse_tool_calls) {
+                if (!builder.add_tool_call(name, "", args->value) || args->is_partial) {
+                    throw common_chat_msg_partial_exception("incomplete tool call");
+                }
+            } else if (args->is_partial) {
+                throw common_chat_msg_partial_exception("incomplete tool call");
+            }
+        }
+    };
+
+    auto regex_match = [](const common_regex & regex, const std::string & input) -> std::optional<common_regex_match> {
+        auto match = regex.search(input, 0, true);
+        if (match.type == COMMON_REGEX_MATCH_TYPE_FULL) {
+            return match;
+        }
+        return std::nullopt;
+    };
+
+    do {
+        auto header_start_pos = builder.pos();
+        auto content_start = builder.try_find_literal("<|message|>");
+        if (!content_start) {
+            throw common_chat_msg_partial_exception("incomplete header");
+        }
+
+        auto header = content_start->prelude;
+
+        if (auto match = regex_match(tool_call1_regex, header)) {
+            auto group = match->groups[1];
+            auto name = header.substr(group.begin, group.end - group.begin);
+            handle_tool_call(name);
+            continue;
+        }
+
+        if (auto match = regex_match(tool_call2_regex, header)) {
+            auto group = match->groups[2];
+            auto name = header.substr(group.begin, group.end - group.begin);
+            handle_tool_call(name);
+            continue;
+        }
+
+        if (regex_match(analysis_regex, header)) {
+            builder.move_to(header_start_pos);
+            if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) {
+                builder.add_content(consume_end(true));
+            } else {
+                builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|end|>");
+            }
+            continue;
+        }
+
+        if(regex_match(final_regex, header) || regex_match(preamble_regex, header)) {
+            builder.add_content(consume_end());
+            continue;
+        }
+
+        // Possibly a malformed message, attempt to recover by rolling
+        // back to pick up the next <|start|>
+        LOG_DBG("%s: unknown header from message: %s\n", __func__, header.c_str());
+        builder.move_to(header_start_pos);
+    } while (builder.try_find_regex(start_regex, std::string::npos, false));
+
+    auto remaining = builder.consume_rest();
+    if (!remaining.empty()) {
+        LOG_DBG("%s: content after last message: %s\n", __func__, remaining.c_str());
+    }
+}
+
+static void common_chat_parse_glm_4_5(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form {
+        /* form.scope_start  = */ "",
+        /* form.tool_start   = */ "<tool_call>",
+        /* form.tool_sep     = */ "",
+        /* form.key_start    = */ "<arg_key>",
+        /* form.key_val_sep  = */ "</arg_key>",
+        /* form.val_end      = */ "</arg_value>",
+        /* form.tool_end     = */ "</tool_call>",
+        /* form.scope_end    = */ "",
+        /* form.key_val_sep2 = */ "<arg_value>",
+    };
+    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
+}
+
+static void common_chat_parse_firefunction_v2(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+    static const common_regex prefix(regex_escape(" functools["));
+    parse_prefixed_json_tool_call_array(builder, prefix, /* rstrip_prefix= */ 1);
+}
+
+static void common_chat_parse_functionary_v3_2(common_chat_msg_parser & builder) {
+    static const common_regex function_regex_start_only(R"((\w+\n\{|python\n|all\n))");
+    static const common_regex function_regex(R"(>>>(\w+\n\{|python\n|all\n))");
+    static const common_regex close_regex(R"(\s*)");
+
+    parse_json_tool_calls(
+        builder,
+        std::nullopt,
+        function_regex_start_only,
+        function_regex,
+        close_regex,
+        std::nullopt,
+        /* allow_raw_python= */ true,
+        /* get_function_name= */ [&](const auto & res) -> std::string {
+            auto at_start = res.groups[0].begin == 0;
+            auto name = builder.str(res.groups[1]);
+            if (!name.empty() && name.back() == '{') {
+                // Unconsume the opening brace '{' to ensure the JSON parsing goes well.
+                builder.move_back(1);
+            }
+            auto idx = name.find_last_not_of("\n{");
+            name = name.substr(0, idx + 1);
+            if (at_start && name == "all") {
+                return "";
+            }
+            return name;
+        });
+}
+
+static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+    // This version of Functionary still supports the llama 3.1 tool call format for the python tool.
+    static const common_regex python_tag_regex(regex_escape("<|python_tag|>"));
+
+    static const common_regex function_regex(R"(<function=(\w+)>)");
+    static const common_regex close_regex(R"(</function>)");
+
+    parse_json_tool_calls(
+        builder,
+        /* block_open= */ std::nullopt,
+        /* function_regex_start_only= */ std::nullopt,
+        function_regex,
+        close_regex,
+        std::nullopt);
+
+    if (auto res = builder.try_find_regex(python_tag_regex)) {
+        auto arguments = wrap_code_as_arguments(builder, builder.consume_rest());
+        builder.add_tool_call("python", "", arguments);
+        return;
+    }
+}
+
+static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
+    builder.try_parse_reasoning("<think>", "</think>");
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    static const common_regex open_regex(
+        "(?:"
+            "(```(?:xml|json)?\\n\\s*)?" // match 1 (block_start)
+            "("                          // match 2 (open_tag)
+                "<tool_call>"
+                "|<function_call>"
+                "|<tool>"
+                "|<tools>"
+                "|<response>"
+                "|<json>"
+                "|<xml>"
+                "|<JSON>"
+            ")?"
+            "(\\s*\\{\\s*\"name\")" // match 3 (named tool call)
+        ")"
+        "|<function=([^>]+)>"            // match 4 (function name)
+        "|<function name=\"([^\"]+)\">"  // match 5 (function name again)
+    );
+
+    while (auto res = builder.try_find_regex(open_regex)) {
+        const auto & block_start = res->groups[1];
+        std::string block_end = block_start.empty() ? "" : "```";
+
+        const auto & open_tag = res->groups[2];
+        std::string close_tag;
+
+        if (!res->groups[3].empty()) {
+            builder.move_to(res->groups[3].begin);
+            close_tag = open_tag.empty() ? "" : "</" + builder.str(open_tag).substr(1);
+
+            if (auto tool_call = builder.try_consume_json_with_dumped_args({{"arguments"}})) {
+                if (!builder.add_tool_call(tool_call->value) || tool_call->is_partial) {
+                    throw common_chat_msg_partial_exception("incomplete tool call");
+                }
+                builder.consume_spaces();
+                builder.consume_literal(close_tag);
+                builder.consume_spaces();
+                if (!block_end.empty()) {
+                    builder.consume_literal(block_end);
+                    builder.consume_spaces();
+                }
+            } else {
+                throw common_chat_msg_partial_exception("failed to parse tool call");
+            }
+        } else {
+            auto function_name = builder.str(res->groups[4]);
+            if (function_name.empty()) {
+                function_name = builder.str(res->groups[5]);
+            }
+            GGML_ASSERT(!function_name.empty());
+
+            close_tag = "</function>";
+
+            if (auto arguments = builder.try_consume_json_with_dumped_args({{}})) {
+                if (!builder.add_tool_call(function_name, "", arguments->value) || arguments->is_partial) {
+                    throw common_chat_msg_partial_exception("incomplete tool call");
+                }
+                builder.consume_spaces();
+                builder.consume_literal(close_tag);
+                builder.consume_spaces();
+                if (!block_end.empty()) {
+                    builder.consume_literal(block_end);
+                    builder.consume_spaces();
+                }
+            }
+        }
+    }
+
+    builder.add_content(builder.consume_rest());
+}
+
+static void common_chat_parse_granite(common_chat_msg_parser & builder) {
+    // Parse thinking tags
+    static const common_regex start_think_regex(regex_escape("<think>"));
+    static const common_regex end_think_regex(regex_escape("</think>"));
+    // Granite models output partial tokens such as "<" and "<think".
+    // By leveraging try_consume_regex()/try_find_regex() throwing
+    // common_chat_msg_partial_exception for these partial tokens,
+    // processing is interrupted and the tokens are not passed to add_content().
+    if (auto res = builder.try_consume_regex(start_think_regex)) {
+        // Restore position for try_parse_reasoning()
+        builder.move_to(res->groups[0].begin);
+        builder.try_find_regex(end_think_regex, std::string::npos, false);
+        // Restore position for try_parse_reasoning()
+        builder.move_to(res->groups[0].begin);
+    }
+    builder.try_parse_reasoning("<think>", "</think>");
+
+    // Parse response tags
+    static const common_regex start_response_regex(regex_escape("<response>"));
+    static const common_regex end_response_regex(regex_escape("</response>"));
+    // Granite models output partial tokens such as "<" and "<response".
+    // Same hack as reasoning parsing.
+    if (builder.try_consume_regex(start_response_regex)) {
+        builder.try_find_regex(end_response_regex);
+    }
+
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    // Look for tool calls
+    static const common_regex tool_call_regex(regex_escape("<|tool_call|>"));
+    if (auto res = builder.try_find_regex(tool_call_regex)) {
+        builder.move_to(res->groups[0].end);
+
+        // Expect JSON array of tool calls
+        if (auto tool_call = builder.try_consume_json_with_dumped_args({{{"arguments"}}})) {
+            if (!builder.add_tool_calls(tool_call->value) || tool_call->is_partial) {
+                throw common_chat_msg_partial_exception("incomplete tool call");
+            }
+        }
+    } else {
+        builder.add_content(builder.consume_rest());
+    }
+}
+
+static void common_chat_parse_nemotron_v2(common_chat_msg_parser & builder) {
+    // Parse thinking tags
+    builder.try_parse_reasoning("<think>", "</think>");
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    // Look for tool calls
+    static const common_regex tool_call_regex(regex_escape("<TOOLCALL>"));
+    if (auto res = builder.try_find_regex(tool_call_regex)) {
+        builder.move_to(res->groups[0].end);
+
+        // Expect JSON array of tool calls
+        auto tool_calls_data = builder.consume_json();
+        if (tool_calls_data.json.is_array()) {
+            if (!builder.try_consume_literal("</TOOLCALL>")) {
+                throw common_chat_msg_partial_exception("Incomplete tool call");
+            }
+            builder.add_tool_calls(tool_calls_data.json);
+        } else {
+            throw common_chat_msg_partial_exception("Incomplete tool call");
+        }
+    }
+    builder.add_content(builder.consume_rest());
+}
+
+static void common_chat_parse_apertus(common_chat_msg_parser & builder) {
+    // Parse thinking tags
+    builder.try_parse_reasoning("<|inner_prefix|>", "<|inner_suffix|>");
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    // Look for tool calls
+    static const common_regex tool_call_regex(regex_escape("<|tools_prefix|>"));
+    if (auto res = builder.try_find_regex(tool_call_regex)) {
+        builder.move_to(res->groups[0].end);
+
+        auto tool_calls_data = builder.consume_json();
+        if (tool_calls_data.json.is_array()) {
+            builder.consume_spaces();
+            if (!builder.try_consume_literal("<|tools_suffix|>")) {
+                throw common_chat_msg_partial_exception("Incomplete tool call");
+            }
+            for (const auto & value : tool_calls_data.json) {
+                if (value.is_object()) {
+                    builder.add_tool_call_short_form(value);
+                }
+            }
+        } else {
+            throw common_chat_msg_partial_exception("Incomplete tool call");
+        }
+    }
+    builder.add_content(builder.consume_rest());
+}
+
+
+static void common_chat_parse_lfm2(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    // LFM2 format: <|tool_call_start|>[{"name": "get_current_time", "arguments": {"location": "Paris"}}]<|tool_call_end|>
+    static const common_regex tool_call_start_regex(regex_escape("<|tool_call_start|>"));
+    static const common_regex tool_call_end_regex(regex_escape("<|tool_call_end|>"));
+
+    // Loop through all tool calls
+    while (auto res = builder.try_find_regex(tool_call_start_regex, std::string::npos, /* add_prelude_to_content= */ true)) {
+        builder.move_to(res->groups[0].end);
+
+        // Parse JSON array format: [{"name": "...", "arguments": {...}}]
+        auto tool_calls_data = builder.consume_json();
+
+        // Consume end marker
+        builder.consume_spaces();
+        if (!builder.try_consume_regex(tool_call_end_regex)) {
+            throw common_chat_msg_partial_exception("Expected <|tool_call_end|>");
+        }
+
+        // Process each tool call in the array
+        if (tool_calls_data.json.is_array()) {
+            for (const auto & tool_call : tool_calls_data.json) {
+                if (!tool_call.is_object()) {
+                    throw common_chat_msg_partial_exception("Tool call must be an object");
+                }
+
+                if (!tool_call.contains("name")) {
+                    throw common_chat_msg_partial_exception("Tool call missing 'name' field");
+                }
+
+                std::string function_name = tool_call.at("name");
+                std::string arguments = "{}";
+
+                if (tool_call.contains("arguments")) {
+                    if (tool_call.at("arguments").is_object()) {
+                        arguments = tool_call.at("arguments").dump();
+                    } else if (tool_call.at("arguments").is_string()) {
+                        arguments = tool_call.at("arguments");
+                    }
+                }
+
+                if (!builder.add_tool_call(function_name, "", arguments)) {
+                    throw common_chat_msg_partial_exception("Incomplete tool call");
+                }
+            }
+        } else {
+            throw common_chat_msg_partial_exception("Expected JSON array for tool calls");
+        }
+
+        // Consume any trailing whitespace after this tool call
+        builder.consume_spaces();
+    }
+
+    // Consume any remaining content after all tool calls
+    auto remaining = builder.consume_rest();
+    if (!string_strip(remaining).empty()) {
+        builder.add_content(remaining);
+    }
+}
+
+static void common_chat_parse_seed_oss(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form {
+        /* form.scope_start = */ "<seed:tool_call>",
+        /* form.tool_start  = */ "<function=",
+        /* form.tool_sep    = */ ">",
+        /* form.key_start   = */ "<parameter=",
+        /* form.key_val_sep = */ ">",
+        /* form.val_end     = */ "</parameter>",
+        /* form.tool_end    = */ "</function>",
+        /* form.scope_end   = */ "</seed:tool_call>",
+    };
+    builder.consume_reasoning_with_xml_tool_calls(form, "<seed:think>", "</seed:think>");
+}
+
+static void common_chat_parse_content_only(common_chat_msg_parser & builder) {
+    builder.try_parse_reasoning("<think>", "</think>");
+    builder.add_content(builder.consume_rest());
+}
+
+static void common_chat_parse(common_chat_msg_parser & builder) {
+    LOG_DBG("Parsing input with format %s: %s\n", common_chat_format_name(builder.syntax().format), builder.input().c_str());
+
+    switch (builder.syntax().format) {
+        case COMMON_CHAT_FORMAT_CONTENT_ONLY:
+            common_chat_parse_content_only(builder);
+            break;
+        case COMMON_CHAT_FORMAT_GENERIC:
+            common_chat_parse_generic(builder);
+            break;
+        case COMMON_CHAT_FORMAT_MISTRAL_NEMO:
+            common_chat_parse_mistral_nemo(builder);
+            break;
+        case COMMON_CHAT_FORMAT_MAGISTRAL:
+            common_chat_parse_magistral(builder);
+            break;
+        case COMMON_CHAT_FORMAT_LLAMA_3_X:
+            common_chat_parse_llama_3_1(builder);
+            break;
+        case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS:
+            common_chat_parse_llama_3_1(builder, /* with_builtin_tools= */ true);
+            break;
+        case COMMON_CHAT_FORMAT_DEEPSEEK_R1:
+            common_chat_parse_deepseek_r1(builder);
+            break;
+        case COMMON_CHAT_FORMAT_DEEPSEEK_V3_1:
+            common_chat_parse_deepseek_v3_1(builder);
+            break;
+        case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2:
+            common_chat_parse_functionary_v3_2(builder);
+            break;
+        case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1:
+            common_chat_parse_functionary_v3_1_llama_3_1(builder);
+            break;
+        case COMMON_CHAT_FORMAT_HERMES_2_PRO:
+            common_chat_parse_hermes_2_pro(builder);
+            break;
+        case COMMON_CHAT_FORMAT_FIREFUNCTION_V2:
+            common_chat_parse_firefunction_v2(builder);
+            break;
+        case COMMON_CHAT_FORMAT_COMMAND_R7B:
+            common_chat_parse_command_r7b(builder);
+            break;
+        case COMMON_CHAT_FORMAT_GRANITE:
+            common_chat_parse_granite(builder);
+            break;
+        case COMMON_CHAT_FORMAT_GPT_OSS:
+            common_chat_parse_gpt_oss(builder);
+            break;
+        case COMMON_CHAT_FORMAT_SEED_OSS:
+            common_chat_parse_seed_oss(builder);
+            break;
+        case COMMON_CHAT_FORMAT_NEMOTRON_V2:
+            common_chat_parse_nemotron_v2(builder);
+            break;
+        case COMMON_CHAT_FORMAT_APERTUS:
+            common_chat_parse_apertus(builder);
+            break;
+        case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS:
+            common_chat_parse_lfm2(builder);
+            break;
+        case COMMON_CHAT_FORMAT_MINIMAX_M2:
+            common_chat_parse_minimax_m2(builder);
+            break;
+        case COMMON_CHAT_FORMAT_GLM_4_5:
+            common_chat_parse_glm_4_5(builder);
+            break;
+        case COMMON_CHAT_FORMAT_KIMI_K2:
+            common_chat_parse_kimi_k2(builder);
+            break;
+        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML:
+            common_chat_parse_qwen3_coder_xml(builder);
+            break;
+        case COMMON_CHAT_FORMAT_APRIEL_1_5:
+            common_chat_parse_apriel_1_5(builder);
+            break;
+        case COMMON_CHAT_FORMAT_XIAOMI_MIMO:
+            common_chat_parse_xiaomi_mimo(builder);
+            break;
+        default:
+            throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
+    }
+    builder.finish();
+}
+
+common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax) {
+    common_chat_msg_parser builder(input, is_partial, syntax);
+    try {
+        common_chat_parse(builder);
+    } catch (const common_chat_msg_partial_exception & ex) {
+        LOG_DBG("Partial parse: %s\n", ex.what());
+        if (!is_partial) {
+            builder.clear_tools();
+            builder.move_to(0);
+            common_chat_parse_content_only(builder);
+        }
+    }
+    auto msg = builder.result();
+    if (!is_partial) {
+        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
+    }
+    return msg;
+}

common/common.cpp

@@ -8,6 +8,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "sampling.h"
 
 #include <algorithm>
 #include <cinttypes>
@@ -949,6 +950,58 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
 // Model utils
 //
 
+static inline void common_init_sampler_from_model(
+    const llama_model * model,
+    common_params_sampling & sparams) {
+
+    const uint64_t config = sparams.user_sampling_config;
+
+    auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
+        if (config & user_config) return;
+
+        char buf[64] = {0};
+        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
+            char * end = nullptr;
+            int32_t v = strtol(buf, &end, 10);
+            if (end && end != buf) dst = v;
+        }
+    };
+
+    auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
+        if (config & user_config) return;
+
+        char buf[128] = {0};
+        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
+            char * end = nullptr;
+            float v = strtof(buf, &end);
+            if (end && end != buf) dst = v;
+        }
+    };
+
+    // Sampling sequence
+    if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
+        char buf[512] = {0};
+        if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
+            const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
+            if (!sampler_names.empty()) {
+                sparams.samplers = common_sampler_types_from_names(sampler_names, true);
+            }
+        }
+    }
+
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K),           sparams.top_k,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P),           sparams.top_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P),           sparams.min_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD),   sparams.xtc_threshold,   common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP),            sparams.temp,            common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N),  sparams.penalty_last_n,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT),  sparams.penalty_repeat,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT),        sparams.mirostat,        common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU),    sparams.mirostat_tau,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
+}
+
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
@@ -960,6 +1013,8 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
+    common_init_sampler_from_model(model, params.sampling);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     auto cparams = common_context_params_to_llama(params);

common/common.h

@@ -140,6 +140,22 @@ struct common_grammar_trigger {
     llama_token token = LLAMA_TOKEN_NULL;
 };
 
+enum common_params_sampling_config : uint64_t {
+    COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS        = 1 << 0,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_K           = 1 << 1,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_P           = 1 << 2,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIN_P           = 1 << 3,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD   = 1 << 5,
+    COMMON_PARAMS_SAMPLING_CONFIG_TEMP            = 1 << 6,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N  = 1 << 7,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT  = 1 << 8,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT        = 1 << 9,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU    = 1 << 10,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
+};
+
+
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -172,6 +188,8 @@ struct common_params_sampling {
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
 
+    uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
+
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
 

common/json-schema-to-grammar.cpp

@@ -268,10 +268,10 @@ static bool is_reserved_name(const std::string & name) {
 }
 
 std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
-std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
 std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
 std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
-    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };
 
 std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};

convert_hf_to_gguf.py

@@ -565,7 +565,7 @@ class ModelBase:
                             gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
                         )
                     )
-                    or not new_name.endswith(".weight")
+                    or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
                 ):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
@@ -4183,6 +4183,51 @@ class Qwen3MoeModel(Qwen2MoeModel):
         super().set_vocab()
 
 
+@ModelBase.register("Qwen3NextForCausalLM")
+class Qwen3NextModel(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3NEXT
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_ssm_conv_kernel(self.hparams["linear_conv_kernel_dim"])
+        self.gguf_writer.add_ssm_state_size(self.hparams["linear_key_head_dim"])
+        self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"])
+        self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"])
+        self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"])
+        if (rope_dim := self.hparams.get("head_dim")) is None:
+            rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25)))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("mtp"):
+            return [] # ignore MTP layers for now
+        if name.endswith(".A_log"):
+            data_torch = -torch.exp(data_torch)
+        elif name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+        elif "conv1d" in name:
+            data_torch = data_torch.squeeze()
+        elif name.endswith("norm.weight") and not name.endswith("linear_attn.norm.weight"):
+            data_torch = data_torch + 1
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("RND1")
+class RND1Model(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.RND1
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # RND1 specific parameters
+        # RND1 uses bidirectional attention
+        self.gguf_writer.add_causal_attention(False)
+
+        if (mask_token_id := self.hparams.get("mask_token_id")) is not None:
+            self.gguf_writer.add_mask_token_id(mask_token_id)
+
+
 @ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration")
 class Qwen3VLVisionModel(MmprojModel):
     def __init__(self, *args, **kwargs):
@@ -10046,6 +10091,25 @@ class LazyTorchTensor(gguf.LazyBase):
         torch.uint8: np.uint8,
     }
 
+    # only used when byteswapping data. Only correct size is needed
+    _dtype_byteswap_map: dict[torch.dtype, type] = {
+        torch.float64: np.float64,
+        torch.float32: np.float32,
+        torch.bfloat16: np.float16,
+        torch.float16: np.float16,
+        torch.int64: np.int64,
+        torch.uint64: np.uint64,
+        torch.int32: np.int32,
+        torch.uint32: np.uint32,
+        torch.int16: np.int16,
+        torch.uint16: np.uint16,
+        torch.int8: np.int8,
+        torch.uint8: np.uint8,
+        torch.bool: np.uint8,
+        torch.float8_e4m3fn: np.uint8,
+        torch.float8_e5m2: np.uint8,
+    }
+
     # used for safetensors slices
     # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046
     # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734
@@ -10089,8 +10153,14 @@ class LazyTorchTensor(gguf.LazyBase):
     @classmethod
     def from_local_tensor(cls, t: gguf.utility.LocalTensor) -> Tensor:
         def load_tensor(tensor: gguf.utility.LocalTensor) -> Tensor:
+            def byteswap_tensor(tensor: np.ndarray, dtype: type) -> np.ndarray:
+                if sys.byteorder == 'big':
+                    # switch data back to big endian
+                    tensor = tensor.view(dtype).byteswap(inplace=False)
+                return tensor
             dtype = cls._dtype_str_map[tensor.dtype]
-            return torch.from_numpy(tensor.mmap_bytes()).view(dtype).reshape(tensor.shape)
+            numpy_dtype = cls._dtype_byteswap_map[dtype]
+            return torch.from_numpy(byteswap_tensor(tensor.mmap_bytes(), numpy_dtype)).view(dtype).reshape(tensor.shape)
         dtype = cls._dtype_str_map[t.dtype]
         shape = t.shape
         lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(t,), func=lambda r: load_tensor(r))
@@ -10098,10 +10168,16 @@ class LazyTorchTensor(gguf.LazyBase):
 
     @classmethod
     def from_remote_tensor(cls, remote_tensor: gguf.utility.RemoteTensor):
+        def byteswap_tensor(tensor: np.ndarray, dtype: type) -> np.ndarray:
+            if sys.byteorder == 'big':
+                # switch data back to big endian
+                tensor = tensor.view(dtype).byteswap(inplace=False)
+            return tensor
         dtype = cls._dtype_str_map[remote_tensor.dtype]
+        numpy_dtype = cls._dtype_byteswap_map[dtype]
         shape = remote_tensor.shape
         meta = cls.meta_with_dtype_and_shape(dtype, shape)
-        lazy = cls(meta=meta, args=(remote_tensor,), func=lambda r: torch.frombuffer(r.data(), dtype=dtype).reshape(shape))
+        lazy = cls(meta=meta, args=(remote_tensor,), func=lambda r: torch.from_numpy(byteswap_tensor(np.frombuffer(r.data(), dtype=numpy_dtype), numpy_dtype)).view(dtype).reshape(shape))
         return cast(torch.Tensor, lazy)
 
     @classmethod

convert_lora_to_gguf.py

@@ -242,7 +242,7 @@ def parse_args() -> argparse.Namespace:
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
     )
     parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f32",
         help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
     )
     parser.add_argument(

docs/backend/SYCL.md

@@ -42,6 +42,9 @@ The following releases are verified and recommended:
 
 ## News
 
+- 2025.11
+  - Support malloc memory on device more than 4GB.
+
 - 2025.2
   - Optimize MUL_MAT Q4_0 on Intel GPU for all dGPUs and built-in GPUs since MTL. Increase the performance of LLM (llama-2-7b.Q4_0.gguf) 21%-87% on Intel GPUs (MTL, ARL-H, Arc, Flex, PVC).
     |GPU|Base tokens/s|Increased tokens/s|Percent|
@@ -789,6 +792,8 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
+| UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Support malloc device memory more than 4GB.|
+
 
 
 ## Known Issues
@@ -835,6 +840,14 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   | The default context is too big. It leads to excessive memory usage.|Set `-c 8192` or a smaller value.|
   | The model is too big and requires more memory than what is available.|Choose a smaller model or change to a smaller quantization, like Q5 -> Q4;<br>Alternatively, use more than one device to load model.|
 
+- `ggml_backend_sycl_buffer_type_alloc_buffer: can't allocate 5000000000 Bytes of memory on device`
+
+  You need to enable to support 4GB memory malloc by:
+  ```
+    export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
+    set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
+  ```
+
 ### **GitHub contribution**:
 Please add the `SYCL :` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
 

examples/batched/README.md

@@ -3,7 +3,7 @@
 The example demonstrates batched generation from a given prompt
 
 ```bash
-./llama-batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4
+./llama-batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4 --kv-unified
 
 ...
 

examples/diffusion/README.md

@@ -6,8 +6,54 @@ More Info:
 - https://github.com/ggml-org/llama.cpp/pull/14644
 - https://github.com/ggml-org/llama.cpp/pull/14771
 
+## Parameters
+The diffusion CLI supports various parameters to control the generation process:
 
-Example of using Dream architechture: `llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual`
+### Core Diffusion Parameters
+- `--diffusion-steps`: Number of diffusion steps (default: 256)
+- `--diffusion-algorithm`: Algorithm for token selection
+  - `0`: ORIGIN - Token will be generated in a purely random order from https://arxiv.org/abs/2107.03006.
+  - `1`: ENTROPY_BASED - Entropy-based selection
+  - `2`: MARGIN_BASED - Margin-based selection
+  - `3`: RANDOM - Random selection
+  - `4`: CONFIDENCE_BASED - Confidence-based selection (default)
+  - More documentation here https://github.com/DreamLM/Dream
+- `--diffusion-visual`: Enable live visualization during generation
 
-Example of using LLaDA architechture: `llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual`
+### Scheduling Parameters
+Choose one of the following scheduling methods:
 
+**Timestep-based scheduling:**
+- `--diffusion-eps`: Epsilon value for timestep scheduling (e.g., 0.001)
+
+**Block-based scheduling:**
+- `--diffusion-block-length`: Block size for block-based scheduling (e.g., 32)
+
+### Sampling Parameters
+- `--temp`: Temperature for sampling (0.0 = greedy/deterministic, higher = more random)
+- `--top-k`: Top-k filtering for sampling
+- `--top-p`: Top-p (nucleus) filtering for sampling
+- `--seed`: Random seed for reproducibility
+
+### Model Parameters
+- `-m`: Path to the GGUF model file
+- `-p`: Input prompt text
+- `-ub`: Maximum sequence length (ubatch size)
+- `-c`: Context size
+- `-b`: Batch size
+
+### Examples
+#### Dream architechture:
+```
+llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual
+```
+
+#### LLaDA architechture:
+```
+llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual
+```
+
+#### RND1 architecture:
+```
+llama-diffusion-cli -m RND1-Base-0910.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-algorithm 1 --diffusion-steps 256 --diffusion-visual --temp 0.5 --diffusion-eps 0.001
+```

examples/embedding/embedding.cpp

@@ -104,12 +104,16 @@ int main(int argc, char ** argv) {
 
     params.embedding = true;
 
+    // get max number of sequences per batch
+    const int n_seq_max = llama_max_parallel_sequences();
+
     // if the number of prompts that would be encoded is known in advance, it's more efficient to specify the
     //   --parallel argument accordingly. for convenience, if not specified, we fallback to unified KV cache
     //   in order to support any number of prompts
     if (params.n_parallel == 1) {
         LOG_INF("%s: n_parallel == 1 -> unified KV cache is enabled\n", __func__);
         params.kv_unified = true;
+        params.n_parallel = n_seq_max;
     }
 
     // utilize the full context
@@ -123,9 +127,6 @@ int main(int argc, char ** argv) {
         params.n_ubatch = params.n_batch;
     }
 
-    // get max number of sequences per batch
-    const int n_seq_max = llama_max_parallel_sequences();
-
     llama_backend_init();
     llama_numa_init(params.numa);
 

examples/json_schema_to_grammar.py

@@ -231,9 +231,9 @@ DOT = '[^\\x0A\\x0D]'
 RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()])
 
 INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
-GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
+GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\\]')
 GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
-GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'}
+GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]', '\\': '\\\\'}
 
 NON_LITERAL_SET = set('|.()[]{}*+?')
 ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('^$.[]()|{}*+?')

examples/model-conversion/scripts/causal/run-converted-model.sh

@@ -4,6 +4,11 @@ set -e
 
 # First try command line argument, then environment variable, then file
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+MODEL_TESTING_PROMPT="${2:-"$MODEL_TESTING_PROMPT"}"
+
+if [ -z "$MODEL_TESTING_PROMPT"]; then
+    MODEL_TESTING_PROMPT="Hello, my name is"
+fi
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -14,7 +19,8 @@ if [ -z "$CONVERTED_MODEL" ]; then
 fi
 
 echo $CONVERTED_MODEL
+echo $MODEL_TESTING_PROMPT
 
 cmake --build ../../build --target llama-logits -j8
 
-../../build/bin/llama-logits -m "$CONVERTED_MODEL" "Hello, my name is"
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" "$MODEL_TESTING_PROMPT"

examples/model-conversion/scripts/causal/run-org-model.py

@@ -184,8 +184,12 @@ model_name = os.path.basename(model_path)
 # of using AutoModelForCausalLM.
 print(f"Model class: {model.__class__.__name__}")
 
-prompt = "Hello, my name is"
-input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+device = next(model.parameters()).device
+if os.getenv("MODEL_TESTING_PROMPT"):
+    prompt = os.getenv("MODEL_TESTING_PROMPT")
+else:
+    prompt = "Hello, my name is"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
 
 print(f"Input tokens: {input_ids}")
 print(f"Input text: {repr(prompt)}")

examples/sycl/run-llama2.sh

@@ -15,6 +15,9 @@ MODEL_FILE=models/llama-2-7b.Q4_0.gguf
 NGL=99
 CONTEXT=4096
 
+#support malloc device memory more than 4GB.
+export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
+
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"

examples/sycl/run-llama3.sh

@@ -6,7 +6,7 @@
 
 # If you want more control, DPC++ Allows selecting a specific device through the
 # following environment variable
-#export ONEAPI_DEVICE_SELECTOR="level_zero:0"
+export ONEAPI_DEVICE_SELECTOR="level_zero:0"
 source /opt/intel/oneapi/setvars.sh
 
 #export GGML_SYCL_DEBUG=1
@@ -18,11 +18,14 @@ MODEL_FILE=models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf
 NGL=99 # Layers offloaded to the GPU. If the device runs out of memory, reduce this value according to the model you are using.
 CONTEXT=4096
 
+#support malloc device memory more than 4GB.
+export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
+
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "Using $GGML_SYCL_DEVICE as the main GPU"
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
 fi

examples/sycl/win-run-llama2.bat

@@ -5,5 +5,7 @@
 set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
+:: support malloc device memory more than 4GB.
+set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 
 .\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 99 -s 0

examples/sycl/win-run-llama3.bat

@@ -5,5 +5,7 @@
 set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
+:: support malloc device memory more than 4GB.
+set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 
-.\build\bin\llama-cli.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -p %INPUT2% -n 400 -e -ngl 99
+.\build\bin\llama-cli.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -p %INPUT2% -n 400 -s 0 -e -ngl 99

ggml/CMakeLists.txt

@@ -25,16 +25,17 @@ if(GIT_EXE)
     )
 endif()
 
-# Build the version string with optional dirty flag
 set(GGML_VERSION "${GGML_VERSION_BASE}")
-if(GGML_GIT_DIRTY AND NOT GGML_GIT_DIRTY EQUAL 0)
-    set(GGML_VERSION "${GGML_VERSION}-dirty")
-endif()
 
 if(NOT GGML_BUILD_COMMIT)
     set(GGML_BUILD_COMMIT "unknown")
 endif()
 
+# Build the commit string with optional dirty flag
+if(DEFINED GGML_GIT_DIRTY AND GGML_GIT_DIRTY EQUAL 1)
+    set(GGML_BUILD_COMMIT "${GGML_BUILD_COMMIT}-dirty")
+endif()
+
 include(CheckIncludeFileCXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@@ -182,6 +183,7 @@ endif()
 # ggml core
 set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
 option(GGML_CPU                             "ggml: enable CPU backend"                        ON)
+option(GGML_SCHED_NO_REALLOC                "ggml: disallow reallocations in ggml-alloc (for debugging)" OFF)
 
 # 3rd party libs / backends
 option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)

ggml/include/ggml-rpc.h

@@ -8,7 +8,7 @@ extern "C" {
 #endif
 
 #define RPC_PROTO_MAJOR_VERSION    3
-#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_MINOR_VERSION    5
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 

ggml/include/ggml.h

@@ -530,6 +530,7 @@ extern "C" {
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
+        GGML_OP_TOP_K,
         GGML_OP_LEAKY_RELU,
         GGML_OP_TRI,
         GGML_OP_FILL,
@@ -2258,18 +2259,25 @@ extern "C" {
             struct ggml_tensor  * a,
             enum ggml_sort_order  order);
 
+    // similar to ggml_top_k but implemented as `argsort` + `view`
+    GGML_API struct ggml_tensor * ggml_argsort_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
+    // top k elements per row
+    // note: the resulting top k indices are in no particular order
+    GGML_API struct ggml_tensor * ggml_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
     GGML_API struct ggml_tensor * ggml_arange(
             struct ggml_context * ctx,
             float                 start,
             float                 stop,
             float                 step);
 
-    // top k elements per row
-    GGML_API struct ggml_tensor * ggml_top_k(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                   k);
-
 #define GGML_KQ_MASK_PAD 64
 
     // q:    [n_embd_k, n_batch,     n_head,    ne3 ]

ggml/src/CMakeLists.txt

@@ -221,6 +221,10 @@ if (GGML_BACKEND_DL)
     target_compile_definitions(ggml-base PUBLIC GGML_BACKEND_DL)
 endif()
 
+if (GGML_SCHED_NO_REALLOC)
+    target_compile_definitions(ggml-base PUBLIC GGML_SCHED_NO_REALLOC)
+endif()
+
 add_library(ggml
             ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)
@@ -328,6 +332,14 @@ function(ggml_add_cpu_backend_variant tag_name)
             set(GGML_INTERNAL_${feat} OFF)
         endforeach()
 
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
+        foreach (feat RVV)
+            set(GGML_INTERNAL_${feat} OFF)
+        endforeach()
+
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
@@ -402,6 +414,13 @@ if (GGML_CPU_ALL_VARIANTS)
         else()
             message(FATAL_ERROR "Unsupported s390x target OS: ${CMAKE_SYSTEM_NAME}")
         endif()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            ggml_add_cpu_backend_variant(riscv64_0)
+            ggml_add_cpu_backend_variant(riscv64_v   RVV)
+        else()
+            message(FATAL_ERROR "Unsupported RISC-V target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
     else()
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()

ggml/src/ggml-alloc.c

@@ -921,10 +921,15 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         }
         if (realloc) {
 #ifndef NDEBUG
-            size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
-            GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+            {
+                size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
+                if (cur_size > 0) {
+                    GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n",
+                        __func__, ggml_backend_buft_name(galloc->bufts[i]),
+                        cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+                }
+            }
 #endif
-
             ggml_vbuffer_free(galloc->buffers[i]);
             galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
             if (galloc->buffers[i] == NULL) {

ggml/src/ggml-backend.cpp

@@ -1395,14 +1395,20 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 
     // allocate graph
     if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
+#ifdef GGML_SCHED_NO_REALLOC
+        GGML_ABORT("%s: failed to allocate graph, but graph re-allocation is disabled by GGML_SCHED_NO_REALLOC\n", __func__);
+#endif
+
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
+#endif
+
         // the re-allocation may cause the split inputs to be moved to a different address
         // synchronize without ggml_backend_sched_synchronize to avoid changing cur_copy
         for (int i = 0; i < sched->n_backends; i++) {
             ggml_backend_synchronize(sched->backends[i]);
         }
-#ifndef NDEBUG
-        GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
-#endif
+
         ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
         if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
             GGML_LOG_ERROR("%s: failed to allocate graph\n", __func__);

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -42,6 +42,7 @@
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_fill_scalar.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
+#include <aclnnop/aclnn_ger.h>
 #include <aclnnop/aclnn_group_norm.h>
 #include <aclnnop/aclnn_grouped_matmul_v3.h>
 #include <aclnnop/aclnn_gt_scalar.h>
@@ -2206,78 +2207,120 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context & ctx,
 }
 
 /**
- * @brief Initializes and caches sine/cosine positional encoding values
- *        (used in RoPE, Rotary Position Embedding) for attention layers.
+ * @brief Initializes and caches all intermediate tensors required for RoPE
+ *        (Rotary Position Embedding), including support for Yarn, mRoPE,
+ *        i-mRoPE, Neox repeat strategy, independent sectors, frequency factors，
+ *        and multi-section rotary groups.
  *
- * This function computes and caches the sin/cos values of
- * θ = position * theta_scale for RoPE encoding. The cache is shared
- * across attention layers, and only the first attention layer will
- * trigger initialization. The cache includes repeated sin/cos values
- * with different repeat methods depending on the @param is_neox flag.
+ * This function computes and caches the per-dimension θ coefficients used for
+ * Q/K rotary embedding. The cache is shared across layers, and recomputed only
+ * when any dependent parameter changes.
  *
- * Steps performed by this function:
- *   1. Identify whether the target tensor belongs to Q/K in attention
- *      and restrict computation to the first layer only.
- *   2. Initialize the theta scale array (arange → power → freq scaling).
- *   3. Allocate sin/cos caches if the max prompt length increases.
- *   4. Compute θ = position * theta_scale.
- *   5. Compute sin(θ), cos(θ) and optionally scale by attn_factor.
- *   6. Expand sin/cos values by repeat or repeat_interleave depending
- *      on whether @param is_neox is enabled.
+ * The function now supports:
+ *   - Yarn RoPE extrapolation (via @param corr_dims and @param ext_factor)
+ *   - Per-dimension independent sector exponent rules (indep_sects + sections[])
+ *   - Multi-section RoPE (mRoPE) index mapping (mrope_used + is_imrope)
+ *   - Frequency factor division (src2)
+ *   - Neox / normal repeat expansion modes
  *
- * @param ctx                The CANN backend context, holding memory pool,
- *                           stream, and persistent buffers for rope init/cache.
- * @param dst                The destination ggml_tensor whose computation
- *                           depends on the RoPE values (usually Qcur/Kcur).
- * @param theta_scale        Scalar exponent base for computing theta scale values.
- * @param freq_scale         Frequency scaling factor, applied to theta scale.
- * @param attn_factor        Attention scaling factor, applied to sin/cos.
- * @param is_neox            Whether to use Neox-style repeat strategy
- *                           (dim expansion vs repeat_interleave).
+ * @param ctx                CANN backend context, containing memory pool,
+ *                           cached buffers, and runtime stream.
+ * @param dst                Destination ggml_tensor whose computation
+ *                           depends on RoPE (typically Qcur or Kcur).
+ * @param corr_dims          [low, high] Yarn correction range.
+ * @param ext_factor         Yarn extrapolation strength. 0 = disabled.
+ * @param theta_scale        Base multiplier for per-dimension θ exponent.
+ * @param freq_scale         Global frequency scaling factor.
+ * @param attn_factor        Optional scaling applied to sin/cos (if needed).
+ * @param is_neox            Whether to use Neox-style dimension interleave.
+ * @param sections           4-way sector sizes for independent-section RoPE
+ *                           and multi-section mRoPE (t/h/w/e).
+ * @param mrope_used         Whether to enable multi-section rotary embedding.
+ * @param is_imrope          Whether to apply interleaved mRoPE rules.
+ * @param indep_sects        Whether each dimension runs independent exponent
+ *                           resets based on @p sections.
  */
-static void aclnn_cache_init(ggml_backend_cann_context & ctx,
-                             ggml_tensor *               dst,
-                             float *                     corr_dims,
-                             float                       ext_factor,
-                             float                       theta_scale,
-                             float                       freq_scale,
-                             float                       attn_factor,
-                             bool                        is_neox) {
+static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
+                                  ggml_tensor *               dst,
+                                  float *                     corr_dims,
+                                  float                       ext_factor,
+                                  float                       theta_scale,
+                                  float                       freq_scale,
+                                  float                       attn_factor,
+                                  bool                        is_neox,
+                                  int                         sections[4],
+                                  bool                        mrope_used,
+                                  bool                        is_imrope,
+                                  bool                        indep_sects) {
     ggml_tensor * src0 = dst->src[0];  // input
     ggml_tensor * src1 = dst->src[1];  // position
     ggml_tensor * src2 = dst->src[2];  // freq_factors
 
-    if (src2 == nullptr && ctx.rope_cache.cached && ctx.rope_cache.ext_factor == ext_factor &&
-        ctx.rope_cache.theta_scale == theta_scale && ctx.rope_cache.freq_scale == freq_scale &&
-        ctx.rope_cache.attn_factor == attn_factor && ctx.rope_cache.is_neox == is_neox) {
+    int64_t theta_scale_length = src0->ne[0] / 2;
+    int64_t position_length    = dst->ne[2];
+
+    // TODO: check theta_scale_length and position_length.
+    if (src2 == nullptr && ctx.rope_cache.cached &&
+        ctx.rope_cache.equal(theta_scale_length, position_length, ext_factor, theta_scale, freq_scale, attn_factor,
+                             is_neox, indep_sects, mrope_used, is_imrope, sections)) {
         // use cache.
         return;
     }
 
-    int64_t theta_scale_length = src0->ne[0] / 2;
-    int64_t theta_scale_ne[]   = { theta_scale_length, 1, 1, 1 };
-    size_t  theta_scale_nb[]   = { sizeof(float), sizeof(float), sizeof(float), theta_scale_length * sizeof(float) };
+    // Step0: calculate tensor shape.
+    int64_t theta_scale_ne[] = { theta_scale_length, 1, 1, 1 };
+    size_t  theta_scale_nb[] = { sizeof(float), theta_scale_length * sizeof(float), theta_scale_length * sizeof(float),
+                                 theta_scale_length * sizeof(float) };
 
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    int64_t position_length = src1->ne[0];
-    int64_t position_ne[]   = { 1, 1, position_length, 1 };
-    size_t  position_nb[]   = { sizeof(int32_t), sizeof(int32_t), sizeof(int32_t), sizeof(int32_t) * position_length };
+    int64_t position_ne[] = { 1, 1, position_length, 1 };
+    size_t  position_nb[] = { sizeof(int32_t), sizeof(int32_t), sizeof(int32_t), sizeof(int32_t) * position_length };
 
-    int64_t theta_ne[] = { theta_scale_length, 1, position_length, 1 };
-    size_t  theta_nb[GGML_MAX_DIMS];
-    theta_nb[0] = sizeof(float);
+    int64_t cache_ne[] = { theta_scale_length, 1, position_length, 1 };
+    size_t  cache_nb[GGML_MAX_DIMS];
+    cache_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
+        cache_nb[i] = cache_nb[i - 1] * cache_ne[i - 1];
     }
 
-    // theta_scale arange, [0,1,...,ne00/2 - 1]
+    // Step1: Compute the coefficient of theta. During the cache_init process, aside from
+    // (1) multiplying by the position,
+    // (2) dividing by freq_factors,
+    // (3) computing the sine and cosine,
+    // the other parameters used in the computation generally do not change in most scenarios.
+    // Therefore, we can first compute this part of the result and then cache it.
+
+    // Step1.1: prepare theta_scale exponent. if this exponent updated, should update theta_scale_tensor.
     acl_tensor_ptr acl_theta_scale_tensor;
-    // cache theta scale
-    if (ctx.rope_cache.theta_scale_length != theta_scale_length ||
-        // theta_scale and freq_scale should not change during the current token inference process,
-        // so we can directly use == here instead of comparing the absolute difference.
-        ctx.rope_cache.theta_scale != theta_scale || ctx.rope_cache.freq_scale != freq_scale) {
-        ctx.rope_cache.theta_scale_length = theta_scale_length;
+    bool           theta_scale_updated = false;
+    if (ctx.rope_cache.theta_scale_length != theta_scale_length || ctx.rope_cache.theta_scale != theta_scale ||
+        ctx.rope_cache.indep_sects != indep_sects) {
+        theta_scale_updated = true;
+        if (ctx.rope_cache.theta_scale_exp_host != nullptr) {
+            free(ctx.rope_cache.theta_scale_exp_host);
+        }
+        ctx.rope_cache.theta_scale_exp_host = (float *) malloc(theta_scale_length * sizeof(float));
+        GGML_ASSERT(ctx.rope_cache.theta_scale_exp_host != nullptr);
+        if (!indep_sects) {
+            ctx.rope_cache.theta_scale_exp_host[0] = 1;
+            for (int i = 1; i < theta_scale_length; i++) {
+                ctx.rope_cache.theta_scale_exp_host[i] = ctx.rope_cache.theta_scale_exp_host[i - 1] * theta_scale;
+            }
+        } else {
+            int sect_dims = sections[0] + sections[1] + sections[2] + sections[3];
+            int sec_w     = sections[1] + sections[0];
+            int sec_e     = sections[2] + sec_w;
+
+            ctx.rope_cache.theta_scale_exp_host[0] = 1;
+            for (int i = 1; i < theta_scale_length; i++) {
+                int sector = i % sect_dims;
+                if (sector == 0 || sector == sections[0] || sector == sec_w || sector == sec_e) {
+                    ctx.rope_cache.theta_scale_exp_host[i] = 1;
+                    continue;
+                }
+                ctx.rope_cache.theta_scale_exp_host[i] = ctx.rope_cache.theta_scale_exp_host[i - 1] * theta_scale;
+            }
+        }
 
         if (ctx.rope_cache.theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(ctx.rope_cache.theta_scale_cache));
@@ -2285,74 +2328,138 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
         ACL_CHECK(aclrtMalloc(&ctx.rope_cache.theta_scale_cache, theta_scale_length * sizeof(float),
                               ACL_MEM_MALLOC_HUGE_FIRST));
 
+        ACL_CHECK(aclrtMemcpyAsync(ctx.rope_cache.theta_scale_cache, theta_scale_length * sizeof(float),
+                                   ctx.rope_cache.theta_scale_exp_host, theta_scale_length * sizeof(float),
+                                   ACL_MEMCPY_HOST_TO_DEVICE, ctx.stream()));
+
         acl_theta_scale_tensor = ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
                                                          theta_scale_ne, theta_scale_nb, 1);
+    }
 
-        float start      = 0;
-        float step       = 1;
-        float stop       = theta_scale_length;
-        float n_elements = theta_scale_length;
-        aclnn_arange(ctx, acl_theta_scale_tensor.get(), start, stop, step, n_elements);
+    // Step1.2: prepare rope_yarn_ramp, if this part updated, should update theta_scale_tensor.
+    bool                 yarn_ramp_tensor_updated = false;
+    ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
+    acl_tensor_ptr       acl_yarn_ramp_tensor;
+    if (ext_factor != 0 &&
+        // TODO: check more parameter.
+        (ctx.rope_cache.theta_scale_length != theta_scale_length || ctx.rope_cache.freq_scale != freq_scale)) {
+        yarn_ramp_tensor_updated = true;
 
-        ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
-        acl_tensor_ptr       acl_yarn_ramp_tensor;
-        if (ext_factor != 0) {
-            // -rope_yarn_ramp
-            // const float y = (i0 / 2 - low) / MAX(0.001f, high - low);
-            // return MIN(1, MAX(0, y)) - 1;
-            yarn_ramp_allocator.alloc(theta_scale_length * sizeof(float));
-            void * yarn_ramp_buffer = yarn_ramp_allocator.get();
-            acl_yarn_ramp_tensor =
-                ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
-            float          zero_value = 0, one_value = 1;
-            float          denom_safe_value = MAX(0.001f, corr_dims[1] - corr_dims[0]);
-            acl_scalar_ptr low              = ggml_cann_create_scalar(&corr_dims[0], aclDataType::ACL_FLOAT);
-            acl_scalar_ptr zero             = ggml_cann_create_scalar(&zero_value, aclDataType::ACL_FLOAT);
-            acl_scalar_ptr one              = ggml_cann_create_scalar(&one_value, aclDataType::ACL_FLOAT);
-            acl_scalar_ptr denom_safe       = ggml_cann_create_scalar(&denom_safe_value, aclDataType::ACL_FLOAT);
-            acl_scalar_ptr ext_factor_sc    = ggml_cann_create_scalar(&ext_factor, aclDataType::ACL_FLOAT);
+        // -rope_yarn_ramp
+        // const float y = (i0 / 2 - low) / MAX(0.001f, high - low);
+        // return MIN(1, MAX(0, y)) - 1;
+        yarn_ramp_allocator.alloc(theta_scale_length * sizeof(float));
+        void * yarn_ramp_buffer = yarn_ramp_allocator.get();
+        acl_yarn_ramp_tensor =
+            ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
+        float          zero_value = 0, one_value = 1;
+        float          denom_safe_value = MAX(0.001f, corr_dims[1] - corr_dims[0]);
+        acl_scalar_ptr low              = ggml_cann_create_scalar(&corr_dims[0], aclDataType::ACL_FLOAT);
+        acl_scalar_ptr zero             = ggml_cann_create_scalar(&zero_value, aclDataType::ACL_FLOAT);
+        acl_scalar_ptr one              = ggml_cann_create_scalar(&one_value, aclDataType::ACL_FLOAT);
+        acl_scalar_ptr denom_safe       = ggml_cann_create_scalar(&denom_safe_value, aclDataType::ACL_FLOAT);
+        acl_scalar_ptr ext_factor_sc    = ggml_cann_create_scalar(&ext_factor, aclDataType::ACL_FLOAT);
 
-            GGML_CANN_CALL_ACLNN_OP(ctx, Subs, acl_theta_scale_tensor.get(), low.get(), one.get(),
-                                    acl_yarn_ramp_tensor.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceDivs, acl_yarn_ramp_tensor.get(), denom_safe.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceThreshold, acl_yarn_ramp_tensor.get(), zero.get(), zero.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceClampMax, acl_yarn_ramp_tensor.get(), one.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSubs, acl_yarn_ramp_tensor.get(), one.get(), one.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), ext_factor_sc.get());
+        aclnn_arange(ctx, acl_yarn_ramp_tensor.get(), 0, theta_scale_length, 1, theta_scale_length);
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSubs, acl_yarn_ramp_tensor.get(), low.get(), one.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceDivs, acl_yarn_ramp_tensor.get(), denom_safe.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceThreshold, acl_yarn_ramp_tensor.get(), zero.get(), zero.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceClampMax, acl_yarn_ramp_tensor.get(), one.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSubs, acl_yarn_ramp_tensor.get(), one.get(), one.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), ext_factor_sc.get());
 
-            // theta_interp = freq_scale * theta_extrap;
-            // theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-            // theta = freq_scale * theta_extrap * (1 - ramp_mix) + theta_extrap * ramp_mix;
-            // theta = freq_scale * theta_extrap - freq_scale * theta_extrap * ramp_mix + theta_extrap * ramp_mix;
-            // theta = theta_extrap * (freq_scale - freq_scale * ramp_mix + ramp_mix);
-            //
-            // we cache (freq_scale - freq_scale * ramp_mix + ramp_mix), Considering that the rope_yarn_ramp here is the inverse
-            // cache freq_scale + (freq_scale - 1) * ramp_mix
-            float          freq_scale_1    = freq_scale - 1;
-            acl_scalar_ptr freq_scale_sc   = ggml_cann_create_scalar(&freq_scale, aclDataType::ACL_FLOAT);
-            acl_scalar_ptr freq_scale_1_sc = ggml_cann_create_scalar(&freq_scale_1, aclDataType::ACL_FLOAT);
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), freq_scale_1_sc.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdds, acl_yarn_ramp_tensor.get(), freq_scale_sc.get(), one.get());
-        }
+        // theta_interp = freq_scale * theta_extrap;
+        // theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+        // theta = freq_scale * theta_extrap * (1 - ramp_mix) + theta_extrap * ramp_mix;
+        // theta = freq_scale * theta_extrap - freq_scale * theta_extrap * ramp_mix + theta_extrap * ramp_mix;
+        // theta = theta_extrap * (freq_scale - freq_scale * ramp_mix + ramp_mix);
+        //
+        // we cache (freq_scale - freq_scale * ramp_mix + ramp_mix), Considering that the rope_yarn_ramp here is the inverse
+        // cache freq_scale + (freq_scale - 1) * ramp_mix
+        float          freq_scale_1    = freq_scale - 1;
+        acl_scalar_ptr freq_scale_sc   = ggml_cann_create_scalar(&freq_scale, aclDataType::ACL_FLOAT);
+        acl_scalar_ptr freq_scale_1_sc = ggml_cann_create_scalar(&freq_scale_1, aclDataType::ACL_FLOAT);
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), freq_scale_1_sc.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdds, acl_yarn_ramp_tensor.get(), freq_scale_sc.get(), one.get());
+    }
 
-        // power
-        acl_scalar_ptr acl_theta_scale = ggml_cann_create_scalar(&theta_scale, aclDataType::ACL_FLOAT);
-        GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale.get(), acl_theta_scale_tensor.get(),
-                                acl_theta_scale_tensor.get());
-
-        if (ext_factor != 0) {
+    // Step 1.3: update theta_scale_tensor according to ext_factor or freq_scale.
+    if (ext_factor != 0) {
+        if (theta_scale_updated || yarn_ramp_tensor_updated) {
+            theta_scale_updated = true;
             aclnn_mul(ctx, acl_theta_scale_tensor.get(), acl_yarn_ramp_tensor.get());
-        } else if (freq_scale != 1) {
-            aclnn_muls(ctx, acl_theta_scale_tensor.get(), freq_scale, nullptr, true);
         }
     } else {
-        // use cache
+        if (freq_scale != 1 && (ctx.rope_cache.freq_scale != freq_scale || theta_scale_updated)) {
+            theta_scale_updated = true;
+            aclnn_muls(ctx, acl_theta_scale_tensor.get(), freq_scale, nullptr, true);
+        }
+    }
+
+    // Nothing changed, use cache.
+    if (!theta_scale_updated) {
         acl_theta_scale_tensor = ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
                                                          theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
     }
 
+    // Step 1.4: prepare select index if mrope
+    acl_tensor_ptr position_select_index_tensor;
+    if (mrope_used) {
+        if (ctx.rope_cache.sections[0] != sections[0] || ctx.rope_cache.sections[1] != sections[1] ||
+            ctx.rope_cache.sections[2] != sections[2] || ctx.rope_cache.sections[3] != sections[3] ||
+            ctx.rope_cache.theta_scale_length != theta_scale_length || ctx.rope_cache.is_imrope != is_imrope) {
+            if (ctx.rope_cache.position_select_index_host != nullptr) {
+                free(ctx.rope_cache.position_select_index_host);
+            }
+            ctx.rope_cache.position_select_index_host = (int *) malloc(theta_scale_length * sizeof(int));
+            GGML_ASSERT(ctx.rope_cache.position_select_index_host != nullptr);
+            int sect_dims = sections[0] + sections[1] + sections[2] + sections[3];
+            int sec_w     = sections[1] + sections[0];
+            int sec_e     = sections[2] + sec_w;
+            // t,h,w,e
+            for (int i = 0; i < theta_scale_length; i++) {
+                int sector = i % sect_dims;
+
+                if (is_imrope) {  // qwen3vl apply interleaved mrope
+                    if (sector % 3 == 1 && sector < 3 * sections[1]) {
+                        ctx.rope_cache.position_select_index_host[i] = 1;
+                    } else if (sector % 3 == 2 && sector < 3 * sections[2]) {
+                        ctx.rope_cache.position_select_index_host[i] = 2;
+                    } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
+                        ctx.rope_cache.position_select_index_host[i] = 0;
+                    } else {
+                        ctx.rope_cache.position_select_index_host[i] = 3;
+                    }
+                } else {
+                    if (sector >= sections[0] && sector < sec_w) {
+                        ctx.rope_cache.position_select_index_host[i] = 1;
+                    } else if (sector >= sec_w && sector < sec_e) {
+                        ctx.rope_cache.position_select_index_host[i] = 2;
+                    } else if (sector >= sec_e) {
+                        ctx.rope_cache.position_select_index_host[i] = 3;
+                    } else {
+                        ctx.rope_cache.position_select_index_host[i] = 0;
+                    }
+                }
+            }
+
+            if (ctx.rope_cache.position_select_index != nullptr) {
+                ACL_CHECK(aclrtFree(ctx.rope_cache.position_select_index));
+            }
+            ACL_CHECK(aclrtMalloc(&ctx.rope_cache.position_select_index, theta_scale_length * sizeof(int),
+                                  ACL_MEM_MALLOC_HUGE_FIRST));
+
+            ACL_CHECK(aclrtMemcpyAsync(ctx.rope_cache.position_select_index, theta_scale_length * sizeof(int),
+                                       ctx.rope_cache.position_select_index_host, theta_scale_length * sizeof(int),
+                                       ACL_MEMCPY_HOST_TO_DEVICE, ctx.stream()));
+        }
+
+        position_select_index_tensor = ggml_cann_create_tensor(ctx.rope_cache.position_select_index, ACL_INT32,
+                                                               sizeof(int), theta_scale_ne, theta_scale_nb, 1);
+    }
+
+    // Step2: divide by freq_factors
     ggml_cann_pool_alloc freq_fac_res_allocator(ctx.pool());
-    // freq_factors
     if (src2) {
         freq_fac_res_allocator.alloc(theta_scale_length * sizeof(float));
         void *         freq_fac_res_ptr = freq_fac_res_allocator.get();
@@ -2365,6 +2472,85 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
         std::swap(acl_theta_scale_tensor, acl_freq_fac_res_tensor);
     }
 
+    // Step3: prepare position_tensor
+    acl_tensor_ptr       acl_position_tensor;
+    ggml_cann_pool_alloc mrope_position_acllocator(ctx.pool());
+    if (mrope_used) {
+        // Step3.1: select current position;
+        // position :
+        // pos1: [[0, 1 ,2 ,3 ],
+        // pos2:  [4, 5 ,6 ,7 ],
+        // pos3:  [8, 9 ,10,11],
+        // pos4:  [12,13,14,15] ]
+        //
+        // select index = [0, 1, 2, 2, 1, 0]
+        //
+        // selected_tensor:
+        // [[0, 1 ,2 ,3 ],
+        //  [4, 5 ,6 ,7 ],
+        //  [8, 9 ,10,11],
+        //  [8, 9 ,10,11],
+        //  [4, 5 ,6 ,7 ],
+        //  [0, 1 ,2 ,3 ]]
+        //
+        // transpose, from [seq_len:dims] to [dims:seq_len]
+        // [0, 4, 8 ,8 ,4, 0],
+        // [1, 5, 9, 9, 5, 1],
+        // [2, 6, 10,10,6 ,2],
+        // [3, 7, 11,11,7 3 ]]
+        //
+        // multipy by theta_scale_tensor
+        // [theta_scale^0, theta_scale^1, ..., theta_scale ^ n]
+
+        int64_t        mrope_position_ne[] = { position_length, 4 };
+        size_t         mrope_position_nb[] = { sizeof(int), position_length * sizeof(int) };
+        acl_tensor_ptr mrope_position =
+            ggml_cann_create_tensor(src1->data, ggml_cann_type_mapping(src1->type), ggml_type_size(src1->type),
+                                    mrope_position_ne, mrope_position_nb, 2);
+
+        // selected position tensor's shape is a transpose of cache tensor.
+        int64_t selected_position_ne[] = { position_length, theta_scale_length };
+        size_t  selected_position_nb[] = { sizeof(float), position_length * sizeof(float) };
+        mrope_position_acllocator.alloc(theta_scale_length * position_length * sizeof(float));
+        void * mrope_position_buffer = mrope_position_acllocator.get();
+        acl_position_tensor =
+            ggml_cann_create_tensor(mrope_position_buffer, ggml_cann_type_mapping(src1->type),
+                                    ggml_type_size(src1->type), selected_position_ne, selected_position_nb, 2);
+        GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, mrope_position.get(), 0, position_select_index_tensor.get(),
+                                acl_position_tensor.get());
+
+        // transpose
+        int64_t transposed_ne[] = { position_length, 1, theta_scale_length, 1 };
+        size_t  transposed_nb[GGML_MAX_DIMS];
+        transposed_nb[0] = sizeof(float);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            transposed_nb[i] = transposed_nb[i - 1] * transposed_ne[i - 1];
+        }
+
+        std::swap(transposed_ne[0], transposed_ne[2]);
+        std::swap(transposed_nb[0], transposed_nb[2]);
+
+        acl_position_tensor =
+            ggml_cann_create_tensor(mrope_position_buffer, ggml_cann_type_mapping(src1->type),
+                                    ggml_type_size(src1->type), transposed_ne, transposed_nb, GGML_MAX_DIMS);
+
+    } else {
+        // auto bcast.
+        acl_position_tensor =
+            ggml_cann_create_tensor(src1->data, ggml_cann_type_mapping(src1->type), ggml_type_size(src1->type),
+                                    position_ne, position_nb, GGML_MAX_DIMS);
+    }
+
+    // Step4: multiply by the position
+    int64_t              theta_length = theta_scale_length * position_length;
+    ggml_cann_pool_alloc theta_allocator(ctx.pool(), theta_length * sizeof(float));
+    void *               theta_buffer = theta_allocator.get();
+
+    acl_tensor_ptr acl_theta_tensor =
+        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float), cache_ne, cache_nb, GGML_MAX_DIMS);
+    aclnn_mul(ctx, acl_position_tensor.get(), acl_theta_scale_tensor.get(), acl_theta_tensor.get());
+
+    // Step5: calculate sin cos.
     // init sin_repeat && cos_repeat, only to accelerate first layer on each device
     if (position_length > ctx.rope_cache.position_length) {
         ctx.rope_cache.position_length = position_length;
@@ -2381,44 +2567,30 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
             aclrtMalloc(&ctx.rope_cache.cos_cache, repeat_theta_length * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST));
     }
 
-    // position
-    acl_tensor_ptr acl_position_tensor =
-        ggml_cann_create_tensor(src1->data, ggml_cann_type_mapping(src1->type), ggml_type_size(src1->type), position_ne,
-                                position_nb, GGML_MAX_DIMS);
-
-    // power * position
-    int64_t              theta_length = theta_scale_length * position_length;
-    ggml_cann_pool_alloc theta_allocator(ctx.pool(), theta_length * sizeof(float));
-    void *               theta_buffer = theta_allocator.get();
-
-    acl_tensor_ptr acl_theta_tensor =
-        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_nb, GGML_MAX_DIMS);
-    aclnn_mul(ctx, acl_position_tensor.get(), acl_theta_scale_tensor.get(), acl_theta_tensor.get());
-
     // sin/cos
     ggml_cann_pool_alloc sin_allocator(ctx.pool(), theta_length * sizeof(float));
     void *               sin_buffer = sin_allocator.get();
     acl_tensor_ptr       acl_sin_tensor =
-        ggml_cann_create_tensor(sin_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
+        ggml_cann_create_tensor(sin_buffer, ACL_FLOAT, sizeof(float), cache_ne, cache_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_sin(ctx, acl_theta_tensor.get(), acl_sin_tensor.get());
 
     ggml_cann_pool_alloc cos_allocator(ctx.pool(), theta_length * sizeof(float));
     void *               cos_buffer = cos_allocator.get();
     acl_tensor_ptr       acl_cos_tensor =
-        ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
+        ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float), cache_ne, cache_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_cos(ctx, acl_theta_tensor.get(), acl_cos_tensor.get());
 
     if (ext_factor != 0) {
         attn_factor *= 1.0f + 0.1f * logf(1.0f / freq_scale);
     }
 
-    // attn_factor
+    // Step 5: multiply by attn_factor
     if (attn_factor != 1) {
         aclnn_muls(ctx, acl_sin_tensor.get(), attn_factor, nullptr, true);
         aclnn_muls(ctx, acl_cos_tensor.get(), attn_factor, nullptr, true);
     }
 
-    int64_t sin_reshape_ne[4] = { src0->ne[0], 1, src0->ne[2], 1 };
+    int64_t sin_reshape_ne[4] = { src0->ne[0], 1, dst->ne[2], 1 };
     size_t  sin_reshape_nb[GGML_MAX_DIMS];
     sin_reshape_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
@@ -2429,8 +2601,9 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
     acl_tensor_ptr acl_cos_repeat_tensor = ggml_cann_create_tensor(ctx.rope_cache.cos_cache, ACL_FLOAT, sizeof(float),
                                                                    sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
 
-    // repeat
+    // Step 6: repeat
     if (is_neox) {
+        // [sinθ1, sinθ1, sinθ2, sinθ2, ..., sinθn, sinθn]
         int64_t repeatsArray[] = { 1, 1, 1, 2 };
         aclnn_repeat(ctx, acl_sin_tensor.get(), acl_sin_repeat_tensor.get(), repeatsArray);
         aclnn_repeat(ctx, acl_cos_tensor.get(), acl_cos_repeat_tensor.get(), repeatsArray);
@@ -2438,17 +2611,15 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
         int64_t num_repeats = 2;
         int64_t dim         = 3;
         int64_t output_size = theta_scale_length * num_repeats;
+        // [sinθ1, sinθ2, ..., sinθn, sinθ1, sinθ2, ..., sinθn]
         aclnn_repeat_interleave(ctx, acl_sin_tensor.get(), acl_sin_repeat_tensor.get(), dim, num_repeats, output_size);
         aclnn_repeat_interleave(ctx, acl_cos_tensor.get(), acl_cos_repeat_tensor.get(), dim, num_repeats, output_size);
     }
 
-    // Other layers use cache except first layer.
-    ctx.rope_cache.cached      = true;
-    ctx.rope_cache.ext_factor  = ext_factor;
-    ctx.rope_cache.theta_scale = theta_scale;
-    ctx.rope_cache.freq_scale  = freq_scale;
-    ctx.rope_cache.attn_factor = attn_factor;
-    ctx.rope_cache.is_neox     = is_neox;
+    // Update cached value.
+    ctx.rope_cache.cached = true;
+    ctx.rope_cache.set(theta_scale_length, position_length, ext_factor, theta_scale, freq_scale, attn_factor, is_neox,
+                       indep_sects, mrope_used, is_imrope, sections);
 }
 
 #ifdef __cplusplus
@@ -2474,6 +2645,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 
     // param
     float     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int sections[4];
     // const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
     const int mode       = ((int32_t *) dst->op_params)[2];
@@ -2482,12 +2654,13 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
-    memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
-    memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
-    memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
-    memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
-    memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     // TODO: n_dims <= ne0
     GGML_ASSERT(n_dims == ne0);
@@ -2498,10 +2671,25 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
+    const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, note: also true for vision (24 & 8 == true) and for imrope
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (mrope_used) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne0/2);
+    }
+
+    if (is_imrope || mrope_used) {
+        is_neox = true;
+    }
 
     // init ctx.rope_cos/rope_sin cache
-    aclnn_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox);
+    aclnn_rope_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox, sections, mrope_used, is_imrope, is_vision);
 
     int64_t sin_reshape_ne[4] = { ne00, 1, ne02, 1 };
     size_t  sin_reshape_nb[GGML_MAX_DIMS];
@@ -2657,8 +2845,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     return;
 #endif
 
-    // ggml_mode = 0 --> aclnn_model = 1
-    int64_t acl_mode = mode == 0 ? 1 : mode;
+    int64_t acl_mode = is_neox ? 0 : 1;
 
     switch (src0->type) {
         case GGML_TYPE_F32:
@@ -3236,3 +3423,64 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context & ctx, ggml_tensor * dst
         GGML_ABORT("Function is not implemented.");
     }
 }
+
+static void ggml_cann_out_prod_fp(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];  // weight
+    ggml_tensor * src1 = dst->src[1];  // input
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, acl_dst.get());
+
+    const int64_t dps2 = ne2 / ne02;
+    const int64_t dps3 = ne3 / ne03;
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            const int64_t i02 = i2 / dps2;
+            const int64_t i03 = i3 / dps3;
+
+            const int64_t  i12 = i2;
+            const int64_t  i13 = i3;
+            acl_tensor_ptr accumulator =
+                ggml_cann_create_tensor((char *) dst->data + i2 * nb2 + i3 * nb3, ggml_cann_type_mapping(dst->type),
+                                        ggml_type_size(dst->type), dst->ne, dst->nb, 2);
+
+            // The outer product needs to be accumulated in this dimension.
+            for (int64_t i1 = 0; i1 < ne11; i1++) {
+                acl_tensor_ptr acl_input = ggml_cann_create_tensor(
+                    (char *) src1->data + i1 * nb11 + i12 * nb12 + i13 * nb13, ggml_cann_type_mapping(src0->type),
+                    ggml_type_size(src0->type), src1->ne, src1->nb, 1);
+
+                acl_tensor_ptr acl_weight = ggml_cann_create_tensor(
+                    (char *) src0->data + i1 * nb01 + i02 * nb02 + i03 * nb03, ggml_cann_type_mapping(src0->type),
+                    ggml_type_size(src0->type), src0->ne, src0->nb, 1);
+
+                ggml_cann_pool_alloc output_allocator(ctx.pool());
+                void *               output_buffer = output_allocator.alloc(ggml_nbytes(dst));
+                acl_tensor_ptr       acl_out = ggml_cann_create_tensor(output_buffer, ggml_cann_type_mapping(dst->type),
+                                                                       ggml_type_size(dst->type), dst->ne, dst->nb, 2);
+
+                GGML_CANN_CALL_ACLNN_OP(ctx, Ger, acl_input.get(), acl_weight.get(), acl_out.get());
+                float       alpha_value = 1.0f;
+                aclScalar * alpha       = aclCreateScalar(&alpha_value, ACL_FLOAT);
+                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdd, accumulator.get(), acl_out.get(), alpha);
+            }
+        }
+    }
+}
+
+void ggml_cann_out_prod(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+
+    const enum ggml_type type = src0->type;
+
+    switch (type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+            ggml_cann_out_prod_fp(ctx, dst);
+            break;
+        default:
+            GGML_ABORT("Unsupport type for GGML_OP_OUT_PROD");
+            break;
+    }
+}

ggml/src/ggml-cann/aclnn_ops.h

@@ -1125,3 +1125,23 @@ void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, ac
     } while (0)
 
 #endif  // CANN_ACLNN_OPS
+
+/**
+ * @brief Performs outer product operation on two ggml tensors using the CANN backend.
+ *
+ * @details This function computes the outer product of two input tensors (src0 and src1)
+ * and stores the result in the destination tensor. The outer product operation is defined as:
+ * dst[i,j,k,l] = sum_m (src0[i,m,k,l] * src1[j,m,k,l])
+ *
+ * The function supports multiple data types including F32, F16. For floating-point
+ * types, it uses batch matrix multiplication for efficient computation.
+ *
+ * The implementation handles 4D tensor broadcasting and batch processing automatically.
+ *
+ * @param ctx The CANN backend context for operation execution and memory management.
+ * @param dst The destination ggml_tensor where the outer product result will be stored.
+ *            The input tensors are assumed to be `dst->src[0]` and `dst->src[1]`.
+ *
+ * @see GGML_CANN_CALL_ACLNN_OP for CANN operator invocation
+ */
+void ggml_cann_out_prod(ggml_backend_cann_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cann/common.h

@@ -300,30 +300,92 @@ struct ggml_cann_graph_lru_cache {
 
 struct ggml_cann_rope_cache {
     ~ggml_cann_rope_cache() {
-        if (theta_scale_cache != nullptr) {
+        if (theta_scale_cache) {
             ACL_CHECK(aclrtFree(theta_scale_cache));
         }
-        if (sin_cache != nullptr) {
+        if (sin_cache) {
             ACL_CHECK(aclrtFree(sin_cache));
         }
-        if (cos_cache != nullptr) {
+        if (cos_cache) {
             ACL_CHECK(aclrtFree(cos_cache));
         }
+        if (position_select_index) {
+            ACL_CHECK(aclrtFree(position_select_index));
+        }
+        if (theta_scale_exp_host) {
+            free(theta_scale_exp_host);
+        }
+        if(position_select_index_host) {
+            free(position_select_index_host);
+        }
     }
 
-    void *  theta_scale_cache  = nullptr;
-    int64_t theta_scale_length = 0;
+    bool equal(int64_t theta_scale_length,
+               int64_t position_length,
+               float   ext_factor,
+               float   theta_scale,
+               float   freq_scale,
+               float   attn_factor,
+               bool    is_neox,
+               bool    indep_sects,
+               bool    mrope_used,
+               bool    is_imrope,
+               int     sections[4]) {
+        return this->theta_scale_length == theta_scale_length && this->position_length == position_length &&
+               this->ext_factor == ext_factor && this->theta_scale == theta_scale && this->freq_scale == freq_scale &&
+               this->attn_factor == attn_factor && this->is_neox == is_neox && this->indep_sects == indep_sects &&
+               this->mrope_used == mrope_used && this->is_imrope == is_imrope && this->sections[0] == sections[0] &&
+               this->sections[1] == sections[1] && this->sections[2] == sections[2] && this->sections[3] == sections[3];
+    }
+
+    void set(int64_t theta_scale_length,
+             int64_t position_length,
+             float    ext_factor,
+             float   theta_scale,
+             float   freq_scale,
+             float   attn_factor,
+             bool    is_neox,
+             bool    indep_sects,
+             bool    mrope_used,
+             bool    is_imrope,
+             int     sections[4]) {
+        this->theta_scale_length = theta_scale_length;
+        this->position_length    = position_length;
+        this->ext_factor         = ext_factor;
+        this->theta_scale        = theta_scale;
+        this->freq_scale         = freq_scale;
+        this->attn_factor        = attn_factor;
+        this->is_neox            = is_neox;
+        this->indep_sects        = indep_sects;
+        this->mrope_used         = mrope_used;
+        this->is_imrope          = is_imrope;
+        this->sections[0]        = sections[0];
+        this->sections[1]        = sections[1];
+        this->sections[2]        = sections[2];
+        this->sections[3]        = sections[3];
+    }
+
+    // memory cache, prepare before inferencing.
+    void *  theta_scale_cache          = nullptr;
+    float * theta_scale_exp_host       = nullptr;
+    int *   position_select_index_host = nullptr;
+    void *  position_select_index      = nullptr;
     // sin/cos cache, used only to accelerate first layer on each device
-    void *  sin_cache          = nullptr;
-    void *  cos_cache          = nullptr;
-    int64_t position_length    = 0;
+    void *  sin_cache                  = nullptr;
+    void *  cos_cache                  = nullptr;
     // Properties to check before reusing the sincos cache
-    bool    cached             = false;
-    float   ext_factor         = 0.0f;
-    float   theta_scale        = 0.0f;
-    float   freq_scale         = 0.0f;
-    float   attn_factor        = 0.0f;
-    bool    is_neox            = false;
+    int64_t theta_scale_length         = 0;
+    int64_t position_length            = 0;
+    bool    cached                     = false;
+    float   ext_factor                 = 0.0f;
+    float   theta_scale                = 0.0f;
+    float   freq_scale                 = 0.0f;
+    float   attn_factor                = 0.0f;
+    bool    is_neox                    = false;
+    bool    indep_sects                = false;
+    bool    mrope_used                 = false;
+    int     sections[4]                = { 0, 0, 0, 0 };
+    bool    is_imrope                  = false;
 };
 
 struct ggml_cann_tensor_cache {

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1886,6 +1886,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
         case GGML_OP_FLASH_ATTN_EXT:
             ggml_cann_flash_attn_ext(ctx, dst);
             break;
+        case GGML_OP_OUT_PROD:
+            ggml_cann_out_prod(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2477,13 +2480,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                     return false;
                 }
 
-                const int mode = ((const int32_t *) op->op_params)[2];
-                if (mode & GGML_ROPE_TYPE_MROPE) {
-                    return false;
-                }
-                if (mode & GGML_ROPE_TYPE_VISION) {
-                    return false;
-                }
                 if (op->src[0]->ne[0] > 896) {
                     return false;
                 }
@@ -2563,6 +2559,16 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_OUT_PROD:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                        return true;
+                    default:
+                        return false;
+                }
+            }
         case GGML_OP_CONV_TRANSPOSE_1D:
             // TODO: ((weightL - 1) * dilationW - padLeft)=1336 should not be larger than 255.
             return (op->src[0]->ne[0] - 1) <= 255;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -224,7 +224,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
             include(CheckCXXSourceCompiles)
             set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-            set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS}")
+            string(REPLACE ";" " " ARCH_FLAGS_STR "${ARCH_FLAGS}")
+            set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS_STR}")
             foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
                 set(ARM_FEATURE "HAVE_${feature}")
                 check_cxx_source_compiles(
@@ -452,22 +453,35 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ggml-cpu/spacemit/ime_kernels.h
             )
         endif()
-        set(MARCH_STR "rv64gc")
-        if (GGML_RV_ZFH)
-            string(APPEND MARCH_STR "_zfh")
-        endif()
-        if (GGML_XTHEADVECTOR)
-            string(APPEND MARCH_STR "_xtheadvector")
-        elseif (GGML_RVV)
-            string(APPEND MARCH_STR "_v")
-            if (GGML_RV_ZVFH)
-                string(APPEND MARCH_STR "_zvfh")
+        if(NOT GGML_CPU_ALL_VARIANTS)
+            set(MARCH_STR "rv64gc")
+            if (GGML_RV_ZFH)
+                string(APPEND MARCH_STR "_zfh")
             endif()
+            if (GGML_XTHEADVECTOR)
+                string(APPEND MARCH_STR "_xtheadvector")
+            elseif (GGML_RVV)
+                string(APPEND MARCH_STR "_v")
+                if (GGML_RV_ZVFH)
+                    string(APPEND MARCH_STR "_zvfh")
+                endif()
+            endif()
+            if (GGML_RV_ZICBOP)
+                string(APPEND MARCH_STR "_zicbop")
+            endif()
+            list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
+        else()
+            # Begin with the lowest baseline
+            set(ARCH_DEFINITIONS "")
+
+            if (GGML_INTERNAL_RVV)
+                message(STATUS "RVV enabled")
+                list(APPEND ARCH_DEFINITIONS GGML_USE_RVV)
+                list(APPEND ARCH_FLAGS -march=rv64gc_v -mabi=lp64d)
+            endif()
+
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} riscv ${ARCH_DEFINITIONS})
         endif()
-        if (GGML_RV_ZICBOP)
-            string(APPEND MARCH_STR "_zicbop")
-        endif()
-        list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
         list(APPEND GGML_CPU_SOURCES

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

@@ -33,10 +33,12 @@
 // repack.cpp
 #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_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
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -44,27 +46,30 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
 // repack.cpp
+#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_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #elif defined(__POWERPC__) || defined(__powerpc__)
 // ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
@@ -76,10 +81,12 @@
 // repack.cpp
 #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_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
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -87,6 +94,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -101,10 +109,12 @@
 // repack.cpp
 #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_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
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -112,6 +122,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -134,15 +145,18 @@
 // repack.cpp
 #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_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
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -163,10 +177,12 @@
 // repack.cpp
 #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_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
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -174,6 +190,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -196,10 +213,12 @@
 // repack.cpp
 #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_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
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -207,6 +226,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0

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

@@ -24,6 +24,29 @@
 
 #define UNUSED GGML_UNUSED
 
+static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
+                                             int16x8_t *     out_mins,
+                                             int8_t *        out_scales) {
+    constexpr uint32_t kmask1 = 0x3f3f3f3f;
+    constexpr uint32_t kmask2 = 0x0f0f0f0f;
+    constexpr uint32_t kmask3 = 0x03030303;
+    constexpr uint8_t  scales_size = 12;
+
+    uint32_t sm[3];
+    memcpy(sm, scales_in, scales_size);
+
+    const uint32_t   mins_0_3 = sm[1] & kmask1;
+    const uint32_t   mins_4_7 = ((sm[2] >> 4) & kmask2) | (((sm[1] >> 6) & kmask3) << 4);
+    const uint32x2_t mins_u32 = { mins_0_3, mins_4_7 };
+
+    *out_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins_u32)));
+
+    uint32_t scales_u32[2];
+    scales_u32[0] = sm[0] & kmask1;
+    scales_u32[1] = (sm[2] & kmask2) | (((sm[0] >> 6) & kmask3) << 4);
+    memcpy(out_scales, scales_u32, 8);
+}
+
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -474,6 +497,295 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemv_q4_K_8x4_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) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_groups = ncols_interleaved / 4; // 0123 and 4567
+    const uint8x16_t m4b        = vdupq_n_u8(0x0f);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[col_groups];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < col_groups; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q4_d_0        = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q4_d_1        = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d          = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0123 = vmulq_f32(q4_d_0, q8_d);
+            float32x4_t sb_scale_4567 = vmulq_f32(q4_d_1, q8_d);
+            float32x4_t q4_dmin_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q4_dmin_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0123   = vmulq_f32(q4_dmin_0, q8_d);
+            float32x4_t sb_min_4567   = vmulq_f32(q4_dmin_1, q8_d);
+
+            // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567
+            int32x4_t bias_acc[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
+            int32x4_t acc_lo[col_groups];
+            int32x4_t acc_hi[col_groups];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_groups; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q4sb_mins[2];
+                int16x8_t q4sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q4sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                }
+
+                int8x16_t q8_qs[64 / 16];
+                for (int i = 0; i < 64 / 16; i++) {
+                    q8_qs[i] = vld1q_s8(q8_ptr[b].qs + sb * 64 + i * 16);
+                }
+
+                for (int c = 0; c < col_groups; c++) {
+                    uint8x16_t q4_cols[8];
+                    for (int i = 0; i < 8; i++) {
+                        q4_cols[i] = vld1q_u8(q4_ptr[b].qs + sb * QK_K + i * 32 + 16 * c);
+                    }
+
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[0], m4b)), q8_qs[0], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[1], m4b)), q8_qs[0], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[2], m4b)), q8_qs[0], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[3], m4b)), q8_qs[0], 3);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[4], m4b)), q8_qs[1], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[5], m4b)), q8_qs[1], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[6], m4b)), q8_qs[1], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[7], m4b)), q8_qs[1], 3);
+
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[0], 4)), q8_qs[2], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[1], 4)), q8_qs[2], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[2], 4)), q8_qs[2], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[3], 4)), q8_qs[2], 3);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[4], 4)), q8_qs[3], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[5], 4)), q8_qs[3], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[6], 4)), q8_qs[3], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[7], 4)), q8_qs[3], 3);
+                }
+
+                // Scales
+                // row c0123 blk0 and blk1
+                const int16x4_t   sc_0123_lo = vget_low_s16(q4sb_scales[0]);
+                const int16x4_t   sc_0123_hi = vget_low_s16(q4sb_scales[1]);
+                const float32x4_t sumf_0123  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[0]),
+                                                                       vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[0])));
+                acc_f32[0]                   = vfmaq_f32(acc_f32[0], sb_scale_0123, sumf_0123);
+                // row c4567 blk0 and blk1
+                const int16x4_t   sc_4567_lo = vget_high_s16(q4sb_scales[0]);
+                const int16x4_t   sc_4567_hi = vget_high_s16(q4sb_scales[1]);
+                const float32x4_t sumf_4567  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[1]),
+                                                                       vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[1])));
+                acc_f32[1]                   = vfmaq_f32(acc_f32[1], sb_scale_4567, sumf_4567);
+
+                // Bias Correction
+                const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+            }  // for sb
+
+            acc_f32[0] = vmlsq_f32(acc_f32[0], vcvtq_f32_s32(bias_acc[0]), sb_min_0123);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_4567);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // #if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q4_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_q4_K_8x8_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) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_pairs = ncols_interleaved / 2;
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[ncols_interleaved / 4];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < ncols_interleaved / 4; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q4_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q4_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q4_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q4_d_1, q8_d);
+            float32x4_t q4_dmin_0  = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q4_dmin_1  = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0   = vmulq_f32(q4_dmin_0, q8_d);
+            float32x4_t sb_min_1   = vmulq_f32(q4_dmin_1, q8_d);
+
+            // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567
+            int32x4_t bias_acc[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
+            // 2 sb each iteration
+            int32x4_t acc_lo[col_pairs];
+            int32x4_t acc_hi[col_pairs];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_pairs; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
+                int16x8_t q4sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q4sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                }
+
+                const uint8_t * q4_base = q4_ptr[b].qs + sb * QK_K;
+
+                // Load the 64 quants from q8K duplicated to use vecdots with the interelaved columns
+                // but still need the qs to use the low and hi bits from q4
+                const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
+                int8x16_t      q8_qs[8];
+                for (int i = 0; i < 8; i++) {
+                    q8_qs[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base + i * 8));
+                }
+
+                // Q4s columns iterated in pairs (01, 23, 45, 67)
+                for (int cp = 0; cp < col_pairs; cp++) {
+                    uint8x16_t q4_qs_cp_0 = vld1q_u8(q4_base + 16 * cp);
+                    uint8x16_t q4_qs_cp_1 = vld1q_u8(q4_base + 16 * cp + 64);
+                    uint8x16_t q4_qs_cp_2 = vld1q_u8(q4_base + 16 * cp + 128);
+                    uint8x16_t q4_qs_cp_3 = vld1q_u8(q4_base + 16 * cp + 192);
+
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_0, m4b)), q8_qs[0]);  // 0 .. 7
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_1, m4b)), q8_qs[1]);  // 8 ..15
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_2, m4b)), q8_qs[2]);  // 16..23
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_3, m4b)), q8_qs[3]);  // 24..31
+
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_0, 4)), q8_qs[4]);  // 32..39
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_1, 4)), q8_qs[5]);  // 40..47
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_2, 4)), q8_qs[6]);  // 48..55
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_3, 4)), q8_qs[7]);  // 56..63
+                }
+
+                // Iterates over a pair of column pairs (4 columns) to use a single 128 register
+                // p = 0 -> 0123  p2 -> 4567
+                for (int i = 0, p = 0; p < col_pairs; i++, p += 2) {
+                    int16x4_t   group_scales_lo = p == 0 ? vget_low_s16(q4sb_scales[0]) : vget_high_s16(q4sb_scales[0]);
+                    int16x4_t   group_scales_hi = p == 0 ? vget_low_s16(q4sb_scales[1]) : vget_high_s16(q4sb_scales[1]);
+                    float32x4_t sb_scale        = p == 0 ? sb_scale_0 : sb_scale_1;
+
+                    // 0123 or 4567
+                    float32x4_t sumf_0 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_lo), vpaddq_s32(acc_lo[p], acc_lo[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_0);
+
+                    float32x4_t sumf_1 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_hi), vpaddq_s32(acc_hi[p], acc_hi[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_1);
+                }
+
+                // Multiply Acc bsum + mins
+                // Each pair of subblocks share the same bsums
+                // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                // cols 0-3 bias
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+
+                // cols 4-7 bias
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+            }  // for sb
+
+            acc_f32[0] = vmlsq_f32(acc_f32[0], vcvtq_f32_s32(bias_acc[0]), sb_min_0);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_1);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, 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) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -1889,3 +2201,412 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
     ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
+
+void ggml_gemm_q4_K_8x4_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) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    acc_size  = 2 * 4;  // 2 row pairs × 4 col pairs
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[acc_size];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < acc_size; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // d4 0 1 2 3, 4 5 6 7
+                float32x4_t q4_d_0123    = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));
+                float32x4_t q4_d_4567    = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));
+                // d8 0 1 2 3
+                float32x4_t q8_d_0123    = vld1q_f32(q8_ptr[b].d);
+                // mins
+                float32x4_t q4_dmin_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));
+                float32x4_t q4_dmin_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));
+
+                // Precomputation of scales and mins
+                float32x4_t sbd_scale_0123[q8_k_blocklen];
+                float32x4_t sbd_scale_4567[q8_k_blocklen];
+                float32x4_t sbd_min_0123[q8_k_blocklen];
+                float32x4_t sbd_min_4567[q8_k_blocklen];
+
+                sbd_scale_0123[0] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 0);
+                sbd_scale_4567[0] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 0);
+                sbd_min_0123[0]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 0);
+                sbd_min_4567[0]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 0);
+
+                sbd_scale_0123[1] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 1);
+                sbd_scale_4567[1] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 1);
+                sbd_min_0123[1]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 1);
+                sbd_min_4567[1]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 1);
+
+                sbd_scale_0123[2] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 2);
+                sbd_scale_4567[2] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 2);
+                sbd_min_0123[2]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 2);
+                sbd_min_4567[2]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 2);
+
+                sbd_scale_0123[3] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 3);
+                sbd_scale_4567[3] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 3);
+                sbd_min_0123[3]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 3);
+                sbd_min_4567[3]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 3);
+
+                // Precomputation of bsums, each vpaddq calcs all the bsums for each row
+                const int16x8_t bsums[q8_k_blocklen] = {
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[QK_K / 64][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                // interleaved bias_acc: [0]->r0 0123, [1]->r1 0123, .., [4]->r0 4567, [5]->r1 4567 ..
+                int32x4_t bias_acc[acc_size];
+                for (int i = 0; i < acc_size; i++) {
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Int accumulators for qs vecdot (4 row x 2 col quartets)
+                    int32x4_t acc_lo[acc_size];
+                    int32x4_t acc_hi[acc_size];
+                    for (int i = 0; i < acc_size; i++) {
+                        acc_lo[i] = vdupq_n_s32(0);
+                        acc_hi[i] = vdupq_n_s32(0);
+                    }
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int16x8_t q4sb_scales[2];
+                    int16x8_t q4sb_mins[2];
+                    for (int i = 0; i < 2; i++) {
+                        int8_t    aux_q4sb[8];
+                        const int offset = sb * 24 + i * 12;
+                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                        q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                    }
+
+                    constexpr int reads_per_sb = 8;  // 8 * 16 bytes each => 32 qs * 4 rows
+                    for (int k = 0; k < reads_per_sb; k++) {
+                        const int8x16_t q8_blk0 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k);
+                        const int8x16_t q8_blk1 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k + 128);
+
+                        // 0..3 & 32..35
+                        const uint8x16_t q4_0123 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 32 * k);
+                        const uint8x16_t q4_4567 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 32 * k + 16);
+
+                        const int8x16_t q4_0123_lo = vreinterpretq_s8_u8(vandq_u8(q4_0123, m4b));
+                        const int8x16_t q4_0123_hi = vreinterpretq_s8_u8(vshrq_n_u8(q4_0123, 4));
+
+                        acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q4_0123_lo, q8_blk0, 0);  //  0..3  r0 c0123
+                        acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q4_0123_lo, q8_blk0, 1);  //  0..3  r1 c0123
+                        acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q4_0123_lo, q8_blk0, 2);  //  0..3  r2 c0123
+                        acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q4_0123_lo, q8_blk0, 3);  //  0..3  r3 c0123
+
+                        acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q4_0123_hi, q8_blk1, 0);  // 32..35 r0 c0123
+                        acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q4_0123_hi, q8_blk1, 1);  // 32..35 r1 c0123
+                        acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q4_0123_hi, q8_blk1, 2);  // 32..35 r2 c0123
+                        acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q4_0123_hi, q8_blk1, 3);  // 32..35 r3 c0123
+
+                        const int8x16_t q4_4567_lo = vreinterpretq_s8_u8(vandq_u8(q4_4567, m4b));
+                        const int8x16_t q4_4567_hi = vreinterpretq_s8_u8(vshrq_n_u8(q4_4567, 4));
+
+                        acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q4_4567_lo, q8_blk0, 0);  //  0..3  r0 c4567
+                        acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q4_4567_lo, q8_blk0, 1);  //  0..3  r1 c4567
+                        acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q4_4567_lo, q8_blk0, 2);  //  0..3  r2 c4567
+                        acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q4_4567_lo, q8_blk0, 3);  //  0..3  r3 c4567
+
+                        acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q4_4567_hi, q8_blk1, 0);  // 32..35 r0 c4567
+                        acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q4_4567_hi, q8_blk1, 1);  // 32..35 r1 c4567
+                        acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q4_4567_hi, q8_blk1, 2);  // 32..35 r2 c4567
+                        acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q4_4567_hi, q8_blk1, 3);  // 32..35 r3 c4567
+                    }
+
+                    // Scale and bias application
+                    // acc is stored interleaved to match output layout
+                    const int16x4_t sc_0123_lo = vget_low_s16(q4sb_scales[0]);
+                    const int16x4_t sc_4567_lo = vget_high_s16(q4sb_scales[0]);
+                    const int16x4_t sc_0123_hi = vget_low_s16(q4sb_scales[1]);
+                    const int16x4_t sc_4567_hi = vget_high_s16(q4sb_scales[1]);
+                    for (int row = 0; row < q8_k_blocklen; row++) {
+                        // Bias correction
+                        // row c0123 blk0 and blk1
+                        const float32x4_t sumf_0123 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[row]),
+                                                    vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[row])));
+                        acc_f32[2 * row] = vfmaq_f32(acc_f32[2 * row], sbd_scale_0123[row], sumf_0123);
+
+                        // row c4567 blk0 and blk1
+                        const float32x4_t sumf_4567 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[row + 4]),
+                                                    vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[row + 4])));
+                        acc_f32[2 * row + 1] = vfmaq_f32(acc_f32[2 * row + 1], sbd_scale_4567[row], sumf_4567);
+
+                        // Bias
+                        const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][row * 2]);
+                        const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][row * 2 + 1]);
+
+                        // row c0123 blk0 and blk1
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+
+                        // row c4567 blk0 and blk1
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+                    }
+                }  // for sb
+
+                for (int row = 0; row < q8_k_blocklen; row++) {
+                    acc_f32[2 * row] = vmlsq_f32(acc_f32[2 * row], vcvtq_f32_s32(bias_acc[2 * row]), sbd_min_0123[row]);
+                    acc_f32[2 * row + 1] =
+                        vmlsq_f32(acc_f32[2 * row + 1], vcvtq_f32_s32(bias_acc[2 * row + 1]), sbd_min_4567[row]);
+                }
+            }  // for b
+
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemm_q4_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemm_q4_K_8x8_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) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    constexpr int    q8_k_blocklen = 4;
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[blocklen];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < blocklen; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // bsums pairs belongs to the same q8_k subblock
+                const int16x8_t bsums[4]{
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[4][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                int32x4_t sb_acc[4];    // Aux accumulators to store subblock (partial) results
+                int32x4_t acc[8];       // rows 01 stored in [0][1][2][3] rows 23 stored in [4][5][6][7]
+                int32x4_t bias_acc[8];  // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567, [2]->r1 0123 ...
+                for (int i = 0; i < 8; i++) {
+                    acc[i]      = vdupq_n_s32(0);
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int8_t    q4sb_scales[2][8];
+                    int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
+                    for (int i = 0; i < 2; i++) {
+                        const int offset = sb * 24 + i * 12;
+                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], q4sb_scales[i]);
+                    }
+
+                    // q8_ptr[b].qs has interleaved Q8 rows (01, 23)
+                    const int8_t * q8_base = q8_ptr[b].qs + sb * 256;
+
+                    int8x16_t q8_qs_01[8];
+                    int8x16_t q8_qs_23[8];
+
+                    // Load 32-byte per row pair, 1 subblock each time
+                    for (int i = 0; i < 8; i++) {
+                        const int offset = i * 32;  // 16 for row 01, 16 for row 23
+                        q8_qs_01[i]      = vld1q_s8(q8_base + offset);
+                        q8_qs_23[i]      = vld1q_s8(q8_base + offset + 16);
+                    }
+
+                    const int8x16_t q8s[2][8] = {
+                        { q8_qs_01[0], q8_qs_01[1], q8_qs_01[2], q8_qs_01[3],
+                          q8_qs_01[4], q8_qs_01[5], q8_qs_01[6], q8_qs_01[7] },
+                        { q8_qs_23[0], q8_qs_23[1], q8_qs_23[2], q8_qs_23[3],
+                          q8_qs_23[4], q8_qs_23[5], q8_qs_23[6], q8_qs_23[7] },
+                    };
+
+                    // Q4s columns iterated in pairs (01, 23, 45, 67)
+                    for (int cp = 0; cp < ncols_interleaved / 2; cp++) {
+                        for (int i = 0; i < 4; i++) {
+                            sb_acc[i] = vdupq_n_s32(0);
+                        }
+
+                        uint8x16_t q4_qs_cp_0 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 0);    // 0 .. 7 & 32..39
+                        uint8x16_t q4_qs_cp_1 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 64);   // 8 ..15 & 40..47
+                        uint8x16_t q4_qs_cp_2 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 128);  // 16..23 & 48..55
+                        uint8x16_t q4_qs_cp_3 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 192);  // 24..31 & 56..63
+                        const int8x16_t q4_nibbles[2][4] = {
+                            {
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_0, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_1, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_2, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_3, m4b)),
+                            },
+                            {
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_0, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_1, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_2, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_3, 4)),
+                            }
+                        };
+
+                        // Calculates the Qs muladd of every row pair (rp) rows 01 and 23 of q8
+                        // for each of the internal 32 qs subblock (blk)
+                        for (int rp = 0; rp < 2; rp++) {
+                            for (int blk = 0; blk < 2; blk++) {
+                                const int8x16_t * q8  = &q8s[rp][4 * blk];
+                                const int8x16_t * q4  = q4_nibbles[blk];
+                                int32x4_t         acc = sb_acc[2 * rp + blk];
+                                // mul add for each qs in the same subblock
+                                for (int qs_offset = 0; qs_offset < 4; qs_offset++) {
+                                    acc = vmmlaq_s32(acc, q4[qs_offset], q8[qs_offset]);
+                                }
+                                sb_acc[2 * rp + blk] = acc;
+                            }
+                        }
+
+                        // Scales[i] corresponds to column i
+                        const int scale_offset = cp * 2;
+                        for (int blk = 0; blk < 2; blk++) {
+                            const int32x4_t block_scale = {
+                                (int32_t) q4sb_scales[blk][scale_offset],
+                                (int32_t) q4sb_scales[blk][scale_offset],
+                                (int32_t) q4sb_scales[blk][scale_offset + 1],
+                                (int32_t) q4sb_scales[blk][scale_offset + 1],
+                            };
+                            acc[cp]     = vmlaq_s32(acc[cp], sb_acc[blk], block_scale);
+                            acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[blk + 2], block_scale);
+                        }
+                    }
+
+                    // Multiply Acc bsum + mins
+                    for (int q8_row = 0; q8_row < 4; q8_row++) {
+                        // Each pair of subblocks share the same bsums
+                        // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                        int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][q8_row * 2]);
+                        int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][q8_row * 2 + 1]);
+
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+                    }
+                }  // for sb
+
+                // Reorder of i8mm output with bias and output layout
+                for (int i = 0; i < 8; i++) {
+                    int32x2x2_t aux = vzip_s32(vget_low_s32(acc[i]), vget_high_s32(acc[i]));
+                    acc[i]          = vcombine_s32(aux.val[0], aux.val[1]);
+                }
+                int32x4_t reorder_acc[8] = {
+                    vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1])),
+                    vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3])),
+                    vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1])),
+                    vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3])),
+                    vcombine_s32(vget_low_s32(acc[4]), vget_low_s32(acc[5])),
+                    vcombine_s32(vget_low_s32(acc[6]), vget_low_s32(acc[7])),
+                    vcombine_s32(vget_high_s32(acc[4]), vget_high_s32(acc[5])),
+                    vcombine_s32(vget_high_s32(acc[6]), vget_high_s32(acc[7])),
+                };
+
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    for (int j = 0; j < 2; j++) {
+                        float32x4_t       q8_d    = vdupq_n_f32(q8_ptr[b].d[i]);
+                        float32x4_t       q4_dmin = vcvt_f32_f16(vld1_f16((const __fp16 *) (q4_ptr[b].dmin + j * 4)));
+                        const float32x4_t dmins   = vmulq_f32(q4_dmin, q8_d);
+
+                        float32x4_t       q4_d  = vcvt_f32_f16(vld1_f16((const __fp16 *) (q4_ptr[b].d + j * 4)));
+                        const float32x4_t scale = vmulq_f32(q4_d, q8_d);
+
+                        acc_f32[2 * i + j] = vmlsq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(bias_acc[2 * i + j]), dmins);
+                        acc_f32[2 * i + j] =
+                            vmlaq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(reorder_acc[2 * i + j]), scale);
+                    }
+                }
+            }  // for b
+
+            // With the previous reorder, the tile is already in the correct memory layout.
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}

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

@@ -0,0 +1,38 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__riscv) && __riscv_xlen == 64
+#include <asm/hwprobe.h>
+#include <asm/unistd.h>
+#include <unistd.h>
+
+struct riscv64_features {
+    bool has_rvv = false;
+
+    riscv64_features() {
+        struct riscv_hwprobe probe;
+        probe.key = RISCV_HWPROBE_KEY_IMA_EXT_0;
+        probe.value = 0;
+
+        int ret = syscall(__NR_riscv_hwprobe, &probe, 1, 0, NULL, 0);
+
+        if (0 == ret) {
+            has_rvv = !!(probe.value & RISCV_HWPROBE_IMA_V);
+        }
+    }
+};
+
+static int ggml_backend_cpu_riscv64_score() {
+    int score = 1;
+    riscv64_features rf;
+
+#ifdef GGML_USE_RVV
+    if (!rf.has_rvv) { return 0; }
+    score += 1 << 1;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_riscv64_score)
+
+#endif  // __riscv && __riscv_xlen == 64

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

@@ -1927,6 +1927,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_argsort(params, tensor);
             } break;
+        case GGML_OP_TOP_K:
+            {
+                ggml_compute_forward_top_k(params, tensor);
+            } break;
         case GGML_OP_LEAKY_RELU:
             {
                 ggml_compute_forward_leaky_relu(params, tensor);
@@ -2311,6 +2315,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
+        case GGML_OP_TOP_K:
         case GGML_OP_FLASH_ATTN_EXT:
         case GGML_OP_FLASH_ATTN_BACK:
         case GGML_OP_SSM_CONV:
@@ -2834,6 +2839,10 @@ struct ggml_cplan ggml_graph_plan(
                         cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
                         cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
                     } break;
+                case GGML_OP_TOP_K:
+                    {
+                        cur += sizeof(int32_t)*node->src[0]->ne[0]*n_tasks;
+                    } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
                         const int64_t ne10 = node->src[1]->ne[0]; // DK

ggml/src/ggml-cpu/ops.cpp

@@ -7794,7 +7794,7 @@ void ggml_compute_forward_timestep_embedding(
 // ggml_compute_forward_argsort
 
 template<enum ggml_sort_order order>
-struct argsort_cmp {
+struct cmp_argsort {
     const float * data;
     bool operator()(int32_t a, int32_t b) const {
         if constexpr (order == GGML_SORT_ORDER_ASC) {
@@ -7833,11 +7833,11 @@ static void ggml_compute_forward_argsort_f32(
 
         switch (order) {
             case GGML_SORT_ORDER_ASC:
-                std::sort(dst_data, dst_data + ne0, argsort_cmp<GGML_SORT_ORDER_ASC>{src_data});
+                std::sort(dst_data, dst_data + ne0, cmp_argsort<GGML_SORT_ORDER_ASC>{src_data});
                 break;
 
             case GGML_SORT_ORDER_DESC:
-                std::sort(dst_data, dst_data + ne0, argsort_cmp<GGML_SORT_ORDER_DESC>{src_data});
+                std::sort(dst_data, dst_data + ne0, cmp_argsort<GGML_SORT_ORDER_DESC>{src_data});
                 break;
 
             default:
@@ -7864,6 +7864,72 @@ void ggml_compute_forward_argsort(
     }
 }
 
+// ggml_compute_forward_top_k
+
+struct cmp_top_k {
+    const float * data;
+    bool operator()(int32_t a, int32_t b) const {
+        return data[a] > data[b];
+    }
+};
+
+static void ggml_compute_forward_top_k_f32(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
+
+    const int top_k = ne0;
+
+    int32_t * tmp = (int32_t *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
+
+    for (int64_t i = ith; i < nr; i += nth) {
+        const float * src_data = (float *)((char *) src0->data + i*nb01);
+
+        for (int64_t j = 0; j < ne00; j++) {
+            tmp[j] = j;
+        }
+
+        std::partial_sort(tmp, tmp + top_k, tmp + ne00, cmp_top_k{src_data});
+
+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+
+        std::copy(tmp, tmp + top_k, dst_data);
+
+        // emphasize that the order is not important
+        if (top_k > 1) {
+            std::swap(dst_data[0], dst_data[1]);
+        }
+    }
+}
+
+void ggml_compute_forward_top_k(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_top_k_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_flash_attn_ext
 
 static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
@@ -9700,7 +9766,8 @@ static void ggml_compute_forward_solve_tri_f32(const struct ggml_compute_params
             }
 
             const float diag = A_batch[i00 * n + i00];
-            GGML_ASSERT(diag != 0.0f && "Zero diagonal in triangular matrix");
+            assert(diag != 0.0f && "Zero diagonal in triangular matrix");
+
             X_batch[i00 * k + i01] = (B_batch[i00 * k + i01] - sum) / diag;
         }
     }

ggml/src/ggml-cpu/ops.h

@@ -81,6 +81,7 @@ void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_top_k(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_fill(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/repack.cpp

@@ -124,6 +124,58 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
     }
 }
 
+
+void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK_K == 256);
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK_K];
+    float iscale[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+            float max = 0;
+
+            for (int j = 0; j < QK_K; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK_K + j];
+                // Update the maximum value of the corresponding super block
+                if(amax < fabsf(srcv[row_iter][j])) {
+                    amax = fabsf(srcv[row_iter][j]);
+                    max = srcv[row_iter][j];
+                }
+            }
+
+            iscale[row_iter] = amax ? -127.f/max : 0;
+
+            y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0;
+        }
+
+        for (int j = 0; j < QK_K / 4; j++) {
+            y[i].bsums[j] = 0;
+        }
+
+        // Quants values are interleaved in sequence of four bytes from corresponding super blocks
+        // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving
+        // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on
+        for (int j = 0; j < QK_K * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id     = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+            int index = (((j & 15) >> 2) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3);
+
+            float x0 = srcv[src_id][src_offset] * iscale[src_id];
+            y[i].qs[j] = nearest_int(x0);
+            y[i].bsums[index] += y[i].qs[j];
+        }
+    }
+}
+
 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK_K == 256);
     assert(k % QK_K == 0);
@@ -192,6 +244,12 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_0>(const float * GGML_RESTR
     ggml_quantize_mat_q8_0_4x8(x, vy, n_per_row);
 }
 
+template <> void ggml_quantize_mat_t<4, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    ggml_quantize_mat_q8_K_4x4(x, vy, n_per_row);
+}
+
 template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
     assert(nrow == 4);
     UNUSED(nrow);
@@ -333,6 +391,77 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemv_q4_K_8x4_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) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 4;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    float sum_minf[8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2] = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
+                uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                        sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
 void ggml_gemv_q4_K_8x8_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) {
     const int qk = QK_K;
     const int nb = n / qk;
@@ -727,6 +856,89 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemm_q4_K_8x4_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) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 4;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][8];
+    float sum_minf[4][8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2] = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
+                    uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                                sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i]);
+                                sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+
 void ggml_gemm_q4_K_8x8_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) {
     const int qk = QK_K;
     const int nb = n / qk;
@@ -1228,9 +1440,10 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block
 
     GGML_UNUSED(data_size);
 }
+
 static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q4_K);
-    GGML_ASSERT(interleave_block == 8);
+    GGML_ASSERT(interleave_block == 8 || interleave_block == 4);
     constexpr int nrows_interleaved = 8;
 
     block_q4_Kx8 * dst = (block_q4_Kx8*)t->data;
@@ -1468,6 +1681,10 @@ template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q4_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_K_to_q4_K_8_bl(t, 4, data, data_size);
+}
+
 template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
 }
@@ -1501,6 +1718,10 @@ template <> void gemv<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -1529,6 +1750,10 @@ template <> void gemm<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -1731,12 +1956,13 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             nchunk0 = (nr0 + min_chunk_size - 1) / min_chunk_size;
         }
 
-        if (nth == 1 || nchunk0 < nth || disable_chunking) {
+        int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
+        // Only increase nchunk0 to nth if it won't make chunks too small
+        if (nth == 1 || ((nchunk0 < nth || disable_chunking) && (nr0 + nth - 1) / nth >= min_chunk_size)) {
             nchunk0 = nth;
+            dr0 = (nr0 + nchunk0 - 1) / nchunk0;
         }
 
-        const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
-
         // Ensure nchunk doesn't exceed the number of rows divided by minimum chunk size
         // This prevents creating too many tiny chunks that could overlap after alignment
         const int64_t max_nchunk = (nr0 + min_chunk_size - 1) / min_chunk_size;
@@ -1930,6 +2156,9 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+
+    // instance for Q4_K
+    static const ggml::cpu::repack::tensor_traits<block_q4_K, 4, 8, GGML_TYPE_Q8_K> q4_K_8x4_q8_K;
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
 
     // instance for Q2
@@ -1961,6 +2190,16 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q4_K_8x8_q8_K;
             }
         }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_K_8x8_q8_K;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_K_8x4_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_Q2_K) {
         if (ggml_cpu_has_avx512()) {
             if (cur->ne[1] % 8 == 0) {

ggml/src/ggml-cpu/repack.h

@@ -80,10 +80,12 @@ extern "C" {
 
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_gemv_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_gemv_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_gemv_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);
+void ggml_gemv_q4_K_8x4_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_q4_K_8x8_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_q2_K_8x8_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_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);
@@ -91,6 +93,7 @@ void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 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);
+void ggml_gemm_q4_K_8x4_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_K_8x8_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_q2_K_8x8_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_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);
@@ -99,10 +102,12 @@ void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_gemv_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_gemv_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_gemv_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);
+void ggml_gemv_q4_K_8x4_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_q4_K_8x8_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_q2_K_8x8_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_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);
@@ -110,6 +115,7 @@ void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
 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);
+void ggml_gemm_q4_K_8x4_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_q4_K_8x8_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_q2_K_8x8_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_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);

ggml/src/ggml-cpu/vec.h

@@ -397,119 +397,118 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 }
 
 inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
-#if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = svcntb() * 8;
-        const int ggml_f16_epr = sve_register_length / 16;
-        const int ggml_f16_step = 8 * ggml_f16_epr;
+#if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
+    const int sve_register_length = svcntb() * 8;
+    const int ggml_f16_epr = sve_register_length / 16;
+    const int ggml_f16_step = 8 * ggml_f16_epr;
 
-        GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
+    GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
 
-        const int np= (n & ~(ggml_f16_step - 1));
+    int np = (n & ~(ggml_f16_step - 1));
 
-        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
+    svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+    svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+    for (int i = 0; i < np; i += ggml_f16_step) {
+        ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
+        ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
+        ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
+        GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
 
-            ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
+        ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
+        ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
+        ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
+        GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
 
-            ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
+        ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
+        ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
+        ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
+        GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
 
-            ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
+        ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
+        ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
+        ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
+        GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
 
-            ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
+        ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
+        ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
+        ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
+        GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
 
-            ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
+        ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
+        ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
+        ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
+        GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
 
-            ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
+        ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
+        ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
+        ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
+        GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
 
-            ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
+        ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
+        ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
+        ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
+        GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
+    }
+    const int np2 = (n & ~(ggml_f16_epr - 1));
+    for (int k = np; k < np2; k += ggml_f16_epr) {
+        svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
+        svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
+        ry = GGML_F16x_VEC_FMA(ry, rx, vx);
+
+        GGML_F16x_VEC_STORE(y + k, ry, 0);
+    }
+
+    if (np2 < n) {
+        svbool_t pg = svwhilelt_b16(np2, n);
+        svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
+        svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
+        hy = svmad_f16_x(pg, hx, vx, hy);
+        svst1_f16(pg, (__fp16 *)(y + np2), hy);
+    }
+    np = n;
+#elif defined(__riscv_zvfh) // implies __riscv_v_intrinsic
+    const int np = n;
+    _Float16 hv = (_Float16)v;
+    for (int i = 0, avl; i < n; i += avl) {
+        avl = __riscv_vsetvl_e16m8(n - i);
+        vfloat16m8_t ax = __riscv_vle16_v_f16m8((const _Float16 *)&x[i], avl);
+        vfloat16m8_t ay = __riscv_vle16_v_f16m8((_Float16 *)&y[i], avl);
+        vfloat16m8_t ny = __riscv_vfmadd_vf_f16m8(ax, hv, ay, avl);
+        __riscv_vse16_v_f16m8((_Float16 *)&y[i], ny, avl);
+    }
+#elif defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
+
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+    GGML_F16_VEC ax[GGML_F16_ARR];
+    GGML_F16_VEC ay[GGML_F16_ARR];
+
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
+
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
-        const int np2 = (n & ~(ggml_f16_epr - 1));
-        for (int k = np; k < np2; k += ggml_f16_epr) {
-            svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
-            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
-            ry = GGML_F16x_VEC_FMA(ry, rx, vx);
-
-            GGML_F16x_VEC_STORE(y + k, ry, 0);
-        }
-
-        if (np2 < n) {
-            svbool_t pg = svwhilelt_b16(np2, n);
-            svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
-            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
-            hy = svmad_f16_x(pg, hx, vx, hy);
-            svst1_f16(pg, (__fp16 *)(y + np2), hy);
-        }
-
-    #elif defined(__riscv_v_intrinsic)
-        // todo: RVV impl
-        // scalar
-        for (int i = 0; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-        }
-    #else
-        const int np = (n & ~(GGML_F16_STEP - 1));
-
-        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
-
-        GGML_F16_VEC ax[GGML_F16_ARR];
-        GGML_F16_VEC ay[GGML_F16_ARR];
-
-        for (int i = 0; i < np; i += GGML_F16_STEP) {
-            for (int j = 0; j < GGML_F16_ARR; j++) {
-                ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
-
-                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
-            }
-        }
-
-        // leftovers
-        for (int i = np; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-        }
-    #endif
+    }
 #else
-    // scalar
-    for (int i = 0; i < n; ++i) {
+    const int np = 0;
+#endif
+
+    // leftovers
+    for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
-#endif
 }
 
 // xs and vs are byte strides of x and v

ggml/src/ggml-cuda/common.cuh

@@ -84,12 +84,12 @@
 
 #define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
 #define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_PH1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // MTT S5000
 
 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
-#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
-#define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
+#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_PH1)
+#define GGML_CUDA_CC_IS_PH1(cc)      (cc >= GGML_CUDA_CC_PH1)
 
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
 #    define GGML_CUDA_USE_CUB
@@ -212,9 +212,9 @@ static const char * cu_get_error_str(CUresult err) {
 #define GGML_USE_VMM
 #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 
-#if defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
@@ -250,12 +250,14 @@ static const char * cu_get_error_str(CUresult err) {
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
 static bool fp16_available(const int cc) {
-    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
+    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
 }
 
 static bool fast_fp16_available(const int cc) {
     return GGML_CUDA_CC_IS_AMD(cc) ||
-        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610);
+        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && fp16_available(cc));
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -272,7 +274,9 @@ static bool fp16_mma_hardware_available(const int cc) {
 }
 
 static bool bf16_mma_hardware_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3 ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
 }
 
 static bool fp32_mma_hardware_available(const int cc) {
@@ -558,8 +562,12 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v
     acc += v.y*u.y;
 }
 
-static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
 #if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+#define V_DOT2_F32_F16_AVAILABLE
+#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+
+static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
+#ifdef V_DOT2_F32_F16_AVAILABLE
     asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
 #else
 #ifdef FAST_FP16_AVAILABLE
@@ -571,7 +579,7 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v,
     acc += tmpv.x * tmpu.x;
     acc += tmpv.y * tmpu.y;
 #endif // FAST_FP16_AVAILABLE
-#endif // defined(GGML_USE_HIP) && (defined(RDNA2)  || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
+#endif // V_DOT2_F32_F16_AVAILABLE
 }
 
 static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v, const half2 u) {

ggml/src/ggml-cuda/cpy.cu

@@ -86,6 +86,9 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
             }
         }
     }
+
+    GGML_UNUSED_VARS(ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+        nb12, nb13);
 }
 
 static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
@@ -202,7 +205,7 @@ static void ggml_cpy_scalar_cuda(
             ne00n = ne00;
             ne01n = ne01;
             ne02n = ne02;
-        } else if (nb00 > nb02) {
+        } else {
             ne00n = ne00;
             ne01n = ne01*ne02;
             ne02n = 1;

ggml/src/ggml-cuda/fattn-common.cuh

@@ -55,11 +55,11 @@ static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_f16(
         ggml_cuda_memcpy_1<sizeof(tmp)>(tmp, K_h2 + k_KQ_0 + (threadIdx.x % nthreads)*cpy_ne);
 #pragma unroll
         for (int k_KQ_1 = 0; k_KQ_1 < cpy_ne; ++k_KQ_1) {
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
             ggml_cuda_mad(sum,                tmp[k_KQ_1] , ((const half2  *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
 #else
             ggml_cuda_mad(sum, __half22float2(tmp[k_KQ_1]), ((const float2 *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
-#endif // FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
     }
 

ggml/src/ggml-cuda/fattn-tile.cuh

@@ -609,7 +609,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
             float KQ_sum_add = 0.0f;
 #pragma unroll
             for (int i0 = 0; i0 < nbatch_fa; i0 += np*warp_size) {
-                const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < k_VKQ_sup ?
+                const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < static_cast<uint32_t>(k_VKQ_sup) ?
                     expf(KQ_acc[(i0/(np*warp_size))*cpw + jc] - KQ_max[jc]) : 0.0f;
                 KQ_sum_add += val;
                 tmp[i0/(np*warp_size)][jc1] = val;

ggml/src/ggml-cuda/fattn-vec.cuh

@@ -86,11 +86,11 @@ static __global__ void flash_attn_ext_vec(
 
     constexpr vec_dot_KQ_t vec_dot_KQ = get_vec_dot_KQ<type_K, D, nthreads_KQ>();
     constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
     constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, half,  V_rows_per_thread>();
 #else
     constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, float, V_rows_per_thread>();
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
@@ -112,13 +112,13 @@ static __global__ void flash_attn_ext_vec(
 
     constexpr int ne_KQ      = ncols*D;
     constexpr int ne_combine = nwarps*V_cols_per_iter*D;
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
     half2            VKQ[ncols][(D/2)/nthreads_V] = {{{0.0f, 0.0f}}};
     __shared__ half   KQ[ne_KQ > ne_combine ? ne_KQ : ne_combine];
 #else
     float2           VKQ[ncols][(D/2)/nthreads_V] = {{{0.0f, 0.0f}}};
     __shared__ float  KQ[ne_KQ > ne_combine ? ne_KQ : ne_combine];
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
 
     float KQ_max[ncols];
     float KQ_sum[ncols];
@@ -129,11 +129,11 @@ static __global__ void flash_attn_ext_vec(
     }
 
     // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
     half2  Q_reg[ncols][(D/2)/nthreads_KQ]; // Will be initialized completely.
 #else
     float2 Q_reg[ncols][(D/2)/nthreads_KQ] = {{{0.0f, 0.0f}}}; // May be only partially initialized.
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
     int    Q_i32[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
     float2  Q_ds[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
     if constexpr (Q_q8_1) {
@@ -155,7 +155,7 @@ static __global__ void flash_attn_ext_vec(
                 for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
                     const int i = i0 + threadIdx.x;
 
-                    if (i0 + WARP_SIZE <= D/sizeof(int) || i < D/sizeof(int)) {
+                    if (i0 + WARP_SIZE <= int(D/sizeof(int)) || i < int(D/sizeof(int))) {
                         tmp_q_i32[i] = 0;
                     }
                 }
@@ -191,7 +191,7 @@ static __global__ void flash_attn_ext_vec(
 
         __syncthreads();
     } else {
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
         const half2 scale_h2 = make_half2(scale, scale);
 #pragma unroll
         for (int j = 0; j < ncols; ++j) {
@@ -233,7 +233,7 @@ static __global__ void flash_attn_ext_vec(
                 Q_reg[j][k].y *= scale;
             }
         }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
     }
 
     const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
@@ -272,7 +272,7 @@ static __global__ void flash_attn_ext_vec(
 
                 KQ_max_new[j] = fmaxf(KQ_max_new[j], sum);
 
-                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == i_KQ_0) {
+                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == uint32_t(i_KQ_0)) {
                     KQ_reg[j] = sum;
                 }
             }
@@ -291,7 +291,7 @@ static __global__ void flash_attn_ext_vec(
             KQ_sum[j] = KQ_sum[j]*KQ_max_scale + KQ_reg[j];
             KQ[j*nthreads + tid] = KQ_reg[j];
 
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
             const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
 #pragma unroll
             for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
@@ -303,7 +303,7 @@ static __global__ void flash_attn_ext_vec(
                 VKQ[j][i_VKQ_0/nthreads_V].x *= KQ_max_scale;
                 VKQ[j][i_VKQ_0/nthreads_V].y *= KQ_max_scale;
             }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
 
 #ifndef GGML_USE_HIP
@@ -314,7 +314,7 @@ static __global__ void flash_attn_ext_vec(
         for (int k0 = 0; k0 < WARP_SIZE; k0 += V_cols_per_iter) {
             const int k = threadIdx.y*WARP_SIZE + k0 + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V);
 
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
             half2 KQ_k[ncols];
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
@@ -353,7 +353,7 @@ static __global__ void flash_attn_ext_vec(
                     }
                 }
             }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
     }
 
@@ -374,7 +374,7 @@ static __global__ void flash_attn_ext_vec(
 
             KQ_sum[j] = KQ_sum[j]*KQ_max_scale + (threadIdx.x == 0 ? expf(sink - KQ_max[j]) : 0.0f);
 
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
             const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
 #pragma unroll
             for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
@@ -386,7 +386,7 @@ static __global__ void flash_attn_ext_vec(
                 VKQ[j][i_VKQ_0/nthreads_V].x *= KQ_max_scale;
                 VKQ[j][i_VKQ_0/nthreads_V].y *= KQ_max_scale;
             }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
     }
 
@@ -421,7 +421,7 @@ static __global__ void flash_attn_ext_vec(
         const float kqmax_scale = expf(KQ_max[j_VKQ] - kqmax_new);
         KQ_max[j_VKQ] = kqmax_new;
 
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
         half2 * VKQ_tmp = (half2 *) KQ + threadIdx.y*(V_cols_per_iter*D/2)
             + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V)*(D/2);
 
@@ -452,7 +452,7 @@ static __global__ void flash_attn_ext_vec(
             ggml_cuda_memcpy_1<V_rows_per_thread/2*sizeof(float)>(VKQ_tmp + i_VKQ,                       &VKQ[j_VKQ][i_VKQ_0/nthreads_V]);
             ggml_cuda_memcpy_1<V_rows_per_thread/2*sizeof(float)>(VKQ_tmp + i_VKQ + V_rows_per_thread/4, &VKQ[j_VKQ][i_VKQ_0/nthreads_V + V_rows_per_thread/4]);
         }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
 
         KQ_sum[j_VKQ] *= kqmax_scale;
         KQ_sum[j_VKQ] = warp_reduce_sum(KQ_sum[j_VKQ]);

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

@@ -53,6 +53,7 @@
 #include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
+#include "ggml-cuda/solve_tri.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -2717,6 +2718,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_OPT_STEP_SGD:
             ggml_cuda_opt_step_sgd(ctx, dst);
             break;
+        case GGML_OP_SOLVE_TRI:
+            ggml_cuda_op_solve_tri(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -3046,7 +3050,12 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     std::initializer_list<enum ggml_op> topk_moe_ops_delayed_softmax =
         ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true);
 
-    if (ops.size() == topk_moe_ops_with_norm.size() &&
+    const auto is_equal = [](const std::initializer_list<enum ggml_op> & list1,
+                             const std::initializer_list<enum ggml_op> & list2) {
+        return std::equal(list1.begin(), list1.end(), list2.begin(), list2.end());
+    };
+
+    if (is_equal(topk_moe_ops_with_norm, ops) &&
         ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx + 9];
@@ -3056,8 +3065,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (ops.size() == topk_moe_ops.size() &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
+    if (is_equal(topk_moe_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx + 4];
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
@@ -3065,7 +3073,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (ops.size() == topk_moe_ops_delayed_softmax.size() &&
+    if (is_equal(topk_moe_ops_delayed_softmax, ops) &&
         ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
         ggml_tensor * weights = cgraph->nodes[node_idx + 5];
@@ -3081,9 +3089,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     std::initializer_list<enum ggml_op> mul_mat_id_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_MUL_MAT_ID, GGML_OP_GLU };
     std::initializer_list<enum ggml_op> mul_mat_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_MUL_MAT,    GGML_OP_GLU };
 
-    if (ops.size() == 5 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}) ||
-                            ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}))) {
-
+    if ((is_equal(mul_mat_bias_glu_ops, ops) || is_equal(mul_mat_id_bias_glu_ops, ops)) &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 4 })) {
         const ggml_tensor * ffn_gate      = cgraph->nodes[node_idx];
         const ggml_tensor * ffn_gate_bias = cgraph->nodes[node_idx + 1];
         const ggml_tensor * ffn_up        = cgraph->nodes[node_idx + 2];
@@ -3095,9 +3102,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (ops.size() == 3 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 2}) ||
-                            ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 2}))) {
-
+    if ((is_equal(mul_mat_id_glu_ops, ops) || is_equal(mul_mat_glu_ops, ops)) &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
         const ggml_tensor * ffn_gate = cgraph->nodes[node_idx];
         const ggml_tensor * ffn_up   = cgraph->nodes[node_idx + 1];
         const ggml_tensor * glu      = cgraph->nodes[node_idx + 2];
@@ -3107,7 +3113,9 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (ops.size() == 3 && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+    std::initializer_list<enum ggml_op> rope_set_rows_ops = { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS };
+
+    if (is_equal(rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
         const ggml_tensor * rope     = cgraph->nodes[node_idx];
         const ggml_tensor * view     = cgraph->nodes[node_idx + 1];
         const ggml_tensor * set_rows = cgraph->nodes[node_idx + 2];
@@ -3837,7 +3845,7 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
 
     // Check if UMA is explicitly enabled via environment variable
     bool uma_env = getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr;
-    bool is_uma = prop.unifiedAddressing > 0 || uma_env;
+    bool is_uma = prop.integrated > 0 || uma_env;
 
     if (is_uma) {
         // For UMA systems (like DGX Spark), use system memory info
@@ -4255,6 +4263,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
             return true;
+        case GGML_OP_SOLVE_TRI:
+            return op->src[0]->ne[0] <= 64 && op->src[1]->ne[0] <= 32;
         default:
             return false;
     }

ggml/src/ggml-cuda/mma.cuh

@@ -73,34 +73,7 @@ namespace ggml_cuda_mma {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
 
-#if defined(GGML_USE_HIP)
-#if defined(RDNA4)
-        static constexpr int ne = I * J / 32;
-        T x[ne] = {0};
-
-        static constexpr __device__ bool supported() {
-            if (I == 16 && J == 16) return true;
-            return false;
-        }
-
-        static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 16 && J == 16) {
-                return 8 * (threadIdx.x / 16) + l;
-            } else {
-                NO_DEVICE_CODE;
-                return -1;
-            }
-        }
-
-        static __device__ __forceinline__ int get_j(const int l) {
-            if constexpr (I == 16 && J == 16) {
-                return threadIdx.x % 16;
-            } else {
-                NO_DEVICE_CODE;
-                return -1;
-            }
-        }
-#else
+#if defined(AMD_MFMA_AVAILABLE)
         static constexpr int ne = I * J / 64;
         T x[ne] = {0};
 
@@ -146,7 +119,6 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif // defined(RDNA4)
 #elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
@@ -177,6 +149,34 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
+#elif defined(AMD_WMMA_AVAILABLE)
+#if defined(RDNA4)
+        static constexpr int ne = I * J / 32;
+        T x[ne] = {0};
+
+        static constexpr __device__ bool supported() {
+            if (I == 16 && J == 16) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 16 && J == 16) {
+                return 8 * (threadIdx.x / 16) + l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 16 && J == 16) {
+                return threadIdx.x % 16;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#endif
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
@@ -437,7 +437,29 @@ namespace ggml_cuda_mma {
             xi[0] = xs[0];
         }
 #elif defined(AMD_WMMA_AVAILABLE)
-        ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
+        if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
+            ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
+
+        } else if constexpr (std::is_same_v<T, int>) {
+            if constexpr (I == 16 && J == 4) {
+                int64_t * xi = (int64_t *) t.x;
+                const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I));
+                xi[0] = xs[0];
+
+            }else if constexpr (I == 16 && J == 8) {
+                int64_t * xi = (int64_t *) t.x;
+                const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I));
+                xi[0] = xs[0];
+
+                const int64_t * xs1 = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I) + 2);
+                xi[1] = xs1[0];
+
+            }else{
+                NO_DEVICE_CODE;
+            }
+        } else {
+            NO_DEVICE_CODE;
+        }
 #else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
@@ -772,6 +794,36 @@ namespace ggml_cuda_mma {
                                                       acc[0],
                                                       0, 0, 0);
 #endif // defined(CDNA3)
+
+#elif defined(AMD_WMMA_AVAILABLE)
+        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+        int32x2_t * a_vec = (int32x2_t *) A.x;
+        int32x2_t * b_vec = (int32x2_t *) B.x;
+
+        using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
+        int32x8_t * acc = (int32x8_t *) D.x;
+
+#if defined(RDNA4)
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+            true,
+            a_vec[0],
+            true,
+            b_vec[0],
+            acc[0],
+            true
+        );
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+            true,
+            a_vec[1],
+            true,
+            b_vec[1],
+            acc[0],
+            true
+        );
+#endif // defined(RDNA4)
+
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -798,6 +850,7 @@ namespace ggml_cuda_mma {
                                                      acc[0],
                                                      0, 0, 0);
 #endif // defined(CDNA3)
+
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -836,10 +889,37 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
             : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
 #else
-        tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D;
-        tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A;
+        tile      <16, 8, float> * D16 = reinterpret_cast<tile      <16, 8, float> *>(&D);
+        const tile<16, 8, half2> * A16 = reinterpret_cast<const tile<16, 8, half2> *>(&A);
         mma(D16[0], A16[0], B);
         mma(D16[1], A16[1], B);
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
     }
+
+static __device__ __forceinline__ void mma(
+            tile<16, 16, int> & D, const tile<16, 4, int> & A, const tile<16, 4, int> & B) {
+#if defined(AMD_WMMA_AVAILABLE)
+    using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+    int32x2_t * a_vec = (int32x2_t *) A.x;
+    int32x2_t * b_vec = (int32x2_t *) B.x;
+
+    using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
+    int32x8_t * acc = (int32x8_t *) D.x;
+
+    acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+        true,
+        a_vec[0],
+        true,
+        b_vec[0],
+        acc[0],
+        false
+    );
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif
+    }
 }
+

ggml/src/ggml-cuda/mmf.cu

@@ -151,7 +151,7 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
             return false;
         }
     } else {
-        if (src1_ncols > 16 || GGML_CUDA_CC_IS_RDNA4(cc)) {
+        if (src1_ncols > 16) {
             return false;
         }
     }

ggml/src/ggml-cuda/mmq.cu

@@ -306,5 +306,11 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return false;
     }
 
-    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    if (amd_wmma_available(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+            return true;
+        }
+    }
+
+    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }

ggml/src/ggml-cuda/solve_tri.cu

@@ -0,0 +1,203 @@
+#include "common.cuh"
+#include "ggml.h"
+#include "solve_tri.cuh"
+
+#define MAX_N_FAST 64
+#define MAX_K_FAST 32
+
+// ======================
+// Fast Kernel (n <= 64, k <= 32) - Warp-based parallel reduction
+// ======================
+// When ncols_template == 0 the bounds for the loops in this function are not
+// known and can't be unrolled. As we want to keep pragma unroll for all other
+// cases we supress the clang transformation warning here.
+#ifdef __clang__
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wpass-failed"
+#endif  // __clang__
+template <int n_template, int k_template>
+static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
+                                          const float * __restrict__ B,
+                                          float * __restrict__ X,
+                                          const uint3  ne02,
+                                          const size_t nb02,
+                                          const size_t nb03,
+                                          const size_t nb12,
+                                          const size_t nb13,
+                                          const size_t nb2,
+                                          const size_t nb3,
+                                          const int    n_arg,
+                                          const int    k_arg) {
+    const int n = n_template == 0 ? n_arg : n_template;
+    const int k = k_template == 0 ? k_arg : k_template;
+
+    const int batch_idx = blockIdx.x;
+    const int lane      = threadIdx.x;
+    const int col_idx   = threadIdx.y;
+
+    if (col_idx >= k) {
+        return;
+    }
+
+    const uint2   i02_i03 = fast_div_modulo(batch_idx, ne02);
+    const int64_t i02     = i02_i03.y;
+    const int64_t i03     = i02_i03.x;
+
+    const float * const A_batch = (const float *) (A + i02 * nb02 + i03 * nb03);
+    const float * const B_batch = (const float *) (B + i02 * nb12 + i03 * nb13);
+    float *             X_batch = (float *) (X + i02 * nb2 + i03 * nb3);
+
+    __shared__ float sA[MAX_N_FAST * MAX_N_FAST];
+    __shared__ float sXt[MAX_N_FAST * (MAX_K_FAST + 1)];
+
+    const int offset = threadIdx.x + threadIdx.y * blockDim.x;
+
+#pragma unroll
+    for (int i = 0; i < n * n; i += k * WARP_SIZE) {
+        int i0 = i + offset;
+        if (i0 < n * n) {
+            sA[i0] = A_batch[i0];
+        }
+    }
+
+    const int rows_per_warp = (n + WARP_SIZE - 1) / WARP_SIZE;
+
+#pragma unroll
+    for (int i = 0; i < rows_per_warp; i++) {
+        const int i0 = lane + i * WARP_SIZE;
+        if (i0 < n) {
+            sXt[col_idx * n + i0] = B_batch[i0 * k + col_idx];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int row = 0; row < n; ++row) {
+        float sum = 0.0f;
+
+        {
+            int j = lane;
+            if (j < row) {
+                sum += sA[row * n + j] * sXt[col_idx * n + j];
+            }
+        }
+        if (row >= WARP_SIZE) {
+            int j = WARP_SIZE + lane;
+            if (j < row) {
+                sum += sA[row * n + j] * sXt[col_idx * n + j];
+            }
+        }
+
+        sum = warp_reduce_sum(sum);
+
+        if (lane == 0) {
+            const float b_val      = sXt[col_idx * n + row];
+            const float a_diag     = sA[row * n + row];
+            // no safeguards for division by zero because that indicates corrupt
+            // data anyway
+            sXt[col_idx * n + row] = (b_val - sum) / a_diag;
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int i = 0; i < rows_per_warp; i++) {
+        const int i0 = lane + i * WARP_SIZE;
+        if (i0 < n) {
+            X_batch[i0 * k + col_idx] = sXt[col_idx * n + i0];
+        }
+    }
+}
+#ifdef __clang__
+#    pragma clang diagnostic pop
+#endif  // __clang__
+
+static void solve_tri_f32_cuda(const float * A,
+                               const float * B,
+                               float *       X,
+                               int           n,
+                               int           k,
+                               int64_t       ne02,
+                               int64_t       ne03,
+                               size_t        nb02,
+                               size_t        nb03,
+                               size_t        nb12,
+                               size_t        nb13,
+                               size_t        nb2,
+                               size_t        nb3,
+                               cudaStream_t  stream) {
+    const uint3 ne02_fd = init_fastdiv_values((uint32_t) ne02);
+    dim3        threads(WARP_SIZE, k);
+    dim3        grid(ne02 * ne03);
+    if (n == 64) {
+        switch (k) {
+            case 32:
+                solve_tri_f32_fast<64, 32>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 16:
+                solve_tri_f32_fast<64, 16>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 14:
+                solve_tri_f32_fast<64, 14>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 12:
+                solve_tri_f32_fast<64, 12>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 10:
+                solve_tri_f32_fast<64, 10>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 8:
+                solve_tri_f32_fast<64, 8>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 6:
+                solve_tri_f32_fast<64, 6>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 4:
+                solve_tri_f32_fast<64, 4>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 2:
+                solve_tri_f32_fast<64, 2>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 1:
+                solve_tri_f32_fast<64, 1>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            default:
+                solve_tri_f32_fast<0, 0>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, n, k);
+        }
+    } else {  // run general case
+        solve_tri_f32_fast<0, 0>
+            <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, n, k);
+    }
+}
+
+void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];  // A (triangular n x x matrix)
+    const ggml_tensor * src1 = dst->src[1];  // B (right hand side of n x k equation columns)
+
+    ggml_is_contiguous(src0);
+    ggml_is_contiguous(src1);
+
+    const int64_t n = src0->ne[0];
+    const int64_t k = src1->ne[0];
+
+    GGML_ASSERT(n <= 64);
+    GGML_ASSERT(k <= 32);
+
+    solve_tri_f32_cuda((const float *) src0->data, (const float *) src1->data, (float *) dst->data, n, k, src0->ne[2],
+                       src0->ne[3], src0->nb[2] / sizeof(float), src0->nb[3] / sizeof(float),
+                       src1->nb[2] / sizeof(float), src1->nb[3] / sizeof(float), dst->nb[2] / sizeof(float),
+                       dst->nb[3] / sizeof(float), ctx.stream());
+}

ggml/src/ggml-cuda/solve_tri.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -2229,7 +2229,7 @@ static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess
 
     int mode = op_params[2];
 
-    if ((mode & GGML_ROPE_TYPE_NEOX) || (mode & GGML_ROPE_TYPE_MROPE) || (mode & GGML_ROPE_TYPE_VISION)) {
+    if ((mode & GGML_ROPE_TYPE_MROPE) || (mode & GGML_ROPE_TYPE_VISION)) {
         return false;
     }
     if (mode & 1) {

ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -24,6 +24,10 @@
 #include "hvx-utils.h"
 #include "ops-utils.h"
 
+// Redefined the types GGML_ROPE_TYPE_NORMAL & GGML_ROPE_TYPE_NEOX as we cant include ggml.h
+#define HTP_ROPE_TYPE_NORMAL 0
+#define HTP_ROPE_TYPE_NEOX   2
+
 #define htp_rope_preamble              \
     const uint32_t ne00 = src0->ne[0]; \
     const uint32_t ne01 = src0->ne[1]; \
@@ -146,6 +150,57 @@ static void init_rope_ctx(struct rope_th_ctx * rope_ctx, struct htp_ops_context
          rope_ctx->ext_factor, rope_ctx->theta_scale, rope_ctx->attn_factor);
 }
 
+static void hvx_calc_rope_neox_f32(const float * restrict src0,
+                              float * restrict dst,
+                              const int num_elems,
+                              const float * restrict theta_cache) {
+    // for (int i = 0; i < num_elems; i += 2) {
+    //const float cos_theta = theta_cache[i + 0];
+    //const float sin_theta = theta_cache[i + 1];
+
+    //const float x0 = src[0];
+    //const float x1 = src[num_elems/2];
+
+    //dst[0] = x0*cos_theta - x1*sin_theta;
+    //dst[num_elems/2] = x0*sin_theta + x1*cos_theta;
+
+    //src += 1;
+    //dst += 1;
+    // }
+
+    const uint8_t * restrict src0_curr  = (const uint8_t *) src0;
+    const uint8_t * restrict theta_curr = (const uint8_t *) theta_cache;
+    uint8_t * restrict dst_curr         = (uint8_t *) dst;
+
+    int step_of_1 = num_elems >> 6;  // 6 because we process two vectors at once
+    int half_size = (sizeof(float) * (num_elems / 2));
+
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v0 = *(HVX_Vector *) src0_curr;
+        HVX_Vector v1 = *(HVX_Vector *) (src0_curr + half_size);
+
+        HVX_Vector v2 = *(HVX_Vector *) theta_curr;
+        HVX_Vector v3 = *(HVX_Vector *) (theta_curr + VLEN);
+
+        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);  // vcos_sin[0] = cos_theta, vcos_sin[1] = sin_theta
+
+        HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_lo_W(vcos_sin));
+        HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_hi_W(vcos_sin));
+        HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(v1, Q6_V_lo_W(vcos_sin));
+        HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(v1, Q6_V_hi_W(vcos_sin));
+
+        HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
+        HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
+
+        *(HVX_Vector *) dst_curr          = Q6_Vsf_equals_Vqf32(v4);
+        *(HVX_Vector *) (dst_curr + half_size) = Q6_Vsf_equals_Vqf32(v5);
+
+        src0_curr += VLEN;
+        theta_curr += 2 * VLEN;
+        dst_curr += VLEN;
+    }
+}
+
 static void hvx_calc_rope_f32(const float * restrict src0,
                               float * restrict dst,
                               const int num_elems,
@@ -212,6 +267,9 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
     const struct htp_tensor * src2 = &octx->src2;
     struct htp_tensor *       dst  = &octx->dst;
 
+    const int32_t mode  = rope_ctx->mode;
+    const bool is_neox  = mode & HTP_ROPE_TYPE_NEOX;
+
     htp_rope_preamble;
 
     const int32_t * pos = (const int32_t *) src1->data;
@@ -247,20 +305,35 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
                 float *       dst_data_loc = dst_data;
 
                 if (1 == opt_path) {
-                    hvx_calc_rope_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
+                    if (is_neox) {
+                        hvx_calc_rope_neox_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
+                    } else {
+                        hvx_calc_rope_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
+                    }
                 } else {
                     for (uint32_t i0 = 0; i0 < rope_ctx->n_dims; i0 += 2) {
                         const float cos_theta = wp0[i0 + 0];
                         const float sin_theta = wp0[i0 + 1];
 
-                        const float x0 = src_loc[0];
-                        const float x1 = src_loc[1];
+                        if (is_neox) {
+                            const float x0 = src_loc[0];
+                            const float x1 = src_loc[rope_ctx->n_dims/2];
 
-                        dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
-                        dst_data_loc[1] = x0 * sin_theta + x1 * cos_theta;
+                            dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
+                            dst_data_loc[rope_ctx->n_dims/2] = x0 * sin_theta + x1 * cos_theta;
 
-                        src_loc += 2;
-                        dst_data_loc += 2;
+                            src_loc += 1;
+                            dst_data_loc += 1;
+                        } else {
+                            const float x0 = src_loc[0];
+                            const float x1 = src_loc[1];
+
+                            dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
+                            dst_data_loc[1] = x0 * sin_theta + x1 * cos_theta;
+
+                            src_loc += 2;
+                            dst_data_loc += 2;
+                        }
                     }
                 }
 

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

@@ -1009,6 +1009,64 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge(ggml_metal_l
     return res;
 }
 
+// note: reuse the argsort kernel for top_k
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_TOP_K);
+
+    char base[256];
+    char name[256];
+
+    // note: the top_k kernel is always descending order
+    ggml_sort_order order = GGML_SORT_ORDER_DESC;
+
+    const char * order_str = "undefined";
+    switch (order) {
+        case GGML_SORT_ORDER_ASC:  order_str = "asc";  break;
+        case GGML_SORT_ORDER_DESC: order_str = "desc"; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_argsort_%s_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->type), order_str);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k_merge(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_TOP_K);
+
+    char base[256];
+    char name[256];
+
+    ggml_sort_order order = GGML_SORT_ORDER_DESC;
+
+    const char * order_str = "undefined";
+    switch (order) {
+        case GGML_SORT_ORDER_ASC:  order_str = "asc";  break;
+        case GGML_SORT_ORDER_DESC: order_str = "desc"; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_argsort_merge_%s_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->type), order_str);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,

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

@@ -128,6 +128,8 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id         (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argmax            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);

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

@@ -905,6 +905,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_LEAKY_RELU:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ARGSORT:
+        case GGML_OP_TOP_K:
         case GGML_OP_ARANGE:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:

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

@@ -832,14 +832,19 @@ typedef struct {
 } ggml_metal_kargs_leaky_relu;
 
 typedef struct {
-    int64_t  ne00;
-    int64_t  ne01;
-    int64_t  ne02;
-    int64_t  ne03;
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
     uint64_t nb00;
     uint64_t nb01;
     uint64_t nb02;
     uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    int32_t  top_k;
 } ggml_metal_kargs_argsort;
 
 typedef struct {
@@ -851,6 +856,11 @@ typedef struct {
     uint64_t nb01;
     uint64_t nb02;
     uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    int32_t  top_k;
     int32_t  len;
 } ggml_metal_kargs_argsort_merge;
 

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

@@ -406,6 +406,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_argsort(ctx, idx);
             } break;
+        case GGML_OP_TOP_K:
+            {
+                n_fuse = ggml_metal_op_top_k(ctx, idx);
+            } break;
         case GGML_OP_LEAKY_RELU:
             {
                 n_fuse = ggml_metal_op_leaky_relu(ctx, idx);
@@ -3678,14 +3682,19 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
     }
 
     ggml_metal_kargs_argsort args = {
-        /*.ne00 =*/ ne00,
-        /*.ne01 =*/ ne01,
-        /*.ne02 =*/ ne02,
-        /*.ne03 =*/ ne03,
-        /*.nb00 =*/ nb00,
-        /*.nb01 =*/ nb01,
-        /*.nb02 =*/ nb02,
-        /*.nb03 =*/ nb03,
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.top_k =*/ nth,
     };
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
@@ -3705,15 +3714,20 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
         ggml_metal_op_concurrency_reset(ctx);
 
         ggml_metal_kargs_argsort_merge args_merge = {
-            .ne00 = ne00,
-            .ne01 = ne01,
-            .ne02 = ne02,
-            .ne03 = ne03,
-            .nb00 = nb00,
-            .nb01 = nb01,
-            .nb02 = nb02,
-            .nb03 = nb03,
-            .len  = len,
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.ne03  =*/ ne03,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne0   =*/ ne0,
+            /*.ne1   =*/ ne1,
+            /*.ne2   =*/ ne2,
+            /*.ne3   =*/ ne3,
+            /*.top_k =*/ ne00,
+            /*.len   =*/ len,
         };
 
         // merges per row
@@ -3737,6 +3751,118 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_top_k(lib, op);
+
+    // bitonic sort requires the number of elements to be power of 2
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    // blocks per row
+    const int npr = (ne00 + nth - 1)/nth;
+
+    const size_t smem = GGML_PAD(nth*sizeof(int32_t), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += sizeof(int32_t)*ggml_nelements(op->src[0]);
+
+    if ((int) ceil(std::log(npr) / std::log(2)) % 2 == 1) {
+        std::swap(bid_dst, bid_tmp);
+    }
+
+    const int top_k = ne0;
+
+    ggml_metal_kargs_argsort args = {
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.top_k =*/ std::min(nth, top_k), // for each block, keep just the top_k indices
+    };
+
+    if (npr > 1) {
+        args.ne0 = (npr - 1)*args.top_k + std::min(ne00 - (npr - 1)*nth, args.top_k);
+    }
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
+
+    ggml_metal_pipeline_t pipeline_merge = ggml_metal_library_get_pipeline_top_k_merge(lib, op);
+
+    int len = args.top_k;
+
+    while (len < args.ne0) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        // merges per row
+        const int nm = (args.ne0 + 2*len - 1) / (2*len);
+
+        const int nth = std::min(512, std::min(len, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge)));
+
+        ggml_metal_kargs_argsort_merge args_merge = {
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.ne03  =*/ ne03,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne0   =*/ args.ne0,
+            /*.ne1   =*/ ne1,
+            /*.ne2   =*/ ne2,
+            /*.ne3   =*/ ne3,
+            /*.top_k =*/ nm == 1 ? top_k : args.ne0, // the final merge outputs top_k elements
+            /*.len   =*/ len,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_merge);
+        ggml_metal_encoder_set_bytes   (enc, &args_merge, sizeof(args_merge), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, nm*ne01, ne02, ne03, nth, 1, 1);
+
+        std::swap(bid_dst, bid_tmp);
+
+        len <<= 1;
+    }
+
+    return 1;
+}
+
 int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

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

@@ -81,6 +81,7 @@ int ggml_metal_op_arange            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argmax            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argsort           (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_top_k             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_leaky_relu        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_opt_step_adamw    (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_opt_step_sgd      (ggml_metal_op_t ctx, int idx);

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

@@ -202,6 +202,10 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
             {
                 res *= 2;
             } break;
+        case GGML_OP_TOP_K:
+            {
+                res = 2*sizeof(int32_t)*ggml_nelements(tensor->src[0]);
+            } break;
         default:
             break;
     }

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

@@ -4670,11 +4670,12 @@ kernel void kernel_argsort_f32_i32(
         ushort3   ntg[[threads_per_threadgroup]]) {
     // bitonic sort
     const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
 
-    const int i00 = (tgpig[0]/args.ne01)*ntg.x;
-    const int i01 =  tgpig[0]%args.ne01;
-    const int i02 =  tgpig[1];
-    const int i03 =  tgpig[2];
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
 
     device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
 
@@ -4710,9 +4711,11 @@ kernel void kernel_argsort_f32_i32(
         }
     }
 
+    const int64_t i0 = ib*args.top_k;
+
     // copy the result to dst without the padding
-    if (i00 + col < args.ne00) {
-        dst += i00 + args.ne00*i01 + args.ne00*args.ne01*i02 + args.ne00*args.ne01*args.ne02*i03;
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
 
         dst[col] = shmem_i32[col];
     }
@@ -4747,22 +4750,22 @@ kernel void kernel_argsort_merge_f32_i32(
 
     const int start = im * (2 * args.len);
 
-    const int len0 = MIN(args.len, MAX(0, args.ne00 - (int)(start)));
-    const int len1 = MIN(args.len, MAX(0, args.ne00 - (int)(start + args.len)));
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
 
     const int total = len0 + len1;
 
     device const int32_t * tmp0 = tmp + start
-        + i01*args.ne00
-        + i02*args.ne00*args.ne01
-        + i03*args.ne00*args.ne01*args.ne02;
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
 
     device const int32_t * tmp1 = tmp0 + args.len;
 
     dst += start
-        + i01*args.ne00
-        + i02*args.ne00*args.ne01
-        + i03*args.ne00*args.ne01*args.ne02;
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
 
     device const float * src0_row = (device const float *)(src0
         + args.nb01*i01
@@ -4776,7 +4779,11 @@ kernel void kernel_argsort_merge_f32_i32(
     const int chunk = (total + ntg.x - 1) / ntg.x;
 
     const int k0 = tpitg.x * chunk;
-    const int k1 = min(k0 + chunk, total);
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
 
     if (k0 >= total) {
         return;

ggml/src/ggml-opencl/CMakeLists.txt

@@ -70,6 +70,7 @@ set(GGML_OPENCL_KERNELS
     group_norm
     im2col_f32
     im2col_f16
+    mean
     mul_mat_Ab_Bi_8x4
     mul_mv_f16_f16
     mul_mv_f16_f32_1row
@@ -109,6 +110,9 @@ set(GGML_OPENCL_KERNELS
     softmax_4_f16
     softmax_f32
     softmax_f16
+    sqr
+    sqrt
+    ssm_conv
     sub
     sum_rows
     transpose

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -449,6 +449,9 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
     cl_kernel kernel_add_id;
     cl_kernel kernel_scale;
+    cl_kernel kernel_sqr_cont_f32, kernel_sqr_cont_f32_4, kernel_sqr_cont_f16, kernel_sqr_cont_f16_4;
+    cl_kernel kernel_sqrt_cont_f32, kernel_sqrt_cont_f32_4, kernel_sqrt_cont_f16, kernel_sqrt_cont_f16_4;
+    cl_kernel kernel_mean_f32;
     cl_kernel kernel_silu, kernel_silu_4;
     cl_kernel kernel_gelu, kernel_gelu_4;
     cl_kernel kernel_gelu_erf, kernel_gelu_erf_4;
@@ -509,6 +512,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_conv_2d_f16;
     cl_kernel kernel_conv_2d_f32;
     cl_kernel kernel_conv_2d_f16_f32;
+    cl_kernel kernel_ssm_conv_f32_f32, kernel_ssm_conv_f32_f32_4;
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_gemv_moe_mxfp4_f32, kernel_gemm_moe_mxfp4_f32;
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
@@ -1552,6 +1556,66 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // sqr
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sqr.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sqr.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f32     = clCreateKernel(prog, "kernel_sqr_cont_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f32_4   = clCreateKernel(prog, "kernel_sqr_cont_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f16     = clCreateKernel(prog, "kernel_sqr_cont_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f16_4   = clCreateKernel(prog, "kernel_sqr_cont_f16_4", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // sqrt
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sqrt.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sqrt.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f32     = clCreateKernel(prog, "kernel_sqrt_cont_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f32_4   = clCreateKernel(prog, "kernel_sqrt_cont_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f16     = clCreateKernel(prog, "kernel_sqrt_cont_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f16_4   = clCreateKernel(prog, "kernel_sqrt_cont_f16_4", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mean
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mean.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mean.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mean_f32 = clCreateKernel(prog, "kernel_mean_f32", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // sub
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1825,6 +1889,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
                 }
     }
 
+    // ssm_conv
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "ssm_conv.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("ssm_conv.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_ssm_conv_f32_f32   = clCreateKernel(prog, "kernel_ssm_conv_f32_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_ssm_conv_f32_f32_4 = clCreateKernel(prog, "kernel_ssm_conv_f32_f32_4", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_id_q4_0_f32_8x_flat
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2959,6 +3041,10 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                    (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
         case GGML_OP_ADD_ID:
             return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                    ggml_is_contiguous(op->src[0]);
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
@@ -3007,6 +3093,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16) ||
                    (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
                    (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
+        case GGML_OP_SSM_CONV:
+            return (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
         case GGML_OP_CONCAT:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -3075,6 +3163,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return cols <= max_workgroup_size && op->src[0]->type == GGML_TYPE_F32;
         }
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
         case GGML_OP_FLASH_ATTN_EXT:
             {
@@ -5193,6 +5282,224 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
     }
 }
 
+static void ggml_cl_sqr(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    // Currently assumes src0 is contiguous
+    int n = ggml_nelements(dst);
+    if (n % 4 == 0) {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqr_cont_f32_4;
+        } else {
+            kernel = backend_ctx->kernel_sqr_cont_f16_4;
+        }
+        n /= 4;
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqr_cont_f32;
+        } else {
+            kernel = backend_ctx->kernel_sqr_cont_f16;
+        }
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_sqrt(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    // Currently assumes src0 is contiguous
+    int n = ggml_nelements(dst);
+    if (n % 4 == 0) {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqrt_cont_f32_4;
+        } else {
+            kernel = backend_ctx->kernel_sqrt_cont_f16_4;
+        }
+        n /= 4;
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqrt_cont_f32;
+        } else {
+            kernel = backend_ctx->kernel_sqrt_cont_f16;
+        }
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_mean(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_UNUSED(src1);
+
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_mean_f32;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb3));
+
+    size_t global_work_size[] = {(size_t)ne01, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_ssm_conv(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    int ne01 = src0->ne[1];
+    cl_ulong nb00 = src0->nb[0];
+    cl_ulong nb01 = src0->nb[1];
+    cl_ulong nb02 = src0->nb[2];
+
+    int ne10 = src1->ne[0];
+    cl_ulong nb11 = src1->nb[1];
+
+    int ne1  = dst->ne[1];
+    int ne2  = dst->ne[2];
+    cl_ulong nb0 = dst->nb[0];
+    cl_ulong nb1 = dst->nb[1];
+    cl_ulong nb2 = dst->nb[2];
+
+    cl_kernel kernel = backend_ctx->kernel_ssm_conv_f32_f32;
+
+    if (ne10 % 4 == 0) {
+        kernel = backend_ctx->kernel_ssm_conv_f32_f32_4;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb2));
+
+    size_t global_work_size[] = {(size_t)ne01, (size_t)ne1, (size_t)ne2};
+    size_t local_work_size[]  = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (ne01 % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
 static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -9091,6 +9398,24 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_sub;
             break;
+        case GGML_OP_SQR:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_sqr;
+            break;
+        case GGML_OP_SQRT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_sqrt;
+            break;
+        case GGML_OP_MEAN:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mean;
+            break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(tensor)) {
                 case GGML_UNARY_OP_GELU:
@@ -9192,6 +9517,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_conv_2d;
             break;
+        case GGML_OP_SSM_CONV:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_ssm_conv;
+            break;
         case GGML_OP_CONCAT:
             if (!any_on_device) {
                 return false;

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

@@ -0,0 +1,39 @@
+
+kernel void kernel_mean_f32(
+    global float *  src0,
+    ulong           offset0,
+    global float *  dst,
+    ulong           offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int i3 = get_global_id(2);
+    int i2 = get_global_id(1);
+    int i1 = get_global_id(0);
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    global float * dst_row = (global float *) ((global char *) dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float row_sum = 0;
+
+    for (int i0 = 0; i0 < ne00; i0++) {
+        row_sum += src_row[i0];
+    }
+
+    dst_row[0] = row_sum / ne00;
+}

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

@@ -0,0 +1,53 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_sqr_cont_f32(
+    global float * src0,
+    ulong          offset0,
+    global float * dst,
+    ulong          offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f32_4(
+    global float4 * src0,
+    ulong           offset0,
+    global float4 * dst,
+    ulong           offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f16(
+    global half * src0,
+    ulong         offset0,
+    global half * dst,
+    ulong         offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f16_4(
+    global half4 * src0,
+    ulong          offset0,
+    global half4 * dst,
+    ulong          offsetd
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}

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

@@ -0,0 +1,53 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_sqrt_cont_f32(
+    global float * src0,
+    ulong          offset0,
+    global float * dst,
+    ulong          offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = sqrt(src0[gid]);
+}
+
+kernel void kernel_sqrt_cont_f32_4(
+    global float4 * src0,
+    ulong           offset0,
+    global float4 * dst,
+    ulong           offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = sqrt(src0[gid]);
+}
+
+kernel void kernel_sqrt_cont_f16(
+    global half * src0,
+    ulong         offset0,
+    global half * dst,
+    ulong         offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = convert_half(sqrt(convert_float(src0[gid])));
+}
+
+kernel void kernel_sqrt_cont_f16_4(
+    global half4 * src0,
+    ulong          offset0,
+    global half4 * dst,
+    ulong          offsetd
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = convert_half4(sqrt(convert_float4(src0[gid])));
+}

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

@@ -0,0 +1,77 @@
+kernel void kernel_ssm_conv_f32_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb00,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne10,
+    ulong         nb11,
+    ulong         nb0,
+    ulong         nb1,
+    ulong         nb2
+){
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ir = get_global_id(0);
+    int i2 = get_global_id(1);
+    int i3 = get_global_id(2);
+
+    int nc  = ne10;
+
+    global float * s = (global float *) (src0 + ir*nb01 + i2*nb00 + i3*nb02);
+    global float * c = (global float *) (src1 + ir*nb11);
+    global float * d = (global float *) (dst  + ir*nb0  + i2*nb1  + i3*nb2);
+
+    float sumf = 0.0f;
+
+    for (int i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    d[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb00,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne10,
+    ulong         nb11,
+    ulong         nb0,
+    ulong         nb1,
+    ulong         nb2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ir = get_global_id(0);
+    int i2 = get_global_id(1);
+    int i3 = get_global_id(2);
+
+    int nc = ne10;
+
+    global float4 * s = (global float4 *) (src0 + ir*nb01 + i2*nb00 + i3*nb02);
+    global float4 * c = (global float4 *) (src1 + ir*nb11);
+    global float  * d = (global float  *) (dst  + ir*nb0  + i2*nb1  + i3*nb2);
+
+    float sumf = 0.0f;
+
+    for (int i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    d[0] = sumf;
+}

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -106,6 +106,7 @@ enum rpc_cmd {
     RPC_CMD_GET_ALLOC_SIZE,
     RPC_CMD_HELLO,
     RPC_CMD_DEVICE_COUNT,
+    RPC_CMD_GRAPH_RECOMPUTE,
     RPC_CMD_COUNT,
 };
 
@@ -205,10 +206,6 @@ struct rpc_msg_copy_tensor_rsp {
     uint8_t result;
 };
 
-struct rpc_msg_graph_compute_rsp {
-    uint8_t result;
-};
-
 struct rpc_msg_get_device_memory_req {
     uint32_t device;
 };
@@ -217,6 +214,11 @@ struct rpc_msg_get_device_memory_rsp {
     uint64_t free_mem;
     uint64_t total_mem;
 };
+
+struct rpc_msg_graph_recompute_req {
+    uint32_t device;
+};
+
 #pragma pack(pop)
 
 // RPC data structures
@@ -234,10 +236,35 @@ struct ggml_backend_rpc_buffer_type_context {
     size_t      max_size;
 };
 
+struct graph_cache {
+
+    bool is_cached(const ggml_cgraph * cgraph) {
+        if ((int)last_graph.size() != cgraph->n_nodes) {
+            return false;
+        }
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            if (memcmp(&last_graph[i], cgraph->nodes[i], sizeof(ggml_tensor)) != 0) {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    void add(const ggml_cgraph * cgraph) {
+        last_graph.resize(cgraph->n_nodes);
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            memcpy(&last_graph[i], cgraph->nodes[i], sizeof(ggml_tensor));
+        }
+    }
+
+    std::vector<ggml_tensor> last_graph;
+};
+
 struct ggml_backend_rpc_context {
     std::string endpoint;
     uint32_t    device;
     std::string name;
+    graph_cache gc;
 };
 
 struct ggml_backend_rpc_buffer_context {
@@ -815,13 +842,24 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve
 
 static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
-    std::vector<uint8_t> input;
-    serialize_graph(rpc_ctx->device, cgraph, input);
-    rpc_msg_graph_compute_rsp response;
-    auto sock = get_socket(rpc_ctx->endpoint);
-    bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size(), &response, sizeof(response));
-    RPC_STATUS_ASSERT(status);
-    return (enum ggml_status)response.result;
+
+    GGML_ASSERT(cgraph->n_nodes > 0);
+    bool reuse = rpc_ctx->gc.is_cached(cgraph);
+    if (reuse) {
+        rpc_msg_graph_recompute_req request;
+        request.device = rpc_ctx->device;
+        auto sock = get_socket(rpc_ctx->endpoint);
+        bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_RECOMPUTE, &request, sizeof(request));
+        RPC_STATUS_ASSERT(status);
+    } else {
+        rpc_ctx->gc.add(cgraph);
+        std::vector<uint8_t> input;
+        serialize_graph(rpc_ctx->device, cgraph, input);
+        auto sock = get_socket(rpc_ctx->endpoint);
+        bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size());
+        RPC_STATUS_ASSERT(status);
+    }
+    return GGML_STATUS_SUCCESS;
 }
 
 static ggml_backend_i ggml_backend_rpc_interface = {
@@ -880,7 +918,8 @@ ggml_backend_t ggml_backend_rpc_init(const char * endpoint, uint32_t device) {
     ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
         /* .endpoint = */ endpoint,
         /* .device   = */ device,
-        /* .name     = */ dev_name
+        /* .name     = */ dev_name,
+        /* .gc       = */ {},
     };
     auto reg = ggml_backend_rpc_add_server(endpoint);
     ggml_backend_t backend = new ggml_backend {
@@ -920,8 +959,9 @@ void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device,
 
 class rpc_server {
 public:
-    rpc_server(std::vector<ggml_backend_t> backends, const char * cache_dir)
-        : backends(std::move(backends)), cache_dir(cache_dir) {
+    rpc_server(std::vector<ggml_backend_t> all_backends, const char * cache_dir)
+        : backends(std::move(all_backends)), cache_dir(cache_dir) {
+        stored_graphs.resize(backends.size());
     }
     ~rpc_server();
 
@@ -936,11 +976,17 @@ public:
     bool set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpc_msg_set_tensor_hash_rsp & response);
     bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response);
     bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response);
-    bool graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response);
+    bool graph_compute(const std::vector<uint8_t> & input);
+    bool graph_recompute(const rpc_msg_graph_recompute_req & request);
     bool init_tensor(const rpc_msg_init_tensor_req & request);
     bool get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_msg_get_alloc_size_rsp & response);
     bool get_device_memory(const rpc_msg_get_device_memory_req & request, rpc_msg_get_device_memory_rsp & response);
 
+    struct stored_graph {
+        ggml_context_ptr ctx_ptr;
+        ggml_cgraph *    graph;
+    };
+
 private:
     bool get_cached_file(uint64_t hash, std::vector<uint8_t> & data);
     ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor);
@@ -953,6 +999,8 @@ private:
     std::vector<ggml_backend_t> backends;
     const char * cache_dir;
     std::unordered_set<ggml_backend_buffer_t> buffers;
+    // store the last computed graph for each backend
+    std::vector<stored_graph> stored_graphs;
 };
 
 void rpc_server::hello(rpc_msg_hello_rsp & response) {
@@ -1394,7 +1442,7 @@ ggml_tensor * rpc_server::create_node(uint64_t id,
     return result;
 }
 
-bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response) {
+bool rpc_server::graph_compute(const std::vector<uint8_t> & input) {
     // serialization format:
     // | device (4 bytes) | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
     if (input.size() < 2*sizeof(uint32_t)) {
@@ -1455,7 +1503,24 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
         }
     }
     ggml_status status = ggml_backend_graph_compute(backends[device], graph);
-    response.result = status;
+    GGML_ASSERT(status == GGML_STATUS_SUCCESS && "Unsuccessful graph computations are not supported with RPC");
+    stored_graphs[device].ctx_ptr.swap(ctx_ptr);
+    stored_graphs[device].graph = graph;
+    return true;
+}
+
+bool rpc_server::graph_recompute(const rpc_msg_graph_recompute_req & request) {
+    uint32_t device = request.device;
+    if (device >= backends.size()) {
+        return false;
+    }
+    if (stored_graphs[device].graph == nullptr) {
+        return false;
+    }
+    ggml_cgraph * graph = stored_graphs[device].graph;
+    LOG_DBG("[%s] device: %u\n", __func__, device);
+    ggml_status status = ggml_backend_graph_compute(backends[device], graph);
+    GGML_ASSERT(status == GGML_STATUS_SUCCESS && "Unsuccessful graph computations are not supported with RPC");
     return true;
 }
 
@@ -1690,11 +1755,17 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
                 if (!recv_msg(sockfd, input)) {
                     return;
                 }
-                rpc_msg_graph_compute_rsp response;
-                if (!server.graph_compute(input, response)) {
+                if (!server.graph_compute(input)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                break;
+            }
+            case RPC_CMD_GRAPH_RECOMPUTE: {
+                rpc_msg_graph_recompute_req request;
+                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                    return;
+                }
+                if (!server.graph_recompute(request)) {
                     return;
                 }
                 break;

ggml/src/ggml-sycl/CMakeLists.txt

@@ -91,7 +91,10 @@ if (GGML_SYCL_F16)
     add_compile_definitions(GGML_SYCL_F16)
 endif()
 
-if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+if (GGML_SYCL_TARGET STREQUAL "INTEL")
+    add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
+    target_link_options(ggml-sycl PRIVATE  -Xs   -ze-intel-greater-than-4GB-buffer-required)
+elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
     add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
 elseif (GGML_SYCL_TARGET STREQUAL "AMD")
     # INFO: Allowed Sub_group_sizes are not consistent through all
@@ -100,7 +103,8 @@ elseif (GGML_SYCL_TARGET STREQUAL "AMD")
     # Target archs tested working: gfx1030, gfx1031, (Only tested sub_group_size = 32)
     add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
 else()
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
+    # default for other target
+    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
 endif()
 
 if (GGML_SYCL_GRAPH)

ggml/src/ggml-sycl/common.hpp

@@ -617,4 +617,30 @@ static __dpct_inline__ float get_alibi_slope(const float    max_bias,
     return dpct::pow(base, exph);
 }
 
+static const sycl::uint3 init_fastdiv_values(uint32_t d) {
+    GGML_ASSERT(d != 0);
+
+    uint32_t L = 0;
+    while (L < 32 && (uint32_t{ 1 } << L) < d) {
+        L++;
+    }
+
+    uint32_t mp = (uint32_t) ((uint64_t{ 1 } << 32) * ((uint64_t{ 1 } << L) - d) / d + 1);
+    return sycl::uint3(mp, L, d);
+}
+
+
+static __dpct_inline__ uint32_t fastdiv(uint32_t n, const sycl::uint3 fastdiv_values) {
+    const uint32_t hi = sycl::mul_hi<unsigned>(n, fastdiv_values.x());
+    return (hi + n) >> fastdiv_values.y();
+}
+
+
+static __dpct_inline__ sycl::uint2 fast_div_modulo(uint32_t n, const sycl::uint3 fastdiv_values) {
+    const uint32_t div_val = fastdiv(n, fastdiv_values);
+    const uint32_t mod_val = n - div_val * fastdiv_values.z();
+    return sycl::uint2(div_val, mod_val);
+}
+
+
 #endif // GGML_SYCL_COMMON_HPP

ggml/src/ggml-sycl/cpy.cpp

@@ -515,9 +515,6 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
-    GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
-    GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
-
     GGML_TENSOR_BINARY_OP_LOCALS01;
 
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));

ggml/src/ggml-sycl/pad_reflect_1d.cpp

@@ -1,72 +1,100 @@
 #include "pad_reflect_1d.hpp"
 
-void pad_reflect_1d_f32(const float* src,float* dst,
-    const int64_t ne0, const int64_t ne02, const int p0, const int p1,
-    const int64_t nb0, const int64_t nb1, const int64_t nb2, const int64_t nb3,
-    const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
-    const sycl::nd_item<3> &item_ct1){
+static void pad_reflect_1d_kernel_f32(
+    const void *__restrict__ src0, void *__restrict__ dst, const int64_t ne0,
+    const int64_t ne00, const sycl::uint3 ne01, const int64_t ne02,
+    const int64_t ne03, const int64_t nb00, const int64_t nb01,
+    const int64_t nb02, const int64_t nb03, const int64_t nb0,
+    const int64_t nb1, const int64_t nb2, const int64_t nb3, const int p0,
+    const int p1, sycl::nd_item<3> item_ct1) {
 
-    const int i0 = item_ct1.get_group(0) * SYCL_CONCAT_BLOCK_SIZE + item_ct1.get_local_id(0);
-    const int i1 = item_ct1.get_group(1);
-    const int g2 = item_ct1.get_group(2);
-    const int i2 = g2 % ne02;
-    const int i3 = g2 / ne02;
+    const int64_t i3 = item_ct1.get_group(0);
+    const int64_t i2 = item_ct1.get_group(1);
 
-    if (i0 >= p0 + ne0 + p1) return;
+    const sycl::uint2 div_mod_packed =
+        fast_div_modulo(item_ct1.get_group(2), ne01);
+    const int64_t tile1 = div_mod_packed.y();
+    const int64_t tile0 = div_mod_packed.x();
+    const int64_t i1 = tile1;
+    const int64_t i0 =
+        item_ct1.get_local_id(2) + tile0 * item_ct1.get_local_range(2);
 
-    int t = i0 - p0;
-    int period = 2 * ne0 -2;
-    int m = t % period;
-    m += (m < 0) * period;
-    int center = ne0 -1;
-    int srci0 = center - abs(center - m);
+    if (i0 >= ne0 || i1 >= ne01.z() || i2 >= ne02 || i3 >= ne03) {
+        return;
+    }
 
-    int offest_src = i3*nb3 + i2*nb2 + i1*nb1 + srci0*nb0;
-    int offest_dst =  i3*nb03 +  i2*nb02 +  i1*nb01 +  i0*nb00;
-    dst[offest_dst] = src[offest_src];
+    const char *src0_ptr =
+        (const char *)src0 + i3 * nb03 + i2 * nb02 + i1 * nb01;
+    char *dst_ptr = (char *)dst + i3 * nb3 + i2 * nb2 + i1 * nb1;
 
+    const int64_t rel_i0 = i0 - p0; // relative i0 in src0
+    int64_t src_idx;
+
+    if (rel_i0 < 0) {
+        // Left padding - reflect
+        src_idx = -rel_i0;
+    } else if (rel_i0 < ne00) {
+        // Middle - copy
+        src_idx = rel_i0;
+    } else {
+        // Right padding - reflect
+        src_idx = 2 * ne00 - 2 - rel_i0;
+    }
+    const float value = *(const float *)(src0_ptr + src_idx * nb00);
+    *(float *)(dst_ptr + i0 * nb0) = value;
+
+    GGML_UNUSED(p1);
 }
 
-void ggml_sycl_op_pad_reflect_1d(ggml_backend_sycl_context& ctx, ggml_tensor* dst){
+void ggml_sycl_op_pad_reflect_1d(ggml_backend_sycl_context &ctx,
+                                 ggml_tensor *dst) {
 
-    const ggml_tensor * src0 = dst->src[0];
-    queue_ptr           stream = ctx.stream();
+    const ggml_tensor *src0 = dst->src[0];
+    dpct::queue_ptr stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    const int32_t * opts = (const int32_t *) dst->op_params;
+    const int32_t *opts = (const int32_t *)dst->op_params;
     const int p0 = opts[0];
     const int p1 = opts[1];
 
-    const int64_t ne0 = src0->ne[0];
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const sycl::uint3 ne01_packed = init_fastdiv_values(ne01);
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
 
-    const int64_t ne00 = dst->ne[0];
-    const int64_t ne01 = dst->ne[1];
-    const int64_t ne02 = dst->ne[2];
-    const int64_t ne03 = dst->ne[3];
+    const int64_t ne0 = dst->ne[0];
 
-    const int64_t nb00 = dst->nb[0];
-    const int64_t nb01 = dst->nb[1];
-    const int64_t nb02 = dst->nb[2];
-    const int64_t nb03 = dst->nb[3];
-    const int64_t nb0 = src0->nb[0];
-    const int64_t nb1 = src0->nb[1];
-    const int64_t nb2 = src0->nb[2];
-    const int64_t nb3 = src0->nb[3];
+    GGML_ASSERT(ne0 == ne00 + p0 + p1);
 
-    int num_blocks = (ne00 + SYCL_CONCAT_BLOCK_SIZE - 1) / SYCL_CONCAT_BLOCK_SIZE;
-    sycl::range<3> global(num_blocks * SYCL_CONCAT_BLOCK_SIZE, ne01, ne02*ne03);
-    sycl::range<3> local(SYCL_CONCAT_BLOCK_SIZE, 1, 1);
+    constexpr int64_t bx = SYCL_PAD_REFLECT_1D_BLOCK_SIZE;
+    const int64_t tiles0 = (ne0 + bx - 1) / bx;
+    const dpct::dim3 grid_dims((unsigned)(ne01 * tiles0), (unsigned)ne02,
+                               (unsigned)ne03);
+    const dpct::dim3 block_dims((unsigned)bx, 1, 1);
 
-    stream->parallel_for(
-        sycl::nd_range<3>(global,
-                            local),
-        [=](sycl::nd_item<3> item_ct1) { pad_reflect_1d_f32(
-            (const float *) src0->data, (float *) dst->data,
-            ne0, ne02, p0, p1,
-            nb0, nb1, nb2, nb3,
-            nb00, nb01, nb02, nb03
-            , item_ct1);
-         });
+    stream->submit([&](sycl::handler &cgh) {
+        auto src0_data_ct0 = src0->data;
+        auto dst_data_ct1 = dst->data;
+        auto src0_nb_ct7 = src0->nb[0];
+        auto src0_nb_ct8 = src0->nb[1];
+        auto src0_nb_ct9 = src0->nb[2];
+        auto src0_nb_ct10 = src0->nb[3];
+        auto dst_nb_ct11 = dst->nb[0];
+        auto dst_nb_ct12 = dst->nb[1];
+        auto dst_nb_ct13 = dst->nb[2];
+        auto dst_nb_ct14 = dst->nb[3];
+
+        cgh.parallel_for(sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             pad_reflect_1d_kernel_f32(
+                                 src0_data_ct0, dst_data_ct1, ne0, ne00,
+                                 ne01_packed, ne02, ne03, src0_nb_ct7,
+                                 src0_nb_ct8, src0_nb_ct9, src0_nb_ct10,
+                                 dst_nb_ct11, dst_nb_ct12, dst_nb_ct13,
+                                 dst_nb_ct14, p0, p1, item_ct1);
+                         });
+    });
 }

ggml/src/ggml-sycl/pad_reflect_1d.hpp

@@ -3,6 +3,8 @@
 
 #include "common.hpp"
 
+#define SYCL_PAD_REFLECT_1D_BLOCK_SIZE 256
+
 void ggml_sycl_op_pad_reflect_1d(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
 #endif // GGML_SYCL_PAD_REFLECT_1D_HPP

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

@@ -0,0 +1,69 @@
+#version 450
+
+#include "types.glsl"
+#include "sum_rows.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 128;
+layout (constant_id = 1) const uint SUBGROUP_SIZE = 32;
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+shared FLOAT_TYPE partial[BLOCK_SIZE / SUBGROUP_SIZE];
+shared FLOAT_TYPE last_sum;
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    const uint i03 = fastdiv(row, p.ne0_12mp, p.ne0_12L);
+    const uint i03_offset = i03 * p.ne01*p.ne02;
+    const uint i02 = fastdiv(row - i03_offset, p.ne0_1mp, p.ne0_1L);
+    const uint i01 = row - i03_offset - i02*p.ne01;
+
+    const uint src_idx = get_aoffset() + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+    const uint dst_idx = get_doffset() + i01 * p.nb11 + i02 * p.nb12 + i03 * p.nb13;
+
+    uint subgroup_id = tid / SUBGROUP_SIZE;
+
+    if (tid == 0) {
+        last_sum = 0;
+    }
+
+    uint col = tid;
+    uint num_iter = CEIL_DIV(p.n_cols, BLOCK_SIZE);
+    for (int i = 0; i < num_iter; ++i) {
+        FLOAT_TYPE v = 0;
+        if (col < p.n_cols) {
+            v = FLOAT_TYPE(data_a[src_idx + col]);
+        }
+        v = subgroupInclusiveAdd(v);
+
+        // Store the largest partial sum for each subgroup, then add the partials for all
+        // lower subgroups and the final partial sum from the previous iteration.
+        if (gl_SubgroupInvocationID == SUBGROUP_SIZE - 1) {
+            partial[subgroup_id] = v;
+        }
+        barrier();
+        for (int j = 0; j < subgroup_id; ++j) {
+            v += partial[j];
+        }
+        v += last_sum;
+        barrier();
+        if (tid == BLOCK_SIZE - 1) {
+            last_sum = v;
+        }
+        if (col < p.n_cols) {
+            data_d[dst_idx + col] = D_TYPE(v);
+        }
+        col += BLOCK_SIZE;
+    }
+}

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

@@ -4,13 +4,6 @@
 
 #include "types.glsl"
 
-#if defined(A_TYPE_PACKED16)
-layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
-#endif
-#if defined(A_TYPE_PACKED32)
-layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
-#endif
-
 #if defined(DATA_A_F32)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);

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

@@ -156,7 +156,7 @@ void main() {
                 tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
                 tensorLayoutM = setTensorLayoutClampValueNV(tensorLayoutM, 0xfc00); // -inf in float16_t
 
-                coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv, mvmax;
+                coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
 
                 coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
 

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

@@ -22,6 +22,13 @@ layout (push_constant) uniform parameter
 
 #if !RMS_NORM_ROPE_FUSION
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #endif

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

@@ -18,6 +18,13 @@ layout (push_constant) uniform parameter
 } p;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 uint get_idx() {

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

@@ -3,6 +3,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
 #include "mul_mat_vec_base.glsl"
+#include "dequant_funcs.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 

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

@@ -13,8 +13,6 @@
 
 #include "mul_mat_vec_iface.glsl"
 
-#include "dequant_funcs.glsl"
-
 layout (push_constant) uniform parameter
 {
     uint ncols;

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

@@ -5,13 +5,15 @@
 #define MAT_VEC_FUSION_FLAGS_SCALE0 0x4
 #define MAT_VEC_FUSION_FLAGS_SCALE1 0x8
 
-#ifndef MMQ
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 #if defined(A_TYPE_VEC4)
 layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
 #endif
-#else
-layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
 #endif
 
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};

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

@@ -10,60 +10,56 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
+#if defined(DATA_A_QUANT_LEGACY) || defined(DATA_A_MXFP4)
 #define K_PER_ITER 8
-
-#include "mul_mmq_funcs.glsl"
+#elif defined(DATA_A_QUANT_K)
+#define K_PER_ITER 16
+#else
+#error unimplemented
+#endif
 
 uint a_offset, b_offset, d_offset;
 
-int32_t cache_b_qs[2];
+int32_t cache_b_qs[K_PER_ITER / 4];
 vec2 cache_b_ds;
 
+#include "mul_mat_vecq_funcs.glsl"
+
 void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i) {
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         const uint col = i*BLOCK_SIZE + tid*K_PER_ITER;
 
         // Preload data_b block
         const uint b_block_idx = (j*p.batch_stride_b + col) / QUANT_K_Q8_1 + b_offset;
-        const uint b_qs_idx = tid % 4;
+        const uint b_qs_idx = tid % (32 / K_PER_ITER);
         const uint b_block_idx_outer = b_block_idx / 4;
         const uint b_block_idx_inner = b_block_idx % 4;
         cache_b_ds = vec2(data_b[b_block_idx_outer].ds[b_block_idx_inner]);
 
 #if QUANT_R == 2
+        // Assumes K_PER_ITER == 8
         cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx];
         cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx + 4];
 #else
+#if K_PER_ITER == 8
         cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2];
         cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2 + 1];
+#elif K_PER_ITER == 16
+        cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 4    ];
+        cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 4 + 1];
+        cache_b_qs[2] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 4 + 2];
+        cache_b_qs[3] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 4 + 3];
+#else
+#error unimplemented
+#endif
 #endif
 
         uint ibi = first_row*p.ncols;
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            const uint a_block_idx = (ibi + col)/QUANT_K + a_offset;
+            const uint a_block_idx = (ibi + col)/QUANT_K_Q8_1 + a_offset;
             ibi += p.ncols;
 
-            int32_t q_sum = 0;
-#if QUANT_R == 2
-            const i32vec2 data_a_qs = repack(a_block_idx, b_qs_idx);
-            q_sum += dotPacked4x8EXT(data_a_qs.x,
-                                     cache_b_qs[0]);
-            q_sum += dotPacked4x8EXT(data_a_qs.y,
-                                     cache_b_qs[1]);
-#else
-            int32_t data_a_qs = repack(a_block_idx, b_qs_idx * 2);
-            q_sum += dotPacked4x8EXT(data_a_qs,
-                                     cache_b_qs[0]);
-            data_a_qs = repack(a_block_idx, b_qs_idx * 2 + 1);
-            q_sum += dotPacked4x8EXT(data_a_qs,
-                                     cache_b_qs[1]);
-#endif
-
-#if QUANT_AUXF == 1
-            temp[j][n] += mul_q8_1(q_sum,  get_d(a_block_idx), cache_b_ds, 4);
-#else
-            temp[j][n] += mul_q8_1(q_sum, get_dm(a_block_idx), cache_b_ds, 4);
-#endif
+            temp[j][n] += mmvq_dot_product(a_block_idx, b_qs_idx);
         }
     }
 }
@@ -72,7 +68,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint tid = gl_LocalInvocationID.x;
 
     get_offsets(a_offset, b_offset, d_offset);
-    a_offset /= QUANT_K;
+    a_offset /= QUANT_K_Q8_1;
     b_offset /= QUANT_K_Q8_1;
 
     FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
@@ -102,14 +98,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     unroll_count = 2;
     unrolled_iters = num_iters & ~(unroll_count - 1);
 
-#if K_PER_ITER == 2
-    if ((p.ncols & 1) != 0 &&
-        unrolled_iters == num_iters &&
-        unrolled_iters > 0) {
-        unrolled_iters -= unroll_count;
-    }
-#endif
-
     while (i < unrolled_iters) {
         // Manually partially unroll the loop
         [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
@@ -128,6 +116,10 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 void main() {
     const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
 
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
+#endif
+
     // do NUM_ROWS at a time, unless there aren't enough remaining rows
     if (first_row + NUM_ROWS <= p.stride_d) {
         compute_outputs(first_row, NUM_ROWS);

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

@@ -0,0 +1,379 @@
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+
+#include "types.glsl"
+
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+FLOAT_TYPE get_dm(uint ib) {
+    return FLOAT_TYPE(data_a[ib].d);
+}
+#endif
+
+#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
+FLOAT_TYPE_VEC2 get_dm(uint ib) {
+    return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
+}
+#endif
+
+#if defined(DATA_A_MXFP4)
+FLOAT_TYPE get_dm(uint ib) {
+    return FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e));
+}
+#endif
+
+#if defined(DATA_A_Q2_K)
+FLOAT_TYPE_VEC2 get_dm(uint ib) {
+    const uint ib_k = ib / 8;
+    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
+}
+#endif
+
+// Each iqs value maps to a 32-bit integer
+#if defined(DATA_A_Q4_0)
+// 2-byte loads for Q4_0 blocks (18 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
+                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
+    const uint32_t vui = pack32(quants);
+    return i32vec2( vui       & 0x0F0F0F0F,
+                   (vui >> 4) & 0x0F0F0F0F);
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * dsb.y));
+}
+#endif
+
+#if defined(DATA_A_Q4_1)
+// 4-byte loads for Q4_1 blocks (20 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const uint32_t vui = data_a_packed32[ib].qs[iqs];
+    return i32vec2( vui       & 0x0F0F0F0F,
+                   (vui >> 4) & 0x0F0F0F0F);
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
+}
+#endif
+
+#if defined(DATA_A_Q5_0)
+// 2-byte loads for Q5_0 blocks (22 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
+                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
+    const uint32_t vui = pack32(quants);
+    const int32_t qh = int32_t((uint32_t(data_a_packed16[ib].qh[1]) << 16 | data_a_packed16[ib].qh[0]) >> (4 * iqs));
+    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
+                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
+
+    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
+                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
+
+    return i32vec2(v0, v1);
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * dsb.y));
+}
+#endif
+
+#if defined(DATA_A_Q5_1)
+// 4-byte loads for Q5_1 blocks (24 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
+                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
+    const uint32_t vui = pack32(quants);
+    const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
+    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
+                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
+
+    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
+                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
+
+    return i32vec2(v0, v1);
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+// 2-byte loads for Q8_0 blocks (34 bytes)
+int32_t repack(uint ib, uint iqs) {
+    return pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2    ],
+                          data_a_packed16[ib].qs[iqs * 2 + 1]));
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(float(q_sum) * da * dsb.x);
+}
+#endif
+
+#if defined(DATA_A_MXFP4)
+// 1-byte loads for mxfp4 blocks (17 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const uint32_t qs = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
+                                      data_a[ib].qs[iqs * 4 + 1],
+                                      data_a[ib].qs[iqs * 4 + 2],
+                                      data_a[ib].qs[iqs * 4 + 3]));
+
+    const u8vec4 i_a0 = unpack8( qs       & 0x0F0F0F0F);
+    const u8vec4 i_a1 = unpack8((qs >> 4) & 0x0F0F0F0F);
+
+    return i32vec2(pack32(i8vec4(kvalues_mxfp4[i_a0.x], kvalues_mxfp4[i_a0.y], kvalues_mxfp4[i_a0.z], kvalues_mxfp4[i_a0.w])),
+                   pack32(i8vec4(kvalues_mxfp4[i_a1.x], kvalues_mxfp4[i_a1.y], kvalues_mxfp4[i_a1.z], kvalues_mxfp4[i_a1.w])));
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(da * dsb.x * float(q_sum) * 0.5);
+}
+#endif
+
+#if defined(DATA_A_QUANT_LEGACY) || defined(DATA_A_MXFP4)
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t q_sum = 0;
+#if QUANT_R == 2
+    const i32vec2 data_a_qs = repack(ib_a, iqs);
+    q_sum += dotPacked4x8EXT(data_a_qs.x,
+                             cache_b_qs[0]);
+    q_sum += dotPacked4x8EXT(data_a_qs.y,
+                             cache_b_qs[1]);
+#else
+    int32_t data_a_qs = repack(ib_a, iqs * 2);
+    q_sum += dotPacked4x8EXT(data_a_qs,
+                             cache_b_qs[0]);
+    data_a_qs = repack(ib_a, iqs * 2 + 1);
+    q_sum += dotPacked4x8EXT(data_a_qs,
+                             cache_b_qs[1]);
+#endif
+
+    // 2 quants per call => divide sums by 8/2 = 4
+    return mul_q8_1(q_sum, get_dm(ib_a), cache_b_ds, 4);
+}
+#endif
+
+#if defined(DATA_A_Q2_K)
+// 4-byte loads for Q2_K blocks (84 bytes)
+i32vec4 repack4(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
+
+    return i32vec4((data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x03030303,
+                   (data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x03030303,
+                   (data_a_packed32[ib_k].qs[qs_idx + 2] >> qs_shift) & 0x03030303,
+                   (data_a_packed32[ib_k].qs[qs_idx + 3] >> qs_shift) & 0x03030303);
+}
+
+uint8_t get_scale(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    return data_a[ib_k].scales[iqs_k / 4];
+}
+
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t sum_d = 0;
+    int32_t sum_m = 0;
+
+    const i32vec4 qs_a = repack4(ib_a, iqs * 4);
+    const uint8_t scale = get_scale(ib_a, iqs * 4);
+    const vec2 dm = vec2(get_dm(ib_a));
+    const int32_t scale_m = int32_t(scale >> 4) * 0x01010101; // Duplicate 8-bit value across 32-bits.
+
+    sum_d += dotPacked4x8EXT(qs_a.x, cache_b_qs[0]) * (scale & 0xF);
+    sum_m += dotPacked4x8EXT(scale_m, cache_b_qs[0]);
+
+    sum_d += dotPacked4x8EXT(qs_a.y, cache_b_qs[1]) * (scale & 0xF);
+    sum_m += dotPacked4x8EXT(scale_m, cache_b_qs[1]);
+
+    sum_d += dotPacked4x8EXT(qs_a.z, cache_b_qs[2]) * (scale & 0xF);
+    sum_m += dotPacked4x8EXT(scale_m, cache_b_qs[2]);
+
+    sum_d += dotPacked4x8EXT(qs_a.w, cache_b_qs[3]) * (scale & 0xF);
+    sum_m += dotPacked4x8EXT(scale_m, cache_b_qs[3]);
+
+    return FLOAT_TYPE(float(cache_b_ds.x) * (float(dm.x) * float(sum_d) - float(dm.y) * float(sum_m)));
+}
+#endif
+
+#if defined(DATA_A_Q3_K)
+// 2-byte loads for Q3_K blocks (110 bytes)
+i32vec4 repack4(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
+    const uint hm_shift = iqs_k / 8;
+
+    // bitwise OR to add 4 if hmask is set, subtract later
+    const i8vec2 vals00 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2    ] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2    ] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals01 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 1] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 1] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals10 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 2] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 2] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals11 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 3] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 3] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals20 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 4] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 4] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals21 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 5] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 5] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals30 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 6] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 6] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals31 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 7] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 7] >> hm_shift) & uint16_t(0x0101)) << 2));
+
+    return i32vec4(pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y) - int8_t(4)),
+                   pack32(i8vec4(vals10.x, vals10.y, vals11.x, vals11.y) - int8_t(4)),
+                   pack32(i8vec4(vals20.x, vals20.y, vals21.x, vals21.y) - int8_t(4)),
+                   pack32(i8vec4(vals30.x, vals30.y, vals31.x, vals31.y) - int8_t(4)));
+}
+
+float get_d_scale(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+    const uint is = iqs_k / 4;
+
+    const int8_t scale = int8_t(((data_a[ib_k].scales[is % 8      ] >> (4 * (is / 8))) & 0x0F0F) |
+                               (((data_a[ib_k].scales[8 + (is % 4)] >> (2 * (is / 4))) & 0x0303) << 4));
+    return float(data_a[ib_k].d) * float(scale - 32);
+}
+
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t q_sum = 0;
+
+    const i32vec4 qs_a = repack4(ib_a, iqs * 4);
+    const float d_scale = get_d_scale(ib_a, iqs * 4);
+
+    q_sum += dotPacked4x8EXT(qs_a.x, cache_b_qs[0]);
+    q_sum += dotPacked4x8EXT(qs_a.y, cache_b_qs[1]);
+    q_sum += dotPacked4x8EXT(qs_a.z, cache_b_qs[2]);
+    q_sum += dotPacked4x8EXT(qs_a.w, cache_b_qs[3]);
+
+    return FLOAT_TYPE(float(cache_b_ds.x) * d_scale * float(q_sum));
+}
+#endif
+
+#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K)
+// 4-byte loads for Q4_K blocks (144 bytes) and Q5_K blocks (176 bytes)
+i32vec4 repack4(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint qs_idx = (iqs_k / 16) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 16) / 8) * 4;
+
+#if defined(DATA_A_Q4_K)
+    const uint32_t vals0 = (data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x0F0F0F0F;
+    const uint32_t vals1 = (data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x0F0F0F0F;
+    const uint32_t vals2 = (data_a_packed32[ib_k].qs[qs_idx + 2] >> qs_shift) & 0x0F0F0F0F;
+    const uint32_t vals3 = (data_a_packed32[ib_k].qs[qs_idx + 3] >> qs_shift) & 0x0F0F0F0F;
+
+    return i32vec4(vals0, vals1, vals2, vals3);
+#else // defined(DATA_A_Q5_K)
+    const uint qh_idx = iqs;
+    const uint qh_shift = iqs_k / 8;
+
+    return i32vec4(((data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x0F0F0F0F) |
+                  (((data_a_packed32[ib_k].qh[qh_idx    ] >> qh_shift) & 0x01010101) << 4),
+                   ((data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x0F0F0F0F) |
+                  (((data_a_packed32[ib_k].qh[qh_idx + 1] >> qh_shift) & 0x01010101) << 4),
+                   ((data_a_packed32[ib_k].qs[qs_idx + 2] >> qs_shift) & 0x0F0F0F0F) |
+                  (((data_a_packed32[ib_k].qh[qh_idx + 2] >> qh_shift) & 0x01010101) << 4),
+                   ((data_a_packed32[ib_k].qs[qs_idx + 3] >> qs_shift) & 0x0F0F0F0F) |
+                  (((data_a_packed32[ib_k].qh[qh_idx + 3] >> qh_shift) & 0x01010101) << 4));
+#endif
+}
+
+vec2 get_dm_scale(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+    const uint is = iqs_k / 8;
+    u8vec2 scale_dm;
+    if (is < 4) {
+        scale_dm = u8vec2(data_a[ib_k].scales[is] & 0x3F, data_a[ib_k].scales[is + 4] & 0x3F);
+    } else {
+        scale_dm = u8vec2((data_a[ib_k].scales[is+4] & 0xF) | ((data_a[ib_k].scales[is-4] & 0xC0) >> 2),
+                          (data_a[ib_k].scales[is+4] >>  4) | ((data_a[ib_k].scales[is  ] & 0xC0) >> 2));
+    }
+
+    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm) * FLOAT_TYPE_VEC2(scale_dm);
+}
+
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t q_sum = 0;
+
+    const i32vec4 qs_a = repack4(ib_a, iqs * 4);
+    const vec2 dm_scale = get_dm_scale(ib_a, iqs * 4);
+
+    q_sum += dotPacked4x8EXT(qs_a.x, cache_b_qs[0]);
+    q_sum += dotPacked4x8EXT(qs_a.y, cache_b_qs[1]);
+    q_sum += dotPacked4x8EXT(qs_a.z, cache_b_qs[2]);
+    q_sum += dotPacked4x8EXT(qs_a.w, cache_b_qs[3]);
+
+    return FLOAT_TYPE(float(cache_b_ds.x) * float(dm_scale.x) * float(q_sum) - float(dm_scale.y) * float(cache_b_ds.y / 2));
+}
+#endif
+
+#if defined(DATA_A_Q6_K)
+// 2-byte loads for Q6_K blocks (210 bytes)
+i32vec4 repack4(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint ql_idx = (iqs_k / 32) * 16 + iqs_k % 16;
+    const uint ql_shift = ((iqs_k % 32) / 16) * 4;
+
+    const uint qh_idx = (iqs_k / 32) * 8 + iqs;
+    const uint qh_shift = ((iqs_k % 32) / 8) * 2;
+
+    const i8vec2 vals00 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals01 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals10 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 2] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 2] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals11 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 3] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 3] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals20 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 4] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 4] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals21 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 5] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 5] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals30 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 6] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 6] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals31 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 7] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 7] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+
+    return i32vec4(pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y)),
+                   pack32(i8vec4(vals10.x, vals10.y, vals11.x, vals11.y)),
+                   pack32(i8vec4(vals20.x, vals20.y, vals21.x, vals21.y)),
+                   pack32(i8vec4(vals30.x, vals30.y, vals31.x, vals31.y)));
+}
+
+float get_d_scale(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+    return float(data_a[ib_k].d) * float(data_a[ib_k].scales[iqs_k / 4]);
+}
+
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t q_sum = 0;
+
+    const i32vec4 qs_a = repack4(ib_a, iqs * 4);
+    const float d_scale = get_d_scale(ib_a, iqs * 4);
+
+    q_sum += dotPacked4x8EXT(qs_a.x, cache_b_qs[0]);
+    q_sum += dotPacked4x8EXT(qs_a.y, cache_b_qs[1]);
+    q_sum += dotPacked4x8EXT(qs_a.z, cache_b_qs[2]);
+    q_sum += dotPacked4x8EXT(qs_a.w, cache_b_qs[3]);
+
+    return FLOAT_TYPE(float(cache_b_ds.x) * float(d_scale) * float(q_sum));
+}
+#endif

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

@@ -78,8 +78,6 @@ layout (constant_id = 10) const uint WARP = 32;
 
 #define BK 32
 
-#define MMQ_SHMEM
-
 #include "mul_mmq_shmem_types.glsl"
 
 #ifdef MUL_MAT_ID

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

@@ -9,31 +9,6 @@
 #if defined(DATA_A_Q4_0) || defined(DATA_A_Q4_1)
 // 2-byte loads for Q4_0 blocks (18 bytes)
 // 4-byte loads for Q4_1 blocks (20 bytes)
-i32vec2 repack(uint ib, uint iqs) {
-#ifdef DATA_A_Q4_0
-    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
-    const uint32_t vui = pack32(quants);
-    return i32vec2( vui       & 0x0F0F0F0F,
-                   (vui >> 4) & 0x0F0F0F0F);
-#else // DATA_A_Q4_1
-    const uint32_t vui = data_a_packed32[ib].qs[iqs];
-    return i32vec2( vui       & 0x0F0F0F0F,
-                   (vui >> 4) & 0x0F0F0F0F);
-#endif
-}
-
-#ifdef DATA_A_Q4_0
-ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * dsb.y));
-}
-#else // DATA_A_Q4_1
-ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-}
-#endif
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
 #ifdef DATA_A_Q4_0
     buf_a[buf_ib].qs[iqs] = pack32(u16vec2(data_a_packed16[ib].qs[iqs * 2],
@@ -73,42 +48,17 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a.y, qs_b1);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
+#ifdef DATA_A_Q4_0
+    return ACC_TYPE(float(cache_a[ib_a].dm) * (float(q_sum) * float(cache_b.ds.x) - 8.0 * float(cache_b.ds.y)));
+#else // DATA_A_Q4_1
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm.x) * float(cache_b.ds.x) + float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
+#endif
 }
-#endif // MMQ_SHMEM
+#endif
 
-#elif defined(DATA_A_Q5_0) || defined(DATA_A_Q5_1)
+#if defined(DATA_A_Q5_0) || defined(DATA_A_Q5_1)
 // 2-byte loads for Q5_0 blocks (22 bytes)
 // 4-byte loads for Q5_1 blocks (24 bytes)
-i32vec2 repack(uint ib, uint iqs) {
-    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
-    const uint32_t vui = pack32(quants);
-#ifdef DATA_A_Q5_0
-    const int32_t qh = int32_t((uint32_t(data_a_packed16[ib].qh[1]) << 16 | data_a_packed16[ib].qh[0]) >> (4 * iqs));
-#else // DATA_A_Q5_1
-    const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
-#endif
-    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
-                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
-
-    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
-                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
-
-    return i32vec2(v0, v1);
-}
-
-#ifdef DATA_A_Q5_0
-ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * dsb.y));
-}
-#else // DATA_A_Q5_1
-ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-}
-#endif
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
 #ifdef DATA_A_Q5_0
     buf_a[buf_ib].qs[iqs] = pack32(u16vec2(data_a_packed16[ib].qs[iqs * 2],
@@ -154,23 +104,16 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a1, qs_b1);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
+#ifdef DATA_A_Q5_0
+    return ACC_TYPE(float(cache_a[ib_a].dm) * (float(q_sum) * float(cache_b.ds.x) - 16.0 * float(cache_b.ds.y)));
+#else // DATA_A_Q5_1
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm.x) * float(cache_b.ds.x) + float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
+#endif
 }
-#endif // MMQ_SHMEM
 #endif
 
 #if defined(DATA_A_Q8_0)
 // 2-byte loads for Q8_0 blocks (34 bytes)
-int32_t repack(uint ib, uint iqs) {
-    return pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-                          data_a_packed16[ib].qs[iqs * 2 + 1]));
-}
-
-ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(float(q_sum) * da * dsb.x);
-}
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     buf_a[buf_ib].qs[iqs] = pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2],
                                            data_a_packed16[ib].qs[iqs * 2 + 1]));
@@ -197,28 +140,12 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a, qs_b);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm) * float(cache_b.ds.x));
 }
-#endif // MMQ_SHMEM
 #endif
 
 #if defined(DATA_A_MXFP4)
 // 1-byte loads for mxfp4 blocks (17 bytes)
-i32vec2 repack(uint ib, uint iqs) {
-    const uint32_t quants = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
-                                          data_a[ib].qs[iqs * 4 + 1],
-                                          data_a[ib].qs[iqs * 4 + 2],
-                                          data_a[ib].qs[iqs * 4 + 3]));
-
-    return i32vec2( quants       & 0x0F0F0F0F,
-                   (quants >> 4) & 0x0F0F0F0F);
-}
-
-ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(da * dsb.x * float(q_sum));
-}
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint32_t qs = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
                                       data_a[ib].qs[iqs * 4 + 1],
@@ -252,37 +179,14 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].d, cache_b.ds, 1);
+    return ACC_TYPE(float(cache_a[ib_a].d) * float(cache_b.ds.x) * float(q_sum));
 }
-#endif // MMQ_SHMEM
 #endif
 
 // For k-quants, ib and iqs still assume 32-wide blocks, but k-quants are 256-wide
 // iqs still refers to a 32-bit integer, meaning 0..7 for 32-wide quants
 #if defined(DATA_A_Q2_K)
 // 4-byte loads for Q2_K blocks (84 bytes)
-int32_t repack(uint ib, uint iqs) {
-    const uint ib_k = ib / 8;
-    const uint iqs_k = (ib % 8) * 8 + iqs;
-
-    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
-    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
-
-    return int32_t((data_a_packed32[ib_k].qs[qs_idx] >> qs_shift) & 0x03030303);
-}
-
-uint8_t get_scale(uint ib, uint iqs) {
-    const uint ib_k = ib / 8;
-    const uint iqs_k = (ib % 8) * 8 + iqs;
-
-    return data_a[ib_k].scales[iqs_k / 4];
-}
-
-ACC_TYPE mul_q8_1(const int32_t sum_d, const int32_t sum_m, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(dsb.x * (dma.x * float(sum_d) - dma.y * float(sum_m)));
-}
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint ib_k = ib / 8;
     const uint iqs_k = (ib % 8) * 8 + iqs * QUANT_R_MMQ;
@@ -326,14 +230,12 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         sum_m += dotPacked4x8EXT(scale_m, cache_b.qs[iqs]);
     }
 
-    return mul_q8_1(sum_d, sum_m, cache_a[ib_a].dm, cache_b.ds, 1);
+    return ACC_TYPE(float(cache_b.ds.x) * (float(cache_a[ib_a].dm.x) * float(sum_d) - float(cache_a[ib_a].dm.y) * float(sum_m)));
 }
-#endif // MMQ_SHMEM
 #endif
 
 #if defined(DATA_A_Q3_K)
 // 2-byte loads for Q3_K blocks (110 bytes)
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint ib_k = ib / 8;
     const uint hm_idx = iqs * QUANT_R_MMQ;
@@ -394,18 +296,12 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
     }
     result += float(cache_a[ib_a].d_scales[1]) * float(q_sum);
 
-    return ACC_TYPE(cache_b.ds.x * result);
+    return ACC_TYPE(float(cache_b.ds.x) * result);
 }
-#endif // MMQ_SHMEM
 #endif
 
 #if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K)
 // 4-byte loads for Q4_K blocks (144 bytes) and Q5_K blocks (176 bytes)
-ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(dsb.x * dma.x * float(q_sum) - dma.y * dsb.y);
-}
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint ib_k = ib / 8;
     const uint iqs_k = (ib % 8) * 8 + iqs * QUANT_R_MMQ;
@@ -427,7 +323,6 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
                                    (((data_a_packed32[ib_k].qh[qh_idx] >> qh_shift) & 0x01010101) << 4));
 #endif
 
-
     if (iqs == 0) {
         // Scale index
         const uint is = iqs_k / 8;
@@ -464,49 +359,12 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
-}
-#endif // MMQ_SHMEM
-#endif
-
-#ifdef MMQ_SHMEM
-void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bool is_in_bounds) {
-    if (is_in_bounds) {
-        const uint ib_outer = ib / 4;
-        const uint ib_inner = ib % 4;
-
-        if (iqs == 0) {
-            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
-        }
-
-        const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
-        buf_b[buf_ib].qs[iqs * 4    ] = values.x;
-        buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
-        buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
-        buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
-    } else {
-        if (iqs == 0) {
-            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(0.0f);
-        }
-
-        buf_b[buf_ib].qs[iqs * 4    ] = 0;
-        buf_b[buf_ib].qs[iqs * 4 + 1] = 0;
-        buf_b[buf_ib].qs[iqs * 4 + 2] = 0;
-        buf_b[buf_ib].qs[iqs * 4 + 3] = 0;
-    }
-}
-
-void block_b_to_registers(const uint ib) {
-    cache_b.ds = buf_b[ib].ds;
-    [[unroll]] for (uint iqs = 0; iqs < BK / 4; iqs++) {
-        cache_b.qs[iqs] = buf_b[ib].qs[iqs];
-    }
+    return ACC_TYPE(float(cache_b.ds.x) * float(cache_a[ib_a].dm.x) * float(q_sum) - float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
 }
 #endif
 
 #if defined(DATA_A_Q6_K)
 // 2-byte loads for Q6_K blocks (210 bytes)
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint ib_k = ib / 8;
     const uint iqs_k = (ib % 8) * 8 + iqs;
@@ -558,32 +416,39 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
     }
     result += float(cache_a[ib_a].d_scales[1]) * float(q_sum);
 
-    return ACC_TYPE(cache_b.ds.x * result);
-}
-#endif // MMQ_SHMEM
-#endif
-
-#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
-FLOAT_TYPE get_d(uint ib) {
-    return FLOAT_TYPE(data_a[ib].d);
+    return ACC_TYPE(float(cache_b.ds.x) * result);
 }
 #endif
 
-#if defined(DATA_A_MXFP4)
-FLOAT_TYPE get_d(uint ib) {
-    return FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e));
-}
-#endif
+void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bool is_in_bounds) {
+    if (is_in_bounds) {
+        const uint ib_outer = ib / 4;
+        const uint ib_inner = ib % 4;
 
-#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
-FLOAT_TYPE_VEC2 get_dm(uint ib) {
-    return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
-}
-#endif
+        if (iqs == 0) {
+            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
+        }
 
-#if defined(DATA_A_Q2_K)
-FLOAT_TYPE_VEC2 get_dm(uint ib) {
-    const uint ib_k = ib / 8;
-    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
+        const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
+        buf_b[buf_ib].qs[iqs * 4    ] = values.x;
+        buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
+        buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
+        buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
+    } else {
+        if (iqs == 0) {
+            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(0.0f);
+        }
+
+        buf_b[buf_ib].qs[iqs * 4    ] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 1] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 2] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 3] = 0;
+    }
+}
+
+void block_b_to_registers(const uint ib) {
+    cache_b.ds = buf_b[ib].ds;
+    [[unroll]] for (uint iqs = 0; iqs < BK / 4; iqs++) {
+        cache_b.qs[iqs] = buf_b[ib].qs[iqs];
+    }
 }
-#endif

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

@@ -0,0 +1,72 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_binary_head.glsl"
+
+layout (constant_id = 1) const uint N = 64;
+layout (constant_id = 2) const uint K = 32;
+
+layout(local_size_x = 128, local_size_y = 1, local_size_z = 1) in;
+
+uint a_base, b_base, x_base;
+
+FLOAT_TYPE get_a(uint r, uint c) {
+    return FLOAT_TYPE(data_a[a_base + r * p.nb01 + c * p.nb00]);
+}
+
+FLOAT_TYPE get_b(uint r, uint c) {
+    return FLOAT_TYPE(data_b[b_base + r * p.nb11 + c * p.nb10]);
+}
+
+void store_x(uint r, uint c, FLOAT_TYPE v) {
+    data_d[x_base + r * p.nb21 + c * p.nb20] = D_TYPE(v);
+}
+
+shared FLOAT_TYPE shA[N * N];
+shared FLOAT_TYPE shB[N * K];
+
+void main() {
+    const uint batch = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    if (batch >= p.ne02 * p.ne03) {
+        return;
+    }
+
+    const uint i3 = batch / p.ne22;
+    const uint i2 = batch % p.ne22;
+    a_base = get_aoffset() + i2 * p.nb02 + i3 * p.nb03;
+    b_base = get_boffset() + i2 * p.nb12 + i3 * p.nb13;
+    x_base = get_doffset() + i2 * p.nb22 + i3 * p.nb23;
+
+    // Load the A matrix into shA
+    [[unroll]] for (uint i = 0; i < N * N; i += gl_WorkGroupSize.x) {
+        uint idx = i + tid;
+        if (((N * N) % gl_WorkGroupSize.x == 0) || idx < N * N) {
+            shA[idx] = get_a(idx / N, idx % N);
+        }
+    }
+    // Load the B matrix into shB
+    [[unroll]] for (uint i = 0; i < N * K; i += gl_WorkGroupSize.x) {
+        uint idx = i + tid;
+        if (((N * K) % gl_WorkGroupSize.x == 0) || idx < N * K) {
+            shB[idx] = get_b(idx / K, idx % K);
+        }
+    }
+    barrier();
+
+    FLOAT_TYPE X[N];
+    // Each thread solves one column
+    if (tid < K) {
+        [[unroll]] for (int r = 0; r < N; ++r) {
+            FLOAT_TYPE b = shB[r * K + tid];
+            // Compute x[r,c] = (b[r,c] - sum(a[r,c]*x[c])) / a[r,r]
+            [[unroll]] for (int c = 0; c < r; ++c) {
+                b -= shA[r * N + c] * X[c];
+            }
+            FLOAT_TYPE x = b / shA[r * N + r];
+            X[r] = x;
+            store_x(r, tid, x);
+        }
+    }
+}

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

@@ -1,6 +1,7 @@
 #version 450
 
 #include "types.glsl"
+#include "sum_rows.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 
@@ -11,30 +12,6 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 
-layout (push_constant) uniform parameter
-{
-    uint n_cols;
-    uint ne01, ne02;
-    uint nb01, nb02, nb03;
-    uint nb11, nb12, nb13;
-    float weight;
-    uint misalign_offsets;
-    uint ne0_12mp, ne0_12L;
-    uint ne0_1mp, ne0_1L;
-} p;
-
-uint get_aoffset() { return p.misalign_offsets >> 16; }
-uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
-
-// see init_fastdiv_values in ggml-vulkan.cpp
-uint fastdiv(uint n, uint mp, uint L) {
-    uint msbs, lsbs;
-    // msbs = mulhi(n, mp)
-    umulExtended(n, mp, msbs, lsbs);
-    return (msbs + n) >> L;
-}
-
-
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {

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

@@ -0,0 +1,25 @@
+
+// vk_op_sum_rows_push_constants
+layout (push_constant) uniform parameter
+{
+    uint n_cols;
+    uint ne01, ne02;
+    uint nb01, nb02, nb03;
+    uint nb11, nb12, nb13;
+    float weight;
+    uint misalign_offsets;
+    uint ne0_12mp, ne0_12L;
+    uint ne0_1mp, ne0_1L;
+} p;
+
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+

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

@@ -0,0 +1,113 @@
+#version 450
+#extension GL_EXT_control_flow_attributes : enable
+
+#include "types.glsl"
+
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int NCOLS_PADDED_LOG2 = 10;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+// Input can either be the source (A) or intermediate values (S).
+// Similarly, output can be either destination (D) or intermediate values (S).
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 0) readonly buffer S {ivec2 data_s[];};
+layout (binding = 1) writeonly buffer D {int data_d[];};
+layout (binding = 1) writeonly buffer T {ivec2 data_t[];};
+
+layout (push_constant) uniform parameter {
+    uint orig_ncols;
+    uint ncols_input;
+    uint ncols_output;
+    uint nrows;
+    uint first_pass;
+    uint last_pass;
+} p;
+
+// pairs of (gid, value)
+shared ivec2 dst_row[BLOCK_SIZE];
+
+void topk(bool needs_bounds_check, const uint row) {
+    const int col = int(gl_LocalInvocationID.x);
+
+    // initialize indices
+    if (gl_GlobalInvocationID.x < p.ncols_input) {
+        if (p.first_pass != 0) {
+            const uint row_offset = row * p.ncols_input;
+            dst_row[col] = ivec2(gl_GlobalInvocationID.x, floatBitsToInt(data_a[row_offset + gl_GlobalInvocationID.x]));
+        } else {
+            const uint row_offset = row * p.orig_ncols;
+            dst_row[col] = data_s[row_offset + gl_GlobalInvocationID.x];
+        }
+    } else {
+        dst_row[col] = ivec2(p.orig_ncols, 0);
+    }
+    barrier();
+
+    if (p.ncols_output == 1) {
+        // Fast path for single output - just do a max reduction
+        [[unroll]] for (int s = BLOCK_SIZE / 2; s >= 1; s /= 2) {
+            if (col < s) {
+                ivec2 a = dst_row[col];
+                ivec2 b = dst_row[col + s];
+                if (a.x >= p.orig_ncols ||
+                    b.x < p.orig_ncols && b.y > a.y) {
+                    dst_row[col] = b;
+                }
+            }
+            barrier();
+        }
+    } else {
+        // bitonic sort on this group of elements
+        uint num_outer_loop_iters = NCOLS_PADDED_LOG2;
+        for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
+            uint num_inner_loop_iters = outer_idx + 1;
+            for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
+                const int ixj = int(col ^ j);
+
+                int idx_0 = (col & k) == 0 ? col : ixj;
+                int idx_1 = (col & k) == 0 ? ixj : col;
+
+                ivec2 sh_idx_0 = dst_row[idx_0];
+                ivec2 sh_idx_1 = dst_row[idx_1];
+                bool idx_0_oob = needs_bounds_check ? sh_idx_0.x >= p.orig_ncols : false;
+                bool idx_1_oob = needs_bounds_check ? sh_idx_1.x >= p.orig_ncols : false;
+
+                if ((idx_0_oob ||
+                    (!idx_1_oob && intBitsToFloat(sh_idx_0.y) < intBitsToFloat(sh_idx_1.y))) && (ixj > col)) {
+                    dst_row[idx_0] = sh_idx_1;
+                    dst_row[idx_1] = sh_idx_0;
+                }
+
+                barrier();
+            }
+        }
+    }
+
+    if (col < p.ncols_output && gl_GlobalInvocationID.x < p.orig_ncols) {
+        if (p.last_pass != 0) {
+            const uint row_offset = row * p.ncols_output;
+            data_d[row_offset + col] = dst_row[col].x;
+        } else {
+            const uint row_offset = row * p.orig_ncols + gl_WorkGroupID.x * p.ncols_output;
+            data_t[row_offset + col] = dst_row[col];
+        }
+    }
+}
+
+void main() {
+    // Fast path for fully occupied workgroups
+    if ((p.ncols_input % BLOCK_SIZE) == 0) {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            topk(false, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    } else {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            topk(true, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    }
+}

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

@@ -0,0 +1,199 @@
+#version 450
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_debug_printf : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_shuffle : enable
+
+#include "types.glsl"
+
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int SUBGROUP_SIZE = 32;
+layout(constant_id = 2) const int SUBGROUP_SIZE_LOG2 = 5;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+// Input can either be the source (A) or intermediate values (S).
+// Similarly, output can be either destination (D) or intermediate values (S).
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 0) readonly buffer S {ivec2 data_s[];};
+layout (binding = 1) writeonly buffer D {int data_d[];};
+layout (binding = 1) writeonly buffer T {ivec2 data_t[];};
+
+layout (push_constant) uniform parameter {
+    uint orig_ncols;
+    uint ncols_input;
+    uint ncols_output;
+    uint nrows;
+    uint first_pass;
+    uint last_pass;
+} p;
+
+// pairs of (gid, value)
+shared ivec2 dst_row[BLOCK_SIZE];
+
+shared int counts[SUBGROUP_SIZE];
+shared int sh_min_idx;
+shared uint sh_total;
+shared uint offset_partials[BLOCK_SIZE / SUBGROUP_SIZE];
+
+// Map float values to uint such that comparisons still work.
+// Positive values set the high bit, negative values are inverted.
+// +0.0 -> 0x80000000, -0.0 -> 0x7FFFFFFF are in the correct places.
+uint f2ui(float x) {
+    uint y = floatBitsToUint(x);
+    if ((y & 0x80000000) != 0) {
+        y ^= ~0;
+    } else {
+        y |= 0x80000000;
+    }
+    return y;
+}
+
+void topk(const uint row) {
+    const int tid = int(gl_LocalInvocationID.x);
+
+    // initialize indices
+    if (gl_GlobalInvocationID.x < p.ncols_input) {
+        if (p.first_pass != 0) {
+            const uint row_offset = row * p.ncols_input;
+            dst_row[tid] = ivec2(gl_GlobalInvocationID.x, floatBitsToInt(data_a[row_offset + gl_GlobalInvocationID.x]));
+        } else {
+            const uint row_offset = row * p.orig_ncols;
+            dst_row[tid] = data_s[row_offset + gl_GlobalInvocationID.x];
+        }
+    } else {
+        dst_row[tid] = ivec2(p.orig_ncols, 0xFF800000); // -inf
+    }
+    barrier();
+
+    if (p.ncols_output == 1) {
+        // Fast path for single output - just do a max reduction
+        [[unroll]] for (int s = BLOCK_SIZE / 2; s >= 1; s /= 2) {
+            if (tid < s) {
+                ivec2 a = dst_row[tid];
+                ivec2 b = dst_row[tid + s];
+                if (a.x >= p.orig_ncols ||
+                    b.x < p.orig_ncols && b.y > a.y) {
+                    dst_row[tid] = b;
+                }
+            }
+            barrier();
+        }
+    } else {
+        // Do an N-ary search to find the K-th largest value.
+        // We remap the float values to be comparable as unsigned integers,
+        // and split the range into 2^N smaller ranges where N is the
+        // subgroup size. Count how many values are in each range, if the K-th
+        // largest value is in the middle of one of thee ranges then repeat
+        // and split again.
+
+        // Mask is the current set of bits we're searching. Shift is the LSB index.
+        int shift = 32 - SUBGROUP_SIZE_LOG2;
+        uint mask = ((1 << SUBGROUP_SIZE_LOG2) - 1) << shift;
+
+        // The current range.
+        uint range_min = 0;
+        uint range_max = 0xFF800000;
+        // How many are above the current range, and how many we need to find.
+        uint total = 0;
+        uint limit = min(p.ncols_output, p.ncols_input - gl_WorkGroupID.x * BLOCK_SIZE);
+
+        while (mask != 0) {
+            barrier();
+            // Initialize bucket counts to zero.
+            if (tid < SUBGROUP_SIZE) {
+                counts[tid] = 0;
+            }
+            barrier();
+            // Count how many values are in each bucket.
+            if (tid < p.ncols_input) {
+                float y = intBitsToFloat(dst_row[tid].y);
+                uint fy = f2ui(y);
+                if (fy >= range_min && fy < range_max) {
+                    uint bucket = (fy & mask) >> shift;
+                    atomicAdd(counts[bucket], 1);
+                }
+            }
+            barrier();
+
+            // On the first subgroup, do a scan to count (from the top down) how
+            // many elements are in the top N buckets. Find the index of the first
+            // that is over the limit. Copy it to the other invocations through
+            // shared memory.
+            if (tid < SUBGROUP_SIZE) {
+                uint partial_sum = counts[SUBGROUP_SIZE - 1 - tid];
+                partial_sum = subgroupInclusiveAdd(partial_sum) + total;
+                uint t = subgroupBallotFindLSB(subgroupBallot(partial_sum >= limit));
+                if (tid == t) {
+                    sh_min_idx = int(SUBGROUP_SIZE - 1 - t);
+                    sh_total = partial_sum;
+                }
+            }
+            barrier();
+            int min_idx = sh_min_idx;
+            total = sh_total;
+
+            // Update the range, and break if we've found the K-th largest.
+            range_max = range_min + ((min_idx + 1) << shift);
+            range_min = range_min + (min_idx << shift);
+
+            if (total == p.ncols_output) {
+                break;
+            }
+            total -= counts[min_idx];
+            mask >>= SUBGROUP_SIZE_LOG2;
+            shift -= SUBGROUP_SIZE_LOG2;
+            if (shift < 0) {
+                shift = 0;
+            }
+        }
+
+        ivec2 v = dst_row[tid];
+
+        // We need to compact these values to the start of the dst_row array.
+        // Have each subgroup count how many items it'll store, so other
+        // subgroups can compute their base offset.
+        bool top = f2ui(intBitsToFloat(v.y)) >= range_min;
+        uvec4 b = subgroupBallot(top);
+        uint bit_count = subgroupBallotBitCount(b);
+        if ((tid % SUBGROUP_SIZE) == 0) {
+            offset_partials[tid / SUBGROUP_SIZE] = bit_count;
+        }
+        barrier();
+
+        uint out_idx = 0;
+        [[unroll]] for (int i = 0; i < BLOCK_SIZE / SUBGROUP_SIZE; ++i) {
+            if (i < tid / SUBGROUP_SIZE) {
+                out_idx += offset_partials[i];
+            }
+        }
+
+        uint bit_count_ex = subgroupBallotExclusiveBitCount(b);
+        if (top) {
+            // TODO: Copy directly to the output?
+            dst_row[out_idx + bit_count_ex] = v;
+        }
+
+        barrier();
+    }
+
+    if (tid < p.ncols_output && gl_GlobalInvocationID.x < p.orig_ncols) {
+        if (p.last_pass != 0) {
+            const uint row_offset = row * p.ncols_output;
+            data_d[row_offset + tid] = dst_row[tid].x;
+        } else {
+            const uint row_offset = row * p.orig_ncols + gl_WorkGroupID.x * p.ncols_output;
+            data_t[row_offset + tid] = dst_row[tid];
+        }
+    }
+}
+
+void main() {
+    uint row = gl_WorkGroupID.y;
+    while (row < p.nrows) {
+        topk(row);
+        row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+    }
+}

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

@@ -0,0 +1,43 @@
+#version 450
+
+#include "rte.glsl"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+#define GGML_TRI_TYPE_UPPER_DIAG 0
+#define GGML_TRI_TYPE_UPPER      1
+#define GGML_TRI_TYPE_LOWER_DIAG 2
+#define GGML_TRI_TYPE_LOWER      3
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i03 = fastdiv(idx, p.ne0_012mp, p.ne0_012L);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, p.ne0_01L);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, p.ne0_0L);
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+
+    int param = floatBitsToInt(p.param1);
+    bool pass = false;
+    switch (param) {
+    case GGML_TRI_TYPE_UPPER_DIAG: pass = i00 >= i01; break;
+    case GGML_TRI_TYPE_UPPER:      pass = i00 >  i01; break;
+    case GGML_TRI_TYPE_LOWER_DIAG: pass = i00 <= i01; break;
+    case GGML_TRI_TYPE_LOWER:      pass = i00 <  i01; break;
+    }
+
+    if (pass) {
+        const float val = float(data_a[get_aoffset() + src0_idx(idx)]);
+        data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val);
+    } else {
+        data_d[get_doffset() + dst_idx(idx)] = D_TYPE(0);
+    }
+}

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

@@ -679,14 +679,20 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
         string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
 
-        string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
 
         // mul mat vec with integer dot product
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-        if (is_legacy_quant(tname)) {
+        if (is_legacy_quant(tname) || tname == "mxfp4" || is_k_quant(tname)) {
             string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
             string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
             string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
+
+            string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
+            string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+            string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
         }
 #endif
 
@@ -846,6 +852,9 @@ void process_shaders() {
     string_to_spv("abs_f16",        "abs.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("abs_f32",        "abs.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
+    string_to_spv("tri_f16",        "tri.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("tri_f32",        "tri.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+
     string_to_spv("softplus_f16",   "softplus.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("softplus_f32",   "softplus.comp",    {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
@@ -913,9 +922,13 @@ void process_shaders() {
     string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
     string_to_spv("argsort_large_f32", "argsort_large.comp", {{"A_TYPE", "float"}});
 
+    string_to_spv("topk_argsort_f32", "topk_argsort.comp", {{"A_TYPE", "float"}});
+    string_to_spv("topk_nary_search_f32", "topk_nary_search.comp", {{"A_TYPE", "float"}});
+
     string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));
     string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
+    string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     for (std::string dim_str : {"", "_3d"}) {
         for (bool bda : {false, true}) {
@@ -940,6 +953,8 @@ void process_shaders() {
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
     string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
+    string_to_spv("solve_tri_f32", "solve_tri.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+
     for (auto transpose : {false, true}) {
         for (auto unroll : {false, true}) {
             for (auto a_f16 : {false, true}) {
@@ -1091,7 +1106,7 @@ void write_output_files() {
 
     for (const std::string& btype : btypes) {
     for (const auto& tname : type_names) {
-        if (btype == "q8_1" && !is_legacy_quant(tname)) {
+        if (btype == "q8_1" && !is_legacy_quant(tname) && tname != "mxfp4" && !is_k_quant(tname)) {
             continue;
         }
         hdr << "extern const void * arr_dmmv_"   << tname << "_" << btype << "_f32_data[3];\n";
@@ -1100,6 +1115,16 @@ void write_output_files() {
             src << "const void * arr_dmmv_"   << tname << "_" << btype << "_f32_data[3] = {mul_mat_vec_" << tname << "_" << btype << "_f32_data, mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_data, mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_no_shmem_data};\n";
             src << "const uint64_t arr_dmmv_" << tname << "_" << btype << "_f32_len[3] =  {mul_mat_vec_" << tname << "_" << btype << "_f32_len,  mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_len, mul_mat_vec_"  << tname << "_" << btype << "_f32_subgroup_no_shmem_len};\n";
         }
+
+        if (btype == "f16") {
+            continue;
+        }
+        hdr << "extern const void * arr_dmmv_id_"   << tname << "_" << btype << "_f32_data[3];\n";
+        hdr << "extern const uint64_t arr_dmmv_id_" << tname << "_" << btype << "_f32_len[3];\n";
+        if (basename(input_filepath) == "mul_mat_vec.comp") {
+            src << "const void * arr_dmmv_id_"   << tname << "_" << btype << "_f32_data[3] = {mul_mat_vec_id_" << tname << "_" << btype << "_f32_data, mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_data, mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_no_shmem_data};\n";
+            src << "const uint64_t arr_dmmv_id_" << tname << "_" << btype << "_f32_len[3] =  {mul_mat_vec_id_" << tname << "_" << btype << "_f32_len,  mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_len, mul_mat_vec_id_"  << tname << "_" << btype << "_f32_subgroup_no_shmem_len};\n";
+        }
     }
     }
 

ggml/src/ggml.c

@@ -990,6 +990,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
+    "TOP_K",
     "LEAKY_RELU",
     "TRI",
     "FILL",
@@ -1023,7 +1024,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 94, "GGML_OP_COUNT != 94");
+static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1098,6 +1099,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
+    "top_k(x)",
     "leaky_relu(x)",
     "tri(x)",
     "fill(x, c)",
@@ -1131,7 +1133,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 94, "GGML_OP_COUNT != 94");
+static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -5036,28 +5038,6 @@ struct ggml_tensor * ggml_roll(
     return result;
 }
 
-// ggml_arange
-
-struct ggml_tensor * ggml_arange(
-        struct ggml_context * ctx,
-        float                 start,
-        float                 stop,
-        float                 step) {
-    GGML_ASSERT(stop > start);
-
-    const int64_t steps = (int64_t) ceilf((stop - start) / step);
-
-    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
-
-    ggml_set_op_params_f32(result, 0, start);
-    ggml_set_op_params_f32(result, 1, stop);
-    ggml_set_op_params_f32(result, 2, step);
-
-    result->op = GGML_OP_ARANGE;
-
-    return result;
-}
-
 // ggml_timestep_embedding
 
 struct ggml_tensor * ggml_timestep_embedding(
@@ -5139,6 +5119,7 @@ struct ggml_tensor * ggml_argsort(
         struct ggml_tensor   * a,
         enum ggml_sort_order   order) {
     GGML_ASSERT(a->ne[0] <= INT32_MAX);
+
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
@@ -5149,6 +5130,24 @@ struct ggml_tensor * ggml_argsort(
     return result;
 }
 
+// ggml_argsort_top_k
+
+struct ggml_tensor * ggml_argsort_top_k(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   k) {
+    GGML_ASSERT(a->ne[0] >= k);
+
+    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
+
+    result = ggml_view_4d(ctx, result,
+                k, result->ne[1], result->ne[2], result->ne[3],
+                   result->nb[1], result->nb[2], result->nb[3],
+                0);
+
+    return result;
+}
+
 // ggml_top_k
 
 struct ggml_tensor * ggml_top_k(
@@ -5157,12 +5156,32 @@ struct ggml_tensor * ggml_top_k(
         int                   k) {
     GGML_ASSERT(a->ne[0] >= k);
 
-    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_I32, k, a->ne[1], a->ne[2], a->ne[3]);
 
-    result = ggml_view_4d(ctx, result,
-                k, result->ne[1], result->ne[2], result->ne[3],
-                   result->nb[1], result->nb[2], result->nb[3],
-                0);
+    result->op     = GGML_OP_TOP_K;
+    result->src[0] = a;
+
+    return result;
+}
+
+// ggml_arange
+
+struct ggml_tensor * ggml_arange(
+        struct ggml_context * ctx,
+        float                 start,
+        float                 stop,
+        float                 step) {
+    GGML_ASSERT(stop > start);
+
+    const int64_t steps = (int64_t) ceilf((stop - start) / step);
+
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
+
+    ggml_set_op_params_f32(result, 0, start);
+    ggml_set_op_params_f32(result, 1, stop);
+    ggml_set_op_params_f32(result, 2, step);
+
+    result->op = GGML_OP_ARANGE;
 
     return result;
 }

gguf-py/gguf/constants.py

@@ -25,6 +25,20 @@ class Keys:
         ALIGNMENT                  = "general.alignment"
         FILE_TYPE                  = "general.file_type"
 
+        # Recommended Sampler Parameters
+        SAMPLING_SEQUENCE           = "general.sampling.sequence"
+        SAMPLING_TOP_K              = "general.sampling.top_k"
+        SAMPLING_TOP_P              = "general.sampling.top_p"
+        SAMPLING_MIN_P              = "general.sampling.min_p"
+        SAMPLING_XTC_PROBABILITY    = "general.sampling.xtc_probability"
+        SAMPLING_XTC_THRESHOLD      = "general.sampling.xtc_threshold"
+        SAMPLING_TEMP               = "general.sampling.temp"
+        SAMPLING_PENALTY_LAST_N     = "general.sampling.penalty_last_n"
+        SAMPLING_PENALTY_REPEAT     = "general.sampling.penalty_repeat"
+        SAMPLING_MIROSTAT           = "general.sampling.mirostat"
+        SAMPLING_MIROSTAT_TAU       = "general.sampling.mirostat_tau"
+        SAMPLING_MIROSTAT_ETA       = "general.sampling.mirostat_eta"
+
         # Authorship Metadata
         NAME                       = "general.name"
         AUTHOR                     = "general.author"
@@ -352,6 +366,7 @@ class MODEL_ARCH(IntEnum):
     QWEN2VL          = auto()
     QWEN3            = auto()
     QWEN3MOE         = auto()
+    QWEN3NEXT        = auto()
     QWEN3VL          = auto()
     QWEN3VLMOE       = auto()
     PHI2             = auto()
@@ -427,6 +442,7 @@ class MODEL_ARCH(IntEnum):
     APERTUS          = auto()
     COGVLM           = auto()
     MINIMAXM2        = auto()
+    RND1             = auto()
     PANGU_EMBED      = auto()
 
 
@@ -516,6 +532,7 @@ class MODEL_TENSOR(IntEnum):
     SSM_D                = auto()
     SSM_NORM             = auto()
     SSM_OUT              = auto()
+    SSM_BETA_ALPHA       = auto() # qwen3next
     TIME_MIX_W0          = auto()
     TIME_MIX_W1          = auto()
     TIME_MIX_W2          = auto()
@@ -721,6 +738,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.QWEN2VL:          "qwen2vl",
     MODEL_ARCH.QWEN3:            "qwen3",
     MODEL_ARCH.QWEN3MOE:         "qwen3moe",
+    MODEL_ARCH.QWEN3NEXT:        "qwen3next",
     MODEL_ARCH.QWEN3VL:          "qwen3vl",
     MODEL_ARCH.QWEN3VLMOE:       "qwen3vlmoe",
     MODEL_ARCH.PHI2:             "phi2",
@@ -797,6 +815,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.APERTUS:          "apertus",
     MODEL_ARCH.MINIMAXM2:        "minimax-m2",
     MODEL_ARCH.COGVLM:           "cogvlm",
+    MODEL_ARCH.RND1:             "rnd1",
     MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
 }
 
@@ -884,6 +903,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.SSM_D:                     "blk.{bid}.ssm_d",
     MODEL_TENSOR.SSM_NORM:                  "blk.{bid}.ssm_norm",
     MODEL_TENSOR.SSM_OUT:                   "blk.{bid}.ssm_out",
+    MODEL_TENSOR.SSM_BETA_ALPHA:            "blk.{bid}.ssm_ba",
     MODEL_TENSOR.TIME_MIX_W0:               "blk.{bid}.time_mix_w0",
     MODEL_TENSOR.TIME_MIX_W1:               "blk.{bid}.time_mix_w1",
     MODEL_TENSOR.TIME_MIX_W2:               "blk.{bid}.time_mix_w2",
@@ -1553,6 +1573,35 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.QWEN3NEXT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.ATTN_GATE,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_INP_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.SSM_A,
+        MODEL_TENSOR.SSM_CONV1D,
+        MODEL_TENSOR.SSM_DT,
+        MODEL_TENSOR.SSM_NORM,
+        MODEL_TENSOR.SSM_IN,
+        MODEL_TENSOR.SSM_BETA_ALPHA,
+        MODEL_TENSOR.SSM_OUT
+    ],
     MODEL_ARCH.QWEN3VL: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2991,6 +3040,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.VISEXP_UP,
         MODEL_TENSOR.VISEXP_DOWN,
     ],
+    MODEL_ARCH.RND1: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
     MODEL_ARCH.PANGU_EMBED: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

gguf-py/gguf/gguf_writer.py

@@ -4,6 +4,7 @@ import logging
 import os
 import shutil
 import struct
+import sys
 import tempfile
 from dataclasses import dataclass
 from enum import Enum, auto
@@ -370,10 +371,15 @@ class GGUFWriter:
 
     def add_tensor(
         self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
-        raw_dtype: GGMLQuantizationType | None = None,
+        raw_dtype: GGMLQuantizationType | None = None, tensor_endianess: GGUFEndian | None = None
     ) -> None:
-        if self.endianess == GGUFEndian.BIG:
-            tensor.byteswap(inplace=True)
+        # if tensor endianness is not passed, assume it's native to system
+        if tensor_endianess is None:
+            tensor_endianess = GGUFEndian.BIG if sys.byteorder == 'big' else GGUFEndian.LITTLE
+
+        if tensor_endianess != self.endianess:
+            # Don't byteswap inplace since lazy copies cannot handle it
+            tensor = tensor.byteswap(inplace=False)
         if self.use_temp_file and self.temp_file is None:
             fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256 * 1024 * 1024)
             fp.seek(0)
@@ -394,13 +400,18 @@ class GGUFWriter:
         if pad != 0:
             fp.write(bytes([0] * pad))
 
-    def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None:
+    def write_tensor_data(self, tensor: np.ndarray[Any, Any], tensor_endianess: GGUFEndian | None = None) -> None:
         if self.state is not WriterState.TI_DATA and self.state is not WriterState.WEIGHTS:
             raise ValueError(f'Expected output file to contain tensor info or weights, got {self.state}')
         assert self.fout is not None
 
-        if self.endianess == GGUFEndian.BIG:
-            tensor.byteswap(inplace=True)
+        # if tensor endianness is not passed, assume it's native to system
+        if tensor_endianess is None:
+            tensor_endianess = GGUFEndian.BIG if sys.byteorder == 'big' else GGUFEndian.LITTLE
+
+        if tensor_endianess != self.endianess:
+            # Don't byteswap inplace since lazy copies cannot handle it
+            tensor = tensor.byteswap(inplace=False)
 
         file_id = -1
         for i, tensors in enumerate(self.tensors):
@@ -496,6 +507,42 @@ class GGUFWriter:
     def add_file_type(self, ftype: int) -> None:
         self.add_uint32(Keys.General.FILE_TYPE, ftype)
 
+    def add_sampling_sequence(self, sequence: str) -> None:
+        self.add_string(Keys.General.SAMPLING_SEQUENCE, sequence)
+
+    def add_sampling_top_k(self, top_k: int) -> None:
+        self.add_int32(Keys.General.SAMPLING_TOP_K, top_k)
+
+    def add_sampling_top_p(self, top_p: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_TOP_P, top_p)
+
+    def add_sampling_min_p(self, min_p: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_MIN_P, min_p)
+
+    def add_sampling_xtc_probability(self, xtc_probability: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_XTC_PROBABILITY, xtc_probability)
+
+    def add_sampling_xtc_threshold(self, xtc_threshold: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_XTC_THRESHOLD, xtc_threshold)
+
+    def add_sampling_temp(self, temp: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_TEMP, temp)
+
+    def add_sampling_penalty_last_n(self, penalty_last_n: int) -> None:
+        self.add_int32(Keys.General.SAMPLING_PENALTY_LAST_N, penalty_last_n)
+
+    def add_sampling_penalty_repeat(self, penalty_repeat: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_PENALTY_REPEAT, penalty_repeat)
+
+    def add_sampling_mirostat(self, mirostat: int) -> None:
+        self.add_int32(Keys.General.SAMPLING_MIROSTAT, mirostat)
+
+    def add_sampling_mirostat_tau(self, mirostat_tau: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_MIROSTAT_TAU, mirostat_tau)
+
+    def add_sampling_mirostat_eta(self, mirostat_eta: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_MIROSTAT_ETA, mirostat_eta)
+
     def add_name(self, name: str) -> None:
         self.add_string(Keys.General.NAME, name)
 

gguf-py/gguf/metadata.py

@@ -17,6 +17,20 @@ logger = logging.getLogger("metadata")
 
 @dataclass
 class Metadata:
+    # Recommended Sampler Parameters to be written to GGUF KV Store
+    sampling_sequence: Optional[str] = None
+    sampling_top_k: Optional[int] = None
+    sampling_top_p: Optional[float] = None
+    sampling_min_p: Optional[float] = None
+    sampling_xtc_probability: Optional[float] = None
+    sampling_xtc_threshold: Optional[float] = None
+    sampling_temp: Optional[float] = None
+    sampling_penalty_last_n: Optional[int] = None
+    sampling_penalty_repeat: Optional[float] = None
+    sampling_mirostat: Optional[int] = None
+    sampling_mirostat_tau: Optional[float] = None
+    sampling_mirostat_eta: Optional[float] = None
+
     # Authorship Metadata to be written to GGUF KV Store
     name: Optional[str] = None
     author: Optional[str] = None
@@ -54,15 +68,43 @@ class Metadata:
 
         model_card = Metadata.load_model_card(model_path)
         hf_params = Metadata.load_hf_parameters(model_path)
+        gen_config = Metadata.load_generation_config(model_path)
         # TODO: load adapter_config.json when possible, it usually contains the base model of the LoRA adapter
 
         # heuristics
         metadata = Metadata.apply_metadata_heuristic(metadata, model_card, hf_params, model_path, total_params)
 
+        if gen_config:
+            metadata.sampling_sequence        = gen_config.get("sequence",        metadata.sampling_sequence)
+            metadata.sampling_top_k           = gen_config.get("top_k",           metadata.sampling_top_k)
+            metadata.sampling_top_p           = gen_config.get("top_p",           metadata.sampling_top_p)
+            metadata.sampling_min_p           = gen_config.get("min_p",           metadata.sampling_min_p)
+            metadata.sampling_xtc_probability = gen_config.get("xtc_probability", metadata.sampling_xtc_probability)
+            metadata.sampling_xtc_threshold   = gen_config.get("xtc_threshold",   metadata.sampling_xtc_threshold)
+            metadata.sampling_temp            = gen_config.get("temperature",     metadata.sampling_temp)
+            metadata.sampling_penalty_last_n  = gen_config.get("penalty_last_n",  metadata.sampling_penalty_last_n)
+            metadata.sampling_penalty_repeat  = gen_config.get("penalty_repeat",  metadata.sampling_penalty_repeat)
+            metadata.sampling_mirostat        = gen_config.get("mirostat",        metadata.sampling_mirostat)
+            metadata.sampling_mirostat_tau    = gen_config.get("mirostat_tau",    metadata.sampling_mirostat_tau)
+            metadata.sampling_mirostat_eta    = gen_config.get("mirostat_eta",    metadata.sampling_mirostat_eta)
+
         # Metadata Override File Provided
         # This is based on LLM_KV_NAMES mapping in llama.cpp
         metadata_override = Metadata.load_metadata_override(metadata_override_path)
 
+        metadata.sampling_sequence        = metadata_override.get(Keys.General.SAMPLING_SEQUENCE,        metadata.sampling_sequence)
+        metadata.sampling_top_k           = metadata_override.get(Keys.General.SAMPLING_TOP_K,           metadata.sampling_top_k)
+        metadata.sampling_top_p           = metadata_override.get(Keys.General.SAMPLING_TOP_P,           metadata.sampling_top_p)
+        metadata.sampling_min_p           = metadata_override.get(Keys.General.SAMPLING_MIN_P,           metadata.sampling_min_p)
+        metadata.sampling_xtc_probability = metadata_override.get(Keys.General.SAMPLING_XTC_PROBABILITY, metadata.sampling_xtc_probability)
+        metadata.sampling_xtc_threshold   = metadata_override.get(Keys.General.SAMPLING_XTC_THRESHOLD,   metadata.sampling_xtc_threshold)
+        metadata.sampling_temp            = metadata_override.get(Keys.General.SAMPLING_TEMP,            metadata.sampling_temp)
+        metadata.sampling_penalty_last_n  = metadata_override.get(Keys.General.SAMPLING_PENALTY_LAST_N,  metadata.sampling_penalty_last_n)
+        metadata.sampling_penalty_repeat  = metadata_override.get(Keys.General.SAMPLING_PENALTY_REPEAT,  metadata.sampling_penalty_repeat)
+        metadata.sampling_mirostat        = metadata_override.get(Keys.General.SAMPLING_MIROSTAT,        metadata.sampling_mirostat)
+        metadata.sampling_mirostat_tau    = metadata_override.get(Keys.General.SAMPLING_MIROSTAT_TAU,    metadata.sampling_mirostat_tau)
+        metadata.sampling_mirostat_eta    = metadata_override.get(Keys.General.SAMPLING_MIROSTAT_ETA,    metadata.sampling_mirostat_eta)
+
         metadata.name            = metadata_override.get(Keys.General.NAME,            metadata.name)
         metadata.author          = metadata_override.get(Keys.General.AUTHOR,          metadata.author)
         metadata.version         = metadata_override.get(Keys.General.VERSION,         metadata.version)
@@ -172,6 +214,23 @@ class Metadata:
         with open(config_path, "r", encoding="utf-8") as f:
             return json.load(f)
 
+    @staticmethod
+    def load_generation_config(model_path: Optional[Path] = None) -> dict[str, Any]:
+        if model_path is None or not model_path.is_dir():
+            return {}
+
+        generation_config_path = model_path / "generation_config.json"
+
+        if not generation_config_path.is_file():
+            return {}
+
+        try:
+            with open(generation_config_path, "r", encoding="utf-8") as f:
+                return json.load(f)
+        except (json.JSONDecodeError, IOError):
+            # not all models have valid generation_config.json
+            return {}
+
     @staticmethod
     def id_to_title(string):
         # Convert capitalization into title form unless acronym or version number
@@ -546,6 +605,32 @@ class Metadata:
 
     def set_gguf_meta_model(self, gguf_writer: gguf.GGUFWriter):
         assert self.name is not None
+
+        if self.sampling_sequence is not None:
+            gguf_writer.add_sampling_sequence(self.sampling_sequence)
+        if self.sampling_top_k is not None:
+            gguf_writer.add_sampling_top_k(self.sampling_top_k)
+        if self.sampling_top_p is not None:
+            gguf_writer.add_sampling_top_p(self.sampling_top_p)
+        if self.sampling_min_p is not None:
+            gguf_writer.add_sampling_min_p(self.sampling_min_p)
+        if self.sampling_xtc_probability is not None:
+            gguf_writer.add_sampling_xtc_probability(self.sampling_xtc_probability)
+        if self.sampling_xtc_threshold is not None:
+            gguf_writer.add_sampling_xtc_threshold(self.sampling_xtc_threshold)
+        if self.sampling_temp is not None:
+            gguf_writer.add_sampling_temp(self.sampling_temp)
+        if self.sampling_penalty_last_n is not None:
+            gguf_writer.add_sampling_penalty_last_n(self.sampling_penalty_last_n)
+        if self.sampling_penalty_repeat is not None:
+            gguf_writer.add_sampling_penalty_repeat(self.sampling_penalty_repeat)
+        if self.sampling_mirostat is not None:
+            gguf_writer.add_sampling_mirostat(self.sampling_mirostat)
+        if self.sampling_mirostat_tau is not None:
+            gguf_writer.add_sampling_mirostat_tau(self.sampling_mirostat_tau)
+        if self.sampling_mirostat_eta is not None:
+            gguf_writer.add_sampling_mirostat_eta(self.sampling_mirostat_eta)
+
         gguf_writer.add_name(self.name)
 
         if self.author is not None: