vulkan: Call ggml_vk_buffer_write_2d from ggml_vk_buffer_copy (#16793)

This lets the copy to the destination device use the host-visible
vidmem optimization.

.github/workflows/build.yml

@@ -1305,6 +1305,81 @@ jobs:
           cd examples/llama.android
           ./gradlew build --no-daemon
 
+  android-ndk-build:
+    runs-on: ubuntu-latest
+
+    env:
+      OPENCL_VERSION: 2025.07.22
+
+    strategy:
+      matrix:
+        include:
+          - build: 'arm64-cpu'
+            defines: '-D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_CURL=OFF -D GGML_OPENMP=OFF'
+          - build: 'arm64-snapdragon'
+            defines: '--preset arm64-android-snapdragon-release'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Install OpenCL Headers and Libs
+        id: install_opencl
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        run: |
+          mkdir opencl
+          curl -L -o opencl/clhpp.tar.gz      https://github.com/KhronosGroup/OpenCL-CLHPP/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          curl -L -o opencl/headers.tar.gz    https://github.com/KhronosGroup/OpenCL-Headers/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          curl -L -o opencl/icd-loader.tar.gz https://github.com/KhronosGroup/OpenCL-ICD-Loader/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          tar -xaf opencl/headers.tar.gz    -C opencl
+          tar -xaf opencl/clhpp.tar.gz      -C opencl
+          tar -xaf opencl/icd-loader.tar.gz -C opencl
+          sudo cp -r opencl/OpenCL-Headers-${OPENCL_VERSION}/CL         ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+          sudo cp -r opencl/OpenCL-CLHPP-${OPENCL_VERSION}/include/CL/* ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include/CL
+          cd opencl/OpenCL-ICD-Loader-${OPENCL_VERSION}
+          cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -DOPENCL_ICD_LOADER_HEADERS_DIR=${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=31 -DANDROID_STL=c++_shared
+          cmake --build build
+          sudo cp build/libOpenCL.so ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+          rm -rf opencl
+
+      - name: Install Hexagon SDK
+        id: install_hexsdk
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        env:
+          HEXSDK_VER: 6.4.0.2
+          HEXTLS_VER: 19.0.04
+        run: |
+          curl -L -o hex-sdk.tar.gz https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v$HEXSDK_VER/hexagon-sdk-v$HEXSDK_VER-amd64-lnx.tar.xz
+          mkdir hex-sdk
+          tar -xaf hex-sdk.tar.gz -C hex-sdk
+          ls -l hex-sdk
+          sudo mv hex-sdk /opt/hexagon
+          echo "HEXAGON_SDK_ROOT=/opt/hexagon/$HEXSDK_VER"                                     >> "$GITHUB_ENV"
+          echo "HEXAGON_TOOLS_ROOT=/opt/hexagon/$HEXSDK_VER/tools/HEXAGON_Tools/$HEXTLS_VER"   >> "$GITHUB_ENV"
+          echo "DEFAULT_HLOS_ARCH=64"                                                          >> "$GITHUB_ENV"
+          echo "DEFAULT_TOOLS_VARIANT=toolv19"                                                 >> "$GITHUB_ENV"
+          echo "DEFAULT_NO_QURT_INC=0"                                                         >> "$GITHUB_ENV"
+          echo "DEFAULT_DSP_ARCH=v73"                                                          >> "$GITHUB_ENV"
+
+      - name: Update CMake presets
+        id: update_presets
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        run: |
+          cp docs/backend/hexagon/CMakeUserPresets.json .
+
+      - name: Build
+        id: ndk_build
+        run: |
+          cmake ${{ matrix.defines }} -B build
+          cmake --build build
+          cmake --install build --prefix pkg-adb/llama.cpp
+
+      - name: Test
+        id: cmake_test
+        run: |
+          echo "FIXME: test on devices"
+
   openEuler-latest-cmake-cann:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:

.github/workflows/release.yml

@@ -134,6 +134,8 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
+          - build: 's390x-z15' # z15 because our CI runners are on z15
+            os: ubuntu-22.04-s390x
           # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
           # - build: 'arm64'
           #   os: ubuntu-22.04-arm

.github/workflows/update-ops-docs.yml

@@ -3,10 +3,12 @@ name: Update Operations Documentation
 on:
     push:
         paths:
+            - 'docs/ops.md'
             - 'docs/ops/**'
             - 'scripts/create_ops_docs.py'
     pull_request:
         paths:
+            - 'docs/ops.md'
             - 'docs/ops/**'
             - 'scripts/create_ops_docs.py'
 

CODEOWNERS

@@ -55,7 +55,7 @@
 /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
+/ggml/src/ggml-cuda/mmf.*               @JohannesGaessler @am17an
 /ggml/src/ggml-cuda/mmq.*               @JohannesGaessler
 /ggml/src/ggml-cuda/mmvf.*              @JohannesGaessler
 /ggml/src/ggml-cuda/mmvq.*              @JohannesGaessler
@@ -65,6 +65,7 @@
 /ggml/src/ggml-impl.h                   @ggerganov @slaren
 /ggml/src/ggml-metal/                   @ggerganov
 /ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
+/ggml/src/ggml-hexagon/                 @max-krasnyansky
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @rgerganov

README.md

@@ -84,6 +84,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
 - [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
+- [x] [Jamba](https://huggingface.co/ai21labs)
 - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
@@ -138,6 +139,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 - [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 - [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
+- [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
 
 #### Multimodal
 
@@ -187,6 +189,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
 - Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
 - Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi)
+- Go (no CGo needed): [hybridgroup/yzma](https://github.com/hybridgroup/yzma)
 
 </details>
 
@@ -278,6 +281,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
+| [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
 
 ## Obtaining and quantizing models
 

ci/run.sh

@@ -75,7 +75,7 @@ if [ ! -z ${GG_BUILD_ROCM} ]; then
         exit 1
     fi
 
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DAMDGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then

common/arg.cpp

@@ -1760,7 +1760,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
     add_opt(common_arg(
         {"-t", "--threads"}, "N",
-        string_format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads),
+        string_format("number of CPU threads to use during generation (default: %d)", params.cpuparams.n_threads),
         [](common_params & params, int value) {
             params.cpuparams.n_threads = value;
             if (params.cpuparams.n_threads <= 0) {
@@ -3248,7 +3248,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--embd-output-format"}, "FORMAT",
-        "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix",
+        "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix, \"raw\" = plain whitespace-delimited output (one embedding per line)",
         [](common_params & params, const std::string & value) {
             params.embd_out = value;
         }
@@ -3435,7 +3435,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.use_jinja = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_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.cpp

@@ -9,8 +9,11 @@
 #include <minja/chat-template.hpp>
 #include <minja/minja.hpp>
 
+#include <algorithm>
 #include <cstdio>
+#include <cctype>
 #include <exception>
+#include <functional>
 #include <iostream>
 #include <optional>
 #include <stdexcept>
@@ -640,6 +643,7 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_SEED_OSS: return "Seed-OSS";
         case COMMON_CHAT_FORMAT_NEMOTRON_V2: return "Nemotron V2";
         case COMMON_CHAT_FORMAT_APERTUS: return "Apertus";
+        case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS: return "LFM2 with JSON tools";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -986,6 +990,126 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     return data;
 }
 
+
+// Case-insensitive find
+static size_t ifind_string(const std::string & haystack, const std::string & needle, size_t pos = 0) {
+    auto it = std::search(
+        haystack.begin() + pos, haystack.end(),
+        needle.begin(), needle.end(),
+        [](char a, char b) { return std::tolower(a) == std::tolower(b); }
+    );
+    return (it == haystack.end()) ? std::string::npos : std::distance(haystack.begin(), it);
+}
+
+static common_chat_params common_chat_params_init_lfm2(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+    const auto is_json_schema_provided = !inputs.json_schema.is_null();
+    const auto is_grammar_provided = !inputs.grammar.empty();
+    const auto are_tools_provided = inputs.tools.is_array() && !inputs.tools.empty();
+
+    // the logic requires potentially modifying the messages
+    auto tweaked_messages = inputs.messages;
+
+    auto replace_json_schema_marker = [](json & messages) -> bool {
+        static std::string marker1 = "force json schema.\n";
+        static std::string marker2 = "force json schema.";
+
+        if (messages.empty() || messages.at(0).at("role") != "system") {
+            return false;
+        }
+
+        std::string content = messages.at(0).at("content");
+
+        for (const auto & marker : {marker1, marker2}) {
+            const auto pos = ifind_string(content, marker);
+            if (pos != std::string::npos) {
+                content.replace(pos, marker.length(), "");
+                // inject modified content back into the messages
+                messages.at(0).at("content") = content;
+                return true;
+            }
+        }
+
+        return false;
+    };
+
+    // Lfm2 model does not natively work with json, but can generally understand the tools structure
+    //
+    // Example of the pytorch dialog structure:
+    //     <|startoftext|><|im_start|>system
+    //     List of tools: <|tool_list_start|>[{"name": "get_candidate_status", "description": "Retrieves the current status of a candidate in the recruitment process", "parameters": {"type": "object", "properties": {"candidate_id": {"type": "string", "description": "Unique identifier for the candidate"}}, "required": ["candidate_id"]}}]<|tool_list_end|><|im_end|>
+    //     <|im_start|>user
+    //     What is the current status of candidate ID 12345?<|im_end|>
+    //     <|im_start|>assistant
+    //     <|tool_call_start|>[get_candidate_status(candidate_id="12345")]<|tool_call_end|>Checking the current status of candidate ID 12345.<|im_end|>
+    //     <|im_start|>tool
+    //     <|tool_response_start|>{"candidate_id": "12345", "status": "Interview Scheduled", "position": "Clinical Research Associate", "date": "2023-11-20"}<|tool_response_end|><|im_end|>
+    //     <|im_start|>assistant
+    //     The candidate with ID 12345 is currently in the "Interview Scheduled" stage for the position of Clinical Research Associate, with an interview date set for 2023-11-20.<|im_end|>
+    //
+    // For the llama server compatibility with json tools semantic,
+    // the client can add "Follow json schema." line into the system message prompt to force the json output.
+    //
+    if (are_tools_provided && (is_json_schema_provided || is_grammar_provided)) {
+        // server/utils.hpp prohibits that branch for the custom grammar anyways
+        throw std::runtime_error("Tools call must not use \"json_schema\" or \"grammar\", use non-tool invocation if you want to use custom grammar");
+    } else if (are_tools_provided && replace_json_schema_marker(tweaked_messages)) {
+        LOG_INF("%s: Using tools to build a grammar\n", __func__);
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            auto schemas = json::array();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                schemas.push_back({
+                    {"type", "object"},
+                    {"properties", {
+                        {"name", {
+                            {"type", "string"},
+                            {"const", function.at("name")},
+                        }},
+                        {"arguments", function.at("parameters")},
+                    }},
+                    {"required", json::array({"name", "arguments", "id"})},
+                });
+            });
+            auto schema = json {
+                {"type", "array"},
+                {"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
+                {"minItems", 1},
+            };
+            if (!inputs.parallel_tool_calls) {
+                schema["maxItems"] = 1;
+            }
+
+            builder.add_rule("root", "\"<|tool_call_start|>\"" + builder.add_schema("tool_calls", schema) + "\"<|tool_call_end|>\"");
+        });
+        // model has no concept of tool selection mode choice,
+        // if the system prompt rendered correctly it will produce a tool call
+        // the grammar goes inside the tool call body
+        data.grammar_lazy = true;
+        data.grammar_triggers = {{COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL, "\\s*<\\|tool_call_start\\|>\\s*\\["}};
+        data.preserved_tokens = {"<|tool_call_start|>", "<|tool_call_end|>"};
+        data.format = COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS;
+    } else if (are_tools_provided && (!is_json_schema_provided && !is_grammar_provided)) {
+        LOG_INF("%s: Using tools without json schema or grammar\n", __func__);
+        // output those tokens
+        data.preserved_tokens = {"<|tool_call_start|>", "<|tool_call_end|>"};
+    } else if (is_json_schema_provided) {
+        LOG_INF("%s: Using provided json schema to build a grammar\n", __func__);
+        data.grammar = json_schema_to_grammar(inputs.json_schema);
+    } else if (is_grammar_provided) {
+        LOG_INF("%s: Using provided grammar\n", __func__);
+        data.grammar = inputs.grammar;
+    } else {
+        LOG_INF("%s: Using content relying on the template\n", __func__);
+    }
+
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
+    LOG_DBG("%s: Prompt: %s\n", __func__, data.prompt.c_str());
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_magistral(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     data.prompt = apply(tmpl, inputs);
@@ -2499,6 +2623,71 @@ static void common_chat_parse_apertus(common_chat_msg_parser & builder) {
     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) {
     // Parse thinking tags first - this handles the main reasoning content
     builder.try_parse_reasoning("<seed:think>", "</seed:think>");
@@ -2748,6 +2937,12 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_apertus(tmpl, params);
     }
 
+    // LFM2 (w/ tools)
+    if (src.find("List of tools: <|tool_list_start|>[") != std::string::npos &&
+        src.find("]<|tool_list_end|>") != std::string::npos) {
+        return common_chat_params_init_lfm2(tmpl, params);
+    }
+
     // Use generic handler when mixing tools + JSON schema.
     // TODO: support that mix in handlers below.
     if ((params.tools.is_array() && params.json_schema.is_object())) {
@@ -2926,6 +3121,9 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         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;
         default:
             throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
     }

common/chat.h

@@ -116,6 +116,7 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_SEED_OSS,
     COMMON_CHAT_FORMAT_NEMOTRON_V2,
     COMMON_CHAT_FORMAT_APERTUS,
+    COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };

common/json-schema-to-grammar.cpp

@@ -41,9 +41,9 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
-static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
-    auto has_min = min_value != std::numeric_limits<int>::min();
-    auto has_max = max_value != std::numeric_limits<int>::max();
+static void _build_min_max_int(int64_t min_value, int64_t max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
+    auto has_min = min_value != std::numeric_limits<int64_t>::min();
+    auto has_max = max_value != std::numeric_limits<int64_t>::max();
 
     auto digit_range = [&](char from, char to) {
         out << "[";
@@ -159,7 +159,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
     if (has_min) {
         if (min_value < 0) {
             out << "\"-\" (";
-            _build_min_max_int(std::numeric_limits<int>::min(), -min_value, out, decimals_left, /* top_level= */ false);
+            _build_min_max_int(std::numeric_limits<int64_t>::min(), -min_value, out, decimals_left, /* top_level= */ false);
             out << ") | [0] | [1-9] ";
             more_digits(0, decimals_left - 1);
         } else if (min_value == 0) {
@@ -194,7 +194,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
             }
             digit_range(c, c);
             out << " (";
-            _build_min_max_int(std::stoi(min_s.substr(1)), std::numeric_limits<int>::max(), out, less_decimals, /* top_level= */ false);
+            _build_min_max_int(std::stoll(min_s.substr(1)), std::numeric_limits<int64_t>::max(), out, less_decimals, /* top_level= */ false);
             out << ")";
             if (c < '9') {
                 out << " | ";
@@ -216,7 +216,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
             _build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
         } else {
             out << "\"-\" (";
-            _build_min_max_int(-max_value, std::numeric_limits<int>::max(), out, decimals_left, /* top_level= */ false);
+            _build_min_max_int(-max_value, std::numeric_limits<int64_t>::max(), out, decimals_left, /* top_level= */ false);
             out << ")";
         }
         return;
@@ -601,7 +601,10 @@ private:
     }
 
     std::string _resolve_ref(const std::string & ref) {
-        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
+        auto it = ref.find('#');
+        std::string ref_fragment = it != std::string::npos ? ref.substr(it + 1) : ref;
+        static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
+        std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
         if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
             _refs_being_resolved.insert(ref);
             json resolved = _refs[ref];
@@ -774,11 +777,24 @@ public:
                         std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
-                            if (target.is_null() || !target.contains(sel)) {
+                            if (target.is_object() && target.contains(sel)) {
+                                target = target[sel];
+                            } else if (target.is_array()) {
+                                size_t sel_index;
+                                try {
+                                    sel_index = std::stoul(sel);
+                                } catch (const std::invalid_argument & e) {
+                                    sel_index = target.size();
+                                }
+                                if (sel_index >= target.size()) {
+                                    _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
+                                    return;
+                                }
+                                target = target[sel_index];
+                            } else {
                                 _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
                                 return;
                             }
-                            target = target[sel];
                         }
                         _refs[ref] = target;
                     }
@@ -925,17 +941,17 @@ public:
             int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
             return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
         } else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
-            int min_value = std::numeric_limits<int>::min();
-            int max_value = std::numeric_limits<int>::max();
+            int64_t min_value = std::numeric_limits<int64_t>::min();
+            int64_t max_value = std::numeric_limits<int64_t>::max();
             if (schema.contains("minimum")) {
-                min_value = schema["minimum"].get<int>();
+                min_value = schema["minimum"].get<int64_t>();
             } else if (schema.contains("exclusiveMinimum")) {
-                min_value = schema["exclusiveMinimum"].get<int>() + 1;
+                min_value = schema["exclusiveMinimum"].get<int64_t>() + 1;
             }
             if (schema.contains("maximum")) {
-                max_value = schema["maximum"].get<int>();
+                max_value = schema["maximum"].get<int64_t>();
             } else if (schema.contains("exclusiveMaximum")) {
-                max_value = schema["exclusiveMaximum"].get<int>() - 1;
+                max_value = schema["exclusiveMaximum"].get<int64_t>() - 1;
             }
             std::stringstream out;
             out << "(";

convert_hf_to_gguf.py

@@ -29,12 +29,29 @@ if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 from gguf.vocab import MistralTokenizerType, MistralVocab
-from mistral_common.tokens.tokenizers.base import TokenizerVersion
-from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN, DATASET_STD
-from mistral_common.tokens.tokenizers.tekken import Tekkenizer
-from mistral_common.tokens.tokenizers.sentencepiece import (
-    SentencePieceTokenizer,
-)
+
+try:
+    from mistral_common.tokens.tokenizers.base import TokenizerVersion # pyright: ignore[reportMissingImports]
+    from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN as _MISTRAL_COMMON_DATASET_MEAN, DATASET_STD as _MISTRAL_COMMON_DATASET_STD # pyright: ignore[reportMissingImports]
+    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # pyright: ignore[reportMissingImports]
+    from mistral_common.tokens.tokenizers.sentencepiece import ( # pyright: ignore[reportMissingImports]
+        SentencePieceTokenizer,
+    )
+
+    _mistral_common_installed = True
+    _mistral_import_error_msg = ""
+except ImportError:
+    _MISTRAL_COMMON_DATASET_MEAN = (0.48145466, 0.4578275, 0.40821073)
+    _MISTRAL_COMMON_DATASET_STD = (0.26862954, 0.26130258, 0.27577711)
+
+    _mistral_common_installed = False
+    TokenizerVersion = None
+    Tekkenizer = None
+    SentencePieceTokenizer = None
+    _mistral_import_error_msg = (
+        "Mistral format requires `mistral-common` to be installed. Please run "
+        "`pip install mistral-common[image,audio]` to install it."
+    )
 
 
 logger = logging.getLogger("hf-to-gguf")
@@ -73,10 +90,8 @@ class ModelBase:
     use_temp_file: bool
     lazy: bool
     dry_run: bool
-    part_names: list[str]
-    is_safetensors: bool
     hparams: dict[str, Any]
-    tensor_names: set[str] | None
+    model_tensors: dict[str, Callable[[], Tensor]]
     gguf_writer: gguf.GGUFWriter
     model_name: str | None
     metadata_override: Path | None
@@ -107,6 +122,9 @@ class ModelBase:
                 type(self) is MmprojModel:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
+        if self.is_mistral_format and not _mistral_common_installed:
+            raise ImportError(_mistral_import_error_msg)
+
         self.dir_model = dir_model
         self.ftype = ftype
         self.fname_out = fname_out
@@ -117,25 +135,8 @@ class ModelBase:
         self.dry_run = dry_run
         self.remote_hf_model_id = remote_hf_model_id
         self.sentence_transformers_dense_modules = sentence_transformers_dense_modules
-        if remote_hf_model_id is not None:
-            self.is_safetensors = True
-
-            def get_remote_tensors() -> Iterator[tuple[str, Tensor]]:
-                logger.info(f"Using remote model with HuggingFace id: {remote_hf_model_id}")
-                remote_tensors = gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id)
-                self.tensor_names = set(name for name in remote_tensors.keys())
-                for name, remote_tensor in remote_tensors.items():
-                    yield (name, LazyTorchTensor.from_remote_tensor(remote_tensor))
-
-            self.get_tensors = get_remote_tensors
-        else:
-            prefix = "model" if not self.is_mistral_format else "consolidated"
-            self.part_names = ModelBase.get_model_part_names(self.dir_model, prefix, ".safetensors")
-            self.is_safetensors = len(self.part_names) > 0
-            if not self.is_safetensors:
-                self.part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
         self.hparams = ModelBase.load_hparams(self.dir_model, self.is_mistral_format) if hparams is None else hparams
-        self.tensor_names = None
+        self.model_tensors = self.index_tensors(remote_hf_model_id=remote_hf_model_id)
         self.metadata_override = metadata_override
         self.model_name = model_name
         self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
@@ -151,6 +152,8 @@ class ModelBase:
                 logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})")
                 self.ftype = gguf.LlamaFileType.MOSTLY_BF16
 
+        self.dequant_model()
+
         # Configure GGUF Writer
         self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file,
                                            split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard)
@@ -172,67 +175,215 @@ class ModelBase:
             return None
         raise KeyError(f"could not find any of: {keys}")
 
-    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
-        tensor_names_from_parts: set[str] = set()
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        tensors: dict[str, Callable[[], Tensor]] = {}
+
+        if remote_hf_model_id is not None:
+            is_safetensors = True
+
+            logger.info(f"Using remote model with HuggingFace id: {remote_hf_model_id}")
+            remote_tensors = gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id)
+            for name, remote_tensor in remote_tensors.items():
+                tensors[name] = lambda r=remote_tensor: LazyTorchTensor.from_remote_tensor(r)
+
+            return tensors
+
+        prefix = "model" if not self.is_mistral_format else "consolidated"
+        part_names: list[str] = ModelBase.get_model_part_names(self.dir_model, prefix, ".safetensors")
+        is_safetensors: bool = len(part_names) > 0
+        if not is_safetensors:
+            part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
+
+        tensor_names_from_index: set[str] = set()
 
         if not self.is_mistral_format:
-            index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
+            index_name = "model.safetensors" if is_safetensors else "pytorch_model.bin"
             index_name += ".index.json"
             index_file = self.dir_model / index_name
 
             if index_file.is_file():
-                self.tensor_names = set()
                 logger.info(f"gguf: loading model weight map from '{index_name}'")
                 with open(index_file, "r", encoding="utf-8") as f:
                     index: dict[str, Any] = json.load(f)
                     weight_map = index.get("weight_map")
                     if weight_map is None or not isinstance(weight_map, dict):
                         raise ValueError(f"Can't load 'weight_map' from {index_name!r}")
-                    self.tensor_names.update(weight_map.keys())
+                    tensor_names_from_index.update(weight_map.keys())
             else:
-                self.tensor_names = tensor_names_from_parts
                 weight_map = {}
         else:
-            self.tensor_names = tensor_names_from_parts
             weight_map = {}
 
-        for part_name in self.part_names:
-            logger.info(f"gguf: loading model part '{part_name}'")
+        for part_name in part_names:
+            logger.info(f"gguf: indexing model part '{part_name}'")
             ctx: ContextManager[Any]
-            if self.is_safetensors:
+            if is_safetensors:
                 from safetensors import safe_open
                 ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu"))
             else:
                 ctx = contextlib.nullcontext(torch.load(str(self.dir_model / part_name), map_location="cpu", mmap=True, weights_only=True))
 
             with ctx as model_part:
-                tensor_names_from_parts.update(model_part.keys())
+                assert model_part is not None
 
                 for name in model_part.keys():
-                    if self.is_safetensors:
+                    if is_safetensors:
                         if self.lazy:
                             data = model_part.get_slice(name)
-                            data = LazyTorchTensor.from_safetensors_slice(data)
+                            data_gen = lambda data=data: LazyTorchTensor.from_safetensors_slice(data)  # noqa: E731
                         else:
                             data = model_part.get_tensor(name)
+                            data_gen = lambda data=data: data  # noqa: E731
                     else:
                         data = model_part[name]
                         if self.lazy:
-                            data = LazyTorchTensor.from_eager(data)
-                    yield name, data
+                            data_gen = lambda data=data: LazyTorchTensor.from_eager(data)  # noqa: E731
+                        else:
+                            data_gen = lambda data=data: data  # noqa: E731
+                    tensors[name] = data_gen
 
         # verify tensor name presence and identify potentially missing files
-        if len(tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
-            missing = sorted(self.tensor_names.difference(tensor_names_from_parts))
-            extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
-            missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
-            if len(extra) == 0 and len(missing_files) > 0:
-                raise ValueError(f"Missing or incomplete model files: {missing_files}\n"
-                                 f"Missing tensors: {missing}")
+        if len(tensor_names_from_index) > 0:
+            tensor_names_from_parts = set(tensors.keys())
+            if len(tensor_names_from_parts.symmetric_difference(tensor_names_from_index)) > 0:
+                missing = sorted(tensor_names_from_index.difference(tensor_names_from_parts))
+                extra = sorted(tensor_names_from_parts.difference(tensor_names_from_index))
+                missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
+                if len(extra) == 0 and len(missing_files) > 0:
+                    raise ValueError(f"Missing or incomplete model files: {missing_files}\n"
+                                     f"Missing tensors: {missing}")
+                else:
+                    raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
+                                     f"Missing tensors: {missing}\n"
+                                     f"Extra tensors: {extra}")
+
+        return tensors
+
+    def dequant_model(self):
+        tensors_to_remove: list[str] = []
+        new_tensors: dict[str, Callable[[], Tensor]] = {}
+
+        if (quant_config := self.hparams.get("quantization_config")) and isinstance(quant_config, dict):
+            quant_method = quant_config.get("quant_method")
+
+            def dequant_bitnet(weight: Tensor, scale: Tensor) -> Tensor:
+                weight = weight.view(torch.uint8)
+                orig_shape = weight.shape
+
+                shift = torch.tensor([0, 2, 4, 6], dtype=torch.uint8).reshape((4, *(1 for _ in range(len(orig_shape)))))
+                data = weight.unsqueeze(0).expand((4, *orig_shape)) >> shift
+                data = data & 3
+                data = (data.float() - 1).reshape((orig_shape[0] * 4, *orig_shape[1:]))
+
+                # The scale is inverted
+                return data / scale.float()
+
+            def dequant_simple(weight: Tensor, scale: Tensor) -> Tensor:
+                scale = scale.float()
+
+                if (weight_block_size := quant_config.get("weight_block_size")):
+                    # TODO: make sure it's a list of integers
+                    for i, size in enumerate(weight_block_size):
+                        scale = scale.repeat_interleave(size, i)
+                # unpad the scale (e.g. when the tensor size isn't a multiple of the block size)
+                scale = scale[tuple(slice(0, size) for size in weight.shape)]
+
+                return weight.float() * scale
+
+            # ref: https://github.com/ModelCloud/GPTQModel/blob/037c5c0f6c9e33c500d975b038d02e7ca437546d/gptqmodel/nn_modules/qlinear/__init__.py#L437-L476
+            def dequant_gptq(g_idx: Tensor, qweight: Tensor, qzeros: Tensor, scales: Tensor) -> Tensor:
+                bits = quant_config["bits"]
+                assert bits in (2, 3, 4, 8)
+                assert qweight.dtype == qzeros.dtype
+                maxq = (2 ** bits) - 1
+                weight = None
+                zeros = None
+                pack_dtype_bits = qweight.dtype.itemsize * 8
+
+                if bits in [2, 4, 8]:
+                    pack_factor = pack_dtype_bits // bits
+                    wf = torch.tensor(list(range(0, pack_dtype_bits, bits)), dtype=torch.int32).unsqueeze(0)
+                    if self.lazy:
+                        wf = LazyTorchTensor.from_eager(wf)
+
+                    zeros = torch.bitwise_right_shift(
+                        qzeros.unsqueeze(2).expand(-1, -1, pack_factor),
+                        wf.unsqueeze(0)
+                    ).to(torch.int16 if bits == 8 else torch.int8)
+                    zeros = torch.bitwise_and(zeros, maxq).reshape(scales.shape)
+
+                    weight = torch.bitwise_and(
+                        torch.bitwise_right_shift(
+                            qweight.unsqueeze(1).expand(-1, pack_factor, -1),
+                            wf.unsqueeze(-1)
+                        ).to(torch.int16 if bits == 8 else torch.int8),
+                        maxq
+                    )
+                elif bits == 3:
+                    raise NotImplementedError("3-bit gptq dequantization is not yet implemented")
+
+                assert weight is not None
+                assert zeros is not None
+
+                weight = weight.reshape(weight.shape[0] * weight.shape[1], weight.shape[2])
+
+                # gptq_v2 doesn't need to offset zeros
+                if quant_config.get("checkpoint_format", "gptq") == "gptq":
+                    zeros += 1
+
+                return (scales[g_idx].float() * (weight - zeros[g_idx]).float()).T
+
+            if quant_method == "bitnet":
+                for name in self.model_tensors.keys():
+                    if name.endswith(".weight_scale"):
+                        weight_name = name.removesuffix("_scale")
+                        w = self.model_tensors[weight_name]
+                        s = self.model_tensors[name]
+                        self.model_tensors[weight_name] = lambda w=w, s=s: dequant_bitnet(w(), s())
+                        tensors_to_remove.append(name)
+            elif quant_method == "fp8":
+                for name in self.model_tensors.keys():
+                    if name.endswith(".weight_scale_inv"):
+                        weight_name = name.removesuffix("_scale_inv")
+                        w = self.model_tensors[weight_name]
+                        s = self.model_tensors[name]
+                        self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s())
+                        tensors_to_remove.append(name)
+            elif quant_method == "gptq":
+                for name in self.model_tensors.keys():
+                    if name.endswith(".qweight"):
+                        base_name = name.removesuffix(".qweight")
+                        g_idx = self.model_tensors[base_name + ".g_idx"]
+                        qweight = self.model_tensors[base_name + ".qweight"]
+                        qzeros = self.model_tensors[base_name + ".qzeros"]
+                        scales = self.model_tensors[base_name + ".scales"]
+                        new_tensors[base_name + ".weight"] = (
+                            lambda g=g_idx, z=qzeros, w=qweight, s=scales: dequant_gptq(
+                                g(), w(), z(), s()
+                            )
+                        )
+                        tensors_to_remove += [
+                            base_name + n
+                            for n in (
+                                ".g_idx",
+                                ".qzeros",
+                                ".qweight",
+                                ".scales",
+                            )
+                        ]
             else:
-                raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
-                                 f"Missing tensors: {missing}\n"
-                                 f"Extra tensors: {extra}")
+                raise NotImplementedError(f"Quant method is not yet supported: {quant_method!r}")
+
+        for name in tensors_to_remove:
+            if name in self.model_tensors:
+                del self.model_tensors[name]
+
+        for name, value in new_tensors.items():
+            self.model_tensors[name] = value
+
+    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
+        for name, gen in self.model_tensors.items():
+            yield name, gen()
 
     def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
         if key not in gguf.MODEL_TENSORS[self.model_arch]:
@@ -591,6 +742,12 @@ class TextModel(ModelBase):
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
             logger.info(f"gguf: experts used count = {n_experts_used}")
+        if (n_expert_groups := self.hparams.get("n_group")) is not None:
+            self.gguf_writer.add_expert_group_count(n_expert_groups)
+            logger.info(f"gguf: expert groups count = {n_expert_groups}")
+        if (n_group_used := self.hparams.get("topk_group")) is not None:
+            self.gguf_writer.add_expert_group_used_count(n_group_used)
+            logger.info(f"gguf: expert groups used count = {n_group_used}")
 
         if (head_dim := self.hparams.get("head_dim")) is not None:
             self.gguf_writer.add_key_length(head_dim)
@@ -892,8 +1049,8 @@ class TextModel(ModelBase):
             # ref: https://huggingface.co/JetBrains/Mellum-4b-base
             res = "mellum"
         if chkhsh == "9b1be57e70d20d9501b2b3186e792d81181ae36ada3903c26f9fea418cf87206":
-            # ref: https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base
-            res = "llada-moe"
+            # ref: https://huggingface.co/inclusionAI/Ling-mini-base-2.0
+            res = "bailingmoe2"
         if chkhsh == "53e325976a6e142379c19b09afcae354f2f496f147afa8f9e189a33fe4e3024e":
             # ref: https://huggingface.co/ibm-granite/granite-docling-258M
             res = "granite-docling"
@@ -1346,6 +1503,17 @@ class MmprojModel(ModelBase):
     def set_type(self):
         self.gguf_writer.add_type(gguf.GGUFType.MMPROJ)
 
+    def prepare_metadata(self, vocab_only: bool):
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        output_type: str = self.ftype.name.partition("_")[2]
+
+        if self.fname_out.is_dir():
+            fname_default: str = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+            self.fname_out = self.fname_out / f"mmproj-{fname_default}.gguf"
+        else:
+            self.fname_out = self.fname_out.parent / gguf.fill_templated_filename(self.fname_out.name, output_type)
+
     def set_gguf_parameters(self):
         self.gguf_writer.add_file_type(self.ftype)
 
@@ -1363,8 +1531,8 @@ class MmprojModel(ModelBase):
             self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads"]))
 
             # preprocessor config
-            image_mean = DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]
-            image_std = DATASET_STD if self.is_mistral_format else self.preprocessor_config["image_std"]
+            image_mean = _MISTRAL_COMMON_DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]
+            image_std = _MISTRAL_COMMON_DATASET_STD if self.is_mistral_format else self.preprocessor_config["image_std"]
 
             self.gguf_writer.add_vision_image_mean(image_mean)
             self.gguf_writer.add_vision_image_std(image_std)
@@ -2033,6 +2201,9 @@ class LlamaModel(TextModel):
             self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
 
     def _set_vocab_mistral(self):
+        if not _mistral_common_installed:
+            raise ImportError(_mistral_import_error_msg)
+
         vocab = MistralVocab(self.dir_model)
         logger.info(
             f"Converting tokenizer {vocab.tokenizer_type} of size {vocab.vocab_size}."
@@ -2289,18 +2460,21 @@ class ArceeModel(LlamaModel):
 )
 class LlavaVisionModel(MmprojModel):
     img_break_tok_id = -1
+    use_break_tok = True
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         if self.hparams.get("model_type") == "pixtral":
             # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py
             self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
-            self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
+            if self.use_break_tok:
+                self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
         elif self.is_mistral_format:
             # hparams is already vision config here so norm_eps is only defined in global_config.
             self.hparams["norm_eps"] = self.global_config.get("norm_eps", None)
             assert self.hparams["norm_eps"] is not None, "norm_eps not found in params.json"
-            self.img_break_tok_id = self.find_vparam(["image_break_token_id"])
+            if self.use_break_tok:
+                self.img_break_tok_id = self.find_vparam(["image_break_token_id"])
         else:
             raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
         logger.info(f"Image break token id: {self.img_break_tok_id}")
@@ -3791,6 +3965,10 @@ class Qwen3Model(Qwen2Model):
         return torch.stack([true_row, false_row], dim=0)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if "model.vision_" in name:
+            # skip multimodal tensors
+            return []
+
         if self.is_rerank:
             is_tied_head = self.is_tied_embeddings and "embed_tokens" in name
             is_real_head = not self.is_tied_embeddings and "lm_head" in name
@@ -4358,27 +4536,6 @@ class CodeShellModel(TextModel):
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
         self.gguf_writer.add_rope_scaling_factor(1.0)
 
-    _has_tok_embd = False
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
-        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
-        tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
-
-        new_name = self.map_tensor_name(name)
-
-        # assuming token_embd.weight is seen before output.weight
-        if not self._has_tok_embd and new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
-            # even though the tensor file(s) does not contain the word embeddings they are still in the weight map
-            if self.tensor_names and "transformer.wte.weight" in self.tensor_names:
-                logger.debug(f"{tok_embd_name} not found before {output_name}, assuming they are tied")
-                self.tensor_names.remove("transformer.wte.weight")
-        elif new_name == tok_embd_name:
-            self._has_tok_embd = True
-
-        return [(new_name, data_torch)]
-
 
 @ModelBase.register("InternLM2ForCausalLM")
 class InternLM2Model(TextModel):
@@ -8055,6 +8212,101 @@ class BailingMoeModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("BailingMoeV2ForCausalLM")
+class BailingMoeV2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.BAILINGMOE2
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if nextn_layers := self.hparams.get("num_nextn_predict_layers", 0):
+            self.block_count = self.hparams["num_hidden_layers"] + nextn_layers
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        if (rope_dim := hparams.get("head_dim")) is None:
+            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
+
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+        else:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_shared_feed_forward_length(hparams.get("moe_shared_expert_intermediate_size", hparams["moe_intermediate_size"] * hparams["num_shared_experts"]))
+        self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_count(hparams["num_experts"])
+        self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
+        self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
+
+        if hparams["score_function"] == "sigmoid":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        elif hparams["score_function"] == "softmax":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+        else:
+            raise ValueError(f"Unsupported score_function value: {hparams['score_function']}")
+
+        if (nextn_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(nextn_layers)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if "mlp.experts" in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            tensors: list[tuple[str, Tensor]] = []
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+
+            return tensors
+
+        if name.endswith(".expert_bias"):
+            name = name.replace(".expert_bias", ".expert_bias.bias")
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register("GroveMoeForCausalLM", "modeling_grove_moe.GroveMoeForCausalLM")
 class GroveMoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GROVEMOE
@@ -8713,6 +8965,13 @@ class SmolLM3Model(LlamaModel):
 class GptOssModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GPT_OSS
 
+    # TODO: remove once MXFP4 is supported more generally
+    def dequant_model(self):
+        quant_config = self.hparams.get("quantization_config")
+        if quant_config is not None and quant_config.get("quant_method") == "mxfp4":
+            return
+        return super().dequant_model()
+
     def transform_nibble_layout(self, tensor):
         assert tensor.dtype == torch.uint8
         assert tensor.shape[-1] == 16
@@ -9115,7 +9374,7 @@ class MistralModel(LlamaModel):
 
     @staticmethod
     def get_community_chat_template(vocab: MistralVocab, templates_dir: Path, is_mistral_format: bool):
-        assert TokenizerVersion is not None, "mistral_common is not installed"
+        assert TokenizerVersion is not None and Tekkenizer is not None and SentencePieceTokenizer is not None, _mistral_import_error_msg
         assert isinstance(vocab.tokenizer, (Tekkenizer, SentencePieceTokenizer)), (
             f"Expected Tekkenizer or SentencePieceTokenizer, got {type(vocab.tokenizer)}"
         )
@@ -9183,6 +9442,21 @@ class PixtralModel(LlavaVisionModel):
         return super().map_tensor_name(name, try_suffixes)
 
 
+@ModelBase.register("LightOnOCRForConditionalGeneration")
+class LightOnOCRVisionModel(LlavaVisionModel):
+    is_mistral_format = False
+    use_break_tok = False
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.LIGHTONOCR)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        name = name.replace("model.vision_encoder.", "vision_tower.")
+        name = name.replace("model.vision_projection.", "multi_modal_projector.")
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("KimiVLForConditionalGeneration")
 class KimiVLModel(MmprojModel):
     def __init__(self, *args, **kwargs):
@@ -9492,11 +9766,9 @@ def main() -> None:
 
     logger.info(f"Loading model: {dir_model.name}")
 
-    if args.mmproj:
-        if "mmproj" not in fname_out.name:
-            fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-")
-
     is_mistral_format = args.mistral_format
+    if is_mistral_format and not _mistral_common_installed:
+        raise ImportError(_mistral_import_error_msg)
     disable_mistral_community_chat_template = args.disable_mistral_community_chat_template
 
     with torch.inference_mode():

convert_hf_to_gguf_update.py

@@ -139,7 +139,7 @@ models = [
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
-    {"name": "llada-moe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base", },
+    {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
     {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
 ]
 

docs/backend/hexagon/CMakeUserPresets.json

@@ -0,0 +1,49 @@
+{
+  "version": 4,
+  "configurePresets": [
+    {
+        "name": "arm64-android-snapdragon",
+        "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "ANDROID_ABI":      "arm64-v8a",
+            "ANDROID_PLATFORM": "android-31",
+            "CMAKE_TOOLCHAIN_FILE": "$env{ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake",
+            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
+            "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
+            "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
+            "CMAKE_CXX_FLAGS_RELWITHDEBINFO": "-O3 -DNDEBUG -g",
+            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
+            "PREBUILT_LIB_DIR": "android_aarch64",
+            "GGML_OPENMP":      "OFF",
+            "GGML_LLAMAFILE":   "OFF",
+            "GGML_OPENCL":      "ON",
+            "GGML_HEXAGON":     "ON",
+            "LLAMA_CURL":       "OFF"
+        }
+    },
+
+    {
+        "name": "arm64-windows-snapdragon",
+        "inherits": [ "base", "arm64-windows-llvm" ],
+        "cacheVariables": {
+            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
+            "PREBUILT_LIB_DIR": "windows_aarch64",
+            "GGML_OPENMP":      "OFF",
+            "GGML_LLAMAFILE":   "OFF",
+            "GGML_OPENCL":      "ON",
+            "GGML_HEXAGON":     "ON",
+            "LLAMA_CURL":       "OFF"
+        }
+    },
+
+    { "name": "arm64-android-snapdragon-debug"  , "inherits": [ "base", "arm64-android-snapdragon", "debug" ] },
+    { "name": "arm64-android-snapdragon-release", "inherits": [ "base", "arm64-android-snapdragon", "release" ] },
+
+    { "name": "arm64-windows-snapdragon-debug"  , "inherits": [ "base", "arm64-windows-snapdragon", "debug" ] },
+    { "name": "arm64-windows-snapdragon-release", "inherits": [ "base", "arm64-windows-snapdragon", "release" ] }
+  ]
+}

docs/backend/hexagon/README.md

@@ -0,0 +1,239 @@
+# Snapdragon-based Android devices
+
+## How to Build
+
+The easiest way to build llama.cpp for a Snapdragon-based Android device is using the toolchain Docker image (see github.com/snapdragon-toolchain).
+This image includes Android NDK, OpenCL SDK, Hexagon SDK, CMake, etc.
+
+This method works on Linux, macOS, and Windows. macOS and Windows users should install Docker Desktop.
+
+```
+~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.3
+[d]/> cd /workspace
+```
+
+The rest of the Android build process assumes that you're running inside the toolchain container.
+Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
+
+```
+[d]/workspace> cp docs/backend/hexagon/CMakeUserPresets.json .
+
+[d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
+Preset CMake variables:
+  ANDROID_ABI="arm64-v8a"
+  ...
+  CMAKE_TOOLCHAIN_FILE="/opt/android-ndk-r28b/build/cmake/android.toolchain.cmake"
+  GGML_HEXAGON="ON"
+  GGML_OPENCL="ON"
+  GGML_OPENMP="OFF"
+  HEXAGON_SDK_ROOT="/opt/hexagon/6.4.0.2"
+...
+-- Including OpenCL backend
+-- Including Hexagon backend
+...
+-- Build files have been written to: /workspace/build-snapdragon
+
+[d]/workspace> cmake --build build-snapdragon
+...
+[144/356] Performing build step for 'htp-v73'
+[1/16] Generating htp_iface_skel.c, htp_iface_stub.c, htp_iface.h
+[2/16] Building C object CMakeFiles/ggml-htp-v73.dir/hvx-sigmoid.c.obj
+[3/16] Building C object CMakeFiles/ggml-htp-v73.dir/htp-dma.c.obj
+[4/16] Building C object CMakeFiles/ggml-htp-v73.dir/worker-pool.c.obj
+...
+-- Installing: /workspace/build-snapdragon/ggml/src/ggml-hexagon/libggml-htp-v73.so
+-- Installing: /workspace/build-snapdragon/ggml/src/ggml-hexagon/libggml-htp-v75.so
+...
+```
+
+To generate an installable "package" simply use cmake --install:
+
+```
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-adb/llama.cpp
+-- Install configuration: "Release"
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-cpu.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-opencl.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-hexagon.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v73.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v75.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v79.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v81.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml.so
+...
+-- Installing: /workspace/pkg-adb/llama.cpp/bin/llama-bench
+-- Installing: /workspace/pkg-adb/llama.cpp/bin/llama-cli
+...
+```
+
+## How to Install
+
+For this step, your device needs to be configured for on-device development.
+Please see https://developer.android.com/studio/debug/dev-options for details.
+
+Once ADB is enabled, use `adb push` to install `pkg-snapdragon` on the device.
+**Note that the toolchain Docker image doesn't have ADB and doesn't set up the ADB bridge. Please use native ADB on the host.**
+
+```
+~/src/llama.cpp$ adb push pkg-adb/llama.cpp /data/local/tmp/
+pkg-adb/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
+pkg-adb/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
+pkg-adb/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
+102 files pushed, 0 skipped. 186.9 MB/s (963151597 bytes in 4.914s)
+```
+
+At this point, you should also install some models:
+
+```
+~/src/llama.cpp$ wget https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf
+...
+2025-10-11 12:04:52 (10.7 MB/s) - ‘Llama-3.2-1B-Instruct-Q4_0.gguf’ saved [773025920/773025920]
+
+~/src/llama.cpp$ adb push Llama-3.2-1B-Instruct-Q4_0.gguf /data/local/tmp/gguf
+Llama-3.2-1B-Instruct-Q4_0.gguf: 1 file pushed, 0 skipped. 38.3 MB/s (773025920 bytes in 19.250s)
+```
+
+## How to Run
+
+The easiest way to run llama.cpp cli tools is using provided wrapper scripts that properly set up all required environment variables.
+
+llama.cpp supports three backends on Snapdragon-based devices: CPU, Adreno GPU (GPUOpenCL), and Hexagon NPU (HTP0-4).
+You can select which backend to run the model on using the `D=` variable, which maps to the `--device` option.
+
+Hexagon NPU behaves as a "GPU" device when it comes to `-ngl` and other offload-related options.
+
+Here are some examples of running various llama.cpp tools via ADB.
+
+Simple question for Llama-3.2-1B
+
+```
+~/src/llama.cpp$ M=Llama-3.2-1B-Instruct-Q4_0.gguf D=HTP0 ./scripts/snapdragon/adb/run-cli.sh -no-cnv -p "what is the most popular cookie in the world?"
+...
+ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
+ggml-hex: Hexagon Arch version v79
+ggml-hex: allocating new session: HTP0
+ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v79.so?htp_iface_skel_handle_invoke&_modver=1.0&_dom=cdsp&_session=0 handle 0xb4000072c7955e50
+...
+load_tensors: offloading output layer to GPU
+load_tensors: offloaded 17/17 layers to GPU
+load_tensors:          CPU model buffer size =   225.49 MiB
+load_tensors:         HTP0 model buffer size =     0.26 MiB
+load_tensors:  HTP0-REPACK model buffer size =   504.00 MiB
+...
+I hope this helps you understand the world's most popular cookies! [end of text]
+...
+llama_perf_sampler_print:    sampling time =      30.08 ms /   487 runs   (    0.06 ms per token, 16191.77 tokens per second)
+llama_perf_context_print:        load time =     617.94 ms
+llama_perf_context_print: prompt eval time =      80.76 ms /    11 tokens (    7.34 ms per token,   136.21 tokens per second)
+llama_perf_context_print:        eval time =    9210.59 ms /   475 runs   (   19.39 ms per token,    51.57 tokens per second)
+llama_perf_context_print:       total time =    9454.92 ms /   486 tokens
+llama_perf_context_print:    graphs reused =        473
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                  439 =   225 +     136 +      77                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                  504 =   504 +       0 +       0                |
+```
+
+Summary request for OLMoE-1B-7B. This is a large model that requires two HTP sessions/devices
+
+```
+~/src/llama.cpp$ M=OLMoE-1B-7B-0125-Instruct-Q4_0.gguf NDEV=2 D=HTP0,HTP1 ./scripts/snapdragon/adb/run-cli.sh -f surfing.txt -no-cnv
+...
+ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
+ggml-hex: Hexagon Arch version v81
+ggml-hex: allocating new session: HTP0
+ggml-hex: allocating new session: HTP1
+...
+load_tensors: offloading output layer to GPU
+load_tensors: offloaded 17/17 layers to GPU
+load_tensors:          CPU model buffer size =   143.86 MiB
+load_tensors:         HTP1 model buffer size =     0.23 MiB
+load_tensors:  HTP1-REPACK model buffer size =  1575.00 MiB
+load_tensors:         HTP0 model buffer size =     0.28 MiB
+load_tensors:  HTP0-REPACK model buffer size =  2025.00 MiB
+...
+llama_context:        CPU  output buffer size =     0.19 MiB
+llama_kv_cache:       HTP1 KV buffer size =   238.00 MiB
+llama_kv_cache:       HTP0 KV buffer size =   306.00 MiB
+llama_kv_cache: size =  544.00 MiB (  8192 cells,  16 layers,  1/1 seqs), K (q8_0):  272.00 MiB, V (q8_0):  272.00 MiB
+llama_context:       HTP0 compute buffer size =    15.00 MiB
+llama_context:       HTP1 compute buffer size =    15.00 MiB
+llama_context:        CPU compute buffer size =    24.56 MiB
+...
+llama_perf_context_print: prompt eval time =    1730.57 ms /   212 tokens (    8.16 ms per token,   122.50 tokens per second)
+llama_perf_context_print:        eval time =    5624.75 ms /   257 runs   (   21.89 ms per token,    45.69 tokens per second)
+llama_perf_context_print:       total time =    7377.33 ms /   469 tokens
+llama_perf_context_print:    graphs reused =        255
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                  742 =   144 +     544 +      54                |
+llama_memory_breakdown_print: |   - HTP1-REPACK        |                 1575 =  1575 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2025 =  2025 +       0 +       0                |
+```
+
+Op test for MUL_MAT
+
+```
+~/src/llama.cpp$ HB=0 ./scripts/snapdragon/adb/run-tool.sh test-backend-ops -b HTP0 -o MUL_MAT
+...
+Backend 2/3: HTP0
+Device description: Hexagon
+Device memory: 2048 MB (2048 MB free)
+MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
+MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
+MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
+
+~/src/llama.cpp-hexagon$ M=Llama-3.2-1B-Instruct-Q4_0.gguf ./scripts/snapdragon/adb/run-bench.sh -p 128 -n 64
+...
+ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
+ggml-hex: Hexagon Arch version v79
+ggml-hex: allocating new session: HTP0
+ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v79.so?htp_iface_skel_handle_invoke&_modver=1.0&_dom=cdsp&_session=0 handle 0xb400007d4b231090
+| model          |       size | params | backend    | ngl | threads | n_batch | mmap |  test |           t/s |
+| ---------------| ---------: | -----: | ---------- | --: | ------: | ------: | ---: | ----: | ------------: |
+| llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 | pp128 | 169.42 ± 1.75 |
+| llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 |  tg64 |  51.54 ± 1.13 |
+
+build: 6a8cf8914 (6733)
+```
+
+## Environment variables
+
+- `GGML_HEXAGON_NDEV=1`
+  Controls the number of devices/sessions to allocate. The default is 1.
+  Most quantized models under 4B fit into a single session; an 8B model needs two, and a 20B model needs four.
+
+- `GGML_HEXAGON_NHVX=0`
+  Controls the number of HVX hardware threads to use. The default is all (actual number varies depending on the hardware version).
+
+- `GGML_HEXAGON_HOSTBUF=1`
+  Controls whether the Hexagon backend allocates host buffers. By default, all buffers except for REPACK are host buffers.
+  This option is required for testing Ops that require REPACK buffers (MUL_MAT and MUL_MAT_ID).
+
+- `GGML_HEXAGON_VERBOSE=1`
+  Enables verbose logging of Ops from the backend. Example output:
+
+  ```
+  ggml-hex: HTP0 graph-compute n_nodes 2
+  ggml-hex: HTP0 matmul : blk.27.ffn_up.weight x ffn_norm-27 -> ffn_up-27 : 3072:8192 x 3072:1 -> 8192:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x1
+  ggml-hex: HTP0 matmul : blk.27.ffn_gate.weight x ffn_norm-27 -> ffn_gate-27 : 3072:8192 x 3072:1 -> 8192:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x3
+  ggml-hex: HTP0 graph-compute n_nodes 1
+  ggml-hex: HTP0 matmul : blk.27.ffn_down.weight x ffn_gate_par-27 -> ffn_out-27 : 8192:3072 x 8192:1 -> 3072:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x0
+  ggml-hex: HTP0 get-tensor result_output : data 0x7592487000 offset 0 size 513024
+  ```
+
+- `GGML_HEXAGON_PROFILE=1`
+  Generates a host-side profile for the ggml-hexagon Ops.
+
+- `GGML_HEXAGON_OPMASK=0x0`
+  Allows enabling specific stages of the processing pipeline:
+
+  - `0x1` Enable Op Queue (i.e., queuing Ops into NPU)
+  - `0x2` Enable Dynamic Quantizer (if needed for the Op)
+  - `0x4` Enable Op Compute (MUL_MAT, etc.)
+
+  Examples:
+
+      `GGML_HEXAGON_OPMASK=0x1 llama-cli ...` - Ops are enqueued but NPU-side processing is stubbed out
+      `GGML_HEXAGON_OPMASK=0x3 llama-cli ...` - NPU performs dynamic quantization and skips the rest
+      `GGML_HEXAGON_OPMASK=0x7 llama-cli ...` - Full queuing and processing of Ops (default)

docs/backend/hexagon/developer.md

@@ -0,0 +1,109 @@
+# Hexagon backend developer details
+
+## Backend libraries
+
+The Hexagon backend consist of two parts:
+
+  - `libggml-hexagon`
+    This is the regular CPU-side GGML backend library, either shared or statically linked
+
+  - `libggml-htp-vNN`
+    This is the NPU-side (HTP stands for Hexagon Tensor Processor) shared library that contains the Op dispatcher and kernels.
+    The correct library is selected automatically at runtime based on the HW version.
+
+Here is an example of the build artifacts
+
+```
+~/src/llama.cpp$ ls -l pkg-adb/llama.cpp/lib/libggml*
+pkg-adb/llama.cpp/lib/libggml-base.so
+pkg-adb/llama.cpp/lib/libggml-cpu.so
+pkg-adb/llama.cpp/lib/libggml-hexagon.so      <<< CPU library
+pkg-adb/llama.cpp/lib/libggml-htp-v73.so      <<< HTP op/kernels for Hexagon v73
+pkg-adb/llama.cpp/lib/libggml-htp-v75.so
+pkg-adb/llama.cpp/lib/libggml-htp-v79.so
+pkg-adb/llama.cpp/lib/libggml-htp-v81.so
+```
+
+## Memory buffers
+
+Hexagon NPU backend takes advantage of the Snapdragon's unified memory model where all buffers are fully accessible by the CPU and GPU.
+The NPU does have a dedicated tightly-coupled memory called VTCM but that memory is used only for intermediate data (e.g. dynamically
+quantized tensors) or temporary data (chunks of the weight tensors fetched via DMA).
+
+Please note that currently the Hexagon backend does not implement SET/GET_ROWS Ops because there is no advantage in offloading those
+to the NPU at this point.
+
+The backend does allocates non-host buffers for the tensors with datatypes that require repacking: Q4_0, Q8_0, MXFP4.
+From the MMU perspective these buffers are still regular buffers (normal access by the CPU) they are marked as non-host simply to force
+the repacking.
+
+## Large model handling
+
+Hexagon NPU session (aka Process Domain (PD) in the Hexagon docs) is limited to a memory mapping of around 3.5GB.
+In llama.cpp/GGML the Hexagon session is mapped to a single GGML backend device (HTP0, HTP1, etc).
+
+In order to map models larger than 3.5GB we need to allocate multiple devices and split the model.
+For this we're taking advantage of the llama.cpp/GGML multi-GPU layer-splitting support.
+Each Hexagon device behaves like a GPU from the offload and model splitting perspective.
+
+Here is an example of running GPT-OSS-20B model on a newer Snapdragon device with 16GB of DDR.
+
+```
+M=gpt-oss-20b-Q4_0.gguf NDEV=4 D=HTP0,HTP1,HTP2,HTP3 P=surfing.txt scripts/snapdragon/adb/run-cli.sh -no-cnv -f surfing.txt -n 32
+...
+LD_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
+ADSP_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
+GGML_HEXAGON_NDEV=4 ./bin/llama-cli --no-mmap -m /data/local/tmp/llama.cpp/../gguf/gpt-oss-20b-Q4_0.gguf
+      -t 4 --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 --device HTP0,HTP1,HTP2,HTP3 -no-cnv -f surfing.txt
+...
+llama_model_loader: - type  f32:  289 tensors
+llama_model_loader: - type q4_0:   96 tensors
+llama_model_loader: - type q8_0:    2 tensors
+llama_model_loader: - type mxfp4:  72 tensors
+...
+load_tensors: offloaded 25/25 layers to GPU
+load_tensors:          CPU model buffer size =  1182.09 MiB
+load_tensors:         HTP1 model buffer size =     6.64 MiB
+load_tensors:  HTP1-REPACK model buffer size =  2505.94 MiB
+load_tensors:         HTP3 model buffer size =     5.55 MiB
+load_tensors:  HTP3-REPACK model buffer size =  2088.28 MiB
+load_tensors:         HTP0 model buffer size =     7.75 MiB
+load_tensors:  HTP0-REPACK model buffer size =  2923.59 MiB
+load_tensors:         HTP2 model buffer size =     6.64 MiB
+load_tensors:  HTP2-REPACK model buffer size =  2505.94 MiB
+...
+llama_context: n_ctx_per_seq (8192) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
+llama_context:        CPU  output buffer size =     0.77 MiB
+llama_kv_cache_iswa: creating non-SWA KV cache, size = 8192 cells
+llama_kv_cache:       HTP1 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP3 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP0 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP2 KV buffer size =    25.50 MiB
+llama_kv_cache: size =  102.00 MiB (  8192 cells,  12 layers,  1/1 seqs), K (q8_0):   51.00 MiB, V (q8_0):   51.00 MiB
+llama_kv_cache_iswa: creating     SWA KV cache, size = 256 cells
+llama_kv_cache:       HTP1 KV buffer size =     0.80 MiB
+llama_kv_cache:       HTP3 KV buffer size =     0.53 MiB
+llama_kv_cache:       HTP0 KV buffer size =     1.06 MiB
+llama_kv_cache:       HTP2 KV buffer size =     0.80 MiB
+llama_kv_cache: size =    3.19 MiB (   256 cells,  12 layers,  1/1 seqs), K (q8_0):    1.59 MiB, V (q8_0):    1.59 MiB
+llama_context:       HTP0 compute buffer size =    16.06 MiB
+llama_context:       HTP1 compute buffer size =    16.06 MiB
+llama_context:       HTP2 compute buffer size =    16.06 MiB
+llama_context:       HTP3 compute buffer size =    16.06 MiB
+llama_context:        CPU compute buffer size =    98.19 MiB
+...
+llama_perf_context_print: prompt eval time =    3843.67 ms /   197 tokens ( 19.51 ms per token, 51.25 tokens per second)
+llama_perf_context_print:        eval time =    1686.13 ms /    31 runs   ( 54.39 ms per token, 18.39 tokens per second)
+llama_perf_context_print:       total time =    6266.30 ms /   228 tokens
+llama_perf_context_print:    graphs reused =         30
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP2 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP3 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                 1476 =  1208 +     105 +     162                |
+llama_memory_breakdown_print: |   - HTP1-REPACK        |                 2505 =  2505 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP3-REPACK        |                 2088 =  2088 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2923 =  2923 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP2-REPACK        |                 2505 =  2505 +       0 +       0                |
+```

docs/build.md

@@ -261,10 +261,12 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build --config Release -- -j 16
   ```
 
+  Note: `GPU_TARGETS` is optional, omitting it will build the code for all GPUs in the current system.
+
   To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
 
   The rocWMMA library is included by default when installing the ROCm SDK using the `rocm` meta package provided by AMD. Alternatively, if you are not using the meta package, you can install the library using the `rocwmma-dev` or `rocwmma-devel` package, depending on your system's package manager.
@@ -282,17 +284,17 @@ You can download it from your Linux distro's package manager or from here: [ROCm
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
   HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
-      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build -- -j 16
   ```
 
 - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
   ```bash
   set PATH=%HIP_PATH%\bin;%PATH%
-  cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+  cmake -S . -B build -G Ninja -DGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
   cmake --build build
   ```
-  Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
+  If necessary, adapt `GPU_TARGETS` to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
   Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
 
 

docs/ops.md

@@ -22,6 +22,7 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
@@ -41,6 +42,7 @@ Legend:
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -70,18 +72,19 @@ Legend:
 |                     OPT_STEP_SGD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
@@ -97,8 +100,8 @@ Legend:
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
 |                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
@@ -108,5 +111,6 @@ Legend:
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                         TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
 |                            XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

docs/ops/CPU.csv

@@ -59,6 +59,14 @@
 "CPU","EXP","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
+"CPU","FLOOR","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","FLOOR","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","CEIL","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","CEIL","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","ROUND","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","ROUND","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","TRUNC","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","TRUNC","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","ABS","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
 "CPU","ABS","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","SGN","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
@@ -119,6 +127,14 @@
 "CPU","EXP","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
+"CPU","FLOOR","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","FLOOR","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","CEIL","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","CEIL","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","ROUND","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","ROUND","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","TRUNC","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","TRUNC","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","1","yes","CPU"

docs/ops/SYCL.csv

@@ -31,6 +31,14 @@
 "SYCL0","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","XIELU","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","SYCL"
 "SYCL0","XIELU","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","SYCL"
+"SYCL0","FLOOR","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","FLOOR","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
+"SYCL0","ROUND","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","ROUND","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
+"SYCL0","TRUNC","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","TRUNC","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","ABS","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","ABS","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
 "SYCL0","SGN","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -95,6 +103,14 @@
 "SYCL0","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","XIELU","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","SYCL"
 "SYCL0","XIELU","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","SYCL"
+"SYCL0","FLOOR","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","FLOOR","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
+"SYCL0","ROUND","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","ROUND","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
+"SYCL0","TRUNC","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","TRUNC","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","ABS","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","ABS","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
 "SYCL0","SGN","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -5621,25 +5637,25 @@
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000000,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000000","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000000,inplace=0","support","1","yes","SYCL"
-"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000000","support","0","no","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000000","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001,inplace=0","support","1","yes","SYCL"
-"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000001","support","0","no","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000001","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100,inplace=0","support","1","yes","SYCL"
-"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000100","support","0","no","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000100","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000,inplace=0","support","1","yes","SYCL"
-"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.100000","support","0","no","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.100000","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=1","support","1","yes","SYCL"
 "SYCL0","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.000000,broadcast=0,multi_add=0","support","1","yes","SYCL"
@@ -9363,8 +9379,8 @@
 "SYCL0","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1]","support","1","yes","SYCL"
 "SYCL0","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1","support","1","yes","SYCL"
 "SYCL0","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0","support","1","yes","SYCL"
-"SYCL0","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","0","no","SYCL"
-"SYCL0","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","0","no","SYCL"
+"SYCL0","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","0","yes","SYCL"
+"SYCL0","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","0","yes","SYCL"
 "SYCL0","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","0","no","SYCL"
 "SYCL0","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","0","no","SYCL"
 "SYCL0","TIMESTEP_EMBEDDING","type=f32,ne_a=[2,1,1,1],dim=320,max_period=10000","support","1","yes","SYCL"

docs/ops/Vulkan.csv

@@ -3263,27 +3263,27 @@
 "Vulkan0","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=1.000000,broadcast=0","support","1","yes","Vulkan"
 "Vulkan0","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=1.000000,broadcast=1","support","1","yes","Vulkan"
 "Vulkan0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=1,n_seqs=1","support","1","yes","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=1","support","1","yes","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"

examples/embedding/README.md

@@ -38,6 +38,7 @@ The above command will output space-separated float values.
 |            | multiple embeddings          | $[[x_1,...,x_n],[x_1,...,x_n],...,[x_1,...,x_n]]$
 | 'json'     | openai style                 |
 | 'json+'    | add cosine similarity matrix |
+| 'raw'      | plain text output            |
 
 ### --embd-separator $"string"$
 | $"string"$   | |

examples/embedding/embedding.cpp

@@ -70,6 +70,29 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     }
 }
 
+// plain, pipe-friendly output: one embedding per line
+static void print_raw_embeddings(const float * emb,
+                                 int n_embd_count,
+                                 int n_embd,
+                                 const llama_model * model,
+                                 enum llama_pooling_type pooling_type,
+                                 int embd_normalize) {
+    const uint32_t n_cls_out = llama_model_n_cls_out(model);
+    const bool is_rank = (pooling_type == LLAMA_POOLING_TYPE_RANK);
+    const int cols = is_rank ? std::min<int>(n_embd, (int) n_cls_out) : n_embd;
+
+    for (int j = 0; j < n_embd_count; ++j) {
+        for (int i = 0; i < cols; ++i) {
+            if (embd_normalize == 0) {
+                LOG("%1.0f%s", emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
+            } else {
+                LOG("%1.7f%s", emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
+            }
+        }
+        LOG("\n");
+    }
+}
+
 int main(int argc, char ** argv) {
     common_params params;
 
@@ -372,6 +395,8 @@ int main(int argc, char ** argv) {
         }
 
         if (notArray) LOG("\n}\n");
+    } else if (params.embd_out == "raw") {
+        print_raw_embeddings(emb, n_embd_count, n_embd, model, pooling_type, params.embd_normalize);
     }
 
     LOG("\n");

examples/json_schema_to_grammar.py

@@ -371,8 +371,17 @@ class SchemaConverter:
                         raise ValueError(f'Unsupported ref {ref}')
 
                     for sel in ref.split('#')[-1].split('/')[1:]:
-                        assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
-                        target = target[sel]
+                        assert target is not None, f'Error resolving ref {ref}: {sel} not in {target}'
+                        if isinstance(target, list):
+                            try:
+                                sel_index = int(sel)
+                            except ValueError:
+                                raise ValueError(f'Error resolving ref {ref}: {sel} not in {target}')
+                            assert 0 <= sel_index < len(target), f'Error resolving ref {ref}: {sel} not in {target}'
+                            target = target[sel_index]
+                        else:
+                            assert sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
+                            target = target[sel]
 
                     self._refs[ref] = target
                 else:
@@ -547,7 +556,8 @@ class SchemaConverter:
 
 
     def _resolve_ref(self, ref):
-        ref_name = ref.split('/')[-1]
+        ref_fragment = ref.split('#')[-1]
+        ref_name = 'ref' + re.sub(r'[^a-zA-Z0-9-]+', '-', ref_fragment)
         if ref_name not in self._rules and ref not in self._refs_being_resolved:
             self._refs_being_resolved.add(ref)
             resolved = self._refs[ref]

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

@@ -138,7 +138,7 @@ if model_path is None:
         "Model path must be specified either via --model-path argument or MODEL_PATH environment variable"
     )
 
-config = AutoConfig.from_pretrained(model_path)
+config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
 
 print("Model type:       ", config.model_type)
 print("Vocab size:       ", config.vocab_size)
@@ -148,8 +148,8 @@ print("BOS token id:     ", config.bos_token_id)
 print("EOS token id:     ", config.eos_token_id)
 
 print("Loading model and tokenizer using AutoTokenizer:", model_path)
-tokenizer = AutoTokenizer.from_pretrained(model_path)
-config = AutoConfig.from_pretrained(model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
 
 if unreleased_model_name:
     model_name_lower = unreleased_model_name.lower()
@@ -171,7 +171,7 @@ if unreleased_model_name:
         exit(1)
 else:
     model = AutoModelForCausalLM.from_pretrained(
-        model_path, device_map="auto", offload_folder="offload"
+        model_path, device_map="auto", offload_folder="offload", trust_remote_code=True
     )
 
 for name, module in model.named_modules():

ggml/CMakeLists.txt

@@ -251,6 +251,8 @@ option(GGML_OPENCL_USE_ADRENO_KERNELS       "ggml: use optimized kernels for Adr
 set   (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING
                                             "gmml: OpenCL API version to target")
 
+option(GGML_HEXAGON                         "ggml: enable Hexagon backend"                    OFF)
+
 # toolchain for vulkan-shaders-gen
 set   (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen")
 

ggml/include/ggml-hexagon.h

@@ -0,0 +1,19 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_BACKEND_API ggml_backend_t ggml_backend_hexagon_init(void);
+
+GGML_BACKEND_API bool ggml_backend_is_hexagon(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_hexagon_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml-rpc.h

@@ -21,8 +21,7 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const c
 GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total);
 
 GGML_BACKEND_API void ggml_backend_rpc_start_server(const char * endpoint, const char * cache_dir,
-                                                    size_t n_threads, size_t n_devices,
-                                                    ggml_backend_dev_t * devices, size_t * free_mem, size_t * total_mem);
+                                                    size_t n_threads, size_t n_devices, ggml_backend_dev_t * devices);
 
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_add_server(const char * endpoint);

ggml/include/ggml.h

@@ -577,6 +577,10 @@ extern "C" {
         GGML_UNARY_OP_EXP,
         GGML_UNARY_OP_GELU_ERF,
         GGML_UNARY_OP_XIELU,
+        GGML_UNARY_OP_FLOOR,
+        GGML_UNARY_OP_CEIL,
+        GGML_UNARY_OP_ROUND,
+        GGML_UNARY_OP_TRUNC,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -1151,6 +1155,46 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_floor(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_floor_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_ceil(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_ceil_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_round(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_round_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+     /**
+     * Truncates the fractional part of each element in the tensor (towards zero).
+     * For example: trunc(3.7) = 3.0, trunc(-2.9) = -2.0
+     * Similar to std::trunc in C/C++.
+     */
+
+    GGML_API struct ggml_tensor * ggml_trunc(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_trunc_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+
+
     // xIELU activation function
     // x = x * (c_a(alpha_n) + c_b(alpha_p, beta) * sigmoid(beta * x)) + eps * (x > 0)
     // where c_a = softplus and c_b(a, b) = softplus(a) + b are constraining functions

ggml/src/CMakeLists.txt

@@ -307,6 +307,10 @@ function(ggml_add_cpu_backend_variant tag_name)
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
     endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -371,6 +375,14 @@ if (GGML_CPU_ALL_VARIANTS)
         else()
             message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
         endif()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            ggml_add_cpu_backend_variant(s390x_z15  Z15 VXE)
+            # ggml_add_cpu_backend_variant(s390x_z16  Z16 VXE)
+            # ggml_add_cpu_backend_variant(s390x_z17  Z17 VXE)
+        else()
+            message(FATAL_ERROR "Unsupported s390x 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()
@@ -390,6 +402,7 @@ ggml_add_backend(Vulkan)
 ggml_add_backend(WebGPU)
 ggml_add_backend(zDNN)
 ggml_add_backend(OpenCL)
+ggml_add_backend(Hexagon)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

ggml/src/ggml-alloc.c

@@ -226,16 +226,23 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
     }
 
     if (best_fit_block == -1) {
-        // no suitable block found, try the last block (this will grow a chunks size)
+        // no suitable block found, try the last block (this may grow a chunks size)
+        int64_t best_reuse = INT64_MIN;
         for (int c = 0; c < alloc->n_chunks; ++c) {
             struct tallocr_chunk * chunk = alloc->chunks[c];
             if (chunk->n_free_blocks > 0) {
                 struct free_block * block = &chunk->free_blocks[chunk->n_free_blocks - 1];
                 max_avail = MAX(max_avail, block->size);
-                if (block->size >= size) {
+                int64_t reuse_factor = chunk->max_size - block->offset - size;
+                // reuse_factor < 0 : amount of extra memory that needs to be allocated
+                // reuse_factor = 0 : allocated free space exactly matches tensor size
+                // reuse_factor > 0 : superfluous memory that will remain unused
+                bool better_reuse = best_reuse < 0 && reuse_factor > best_reuse;
+                bool better_fit = reuse_factor >= 0 && reuse_factor < best_reuse;
+                if (block->size >= size && (better_reuse || better_fit)) {
                     best_fit_chunk = c;
                     best_fit_block = chunk->n_free_blocks - 1;
-                    break;
+                    best_reuse = reuse_factor;
                 }
             }
         }
@@ -268,7 +275,7 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, addr, tensor);
     size_t cur_max = addr.offset + size;
-    if (cur_max > alloc->max_size[addr.chunk]) {
+    if (cur_max > chunk->max_size) {
         // sort allocated_tensors by chunk/offset
         for (int i = 0; i < 1024; i++) {
             for (int j = i + 1; j < 1024; j++) {
@@ -598,6 +605,26 @@ static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor
     return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
+// free the extra space at the end if the new tensor is smaller
+static void ggml_gallocr_free_extra_space(ggml_gallocr_t galloc, struct ggml_tensor * node, struct ggml_tensor * parent) {
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+    struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
+
+    size_t parent_size = ggml_backend_buft_get_alloc_size(galloc->bufts[p_hn->buffer_id], parent);
+    size_t node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
+
+    GGML_ASSERT(parent_size >= node_size);
+
+    if (parent_size > node_size) {
+        struct ggml_dyn_tallocr * p_alloc = galloc->buf_tallocs[p_hn->buffer_id];
+        struct buffer_address p_addr = p_hn->addr;
+        p_addr.offset += node_size;
+        size_t extra_size = parent_size - node_size;
+        AT_PRINTF("freeing extra %zu bytes from parent %s for %s\n", extra_size, parent->name, node->name);
+        ggml_dyn_tallocr_free_tensor(p_alloc, p_addr, extra_size, parent);
+    }
+}
+
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
     GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
@@ -643,6 +670,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
                             hn->addr = p_hn->addr;
                             p_hn->allocated = false; // avoid freeing the parent
                             view_src_hn->allocated = false;
+                            ggml_gallocr_free_extra_space(galloc, node, view_src);
                             return;
                         }
                     } else {
@@ -650,6 +678,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
                         hn->buffer_id = p_hn->buffer_id;
                         hn->addr = p_hn->addr;
                         p_hn->allocated = false; // avoid freeing the parent
+                        ggml_gallocr_free_extra_space(galloc, node, parent);
                         return;
                     }
                 }

ggml/src/ggml-backend-reg.cpp

@@ -57,6 +57,10 @@
 #include "ggml-opencl.h"
 #endif
 
+#ifdef GGML_USE_HEXAGON
+#include "ggml-hexagon.h"
+#endif
+
 #ifdef GGML_USE_BLAS
 #include "ggml-blas.h"
 #endif
@@ -199,6 +203,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
+#ifdef GGML_USE_HEXAGON
+        register_backend(ggml_backend_hexagon_reg());
+#endif
 #ifdef GGML_USE_CANN
         register_backend(ggml_backend_cann_reg());
 #endif
@@ -598,6 +605,7 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("sycl", silent, dir_path);
     ggml_backend_load_best("vulkan", silent, dir_path);
     ggml_backend_load_best("opencl", silent, dir_path);
+    ggml_backend_load_best("hexagon", silent, dir_path);
     ggml_backend_load_best("musa", silent, dir_path);
     ggml_backend_load_best("cpu", silent, dir_path);
     // check the environment variable GGML_BACKEND_PATH to load an out-of-tree backend

ggml/src/ggml-cann/acl_tensor.cpp

@@ -51,28 +51,31 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
     return ACL_DT_UNDEFINED;
 }
 
-aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
-                                   size_t* nb, int64_t dims, aclFormat format,
-                                   size_t offset) {
+aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
+                                    int64_t *           ne,
+                                    size_t *            nb,
+                                    int64_t             dims,
+                                    aclFormat           format,
+                                    size_t              offset) {
     // If tensor is bcasted, Up to GGML_MAX_DIMS additional dimensions will be
     // added.
     int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
 
     if (ne == nullptr) {
         for (int i = 0; i < GGML_MAX_DIMS; i++) {
-            acl_ne[i] = tensor->ne[i];
+            acl_ne[i]     = tensor->ne[i];
             // The step size of acl is in elements.
             acl_stride[i] = tensor->nb[i] / ggml_element_size(tensor);
         }
     } else {
         // With bcast
         for (int i = 0; i < dims; i++) {
-            acl_ne[i] = ne[i];
+            acl_ne[i]     = ne[i];
             acl_stride[i] = nb[i] / ggml_element_size(tensor);
         }
     }
 
-    int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
+    int64_t final_dims      = (dims == 0 ? GGML_MAX_DIMS : dims);
     int64_t acl_storage_len = 1;
     for (int i = 0; i < final_dims; i++) {
         acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
@@ -84,15 +87,13 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
     std::reverse(acl_ne, acl_ne + final_dims);
     std::reverse(acl_stride, acl_stride + final_dims);
 
-    aclTensor* acl_tensor = aclCreateTensor(
-        acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
-        elem_offset, format, &acl_storage_len, 1,
-        tensor->data);
+    aclTensor * acl_tensor = aclCreateTensor(acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
+                                             elem_offset, format, &acl_storage_len, 1, tensor->data);
 
     return acl_tensor;
 }
 
-bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
+bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1) {
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         if (t1->ne[i] != t0->ne[i] && t1->ne[i] != 1) {
             return true;
@@ -101,15 +102,16 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
     return false;
 }
 
-int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
-                                  const ggml_tensor* src1,
-                                  int64_t* bcast_src0_ne,
-                                  int64_t* bcast_src1_ne, size_t* bcast_src0_nb,
-                                  size_t* bcast_src1_nb) {
+int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
+                                  const ggml_tensor * src1,
+                                  int64_t *           bcast_src0_ne,
+                                  int64_t *           bcast_src1_ne,
+                                  size_t *            bcast_src0_nb,
+                                  size_t *            bcast_src1_nb) {
     GGML_ASSERT(ggml_can_repeat(src1, src0));
     int bcast_dim_cnt = 0;
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        int64_t nr = src0->ne[i] / src1->ne[i];
+        int64_t nr                   = src0->ne[i] / src1->ne[i];
         bcast_src0_ne[bcast_dim_cnt] = src0->ne[i] / nr;
         bcast_src1_ne[bcast_dim_cnt] = src1->ne[i];
         bcast_src0_nb[bcast_dim_cnt] = src0->nb[i];
@@ -119,21 +121,26 @@ int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
             // Need to add an extra dim.
             bcast_src0_ne[bcast_dim_cnt] = nr;
             bcast_src1_ne[bcast_dim_cnt] = 1;
-            bcast_src0_nb[bcast_dim_cnt] = bcast_src0_nb[bcast_dim_cnt - 1] *
-                                           bcast_src0_ne[bcast_dim_cnt - 1];
-            bcast_src1_nb[bcast_dim_cnt] = bcast_src1_nb[bcast_dim_cnt - 1] *
-                                           bcast_src1_ne[bcast_dim_cnt - 1];
+            bcast_src0_nb[bcast_dim_cnt] = bcast_src0_nb[bcast_dim_cnt - 1] * bcast_src0_ne[bcast_dim_cnt - 1];
+            bcast_src1_nb[bcast_dim_cnt] = bcast_src1_nb[bcast_dim_cnt - 1] * bcast_src1_ne[bcast_dim_cnt - 1];
             bcast_dim_cnt++;
         }
     }
     return bcast_dim_cnt;
 }
 
-int64_t ggml_cann_get_mulmat_bcast_shape(
-    const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
-    const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
-    int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
-    size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb) {
+int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
+                                         const int64_t * weight_ne,
+                                         const int64_t * dst_ne,
+                                         const size_t *  input_nb,
+                                         const size_t *  weight_nb,
+                                         const size_t *  dst_nb,
+                                         int64_t *       bcast_input_ne,
+                                         int64_t *       bcast_weight_ne,
+                                         int64_t *       bcast_dst_ne,
+                                         size_t *        bcast_input_nb,
+                                         size_t *        bcast_weight_nb,
+                                         size_t *        bcast_dst_nb) {
     // input and dst shoule in same shape, except first two dims.
     GGML_ASSERT(input_ne[2] == dst_ne[2]);
     GGML_ASSERT(input_ne[3] == dst_ne[3]);
@@ -148,34 +155,30 @@ int64_t ggml_cann_get_mulmat_bcast_shape(
         // Do not use bcast in the first two dimensions because we only support
         // the bcast batch dimension. Just copy them.
         if (i < 2 || nr == 1) {
-            bcast_input_ne[bcast_dim_cnt] = input_ne[i];
+            bcast_input_ne[bcast_dim_cnt]  = input_ne[i];
             bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
-            bcast_dst_ne[bcast_dim_cnt] = dst_ne[i];
+            bcast_dst_ne[bcast_dim_cnt]    = dst_ne[i];
 
-            bcast_input_nb[bcast_dim_cnt] = input_nb[i];
+            bcast_input_nb[bcast_dim_cnt]  = input_nb[i];
             bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
-            bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
+            bcast_dst_nb[bcast_dim_cnt]    = dst_nb[i];
             bcast_dim_cnt++;
         } else {
             // Need to add an extra dim.
-            bcast_input_ne[bcast_dim_cnt] = nr;
-            bcast_dst_ne[bcast_dim_cnt] = nr;
+            bcast_input_ne[bcast_dim_cnt]  = nr;
+            bcast_dst_ne[bcast_dim_cnt]    = nr;
             bcast_weight_ne[bcast_dim_cnt] = 1;
-            bcast_input_nb[bcast_dim_cnt] = input_nb[i];
-            bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
+            bcast_input_nb[bcast_dim_cnt]  = input_nb[i];
+            bcast_dst_nb[bcast_dim_cnt]    = dst_nb[i];
             bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
             bcast_dim_cnt++;
 
-            bcast_input_ne[bcast_dim_cnt] = input_ne[i] / nr;
-            bcast_dst_ne[bcast_dim_cnt] = dst_ne[i] / nr;
+            bcast_input_ne[bcast_dim_cnt]  = input_ne[i] / nr;
+            bcast_dst_ne[bcast_dim_cnt]    = dst_ne[i] / nr;
             bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
-            bcast_input_nb[bcast_dim_cnt] = bcast_input_nb[bcast_dim_cnt - 1] *
-                                            bcast_input_ne[bcast_dim_cnt - 1];
-            bcast_dst_nb[bcast_dim_cnt] = bcast_dst_nb[bcast_dim_cnt - 1] *
-                                          bcast_dst_ne[bcast_dim_cnt - 1];
-            bcast_weight_nb[bcast_dim_cnt] =
-                bcast_weight_nb[bcast_dim_cnt - 1] *
-                bcast_weight_ne[bcast_dim_cnt - 1];
+            bcast_input_nb[bcast_dim_cnt]  = bcast_input_nb[bcast_dim_cnt - 1] * bcast_input_ne[bcast_dim_cnt - 1];
+            bcast_dst_nb[bcast_dim_cnt]    = bcast_dst_nb[bcast_dim_cnt - 1] * bcast_dst_ne[bcast_dim_cnt - 1];
+            bcast_weight_nb[bcast_dim_cnt] = bcast_weight_nb[bcast_dim_cnt - 1] * bcast_weight_ne[bcast_dim_cnt - 1];
             bcast_dim_cnt++;
         }
     }

ggml/src/ggml-cann/acl_tensor.h

@@ -62,10 +62,12 @@ aclDataType ggml_cann_type_mapping(ggml_type type);
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
-aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = nullptr,
-                             size_t* nb = nullptr, int64_t dims = 0,
-                             aclFormat format = ACL_FORMAT_ND,
-                             size_t offset = 0);
+aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
+                                    int64_t *           ne     = nullptr,
+                                    size_t *            nb     = nullptr,
+                                    int64_t             dims   = 0,
+                                    aclFormat           format = ACL_FORMAT_ND,
+                                    size_t              offset = 0);
 
 /**
  * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
@@ -87,12 +89,15 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
-template<typename TYPE>
-aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                                   TYPE type_size, int64_t* ne, TYPE* nb,
-                                   int64_t dims,
-                                   aclFormat format = ACL_FORMAT_ND,
-                                   size_t offset = 0) {
+template <typename TYPE>
+aclTensor * ggml_cann_create_tensor(void *      data_ptr,
+                                    aclDataType dtype,
+                                    TYPE        type_size,
+                                    int64_t *   ne,
+                                    TYPE *      nb,
+                                    int64_t     dims,
+                                    aclFormat   format = ACL_FORMAT_ND,
+                                    size_t      offset = 0) {
     int64_t tmp_ne[GGML_MAX_DIMS * 2];
     int64_t tmp_stride[GGML_MAX_DIMS * 2];
 
@@ -109,9 +114,8 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
     std::reverse(tmp_ne, tmp_ne + dims);
     std::reverse(tmp_stride, tmp_stride + dims);
 
-    aclTensor* acl_tensor =
-        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
-                        format, &acl_storage_len, 1, data_ptr);
+    aclTensor * acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size, format, &acl_storage_len, 1, data_ptr);
 
     return acl_tensor;
 }
@@ -132,7 +136,7 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
  *          to 1. If such a dimension is found, broadcasting is required to align t1
  *          with t0 for element-wise operations.
  */
-bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);
+bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1);
 
 /**
  * @brief   Computes broadcast shapes and strides for two ggml_tensors.
@@ -187,19 +191,21 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);
  *  dim1 in a inserted dim, should add nb for dim1,
  *  and all other nb moves to next in order.
  */
-int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
-                        int64_t* bcast_ne_src0, int64_t* bcast_ne_src1,
-                        size_t* bcast_nb_src0, size_t* bcast_nb_src1);
+int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
+                                  const ggml_tensor * src1,
+                                  int64_t *           bcast_ne_src0,
+                                  int64_t *           bcast_ne_src1,
+                                  size_t *            bcast_nb_src0,
+                                  size_t *            bcast_nb_src1);
 
 // Bcast macro to avoid duplicate code.
-#define BCAST_SHAPE(src0, src1)                                              \
-    int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2];                            \
-    int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2];                            \
-    size_t bcast_##src0##_nb[GGML_MAX_DIMS * 2];                             \
-    size_t bcast_##src1##_nb[GGML_MAX_DIMS * 2];                             \
-    int64_t bcast_dims = ggml_cann_get_bcast_shape(                          \
-        src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, bcast_##src0##_nb, \
-        bcast_##src1##_nb);
+#define BCAST_SHAPE(src0, src1)                                                                      \
+    int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2];                                                    \
+    int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2];                                                    \
+    size_t  bcast_##src0##_nb[GGML_MAX_DIMS * 2];                                                    \
+    size_t  bcast_##src1##_nb[GGML_MAX_DIMS * 2];                                                    \
+    int64_t bcast_dims = ggml_cann_get_bcast_shape(src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, \
+                                                   bcast_##src0##_nb, bcast_##src1##_nb);
 
 #define BCAST_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
 
@@ -233,26 +239,31 @@ int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* sr
  *       before cast dim.
  * @sa ggml_cann_get_bcast_shape
  */
-int64_t ggml_cann_get_mulmat_bcast_shape(
-    const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
-    const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
-    int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
-    size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb);
+int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
+                                         const int64_t * weight_ne,
+                                         const int64_t * dst_ne,
+                                         const size_t *  input_nb,
+                                         const size_t *  weight_nb,
+                                         const size_t *  dst_nb,
+                                         int64_t *       bcast_input_ne,
+                                         int64_t *       bcast_weight_ne,
+                                         int64_t *       bcast_dst_ne,
+                                         size_t *        bcast_input_nb,
+                                         size_t *        bcast_weight_nb,
+                                         size_t *        bcast_dst_nb);
 
 // Bcast macro to avoid duplicate code.
-#define BCAST_MUL_MAT_SHAPE(input, weight, dst)                         \
-    int64_t bcast_##input##_ne[GGML_MAX_DIMS * 2];                      \
-    int64_t bcast_##weight##_ne[GGML_MAX_DIMS * 2];                     \
-    int64_t bcast_##dst##_ne[GGML_MAX_DIMS * 2];                        \
-    size_t bcast_##input##_nb[GGML_MAX_DIMS * 2];                       \
-    size_t bcast_##weight##_nb[GGML_MAX_DIMS * 2];                      \
-    size_t bcast_##dst##_nb[GGML_MAX_DIMS * 2];                         \
-    int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape(              \
-        input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, \
-        bcast_##input##_ne, bcast_##weight##_ne, bcast_##dst##_ne,      \
-        bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
+#define BCAST_MUL_MAT_SHAPE(input, weight, dst)                                                                  \
+    int64_t bcast_##input##_ne[GGML_MAX_DIMS * 2];                                                               \
+    int64_t bcast_##weight##_ne[GGML_MAX_DIMS * 2];                                                              \
+    int64_t bcast_##dst##_ne[GGML_MAX_DIMS * 2];                                                                 \
+    size_t  bcast_##input##_nb[GGML_MAX_DIMS * 2];                                                               \
+    size_t  bcast_##weight##_nb[GGML_MAX_DIMS * 2];                                                              \
+    size_t  bcast_##dst##_nb[GGML_MAX_DIMS * 2];                                                                 \
+    int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape(                                                       \
+        input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, bcast_##input##_ne, bcast_##weight##_ne, \
+        bcast_##dst##_ne, bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
 
-#define BCAST_MUL_MAT_PARAM(tensor) \
-    bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
+#define BCAST_MUL_MAT_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
 
 #endif  // CANN_ACL_TENSOR_H

ggml/src/ggml-cann/aclnn_ops.h

@@ -62,7 +62,7 @@
  * @param   dst The ggml tensor representing the destination, which op is
  *              GGML_OP_REPEAT and specifies the desired dimensions.
  */
-void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_repeat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies the Leaky ReLU activation function to a tensor using the CANN
@@ -82,7 +82,7 @@ void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result of the Leaky ReLU
  *            activation is stored, which op is `GGML_OP_LEAKY_RELU`
  */
-void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_leaky_relu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief    Concatenates multiple tensors along a specified dimension using the
@@ -97,7 +97,7 @@ void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @attention tensorList length should be 2 and the dimension using for concat
  *            default to 1.
  */
-void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_concat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Generates a sequence of evenly spaced values within a specified
@@ -113,7 +113,7 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *            `start`, 'stop' and 'step' are in dst->op_params and dst->op is
  *            `GGML_OP_ARANGE`.
  */
-void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_arange(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies a clamp operation to the elements of a ggml tensor using the
@@ -131,7 +131,7 @@ void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the clamped values will be stored.
  *            dst->op is `GGML_OP_CLAMP`, `min` and `max` value is in dst->params.
  */
-void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_clamp(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Scales the elements of a ggml tensor by a constant factor using the
@@ -148,7 +148,7 @@ void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the scaled values will be stored.
  *            dst->op is `GGML_OP_SCALE` and `scale` value is in dst->params.
  */
-void ggml_cann_scale(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_scale(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Sorts the elements of a ggml tensor and returns the indices that
@@ -163,7 +163,7 @@ void ggml_cann_scale(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the sorted indices will be stored.
  *            dst->op is `GGML_OP_ARGSORT`.
  */
-void ggml_cann_argsort(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_argsort(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the Layer Normalization for a ggml tensor using the CANN
@@ -185,7 +185,7 @@ void ggml_cann_argsort(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the normalized values will be stored.
  * @attention `Var` defaults to dst->ne[0].
  */
-void ggml_cann_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief  Computes the Group Normalization for a ggml tensor using the CANN
@@ -209,7 +209,7 @@ void ggml_cann_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  * @attention eps defaults to 1e-6f.
  */
-void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_group_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the accumulation of tensors using the CANN backend.
@@ -228,7 +228,7 @@ void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the accumulated values will be stored.
  *            `inplace` is in dst->params, and dst->op is `GGML_OP_ACC`.
  */
-void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_acc(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the sum of elements along the last dimension of a ggml tensor
@@ -244,7 +244,7 @@ void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  * @attention `reduce_dims` defaults to 3, which means the last dimension.
  */
-void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_sum_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the sum of elements in a ggml tensor.
@@ -258,7 +258,7 @@ void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  */
 
-void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Upsamples a ggml tensor using nearest neighbor interpolation using
@@ -274,8 +274,7 @@ void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the upsampled values will be stored.
  *            dst->op is `GGML_OP_UPSCALE`.
  */
-void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx,
-                                  ggml_tensor* dst);
+void ggml_cann_upsample_nearest2d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Pads a ggml tensor to match the dimensions of the destination tensor
@@ -290,7 +289,7 @@ void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx,
  * @param dst The destination tensor, which specifies the target dimensions for
  *            padding. dst->op is `GGML_OP_PAD`.
  */
-void ggml_cann_pad(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_pad(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Executes a 2D pooling operation on a ggml tensor using the CANN
@@ -307,7 +306,7 @@ void ggml_cann_pad(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor on which the pooling operation is to be
  *            performed. dst->op is `GGML_OP_POOL_2D`.
  */
-void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_pool2d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Duplicates a ggml tensor using the CANN backend.
@@ -326,7 +325,7 @@ void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *            different shape and dst is no-contiguous.
  * @note:     This func need to simplify.
  */
-void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_dup(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the Root Mean Square (RMS) normalization of a ggml tensor
@@ -348,7 +347,7 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the normalized values will be stored.
  *            dst->op is `GGML_OP_RMS_NORM`.
  */
-void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_rms_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies a diagonal mask to the tensor with a specified value.
@@ -363,7 +362,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *            `GGML_OP_DIAG_MASK`
  * @param value The value to use for masking.
  */
-void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst, float value);
+void ggml_cann_diag_mask(ggml_backend_cann_context & ctx, ggml_tensor * dst, float value);
 
 /**
  * @brief   Performs an image-to-column transformation on the input tensor.
@@ -378,7 +377,7 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst, float
  * @param dst The destination tensor that stores the result of the operation.
  *            dst->op is `GGML_OP_IM2COL`.
  */
-void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_im2col(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes time step embeddings using sine and cosine functions.
@@ -392,10 +391,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result of the embedding operation
  *            will be stored. dst->op is `GGML_OP_TIMESTEP_EMBEDDING`.
  */
-void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_timestep_embedding(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 // @see ggml_cann_dup.
-void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the softmax activation with optional masking.
@@ -417,7 +416,7 @@ void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result will be stored. dst->op is
  *            `GGML_OP_SOFTMAX`.
  */
-void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Extracts specific rows from a tensor based on indices.
@@ -429,7 +428,7 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param ctx The backend CANN context for executing operations.
  * @param dst The destination tensor where the extracted rows will be stored.
  */
-void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Writes specific rows into a tensor at positions specified by indices.
@@ -441,7 +440,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param ctx The backend CANN context for executing operations.
  * @param dst The destination tensor where the specified rows will be updated.
  */
-void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_set_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Executes matrix multiplication for the given tensor.
@@ -454,7 +453,7 @@ void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor for storing the result of the matrix
  *            multiplication. dst->op is `GGML_OP_MUL_MAT`.
  */
-void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_mul_mat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief Applies Rotary Positional Embedding (RoPE) to the input tensor.
@@ -477,7 +476,7 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @note The function currently does not support cases where the freq_scale is
  *       not equal 1.
  */
-void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the index of the maximum value along the specified dimension
@@ -492,7 +491,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the indices of the maximum values will
  *            be stored. dst->op is `GGML_OP_ARGMAX`.
  */
-void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_argmax(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief Adds two tensors element-wise and stores the result in a destination
@@ -509,8 +508,10 @@ void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param acl_src1 The second source tensor.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
-    aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
+void aclnn_add(ggml_backend_cann_context & ctx,
+               aclTensor *                 acl_src0,
+               aclTensor *                 acl_src1,
+               aclTensor *                 acl_dst = nullptr);
 
 /**
  * @brief Sub two tensors element-wise and stores the result in a destination
@@ -527,8 +528,10 @@ void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
  * @param acl_src1 The second source tensor.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
-    aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
+void aclnn_sub(ggml_backend_cann_context & ctx,
+               aclTensor *                 acl_src0,
+               aclTensor *                 acl_src1,
+               aclTensor *                 acl_dst = nullptr);
 
 /**
  * @brief Performs element-wise multiplication of two tensors and stores the
@@ -546,8 +549,10 @@ void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
  * @param acl_other The second tensor for element-wise multiplication.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-    aclTensor* acl_other, aclTensor* acl_dst = nullptr);
+void aclnn_mul(ggml_backend_cann_context & ctx,
+               aclTensor *                 acl_src,
+               aclTensor *                 acl_other,
+               aclTensor *                 acl_dst = nullptr);
 
 /**
  * @brief Matrix division, optionally in-place.
@@ -567,8 +572,10 @@ void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * @param inplace Flag indicating whether to perform the operation in-place on
  * `acl_src`.
  */
-void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-    aclTensor* acl_other, aclTensor* acl_dst = nullptr);
+void aclnn_div(ggml_backend_cann_context & ctx,
+               aclTensor *                 acl_src,
+               aclTensor *                 acl_other,
+               aclTensor *                 acl_dst = nullptr);
 
 /**
  * @brief Applies element-wise cosine function to the elements of a tensor.
@@ -584,8 +591,7 @@ void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * @param acl_dst The destination tensor where the cosine results will be
  * stored.
  */
-void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-    aclTensor* acl_dst);
+void aclnn_cos(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst);
 
 /**
  * @brief Applies element-wise sine function to the elements of a tensor.
@@ -602,8 +608,7 @@ void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * @param acl_src The source tensor on which the sine function will be applied.
  * @param acl_dst The destination tensor where the sine results will be stored.
  */
-void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-    aclTensor* acl_dst);
+void aclnn_sin(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst);
 
 /**
  * @brief Prepares broadcast-compatible ACL tensors for two input tensors and one
@@ -621,8 +626,12 @@ void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * @param acl_src1 Output pointer to the created ACL tensor corresponding to src1.
  * @param acl_dst  Output pointer to the created ACL tensor corresponding to dst.
  */
-void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst,
-    aclTensor ** acl_src0, aclTensor ** acl_src1, aclTensor ** acl_dst);
+void bcast_shape(ggml_tensor * src0,
+                 ggml_tensor * src1,
+                 ggml_tensor * dst,
+                 aclTensor **  acl_src0,
+                 aclTensor **  acl_src1,
+                 aclTensor **  acl_dst);
 
 /**
  * @brief   Computes the 1D transposed convolution (deconvolution) of a ggml
@@ -637,7 +646,7 @@ void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst,
  * @param dst The destination tensor where the transposed convolution result
  * will be stored. dst->op is `GGML_OP_CONV_TRANSPOSE_1D`.
  */
-void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_conv_transpose_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies the ELU (Exponential Linear Unit) activation to a ggml tensor
@@ -662,7 +671,7 @@ void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* ds
  * @param dst The destination tensor where the ELU-activated result will be stored.
  *            dst->op is expected to be `GGML_OP_ELU`.
  */
-void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_elu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the mean of a ggml tensor element-wise using the CANN backend.
@@ -677,7 +686,7 @@ void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the mean result will be stored.
  *            dst->op is expected to be `GGML_OP_MEAN`.
  */
-void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_mean(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies 1D reflect padding to a ggml tensor using the CANN backend.
@@ -692,7 +701,7 @@ void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the padded result will be stored.
  *            dst->op is expected to be `GGML_OP_PAD_REFLECT_1D`.
  */
-void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_pad_reflect_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Counts the number of equal elements in two ggml tensors using the CANN backend.
@@ -708,7 +717,7 @@ void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_COUNT_EQUAL`.
  */
-void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies the Step activation function to a ggml tensor using the CANN backend.
@@ -723,7 +732,7 @@ void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_STEP`.
  */
-void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Performs the Flash Attention extended operator using the CANN backend.
@@ -738,59 +747,46 @@ void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_FLASH_ATTN_EXT`.
  */
-void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_flash_attn_ext(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /*
  * @brief A generic wrapper for ACL resources with custom deleter support.
  */
-using any_acl_resource = std::unique_ptr<void, std::function<void(void*)>>;
+using any_acl_resource = std::unique_ptr<void, std::function<void(void *)>>;
 
 /**
  * @brief Trait structure used to define how to destroy a given ACL resource type.
  *
  * @tparam T ACL resource type.
  */
-template<typename T>
-struct acl_resource_traits;
+template <typename T> struct acl_resource_traits;
 
 /**
  * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
  */
-template<>
-struct acl_resource_traits<aclTensor> {
-    static void destroy(void* p) {
-        ACL_CHECK(aclDestroyTensor(static_cast<aclTensor*>(p)));
-    }
+template <> struct acl_resource_traits<aclTensor> {
+    static void destroy(void * p) { ACL_CHECK(aclDestroyTensor(static_cast<aclTensor *>(p))); }
 };
 
 /**
  * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
  */
-template<>
-struct acl_resource_traits<aclIntArray> {
-    static void destroy(void* p) {
-        ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray*>(p)));
-    }
+template <> struct acl_resource_traits<aclIntArray> {
+    static void destroy(void * p) { ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray *>(p))); }
 };
 
 /**
  * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
  */
-template<>
-struct acl_resource_traits<aclScalar> {
-    static void destroy(void* p) {
-        ACL_CHECK(aclDestroyScalar(static_cast<aclScalar*>(p)));
-    }
+template <> struct acl_resource_traits<aclScalar> {
+    static void destroy(void * p) { ACL_CHECK(aclDestroyScalar(static_cast<aclScalar *>(p))); }
 };
 
 /**
  * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
  */
-template<>
-struct acl_resource_traits<aclTensorList> {
-    static void destroy(void* p) {
-        ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList*>(p)));
-    }
+template <> struct acl_resource_traits<aclTensorList> {
+    static void destroy(void * p) { ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList *>(p))); }
 };
 
 /**
@@ -800,14 +796,8 @@ struct acl_resource_traits<aclTensorList> {
  * @param ptr Raw pointer to ACL resource.
  * @return any_acl_resource Smart pointer that handles destruction.
  */
-template<typename T>
-any_acl_resource make_acl_resource(T* ptr) {
-    return any_acl_resource(
-        static_cast<void*>(ptr),
-        [](void* p) {
-            acl_resource_traits<T>::destroy(p);
-        }
-    );
+template <typename T> any_acl_resource make_acl_resource(T * ptr) {
+    return any_acl_resource(static_cast<void *>(ptr), [](void * p) { acl_resource_traits<T>::destroy(p); });
 }
 
 /**
@@ -817,8 +807,7 @@ any_acl_resource make_acl_resource(T* ptr) {
  * @param vec Target vector to hold ACL resources.
  * @param args Raw pointers to ACL resources.
  */
-template<typename... Args>
-void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
+template <typename... Args> void register_acl_resources(std::vector<any_acl_resource> & vec, Args *... args) {
     (vec.emplace_back(make_acl_resource(args)), ...);
 }
 
@@ -826,39 +815,36 @@ void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
  * @brief Task class that wraps the execution of an aclnn function call.
  */
 class aclnn_task : public cann_task {
-    public:
-        aclnn_task(aclnn_func_t aclnn_func, void * workspace_addr,
-                   uint64_t workspace_size, aclOpExecutor * executor,
-                   aclrtStream stream) :
-            aclnn_func_(aclnn_func),
-            workspace_addr_(workspace_addr),
-            workspace_size_(workspace_size),
-            executor_(executor),
-            stream_(stream) {}
-        virtual void run_task() override {
-            ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_));
-        }
-    private:
-        aclnn_func_t aclnn_func_;
-        void *          workspace_addr_;
-        uint64_t        workspace_size_;
-        aclOpExecutor * executor_;
-        aclrtStream     stream_;
+  public:
+    aclnn_task(aclnn_func_t    aclnn_func,
+               void *          workspace_addr,
+               uint64_t        workspace_size,
+               aclOpExecutor * executor,
+               aclrtStream     stream) :
+        aclnn_func_(aclnn_func),
+        workspace_addr_(workspace_addr),
+        workspace_size_(workspace_size),
+        executor_(executor),
+        stream_(stream) {}
+
+    virtual void run_task() override { ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_)); }
+  private:
+    aclnn_func_t    aclnn_func_;
+    void *          workspace_addr_;
+    uint64_t        workspace_size_;
+    aclOpExecutor * executor_;
+    aclrtStream     stream_;
 };
 
 /**
  * @brief Task class that releases ACL resources after usage.
  */
 class release_resource_task : public cann_task {
-public:
-    release_resource_task(std::vector<any_acl_resource>&& resources){
-        resource_ = std::move(resources);
-    }
+  public:
+    release_resource_task(std::vector<any_acl_resource> && resources) { resource_ = std::move(resources); }
 
-    virtual void run_task() override {
-        resource_.clear();
-    }
-private:
+    virtual void run_task() override { resource_.clear(); }
+  private:
     std::vector<any_acl_resource> resource_;
 };
 
@@ -866,38 +852,40 @@ private:
  * @brief Task class for performing asynchronous memory copy operations.
  */
 class async_memcpy_task : public cann_task {
-public:
-    async_memcpy_task(void* dst, const void* src, size_t size,
-                      aclrtMemcpyKind kind, aclrtStream stream)
-        : dst_(dst), src_(src), size_(size), kind_(kind), stream_(stream) {}
+  public:
+    async_memcpy_task(void * dst, const void * src, size_t size, aclrtMemcpyKind kind, aclrtStream stream) :
+        dst_(dst),
+        src_(src),
+        size_(size),
+        kind_(kind),
+        stream_(stream) {}
 
-    virtual void run_task() override {
-        ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_));
-    }
-private:
-    void* dst_;
-    const void* src_;
-    size_t size_;
+    virtual void run_task() override { ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_)); }
+  private:
+    void *          dst_;
+    const void *    src_;
+    size_t          size_;
     aclrtMemcpyKind kind_;
-    aclrtStream stream_;
+    aclrtStream     stream_;
 };
 
 /**
  * @brief Task class for performing asynchronous memory set operations.
  */
 class async_memset_task : public cann_task {
-    public:
-    async_memset_task(void* buffer, size_t size, int32_t value, aclrtStream stream)
-            : buffer_(buffer), size_(size), value_(value), stream_(stream) {}
+  public:
+    async_memset_task(void * buffer, size_t size, int32_t value, aclrtStream stream) :
+        buffer_(buffer),
+        size_(size),
+        value_(value),
+        stream_(stream) {}
 
-        virtual void run_task() override {
-            ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_));
-        }
-    private:
-        void* buffer_;
-        size_t size_;
-        int32_t value_;
-        aclrtStream stream_;
+    virtual void run_task() override { ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_)); }
+  private:
+    void *      buffer_;
+    size_t      size_;
+    int32_t     value_;
+    aclrtStream stream_;
 };
 
 /**
@@ -918,25 +906,24 @@ class async_memset_task : public cann_task {
  * same stream are executed in queue order.
  */
 
-#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                          \
-    do {                                                                                    \
-        uint64_t        workspaceSize = 0;                                                  \
-        aclOpExecutor * executor;                                                           \
-        void *          workspaceAddr = nullptr;                                            \
-        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));\
-        /* workspace should alloced in main thread to keep malloc order when using vmm. */  \
-        if (workspaceSize > 0) {                                                            \
-            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);            \
-            workspaceAddr = workspace_allocator.get();                                      \
-        }                                                                                   \
-        if (CTX.async_mode) {                                                               \
-            auto task =                                                                     \
-                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize,  \
-                    executor, CTX.stream()); \
-            CTX.task_queue.submit_task(std::move(task));                                    \
-        } else {                                                                            \
-            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));\
-        }                                                                                   \
+#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                                                  \
+    do {                                                                                                            \
+        uint64_t        workspaceSize = 0;                                                                          \
+        aclOpExecutor * executor;                                                                                   \
+        void *          workspaceAddr = nullptr;                                                                    \
+        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));                        \
+        /* workspace should alloced in main thread to keep malloc order when using vmm. */                          \
+        if (workspaceSize > 0) {                                                                                    \
+            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);                                    \
+            workspaceAddr = workspace_allocator.get();                                                              \
+        }                                                                                                           \
+        if (CTX.async_mode) {                                                                                       \
+            auto task =                                                                                             \
+                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize, executor, CTX.stream()); \
+            CTX.task_queue.submit_task(std::move(task));                                                            \
+        } else {                                                                                                    \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));                        \
+        }                                                                                                           \
     } while (0)
 
 /**
@@ -947,11 +934,10 @@ class async_memset_task : public cann_task {
  * @param ctx Backend context which manages task submission and async mode.
  * @param args Pointers to ACL resources to be released.
  */
-template <typename... Args>
-void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
+template <typename... Args> void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
     std::vector<any_acl_resource> resources;
     register_acl_resources(resources, std::forward<Args>(args)...);
-    if(ctx.async_mode) {
+    if (ctx.async_mode) {
         auto task = std::make_unique<release_resource_task>(std::move(resources));
         ctx.task_queue.submit_task(std::move(task));
     }
@@ -966,8 +952,11 @@ void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... arg
  * @param len Size of memory to copy (in bytes).
  * @param kind Type of memory copy (host-to-device, device-to-host, etc).
  */
-inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
-                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx,
+                                   void *                      dst,
+                                   const void *                src,
+                                   size_t                      len,
+                                   aclrtMemcpyKind             kind) {
     if (ctx.async_mode) {
         auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx.stream());
         ctx.task_queue.submit_task(std::move(task));
@@ -976,8 +965,11 @@ inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
     }
 }
 
-inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
-                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx,
+                                   void *                      dst,
+                                   const void *                src,
+                                   size_t                      len,
+                                   aclrtMemcpyKind             kind) {
     if (ctx->async_mode) {
         auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx->stream());
         ctx->task_queue.submit_task(std::move(task));
@@ -994,8 +986,7 @@ inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
  * @param size Size of the memory buffer (in bytes).
  * @param value Value to set in the buffer.
  */
-inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer,
-                                   size_t size, int value) {
+inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer, size_t size, int value) {
     if (ctx.async_mode) {
         auto task = std::make_unique<async_memset_task>(buffer, size, value, ctx.stream());
         ctx.task_queue.submit_task(std::move(task));
@@ -1029,7 +1020,7 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
  * @param dst The destination tensor where the expert-weighted token outputs are stored.
  *            Expected to be of shape [M, K, N, 1].
  */
-void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Check whether a tensor is a weight tensor for matrix multiplication.
@@ -1041,20 +1032,14 @@ void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  * @param tensor Pointer to the target ggml_tensor object (const-qualified).
  */
-static bool is_matmul_weight(const ggml_tensor* tensor) {
-    std::string name = ggml_get_name(tensor);
-    static const std::unordered_set<std::string> weight_suffixes{
-        "output.weight",
-        "attn_q.weight",
-        "attn_k.weight",
-        "attn_v.weight",
-        "attn_output.weight",
-        "ffn_gate.weight",
-        "ffn_up.weight",
-        "ffn_down.weight"
-    };
+static bool is_matmul_weight(const ggml_tensor * tensor) {
+    std::string                                  name = ggml_get_name(tensor);
+    static const std::unordered_set<std::string> weight_suffixes{ "output.weight",      "attn_q.weight",
+                                                                  "attn_k.weight",      "attn_v.weight",
+                                                                  "attn_output.weight", "ffn_gate.weight",
+                                                                  "ffn_up.weight",      "ffn_down.weight" };
 
-    for (const auto& suffix : weight_suffixes) {
+    for (const auto & suffix : weight_suffixes) {
         if (name.find(suffix) != std::string::npos) {
             return true;
         }
@@ -1078,14 +1063,13 @@ static bool is_matmul_weight(const ggml_tensor* tensor) {
  * @param ctx The CANN backend context used to manage execution and resources.
  * @param dst The destination tensor.
  */
-template <auto binary_op>
-void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    ggml_tensor* src0 = dst->src[0];
-    ggml_tensor* src1 = dst->src[1];
+template <auto binary_op> void ggml_cann_binary_op(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
 
-    aclTensor* acl_src0;
-    aclTensor* acl_src1;
-    aclTensor* acl_dst;
+    aclTensor * acl_src0;
+    aclTensor * acl_src1;
+    aclTensor * acl_dst;
 
     // Need bcast
     bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
@@ -1094,7 +1078,6 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }
 
-
 /**
  * @brief Applies a unary operation to an input tensor using the CANN backend.
  *
@@ -1107,12 +1090,12 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
  * @param ctx The CANN backend context for managing resources and execution.
  * @param dst The destination tensor. Its src[0] is treated as the input tensor.
  */
-template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
-    void ggml_cann_op_unary(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    ggml_tensor* src = dst->src[0];
+template <void unary_op(ggml_backend_cann_context &, aclTensor *, aclTensor *)>
+void ggml_cann_op_unary(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
 
-    aclTensor* acl_src = ggml_cann_create_tensor(src);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    aclTensor * acl_src = ggml_cann_create_tensor(src);
+    aclTensor * acl_dst = ggml_cann_create_tensor(dst);
 
     unary_op(ctx, acl_src, acl_dst);
     ggml_cann_release_resources(ctx, acl_src, acl_dst);
@@ -1138,9 +1121,9 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
  *
  * @see GGML_CANN_CALL_OP_UNARY
  */
-void ggml_cann_op_unary(
-    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
-    ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_op_unary(std::function<void(ggml_backend_cann_context &, aclTensor *, aclTensor *)> unary_op,
+                        ggml_backend_cann_context &                                                ctx,
+                        ggml_tensor *                                                              dst);
 
 /**
  * @brief Applies a gated (GLU-style) unary operation using the CANN backend.
@@ -1172,9 +1155,9 @@ void ggml_cann_op_unary(
  *
  * @see GGML_CANN_CALL_OP_UNARY_GATED
  */
-void ggml_cann_op_unary_gated(
-    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
-    ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, aclTensor *, aclTensor *)> unary_op,
+                              ggml_backend_cann_context &                                                ctx,
+                              ggml_tensor *                                                              dst);
 
 /**
  * @brief Helper macro to call a unary ACL operator via ggml_cann_op_unary.
@@ -1197,16 +1180,13 @@ void ggml_cann_op_unary_gated(
  * @see ggml_cann_op_unary
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                              \
-    do {                                                              \
-        auto lambda = [](ggml_backend_cann_context& ctx,              \
-            aclTensor* acl_src,                                       \
-            aclTensor* acl_dst) {                                     \
-            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
-        };                                                            \
-        ggml_cann_op_unary(lambda, ctx, dst);                         \
-    }                                                                 \
-    while (0)
+#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                                                              \
+    do {                                                                                              \
+        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
+        };                                                                                            \
+        ggml_cann_op_unary(lambda, ctx, dst);                                                         \
+    } while (0)
 
 /**
  * @brief Helper macro to call a gated unary ACL operator via ggml_cann_op_unary_gated.
@@ -1229,15 +1209,12 @@ void ggml_cann_op_unary_gated(
  * @see ggml_cann_op_unary_gated
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                        \
-    do {                                                              \
-        auto lambda = [](ggml_backend_cann_context& ctx,              \
-            aclTensor* acl_src,                                       \
-            aclTensor* acl_dst) {                                     \
-            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
-        };                                                            \
-        ggml_cann_op_unary_gated(lambda, ctx, dst);                   \
-    }                                                                 \
-    while (0)
+#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                                                        \
+    do {                                                                                              \
+        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
+        };                                                                                            \
+        ggml_cann_op_unary_gated(lambda, ctx, dst);                                                   \
+    } while (0)
 
 #endif  // CANN_ACLNN_OPS

ggml/src/ggml-cann/common.h

@@ -44,7 +44,7 @@
 #include "../include/ggml.h"
 #include "../ggml-impl.h"
 
-#define MATRIX_ROW_PADDING 512
+#define MATRIX_ROW_PADDING    512
 #define GGML_CANN_MAX_STREAMS 8
 
 /**
@@ -56,8 +56,7 @@
  * @param line The line number at which the error occurred.
  * @param msg The error message.
  */
-[[noreturn]] void ggml_cann_error(const char* stmt, const char* func,
-                                  const char* file, int line, const char* msg);
+[[noreturn]] void ggml_cann_error(const char * stmt, const char * func, const char * file, int line, const char * msg);
 
 /**
  * @brief Checks the result of a CANN function call and invokes the error
@@ -89,25 +88,24 @@ struct ggml_cann_device_info {
      * @brief Information about a single CANN device.
      */
     struct cann_device_info {
-        int cc;                 /**< Compute capability.                   */
+        int    cc;              /**< Compute capability.                   */
         size_t smpb;            /**< Maximum shared memory per block.      */
-        bool vmm;               /**< Virtual memory support.               */
+        bool   vmm;             /**< Virtual memory support.               */
         size_t vmm_granularity; /**< Granularity of virtual memory.        */
         size_t total_vram;      /**< Total video RAM available on the device. */
     };
 
-    cann_device_info devices[GGML_CANN_MAX_DEVICES] =
-        {}; /**< Array of CANN device information. */
+    cann_device_info devices[GGML_CANN_MAX_DEVICES] = {}; /**< Array of CANN device information. */
 };
 
-const ggml_cann_device_info& ggml_cann_info();
+const ggml_cann_device_info & ggml_cann_info();
 
-void ggml_cann_set_device(int32_t device);
+void    ggml_cann_set_device(int32_t device);
 int32_t ggml_cann_get_device();
 
-std::optional<std::string> get_env(const std::string& name);
-bool parse_bool(const std::string& value);
-int parse_integer(const std::string& value);
+std::optional<std::string> get_env(const std::string & name);
+bool                       parse_bool(const std::string & value);
+int                        parse_integer(const std::string & value);
 
 /**
  * @brief Abstract base class for memory pools used by CANN.
@@ -126,7 +124,7 @@ struct ggml_cann_pool {
      *                     will be stored.
      * @return             Pointer to the allocated memory block.
      */
-    virtual void* alloc(size_t size, size_t* actual_size) = 0;
+    virtual void * alloc(size_t size, size_t * actual_size) = 0;
 
     /**
      * @brief Frees a previously allocated memory block.
@@ -136,16 +134,16 @@ struct ggml_cann_pool {
      * @note Note that all CANN opertors are running async. Make sure memory is
      *       still avaiable before this operator finished.
      */
-    virtual void free(void* ptr, size_t size) = 0;
+    virtual void free(void * ptr, size_t size) = 0;
 };
 
 /**
  * @brief RAII wrapper for managing memory allocations from a CANN memory pool.
  */
 struct ggml_cann_pool_alloc {
-    ggml_cann_pool* pool = nullptr; /**< Pointer to the memory pool. */
-    void* ptr = nullptr;    /**< Pointer to the allocated memory block. */
-    size_t actual_size = 0; /**< Actual size of the allocated memory block. */
+    ggml_cann_pool * pool        = nullptr; /**< Pointer to the memory pool. */
+    void *           ptr         = nullptr; /**< Pointer to the allocated memory block. */
+    size_t           actual_size = 0;       /**< Actual size of the allocated memory block. */
 
     /**
      * @brief Default constructor.
@@ -156,16 +154,14 @@ struct ggml_cann_pool_alloc {
      * @brief Constructor that initializes the memory pool.
      * @param pool Reference to the memory pool.
      */
-    explicit ggml_cann_pool_alloc(ggml_cann_pool& pool) : pool(&pool) {}
+    explicit ggml_cann_pool_alloc(ggml_cann_pool & pool) : pool(&pool) {}
 
     /**
      * @brief Constructor that initializes the memory pool and allocates memory.
      * @param pool Reference to the memory pool.
      * @param size Size of the memory block to allocate.
      */
-    ggml_cann_pool_alloc(ggml_cann_pool& pool, size_t size) : pool(&pool) {
-        alloc(size);
-    }
+    ggml_cann_pool_alloc(ggml_cann_pool & pool, size_t size) : pool(&pool) { alloc(size); }
 
     /**
      * @brief Destructor that frees the allocated memory block.
@@ -181,7 +177,7 @@ struct ggml_cann_pool_alloc {
      * @param size Size of the memory block to allocate.
      * @return Pointer to the allocated memory block.
      */
-    void* alloc(size_t size) {
+    void * alloc(size_t size) {
         GGML_ASSERT(pool != nullptr);
         GGML_ASSERT(ptr == nullptr);
         ptr = pool->alloc(size, &this->actual_size);
@@ -194,7 +190,7 @@ struct ggml_cann_pool_alloc {
      * @param size Size of the memory block to allocate.
      * @return Pointer to the allocated memory block.
      */
-    void* alloc(ggml_cann_pool& pool, size_t size) {
+    void * alloc(ggml_cann_pool & pool, size_t size) {
         this->pool = &pool;
         return alloc(size);
     }
@@ -203,25 +199,25 @@ struct ggml_cann_pool_alloc {
      * @brief Gets the pointer to the allocated memory block.
      * @return Pointer to the allocated memory block.
      */
-    void* get() { return ptr; }
+    void * get() { return ptr; }
 
     // Deleted copy constructor
-    ggml_cann_pool_alloc(const ggml_cann_pool_alloc&) = delete;
+    ggml_cann_pool_alloc(const ggml_cann_pool_alloc &) = delete;
 
     // Deleted move constructor
-    ggml_cann_pool_alloc(ggml_cann_pool_alloc&&) = delete;
+    ggml_cann_pool_alloc(ggml_cann_pool_alloc &&) = delete;
 
     // Deleted copy assignment operator
-    ggml_cann_pool_alloc& operator=(const ggml_cann_pool_alloc&) = delete;
+    ggml_cann_pool_alloc & operator=(const ggml_cann_pool_alloc &) = delete;
 
     // Deleted move assignment operator
-    ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
+    ggml_cann_pool_alloc & operator=(ggml_cann_pool_alloc &&) = delete;
 };
 
 /**
  * @brief Function pointer type for ACLNN operator calls.
  */
-using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStream);
+using aclnn_func_t = aclnnStatus (*)(void *, uint64_t, aclOpExecutor *, aclrtStream);
 
 /**
  * @brief Base class for all CANN tasks to be submitted to the task queue.
@@ -229,7 +225,7 @@ using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStrea
  * Users should override the run_task() method with actual task logic.
  */
 class cann_task {
-public:
+  public:
     virtual void run_task() {}
 };
 
@@ -237,16 +233,20 @@ public:
  * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
  */
 class cann_task_queue {
-public:
+  public:
     /**
      * @brief Constructs a task queue with a fixed power-of-two capacity for a specific device.
      *
      * @param capacity Queue capacity. Must be a power of 2.
      * @param device Target device ID (used for context setting).
      */
-    explicit cann_task_queue(size_t capacity, int32_t device)
-        : buffer_(capacity), capacity_(capacity), head_(0), tail_(0),
-          running_(false), device_(device) {
+    explicit cann_task_queue(size_t capacity, int32_t device) :
+        buffer_(capacity),
+        capacity_(capacity),
+        head_(0),
+        tail_(0),
+        running_(false),
+        device_(device) {
         GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
         mask_ = capacity_ - 1;
     }
@@ -257,7 +257,7 @@ public:
      * @param item Unique pointer to the task.
      * @return true if the task was successfully enqueued, false if the queue was full.
      */
-    bool enqueue(std::unique_ptr<cann_task>&& item) {
+    bool enqueue(std::unique_ptr<cann_task> && item) {
         size_t next_tail = (tail_ + 1) & mask_;
 
         if (next_tail == head_) {
@@ -276,17 +276,16 @@ public:
      *
      * @param task Task to be submitted.
      */
-    void submit_task(std::unique_ptr<cann_task>&& task) {
-        while(!enqueue(std::move(task))) {
+    void submit_task(std::unique_ptr<cann_task> && task) {
+        while (!enqueue(std::move(task))) {
             std::this_thread::yield();
             continue;
         }
 
         if (!running_) {
             running_ = true;
-            thread_ = std::thread(&cann_task_queue::execute, this);
+            thread_  = std::thread(&cann_task_queue::execute, this);
         }
-
     }
 
     /**
@@ -309,7 +308,7 @@ public:
         }
     }
 
-private:
+  private:
     /**
      * @brief Worker thread function that continuously dequeues and executes tasks.
      */
@@ -317,7 +316,7 @@ private:
         ggml_cann_set_device(device_);
 
         while (running_) {
-            if(head_ == tail_) {
+            if (head_ == tail_) {
                 std::this_thread::yield();
                 continue;
             }
@@ -330,24 +329,24 @@ private:
     }
 
     std::vector<std::unique_ptr<cann_task>> buffer_;
-    const size_t capacity_;
-    size_t mask_;
-    size_t head_;
-    size_t tail_;
-    bool running_;
-    std::thread thread_;
-    int32_t device_;
+    const size_t                            capacity_;
+    size_t                                  mask_;
+    size_t                                  head_;
+    size_t                                  tail_;
+    bool                                    running_;
+    std::thread                             thread_;
+    int32_t                                 device_;
 };
 
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
     // dst tensor
-    void * node_address;
+    void *  node_address;
     int64_t ne[GGML_MAX_DIMS];
-    size_t nb[GGML_MAX_DIMS];
+    size_t  nb[GGML_MAX_DIMS];
 
     // src tensor
-    void * src_address[GGML_MAX_SRC];
+    void *  src_address[GGML_MAX_SRC];
     int64_t src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
     size_t  src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
 
@@ -376,13 +375,11 @@ struct ggml_cann_graph {
  * move existing graphs to the front (most recently used), and clear the cache.
  */
 struct ggml_cann_graph_lru_cache {
-    size_t capacity;  /**< Maximum number of graphs in the cache. */
+    size_t capacity;                         /**< Maximum number of graphs in the cache. */
 
-    std::list<ggml_cann_graph*> cache_list; /**< List storing cached graphs as raw pointers. */
+    std::list<ggml_cann_graph *> cache_list; /**< List storing cached graphs as raw pointers. */
 
-    ggml_cann_graph_lru_cache() {
-        capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12"));
-    }
+    ggml_cann_graph_lru_cache() { capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12")); }
 
     /**
      * @brief Push a new graph to the front of the cache.
@@ -390,11 +387,11 @@ struct ggml_cann_graph_lru_cache {
      * @param new_node Pointer to the new ggml_cann_graph to cache.
      *        Ownership is transferred to the cache (cache will delete it).
      */
-    void push(ggml_cann_graph* new_node) {
+    void push(ggml_cann_graph * new_node) {
         if (cache_list.size() >= capacity) {
-            ggml_cann_graph* old = cache_list.back();
+            ggml_cann_graph * old = cache_list.back();
             cache_list.pop_back();
-            delete old; // free the old graph
+            delete old;  // free the old graph
         }
         cache_list.push_front(new_node);
     }
@@ -403,7 +400,7 @@ struct ggml_cann_graph_lru_cache {
      * @brief Move an existing graph to the front of the cache.
      * @param node Pointer to the ggml_cann_graph to move.
      */
-    void move_to_front(ggml_cann_graph* node) {
+    void move_to_front(ggml_cann_graph * node) {
         cache_list.remove(node);
         cache_list.push_front(node);
     }
@@ -421,92 +418,89 @@ struct ggml_cann_graph_lru_cache {
     /**
      * @brief Destructor that clears the cache and frees all cached graphs.
      */
-    ~ggml_cann_graph_lru_cache() {
-        clear();
-    }
+    ~ggml_cann_graph_lru_cache() { clear(); }
 };
 #endif  // USE_ACL_GRAPH
 
 struct ggml_cann_rope_cache {
     ~ggml_cann_rope_cache() {
-        if(theta_scale_cache != nullptr) {
+        if (theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(theta_scale_cache));
         }
-        if(sin_cache != nullptr) {
+        if (sin_cache != nullptr) {
             ACL_CHECK(aclrtFree(sin_cache));
         }
-        if(cos_cache != nullptr) {
+        if (cos_cache != nullptr) {
             ACL_CHECK(aclrtFree(cos_cache));
         }
     }
 
-    void* theta_scale_cache = nullptr;
+    void *  theta_scale_cache  = nullptr;
     int64_t theta_scale_length = 0;
     // 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;
+    int64_t position_length    = 0;
     // 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;
+    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;
 };
 
 struct ggml_cann_tensor_cache {
     ~ggml_cann_tensor_cache() {
-        if(cache != nullptr) {
+        if (cache != nullptr) {
             ACL_CHECK(aclrtFree(cache));
         }
     }
 
-    void* cache = nullptr;
-    int64_t size = 0;
+    void *  cache = nullptr;
+    int64_t size  = 0;
 };
 
 /**
  * @brief Context for managing CANN backend operations.
  */
 struct ggml_backend_cann_context {
-    int32_t device;                  /**< Device ID. */
-    std::string name;                /**< Name of the device. */
-    std::string description;         /**< Description of the device. */
-    aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+    int32_t     device;               /**< Device ID. */
+    std::string name;                 /**< Name of the device. */
+    std::string description;          /**< Description of the device. */
+    aclrtEvent  copy_event = nullptr; /**< Event for managing copy operations. */
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
     ggml_cann_graph_lru_cache graph_lru_cache;
-    bool acl_graph_mode = true;
+    bool                      acl_graph_mode = true;
 #endif
-    cann_task_queue task_queue;
-    bool async_mode;
+    cann_task_queue        task_queue;
+    bool                   async_mode;
     // Rope Cache
-    ggml_cann_rope_cache rope_cache;
+    ggml_cann_rope_cache   rope_cache;
     // Constant Pool
     ggml_cann_tensor_cache rms_norm_one_tensor_cache;
     ggml_cann_tensor_cache rms_norm_zero_tensor_cache;
 
-    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
+    aclrtStream streams[GGML_CANN_MAX_STREAMS] = { nullptr }; /**< Array of streams for the device. */
 
     /**
      * @brief Constructor for initializing the context with a given device.
      * @param device Device ID.
      */
-    explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
+    explicit ggml_backend_cann_context(int device) :
+        device(device),
+        name("CANN" + std::to_string(device)),
+        task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
 
         async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
-        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
-            device, async_mode ? "ON" : "OFF");
+        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__, device, async_mode ? "ON" : "OFF");
 #ifdef USE_ACL_GRAPH
         acl_graph_mode = parse_bool(get_env("GGML_CANN_ACL_GRAPH").value_or("on"));
-        GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n",
-              __func__, device,
-              acl_graph_mode ? "GRAPH" : "EAGER",
-              acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
+        GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n", __func__, device, acl_graph_mode ? "GRAPH" : "EAGER",
+                      acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
 #endif
     }
 
@@ -549,8 +543,7 @@ struct ggml_backend_cann_context {
     aclrtStream stream() { return stream(0); }
 
     // TODO: each stream should have a memory pool.
-    std::unique_ptr<ggml_cann_pool>
-        mem_pool; /**< Memory pool for the device. */
+    std::unique_ptr<ggml_cann_pool> mem_pool; /**< Memory pool for the device. */
 
     /**
      * @brief Create a new memory pool for a given device.
@@ -563,7 +556,7 @@ struct ggml_backend_cann_context {
      * @brief Get or create the memory pool for the context.
      * @return Reference to the memory pool.
      */
-    ggml_cann_pool& pool() {
+    ggml_cann_pool & pool() {
         if (mem_pool == nullptr) {
             mem_pool = new_pool_for_device(device);
         }

ggml/src/ggml-cpu/CMakeLists.txt

@@ -466,29 +466,45 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         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-cpu/arch/s390/quants.c)
-        file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
-        string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/s390/quants.c)
 
-        # TODO: Separation to determine activation of VX/VXE/VXE2
-        if (${S390X_M} MATCHES "8561|8562")
-            message(STATUS "z15 target")
-            list(APPEND ARCH_FLAGS -march=z15)
-        elseif (${S390X_M} MATCHES "3931")
-            message(STATUS "z16 target")
-            list(APPEND ARCH_FLAGS -march=z16)
-        elseif (${S390X_M} MATCHES "9175|9176")
-            # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
-            #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
-            message(STATUS "z17 target")
-            list(APPEND ARCH_FLAGS -march=arch15)
-        else()
-            message(STATUS "Unknown target")
-            message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
-            list(APPEND ARCH_FLAGS -march=native -mtune=native)
+        # for native compilation
+        if (GGML_NATIVE)
+            # check machine level to determine target
+            file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
+            string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
+
+            # TODO: Separation to determine activation of VX/VXE/VXE2
+            if (${S390X_M} MATCHES "8561|8562")
+                message(STATUS "z15 target")
+                list(APPEND ARCH_FLAGS -march=z15)
+            elseif (${S390X_M} MATCHES "3931")
+                message(STATUS "z16 target")
+                list(APPEND ARCH_FLAGS -march=z16)
+            elseif (${S390X_M} MATCHES "9175|9176")
+                # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+                #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
+                message(STATUS "z17 target")
+                list(APPEND ARCH_FLAGS -march=arch15)
+            else()
+                message(STATUS "Unknown target")
+                message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
+                list(APPEND ARCH_FLAGS -march=native -mtune=native)
+            endif()
+        # for cross-compilation
+        elseif(GGML_CPU_ALL_VARIANTS)
+            # range through IBM z15 to z17
+            # NOTE: update when a new hardware level is released
+            foreach (ZHW RANGE 15 17)
+                if(DEFINED GGML_INTERNAL_Z${ZHW})
+                    message(STATUS "z${ZHW} cross-compile target")
+                    list(APPEND ARCH_FLAGS -march=z${ZHW})
+                endif()
+            endforeach()
         endif()
 
-        if (GGML_VXE)
+        if (GGML_VXE OR GGML_INTERNAL_VXE)
             message(STATUS "VX/VXE/VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
             list(APPEND ARCH_DEFINITIONS GGML_VXE)

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -68,7 +68,7 @@ struct ggml_compute_params {
 #endif  // __VXE2__
 #endif  // __s390x__ && __VEC__
 
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_FEATURE_SVE) && defined(__linux__)
 #include <sys/prctl.h>
 #endif
 

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

@@ -689,8 +689,13 @@ bool ggml_is_numa(void) {
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__linux__)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+#else
+    // TODO: add support of SVE for non-linux systems
+#error "TODO: SVE is not supported on this platform. To use SVE, sve_cnt needs to be initialized here."
+#endif
 #endif
 }
 
@@ -2179,6 +2184,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     {
                         n_tasks = 1;
                     } break;
@@ -3558,13 +3567,17 @@ void ggml_cpu_init(void) {
 #ifdef GGML_USE_OPENMP
             //if (!getenv("OMP_WAIT_POLICY")) {
             //    // set the wait policy to active, so that OpenMP threads don't sleep
-            //    putenv("OMP_WAIT_POLICY=active");
+            //    setenv("OMP_WAIT_POLICY", "active", 0)
             //}
 
             if (!getenv("KMP_BLOCKTIME")) {
                 // set the time to wait before sleeping a thread
                 // this is less aggressive than setting the wait policy to active, but should achieve similar results in most cases
-                putenv("KMP_BLOCKTIME=200"); // 200ms
+#ifdef _WIN32
+                _putenv_s("KMP_BLOCKTIME", "200"); // 200ms
+#else
+                setenv("KMP_BLOCKTIME", "200", 0); // 200ms
+#endif
             }
 #endif
         }

ggml/src/ggml-cpu/ops.cpp

@@ -7519,8 +7519,8 @@ static void ggml_compute_forward_upscale_f32(
         float pixel_offset = 0.5f;
         if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
             pixel_offset = 0.0f;
-            sf0 = (float)(ne0 - 1) / (src0->ne[0] - 1);
-            sf1 = (float)(ne1 - 1) / (src0->ne[1] - 1);
+            sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+            sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
         }
 
         for (int64_t i3 = 0; i3 < ne3; i3++) {
@@ -8993,6 +8993,22 @@ void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_exp(params, dst);
             } break;
+        case GGML_UNARY_OP_FLOOR:
+            {
+                ggml_compute_forward_floor(params, dst);
+            } break;
+        case GGML_UNARY_OP_CEIL:
+            {
+                ggml_compute_forward_ceil(params, dst);
+            } break;
+        case GGML_UNARY_OP_ROUND:
+            {
+                ggml_compute_forward_round(params, dst);
+            } break;
+        case GGML_UNARY_OP_TRUNC:
+            {
+                ggml_compute_forward_trunc(params, dst);
+            } break;
         case GGML_UNARY_OP_XIELU:
             {
                 ggml_compute_forward_xielu(params, dst);

ggml/src/ggml-cpu/spacemit/ime.cpp

@@ -485,8 +485,9 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
             int32_t          start                  = ith * task_per_thread;
             int32_t          end                    = std::min((ith + 1) * task_per_thread, task_count);
             for (int32_t compute_idx = start; compute_idx < end; compute_idx++) {
-                int32_t                             gemm_idx = compute_idx / block_size_m;
-                int32_t                             m_idx    = compute_idx % block_size_m * block_size_m;
+                int32_t                             gemm_idx = compute_idx / per_gemm_block_count_m;
+                int32_t                             block_idx_in_gemm = compute_idx % per_gemm_block_count_m;
+                int32_t                             m_idx    = block_idx_in_gemm * block_size_m;
                 const qnbitgemm_spacemit_ime_args & data     = qnbitgemm_args[gemm_idx];
                 int32_t rows_tobe_handled = (gemm_m - m_idx) > block_size_m ? block_size_m : (gemm_m - m_idx);
 

ggml/src/ggml-cpu/unary-ops.cpp

@@ -73,6 +73,22 @@ static inline float op_log(float x) {
     return logf(x);
 }
 
+static inline float op_floor(float x) {
+    return floorf(x);
+}
+
+static inline float op_ceil(float x) {
+    return ceilf(x);
+}
+
+static inline float op_round(float x) {
+    return roundf(x);
+}
+
+static inline float op_trunc(float x) {
+    return truncf(x);
+}
+
 template <float (*op)(float), typename src0_t, typename dst_t>
 static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
     constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
@@ -274,6 +290,22 @@ void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor *
     unary_op<op_log>(params, dst);
 }
 
+void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_floor>(params, dst);
+}
+
+void ggml_compute_forward_ceil(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_ceil>(params, dst);
+}
+
+void ggml_compute_forward_round(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_round>(params, dst);
+}
+
+void ggml_compute_forward_trunc(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_trunc>(params, dst);
+}
+
 void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor * dst) {
     const float alpha_n = ggml_get_op_params_f32(dst, 1);
     const float alpha_p = ggml_get_op_params_f32(dst, 2);

ggml/src/ggml-cpu/unary-ops.h

@@ -22,6 +22,10 @@ void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_floor(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_ceil(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_round(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_trunc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_xielu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 
 #ifdef __cplusplus

ggml/src/ggml-cuda/argsort.cu

@@ -1,5 +1,81 @@
 #include "argsort.cuh"
 
+#ifdef GGML_CUDA_USE_CUB
+#    include <cub/cub.cuh>
+using namespace cub;
+#endif  // GGML_CUDA_USE_CUB
+
+static __global__ void init_indices(int * indices, const int ncols, const int nrows) {
+    const int col = blockIdx.x * blockDim.x + threadIdx.x;
+    const int row = blockIdx.y;
+
+    if (col < ncols && row < nrows) {
+        indices[row * ncols + col] = col;
+    }
+}
+
+static __global__ void init_offsets(int * offsets, const int ncols, const int nrows) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx <= nrows) {
+        offsets[idx] = idx * ncols;
+    }
+}
+
+#ifdef GGML_CUDA_USE_CUB
+static void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
+                                     const float *    x,
+                                     int *            dst,
+                                     const int        ncols,
+                                     const int        nrows,
+                                     ggml_sort_order  order,
+                                     cudaStream_t     stream) {
+    ggml_cuda_pool_alloc<int>   temp_indices_alloc(pool, ncols * nrows);
+    ggml_cuda_pool_alloc<float> temp_keys_alloc(pool, ncols * nrows);
+    ggml_cuda_pool_alloc<int>   offsets_alloc(pool, nrows + 1);
+
+    int *   temp_indices = temp_indices_alloc.get();
+    float * temp_keys    = temp_keys_alloc.get();
+    int *   d_offsets    = offsets_alloc.get();
+
+    static const int block_size = 256;
+    const dim3 grid_size((ncols + block_size - 1) / block_size, nrows);
+    init_indices<<<grid_size, block_size, 0, stream>>>(temp_indices, ncols, nrows);
+
+    const dim3 offset_grid((nrows + block_size - 1) / block_size);
+    init_offsets<<<offset_grid, block_size, 0, stream>>>(d_offsets, ncols, nrows);
+
+    cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream);
+
+    size_t temp_storage_bytes = 0;
+
+    if (order == GGML_SORT_ORDER_ASC) {
+        DeviceSegmentedRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+                                            temp_indices, dst,                                  // values (indices)
+                                            ncols * nrows, nrows,                            // num items, num segments
+                                            d_offsets, d_offsets + 1, 0, sizeof(float) * 8,  // all bits
+                                            stream);
+    } else {
+        DeviceSegmentedRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices,
+                                                      dst, ncols * nrows, nrows, d_offsets, d_offsets + 1, 0,
+                                                      sizeof(float) * 8, stream);
+    }
+
+    ggml_cuda_pool_alloc<uint8_t> temp_storage_alloc(pool, temp_storage_bytes);
+    void *                        d_temp_storage = temp_storage_alloc.get();
+
+    if (order == GGML_SORT_ORDER_ASC) {
+        DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst,
+                                            ncols * nrows, nrows, d_offsets, d_offsets + 1, 0, sizeof(float) * 8,
+                                            stream);
+    } else {
+        DeviceSegmentedRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
+                                                      temp_indices, dst, ncols * nrows, nrows, d_offsets, d_offsets + 1,
+                                                      0, sizeof(float) * 8, stream);
+    }
+}
+#endif  // GGML_CUDA_USE_CUB
+
+// Bitonic sort implementation
 template<typename T>
 static inline __device__ void ggml_cuda_swap(T & a, T & b) {
     T tmp = a;
@@ -65,7 +141,12 @@ static int next_power_of_2(int x) {
     return n;
 }
 
-static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
+static void argsort_f32_i32_cuda_bitonic(const float *   x,
+                                         int *           dst,
+                                         const int       ncols,
+                                         const int       nrows,
+                                         ggml_sort_order order,
+                                         cudaStream_t    stream) {
     // bitonic sort requires ncols to be power of 2
     const int ncols_pad = next_power_of_2(ncols);
 
@@ -77,9 +158,11 @@ static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, co
     GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
 
     if (order == GGML_SORT_ORDER_ASC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+        k_argsort_f32_i32<GGML_SORT_ORDER_ASC>
+            <<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else if (order == GGML_SORT_ORDER_DESC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+        k_argsort_f32_i32<GGML_SORT_ORDER_DESC>
+            <<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else {
         GGML_ABORT("fatal error");
     }
@@ -100,5 +183,18 @@ void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
 
-    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
+#ifdef GGML_CUDA_USE_CUB
+    const int    ncols_pad      = next_power_of_2(ncols);
+    const size_t shared_mem     = ncols_pad * sizeof(int);
+    const size_t max_shared_mem = ggml_cuda_info().devices[ggml_cuda_get_device()].smpb;
+
+    if (shared_mem > max_shared_mem || ncols > 1024) {
+        ggml_cuda_pool & pool = ctx.pool();
+        argsort_f32_i32_cuda_cub(pool, src0_d, (int *) dst_d, ncols, nrows, order, stream);
+    } else {
+        argsort_f32_i32_cuda_bitonic(src0_d, (int *) dst_d, ncols, nrows, order, stream);
+    }
+#else
+    argsort_f32_i32_cuda_bitonic(src0_d, (int *) dst_d, ncols, nrows, order, stream);
+#endif
 }

ggml/src/ggml-cuda/binbcast.cu

@@ -272,7 +272,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         const uint3 ne12 = init_fastdiv_values((uint32_t) cne1[2]);
         const uint3 ne13 = init_fastdiv_values((uint32_t) cne1[3]);
 
-        if (block_nums.z > 65535) {
+        if (block_nums.z > 65535 || block_nums.y > 65535) {
             int         block_num  = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
             const uint3 prod_012    = init_fastdiv_values((uint32_t) (ne0 * ne1 * ne2));
             const uint3 prod_01     = init_fastdiv_values((uint32_t) (ne0 * ne1));

ggml/src/ggml-cuda/common.cuh

@@ -944,13 +944,6 @@ struct ggml_cuda_graph {
     bool disable_due_to_failed_graph_capture = false;
     int number_consecutive_updates = 0;
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
-    bool use_cpy_indirection = false;
-    std::vector<char *> cpy_dest_ptrs;
-    char ** dest_ptrs_d;
-    int dest_ptrs_size = 0;
-    // Index to allow each cpy kernel to be aware of it's position within the graph
-    // relative to other cpy nodes.
-    int graph_cpynode_index = -1;
 #endif
 };
 
@@ -1012,3 +1005,16 @@ struct ggml_backend_cuda_context {
         return pool(device);
     }
 };
+
+struct ggml_cuda_mm_fusion_args_host {
+    const ggml_tensor * x_bias = nullptr;
+    const ggml_tensor * gate = nullptr;
+    const ggml_tensor * gate_bias = nullptr;
+    ggml_glu_op glu_op;
+};
+struct ggml_cuda_mm_fusion_args_device {
+    const void * x_bias = nullptr;
+    const void * gate = nullptr;
+    const void * gate_bias = nullptr;
+    ggml_glu_op glu_op;
+};

ggml/src/ggml-cuda/convert.cuh

@@ -1,3 +1,4 @@
+#pragma once
 #include "common.cuh"
 
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256

ggml/src/ggml-cuda/cpy.cu

@@ -8,18 +8,16 @@
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
 template <cpy_kernel_t cpy_1>
-static __global__ void cpy_flt(const char * cx, char * cdst_direct, const int ne,
+static __global__ void cpy_flt(const char * cx, char * cdst, const int ne,
                                const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                               const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+                               const int nb12, const int nb13) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
         return;
     }
 
-    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
-
     // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
     // then combine those indices with the corresponding byte offsets to get the total offsets
     const int64_t i03 = i/(ne00 * ne01 * ne02);
@@ -63,18 +61,16 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
+static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+                                 const int nb12, const int nb13) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
-    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
-
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -91,18 +87,16 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int ne,
+static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+                                 const int nb12, const int nb13) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
-    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
-
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -118,67 +112,71 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
-// Copy destination pointers to GPU to be available when pointer indirection is in use
+template<typename src_t, typename dst_t>
+static __global__ void cpy_flt_contiguous(const char * cx, char * cdst, const int64_t ne) {
+    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
-void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream) {
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
-    if (cuda_graph->dest_ptrs_size < host_dest_ptrs_size) { // (re-)allocate GPU memory for destination pointers
-        CUDA_CHECK(cudaStreamSynchronize(stream));
-        if (cuda_graph->dest_ptrs_d != nullptr) {
-            CUDA_CHECK(cudaFree(cuda_graph->dest_ptrs_d));
-        }
-        CUDA_CHECK(cudaMalloc(&cuda_graph->dest_ptrs_d, host_dest_ptrs_size*sizeof(char *)));
-        cuda_graph->dest_ptrs_size = host_dest_ptrs_size;
+    if (i >= ne) {
+        return;
     }
-    // copy destination pointers to GPU
-    CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream));
-    cuda_graph->graph_cpynode_index = 0; // reset index
-#else
-    GGML_UNUSED_VARS(cuda_graph, host_dest_ptrs, host_dest_ptrs_size, stream);
-#endif
+
+    const src_t * x = (const src_t *) cx;
+    dst_t *     dst = (dst_t *) cdst;
+
+    dst[i] = ggml_cuda_cast<dst_t>(x[i]);
+}
+
+template<typename src_t, typename dst_t>
+static void ggml_cpy_flt_contiguous_cuda(
+    const char * cx, char * cdst, const int64_t ne,
+cudaStream_t stream) {
+
+    const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_flt_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne);
 }
 
 template<typename src_t, typename dst_t>
 static void ggml_cpy_flt_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q8_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK8_0 == 0);
     const int num_blocks = ne / QK8_0;
     cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q8_0_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q4_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
     cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q4_0_f32_cuda(
@@ -187,22 +185,22 @@ static void ggml_cpy_q4_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q4_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
     cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q4_1_f32_cuda(
@@ -211,22 +209,22 @@ static void ggml_cpy_q4_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q5_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
     cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q5_0_f32_cuda(
@@ -235,22 +233,22 @@ static void ggml_cpy_q5_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q5_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK5_1 == 0);
     const int num_blocks = ne / QK5_1;
     cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q5_1_f32_cuda(
@@ -259,25 +257,25 @@ static void ggml_cpy_q5_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_iq4_nl_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK4_NL == 0);
     const int num_blocks = ne / QK4_NL;
     cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -311,17 +309,9 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
 
-    char ** dest_ptrs_d = nullptr;
-    int graph_cpynode_index = -1;
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
-        dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
-        graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
-    }
-#else
-    GGML_UNUSED(disable_indirection_for_this_node);
-#endif
-    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+    const bool contiguous_srcs = ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
+
+    if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
 #if defined(GGML_USE_MUSA) && defined(GGML_MUSA_MUDNN_COPY)
         if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) {
@@ -329,134 +319,94 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         } else
 #endif // GGML_USE_MUSA && GGML_MUSA_MUDNN_COPY
         {
-            if (src0->type == GGML_TYPE_F32) {
-                ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
-            } else {
-                CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
-            }
+            CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, float>           (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<float, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<float, half>        (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<float, half>        (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
-        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
-        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
-        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<half, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<half, half>               (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<half, nv_bfloat16>  (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<half, nv_bfloat16>    (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<half, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<half, float>        (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<half, float>          (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<nv_bfloat16, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, half>  (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<nv_bfloat16, half>    (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<nv_bfloat16, float>   (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
-        ggml_cpy_flt_cuda<float, int32_t> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<float, int32_t>     (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<float, int32_t>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<int32_t, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<int32_t, float>     (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<int32_t, float>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
-        ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
-    }
-#else
-    GGML_UNUSED(disable_indirection_for_this_node);
-#endif
-
 }
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
-    bool disable_indirection = true;
-    ggml_cuda_cpy(ctx, src0, dst, disable_indirection);
-}
-
-void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
-    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
-        // Prioritize CUDA graph compatibility over direct memory copy optimization.
-        // Using copy kernels here maintains graph indirection support, preventing performance regression from disabled CUDA graphs.
-        if (src0->type == GGML_TYPE_F32) {
-            return (void*) cpy_flt<cpy_1_flt<float, float>>;
-        } else {
-            return nullptr;
-        }
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_flt<cpy_1_flt<float, float>>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_flt<cpy_1_flt<float, nv_bfloat16>>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_flt<cpy_1_flt<float, half>>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
-    } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>;
-    } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>;
-    } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>;
-    } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
-        return (void*) cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>;
-    } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>;
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_flt<cpy_1_flt<half, half>>;
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_flt<cpy_1_flt<half, nv_bfloat16>>;
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_flt<cpy_1_flt<half, float>>;
-    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, half>>;
-    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, nv_bfloat16>>;
-    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, float>>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
-        return (void*) cpy_flt<cpy_1_flt<float, int32_t>>;
-    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_flt<cpy_1_flt<int32_t, float>>;
-    } else {
-        GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
-                ggml_type_name(src0->type), ggml_type_name(src1->type));
-    }
+    ggml_cuda_cpy(ctx, src0, dst);
 }

ggml/src/ggml-cuda/cpy.cuh

@@ -2,10 +2,6 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
-
-void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1);
-
-void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream);

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

@@ -895,6 +895,7 @@ void launch_fattn(
     const dim3 block_dim(warp_size, nwarps, 1);
     int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
     CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
+    GGML_ASSERT(max_blocks_per_sm > 0);
     int parallel_blocks = max_blocks_per_sm;
 
     dim3 blocks_num;

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

@@ -540,10 +540,12 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
                 KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] = logit_softcap * tanhf(KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
             }
 
-            KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] += (ncols2 > 1 || mask) && (!oob_check || i_KQ < k_VKQ_sup) ?
-                slope*__half2float(mask[j*stride_mask + k_VKQ_0 + i_KQ]) : 0.0f;
+            if (!oob_check || i_KQ < k_VKQ_sup) {
+                KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] += (ncols2 > 1 || mask) ?
+                    slope*__half2float(mask[j*stride_mask + k_VKQ_0 + i_KQ]) : 0.0f;
 
-            KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
+                KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
+            }
         }
 
         KQ_max_new[jc0] = warp_reduce_max<warp_size>(KQ_max_new[jc0]);
@@ -581,10 +583,9 @@ 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 = expf(KQ_acc[(i0/(np*warp_size))*cpw + jc] - KQ_max[jc]);
-                if (!oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < k_VKQ_sup) {
-                    KQ_sum_add += val;
-                }
+                const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < 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;
             }
             KQ_sum[jc] = KQ_sum[jc]*KQ_max_scale + KQ_sum_add;
@@ -975,26 +976,6 @@ static __global__ void flash_attn_tile(
         }
     }
 
-    if (gridDim.y == 1) {
-#pragma unroll
-        for (int jc0 = 0; jc0 < cpw; ++jc0) {
-#ifdef FAST_FP16_AVAILABLE
-            const half2 KQ_sum_jc_inv = make_half2(1.0f/KQ_sum[jc0], 1.0f/KQ_sum[jc0]);
-#pragma unroll
-            for (int i = 0; i < (DVp/2)/warp_size; ++i) {
-                VKQ[jc0*((DVp/2)/warp_size) + i] *= KQ_sum_jc_inv;
-            }
-#else
-            const float KQ_sum_jc_inv = 1.0f/KQ_sum[jc0];
-#pragma unroll
-            for (int i = 0; i < (DVp/2)/warp_size; ++i) {
-                VKQ[jc0*((DVp/2)/warp_size) + i].x *= KQ_sum_jc_inv;
-                VKQ[jc0*((DVp/2)/warp_size) + i].y *= KQ_sum_jc_inv;
-            }
-#endif // FAST_FP16_AVAILABLE
-        }
-    }
-
     // Write back results:
 #pragma unroll
     for (int jc0 = 0; jc0 < cpw; ++jc0) {
@@ -1007,6 +988,8 @@ static __global__ void flash_attn_tile(
             return;
         }
 
+        const float scale = gridDim.y == 1 ? 1.0f/KQ_sum[jc0] : 1.0f;
+
         const int j_dst_unrolled = ((sequence*ne01 + col_Q_0 + j)*ne02 + head0 + c)*gridDim.y + blockIdx.y;
 
 #ifdef FAST_FP16_AVAILABLE
@@ -1017,6 +1000,8 @@ static __global__ void flash_attn_tile(
 #pragma unroll
             for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
                 tmp[i1] = __half22float2(VKQ[jc0*((DVp/2)/warp_size) + i0/warp_size + i1]);
+                tmp[i1].x *= scale;
+                tmp[i1].y *= scale;
             }
             if (i0 + warp_size*cpy_ne_D <= DV/2 || i0 + threadIdx.x*cpy_ne_D < DV/2) {
                 ggml_cuda_memcpy_1<sizeof(tmp)>(&dst[j_dst_unrolled*DV + 2*i0 + threadIdx.x*(2*cpy_ne_D)], tmp);
@@ -1027,6 +1012,11 @@ static __global__ void flash_attn_tile(
 #pragma unroll
         for (int i0 = 0; i0 < DVp; i0 += warp_size*cpy_ne_D) {
             if (i0 + warp_size*cpy_ne_D <= DV || i0 + threadIdx.x*cpy_ne_D < DV) {
+#pragma unroll
+                for (int i1 = 0; i1 < cpy_ne_D/2; ++i1) {
+                    VKQ[jc0*((DVp/2)/warp_size) + i0/(2*warp_size) + i1].x *= scale;
+                    VKQ[jc0*((DVp/2)/warp_size) + i0/(2*warp_size) + i1].y *= scale;
+                }
                 ggml_cuda_memcpy_1<cpy_ne_D*4>(
                     &dst[j_dst_unrolled*DV + i0 + threadIdx.x*cpy_ne_D],
                     &VKQ[jc0*((DVp/2)/warp_size) + i0/(2*warp_size)]);

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

@@ -516,8 +516,8 @@ void ggml_cuda_flash_attn_ext_vec_case_impl(ggml_backend_cuda_context & ctx, ggm
     const int nthreads = ggml_cuda_fattn_vec_get_nthreads_host(cc);
     const int nwarps   = nthreads / WARP_SIZE;
     fattn_kernel_t fattn_kernel = flash_attn_ext_vec<D, cols_per_block, type_K, type_V, use_logit_softcap>;
-    constexpr bool need_f16_K = false;
-    constexpr bool need_f16_V = false;
+    const bool need_f16_K = type_K == GGML_TYPE_F16;
+    const bool need_f16_V = type_V == GGML_TYPE_F16;
     constexpr size_t nbytes_shared = 0;
     launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
@@ -526,11 +526,6 @@ template <int D, ggml_type type_K, ggml_type type_V>
 void ggml_cuda_flash_attn_ext_vec_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * KQV = dst;
     const ggml_tensor * Q   = dst->src[0];
-    const ggml_tensor * K   = dst->src[1];
-    const ggml_tensor * V   = dst->src[2];
-
-    GGML_ASSERT(K->type == type_K);
-    GGML_ASSERT(V->type == type_V);
 
     float logit_softcap;
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));

ggml/src/ggml-cuda/fattn.cu

@@ -116,11 +116,15 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
     }
 }
 
-#define FATTN_VEC_CASE(D, type_K, type_V)                                \
-    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) { \
-        ggml_cuda_flash_attn_ext_vec_case<D, type_K, type_V>(ctx, dst);  \
-        return;                                                          \
-    }                                                                    \
+#define FATTN_VEC_CASE(D, type_K, type_V)                                                                        \
+    {                                                                                                            \
+        const bool type_K_okay = K->type == (type_K) || (K->type == GGML_TYPE_F32 && (type_K) == GGML_TYPE_F16); \
+        const bool type_V_okay = V->type == (type_V) || (V->type == GGML_TYPE_F32 && (type_V) == GGML_TYPE_F16); \
+        if (Q->ne[0] == (D) && type_K_okay && type_V_okay) {                                                     \
+            ggml_cuda_flash_attn_ext_vec_case<D, type_K, type_V>(ctx, dst);                                      \
+            return;                                                                                              \
+        }                                                                                                        \
+    }                                                                                                            \
 
 #define FATTN_VEC_CASES_ALL_D(type_K, type_V) \
     FATTN_VEC_CASE( 64, type_K, type_V)       \
@@ -247,6 +251,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 #endif // GGML_CUDA_FA_ALL_QUANTS
 
     switch (K->type) {
+        case GGML_TYPE_F32:
         case GGML_TYPE_F16:
             break;
         case GGML_TYPE_Q4_1:
@@ -272,7 +277,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     // If Turing tensor cores available, use them:
     if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40) {
         if (can_use_vector_kernel) {
-            if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+            if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
                     return BEST_FATTN_KERNEL_VEC;
                 }
@@ -305,7 +310,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 
     // If there are no tensor cores available, use the generic tile kernel:
     if (can_use_vector_kernel) {
-        if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+        if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
             if (Q->ne[1] == 1) {
                 if (!gqa_opt_applies) {
                     return BEST_FATTN_KERNEL_VEC;

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

@@ -50,6 +50,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
 #include "ggml.h"
@@ -273,6 +274,15 @@ static ggml_cuda_device_info ggml_cuda_init() {
         } else if (device_name.substr(0, 21) == "NVIDIA GeForce GTX 16") {
             turing_devices_without_mma.push_back({ id, device_name });
         }
+
+        // Temporary performance fix:
+        // Setting device scheduling strategy for iGPUs with cc121 to "spinning" to avoid delays in cuda synchronize calls.
+        // TODO: Check for future drivers the default scheduling strategy and
+        // remove this call again when cudaDeviceScheduleSpin is default.
+        if (prop.major == 12 && prop.minor == 1) {
+            CUDA_CHECK(cudaSetDeviceFlags(cudaDeviceScheduleSpin));
+        }
+
 #endif  // defined(GGML_USE_HIP)
     }
 
@@ -1948,8 +1958,15 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
 
         size_t src1_stride_size = sizeof(cuda_t);
 
-        dim3 block_dims(ne13, ne12);
-        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
+        const int threads_x = 16;
+        const int threads_y = 16;
+        dim3 block_dims(threads_x, threads_y);
+
+        dim3 grid_dims(
+            (ne13 + threads_x - 1) / threads_x,
+            (ne12 + threads_y - 1) / threads_y
+        );
+        k_compute_batched_ptrs<<<grid_dims, block_dims, 0, main_stream>>>(
                 src0_ptr, src1_ptr, dst_t,
                 ptrs_src.get(), ptrs_dst.get(),
                 ne12, ne13,
@@ -1998,6 +2015,147 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     }
 }
 
+static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
+                                          const ggml_tensor * ffn_gate,
+                                          const ggml_tensor * glu,
+                                          const ggml_tensor * ffn_up_bias = nullptr,
+                                          const ggml_tensor * ffn_gate_bias = nullptr) {
+    const bool has_bias = ffn_up_bias != nullptr || ffn_gate_bias != nullptr;
+
+    if (has_bias && (!ffn_up_bias || !ffn_gate_bias)) {
+        return false;
+    }
+
+    const bool is_mul_mat     = ffn_up->op == GGML_OP_MUL_MAT     && ffn_gate->op == GGML_OP_MUL_MAT     && glu->op == GGML_OP_GLU;
+    const bool is_mul_mat_id  = ffn_up->op == GGML_OP_MUL_MAT_ID  && ffn_gate->op == GGML_OP_MUL_MAT_ID  && glu->op == GGML_OP_GLU;
+
+    GGML_ASSERT(ffn_up && ffn_gate && glu);
+
+    if (!is_mul_mat && !is_mul_mat_id) {
+        return false;
+    }
+
+    const ggml_op expected_bias_op = is_mul_mat ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+    if (has_bias) {
+        if (ffn_up_bias->op != expected_bias_op || ffn_gate_bias->op != expected_bias_op) {
+            return false;
+        }
+
+        if (glu->src[0] != ffn_gate_bias || glu->src[1] != ffn_up_bias) {
+            return false;
+        }
+
+        if (expected_bias_op == GGML_OP_ADD) {
+            const bool up_has_mul   = ffn_up_bias->src[0] == ffn_up || ffn_up_bias->src[1] == ffn_up;
+            const bool gate_has_mul = ffn_gate_bias->src[0] == ffn_gate || ffn_gate_bias->src[1] == ffn_gate;
+            if (!up_has_mul || !gate_has_mul) {
+                return false;
+            }
+        } else { // GGML_OP_ADD_ID
+            if (ffn_up_bias->src[0] != ffn_up || ffn_gate_bias->src[0] != ffn_gate) {
+                return false;
+            }
+            if (ffn_up_bias->src[2] != ffn_up->src[2] || ffn_gate_bias->src[2] != ffn_gate->src[2]) {
+                return false;
+            }
+        }
+    } else {
+        if (glu->src[0] != ffn_gate && glu->src[1] != ffn_up) {
+            return false;
+        }
+    }
+
+    if (ffn_up->src[0]->type != ffn_gate->src[0]->type || !ggml_are_same_shape(ffn_up->src[0], ffn_gate->src[0]) ||
+        !ggml_are_same_stride(ffn_up->src[0], ffn_gate->src[0])) {
+        return false;
+    }
+
+    if (ffn_up->src[1] != ffn_gate->src[1]) {
+        return false;
+    }
+
+    if (ffn_up->src[2] && (ffn_up->src[2] != ffn_gate->src[2])) {
+        return false;
+    }
+
+    static constexpr std::array<ggml_glu_op, 3> valid_glu_ops = { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_OAI };
+
+    if (std::find(valid_glu_ops.begin(), valid_glu_ops.end(), ggml_get_glu_op(glu)) == valid_glu_ops.end()) {
+        return false;
+    }
+
+    if (const bool swapped = ggml_get_op_params_i32(glu, 1); swapped) {
+        return false;
+    }
+
+    const bool split = ggml_backend_buft_is_cuda_split(ffn_up->src[0]->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(ffn_gate->src[0]->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    return true;
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
+    ggml_tensor *       src0 = tensor->src[0];
+    ggml_tensor *       src1 = tensor->src[1];
+    const ggml_tensor * dst  = tensor;
+
+    const bool is_mul_mat_id = tensor->op == GGML_OP_MUL_MAT_ID;
+
+    bool use_mul_mat_vec_f =
+        (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16) &&
+        src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    const int cc      = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
+
+    //we only support fusion for ncols_dst = 1
+    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
+        return false;
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
+        return false;
+    }
+
+
+    return use_mul_mat_vec_f;
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
+    ggml_tensor *       src0 = tensor->src[0];
+    ggml_tensor *       src1 = tensor->src[1];
+    const ggml_tensor * dst  = tensor;
+
+    const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE &&
+                                   ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) &&
+                                   src0->view_src;
+
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 &&
+                             dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+
+    // fusion is not universally faster on Pascal
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    if (cc <= GGML_CUDA_CC_PASCAL) {
+        return false;
+    }
+    //we only support fusion for ncols_dst = 1
+    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
+        return false;
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
+        return false;
+    }
+
+    return use_mul_mat_vec_q;
+}
+
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
@@ -2259,6 +2417,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SET_ROWS:
             ggml_cuda_op_set_rows(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_cuda_op_set(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2633,11 +2794,10 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
 }
 
 #ifdef USE_CUDA_GRAPH
-static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
+static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
     bool use_cuda_graph) {
 
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
-    cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
 
     const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
     const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
@@ -2688,33 +2848,11 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_CPY) {
-
-            // Store the pointers which are updated for each token, such that these can be sent
-            // to the device and accessed using indirection from CUDA graph
-            cuda_ctx->cuda_graph->cpy_dest_ptrs.push_back((char *) node->src[1]->data);
-
-            // store a pointer to each copy op CUDA kernel to identify it later
-            void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
-            if (!ptr) {
-                use_cuda_graph = false;
-#ifndef NDEBUG
-                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
-#endif
-            }
-        }
-
         if (!use_cuda_graph) {
             break;
         }
     }
 
-    if (use_cuda_graph) {
-        cuda_ctx->cuda_graph->use_cpy_indirection = true;
-        // copy pointers to GPU so they can be accessed via indirection within CUDA graph
-        ggml_cuda_cpy_dest_ptrs_copy(cuda_ctx->cuda_graph.get(), cuda_ctx->cuda_graph->cpy_dest_ptrs.data(), cuda_ctx->cuda_graph->cpy_dest_ptrs.size(), cuda_ctx->stream());
-    }
-
     return use_cuda_graph;
 }
 
@@ -2733,7 +2871,6 @@ static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_p
 
 static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
     if (node->data != graph_node_properties->node_address &&
-          node->op != GGML_OP_CPY &&
           node->op != GGML_OP_VIEW) {
         return false;
     }
@@ -2754,14 +2891,13 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (node->src[i] &&
             node->src[i]->data != graph_node_properties->src_address[i] &&
-            node->op != GGML_OP_CPY &&
             node->op != GGML_OP_VIEW
         ) {
             return false;
         }
     }
 
-    if (node->op == GGML_OP_SCALE &&
+    if ((node->op == GGML_OP_SCALE || node->op == GGML_OP_GLU) &&
         memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
         return false;
     }
@@ -2834,43 +2970,74 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
 #endif
 
     //TODO: remove special case once ggml_can_fuse can handle empty nodes
-    std::initializer_list<enum ggml_op> topk_moe_ops           = ggml_cuda_topk_moe_ops(false);
-    std::initializer_list<enum ggml_op> topk_moe_ops_with_norm = ggml_cuda_topk_moe_ops(true);
+    std::initializer_list<enum ggml_op> topk_moe_ops =
+        ggml_cuda_topk_moe_ops(/*with_norm*/ false, /*delayed_softmax=*/false);
+    std::initializer_list<enum ggml_op> topk_moe_ops_with_norm =
+        ggml_cuda_topk_moe_ops(/*with_norm=*/true, /*delayed_softmax=*/false);
+    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() && std::equal(ops.begin(), ops.end(), topk_moe_ops_with_norm.begin())) {
-
-        if (node_idx + topk_moe_ops_with_norm.size() > (size_t)cgraph->n_nodes) {
-            return false;
-        }
-
-        for (size_t i = 0; i < topk_moe_ops_with_norm.size(); i++) {
-            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops_with_norm.begin()[i]) return false;
-        }
+    if (ops.size() == topk_moe_ops_with_norm.size() &&
+        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+8];
+        ggml_tensor * weights = cgraph->nodes[node_idx + 9];
 
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
             return true;
         }
     }
 
-    if (ops.size() == topk_moe_ops.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops.begin())) {
-
-        if (node_idx + topk_moe_ops.size() > (size_t)cgraph->n_nodes) {
-            return false;
-        }
-
-        for (size_t i = 0; i < topk_moe_ops.size(); i++) {
-            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops.begin()[i]) return false;
-        }
-
+    if (ops.size() == topk_moe_ops.size() &&
+        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];
+        ggml_tensor * weights = cgraph->nodes[node_idx + 4];
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
             return true;
         }
     }
 
+    if (ops.size() == topk_moe_ops_delayed_softmax.size() &&
+        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];
+
+        if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+            return true;
+        }
+    }
+
+    std::initializer_list<enum ggml_op> mul_mat_bias_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_GLU };
+    std::initializer_list<enum ggml_op> mul_mat_id_bias_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_GLU };
+
+    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}))) {
+
+        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];
+        const ggml_tensor * ffn_up_bias   = cgraph->nodes[node_idx + 3];
+        const ggml_tensor * glu           = cgraph->nodes[node_idx + 4];
+
+        if (ggml_cuda_should_fuse_mul_mat(ffn_up, ffn_gate, glu, ffn_up_bias, ffn_gate_bias)) {
+            return true;
+        }
+    }
+
+    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}))) {
+
+        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];
+
+        if (ggml_cuda_should_fuse_mul_mat(ffn_up, ffn_gate, glu)) {
+            return true;
+        }
+    }
+
     if (!ggml_can_fuse(cgraph, node_idx, ops)) {
         return false;
     }
@@ -2901,7 +3068,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
 
         //if rms norm is the B operand, then we don't handle broadcast
-        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm->src[1])) {
+        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm)) {
             return false;
         }
 
@@ -2951,9 +3118,20 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
         // With the use of CUDA graphs, the execution will be performed by the graph launch.
         if (!use_cuda_graph || cuda_graph_update_required) {
 
+            [[maybe_unused]] int prev_i = 0;
+
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
 
+
+#ifdef GGML_CUDA_DEBUG
+                const int nodes_fused = i - prev_i - 1;
+                prev_i = i;
+                if (nodes_fused > 0) {
+                    GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
+                }
+#endif
+
                 if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
                     continue;
                 }
@@ -2962,21 +3140,35 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                 if (!disable_fusion) {
 
                     if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ true), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i+8];
-                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node, weights, selected_experts, /*with norm*/ true);
-                        i += 8;
+                        ggml_tensor * weights          = cgraph->nodes[i + 9];
+                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
+                        ggml_tensor * clamp            = cgraph->nodes[i + 7];
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ true,
+                                              /*delayed softmax*/ false, clamp);
+                        i += 9;
                         continue;
                     }
 
                     if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ false), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i+4];
-                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node, weights, selected_experts, /*with norm*/ false);
+                        ggml_tensor * weights          = cgraph->nodes[i + 4];
+                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ false,
+                                              /*delayed softmax*/ false);
                         i += 4;
                         continue;
                     }
 
+                    if (ggml_cuda_can_fuse(cgraph, i,
+                                           ggml_cuda_topk_moe_ops(/*with norm*/ false, /*delayed softmax*/ true), {})) {
+                        ggml_tensor * weights = cgraph->nodes[i + 5];
+                        ggml_tensor * ids     = cgraph->nodes[i + 1];
+
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, ids, /*with norm*/ false,
+                                              /*delayed_softmax*/ true);
+                        i += 5;
+                        continue;
+                    }
+
                     if (node->op == GGML_OP_ADD) {
                         int n_fuse = 0;
                         ggml_op ops[8];
@@ -3008,6 +3200,184 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         }
                     }
 
+                    bool fused_mul_mat_vec = false;
+                    int fused_node_count = 0;
+
+                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+                        if (ggml_cuda_can_fuse(cgraph, i, { op, bias_op, op, bias_op, GGML_OP_GLU }, {})) {
+                            ggml_tensor * glu         = cgraph->nodes[i + 4];
+                            ggml_tensor * gate_bias_n = glu->src[0];
+                            ggml_tensor * up_bias_n   = glu->src[1];
+
+                            //we don't assume the order for {gate, up}. Instead infer it from the bias tensor
+                            ggml_tensor * gate_n      = nullptr;
+                            ggml_tensor * up_n        = nullptr;
+
+                            if (gate_bias_n->src[0] == cgraph->nodes[i] || gate_bias_n->src[1] == cgraph->nodes[i]) {
+                                gate_n = cgraph->nodes[i];
+                                up_n   = cgraph->nodes[i + 2];
+                            } else if (gate_bias_n->src[0] == cgraph->nodes[i + 2] || gate_bias_n->src[1] == cgraph->nodes[i + 2]) {
+                                gate_n = cgraph->nodes[i + 2];
+                                up_n   = cgraph->nodes[i];
+                            } else {
+                                continue;
+                            }
+
+                            auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) {
+                                if (op_bias == GGML_OP_ADD) {
+                                    if (bias_node->src[0] == mul_node) {
+                                        return bias_node->src[1];
+                                    }
+                                    if (bias_node->src[1] == mul_node) {
+                                        return bias_node->src[0];
+                                    }
+                                    return (ggml_tensor *) nullptr;
+                                }
+                                GGML_ASSERT(op_bias == GGML_OP_ADD_ID);
+                                GGML_ASSERT(bias_node->src[0] == mul_node);
+                                return bias_node->src[1];
+                            };
+
+                            ggml_tensor * up_bias_tensor   = get_bias_tensor(up_bias_n, up_n, bias_op);
+                            ggml_tensor * gate_bias_tensor = get_bias_tensor(gate_bias_n, gate_n, bias_op);
+
+                            if (!up_bias_tensor || !gate_bias_tensor) {
+                                continue;
+                            }
+
+                            const ggml_tensor * src0 = up_n->src[0];
+                            const ggml_tensor * src1 = up_n->src[1];
+                            const ggml_tensor * ids  = up_n->src[2];
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate      = gate_n->src[0];
+                                fusion_data.x_bias    = up_bias_tensor;
+                                fusion_data.gate_bias = gate_bias_tensor;
+                                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 5;
+                                break;
+                            }
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate      = gate_n->src[0];
+                                fusion_data.x_bias    = up_bias_tensor;
+                                fusion_data.gate_bias = gate_bias_tensor;
+                                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 5;
+                                break;
+                            }
+                        } else if (ggml_cuda_can_fuse(cgraph, i, { op, op, GGML_OP_GLU }, {})) {
+                            ggml_tensor * glu  = cgraph->nodes[i + 2];
+                            ggml_tensor * gate = glu->src[0];
+                            ggml_tensor * up   = glu->src[1];
+
+                            bool ok = (gate == cgraph->nodes[i] && up == cgraph->nodes[i + 1])
+                                || (gate == cgraph->nodes[i + 1] && up == cgraph->nodes[i]);
+
+                            if (!ok) continue;
+
+                            const ggml_tensor * src0 = up->src[0];
+                            const ggml_tensor * src1 = up->src[1];
+                            const ggml_tensor * ids  = up->src[2];
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate   = gate->src[0];
+                                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 3;
+                                break;
+                            }
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate   = gate->src[0];
+                                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 3;
+                                break;
+                            }
+                        }
+                    }
+
+                    if (fused_mul_mat_vec) {
+                        i += fused_node_count - 1;
+                        continue;
+                    }
+
+                    fused_mul_mat_vec = false;
+                    fused_node_count = 0;
+
+                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+                        if (!ggml_can_fuse(cgraph, i, { op, bias_op })) {
+                            continue;
+                        }
+
+                        ggml_tensor * mm_node   = cgraph->nodes[i];
+                        ggml_tensor * bias_node = cgraph->nodes[i + 1];
+
+                        ggml_tensor * bias_tensor = nullptr;
+                        if (bias_op == GGML_OP_ADD) {
+                            if (bias_node->src[0] == mm_node) {
+                                bias_tensor = bias_node->src[1];
+                            } else if (bias_node->src[1] == mm_node) {
+                                bias_tensor = bias_node->src[0];
+                            } else {
+                                continue;
+                            }
+                        } else {
+                            if (bias_node->src[0] != mm_node) {
+                                continue;
+                            }
+                            bias_tensor = bias_node->src[1];
+                        }
+
+                        const ggml_tensor * src0 = mm_node->src[0];
+                        const ggml_tensor * src1 = mm_node->src[1];
+                        const ggml_tensor * ids  = mm_node->src[2];
+
+                        if (bias_op == GGML_OP_ADD_ID && bias_node->src[2] != ids) {
+                            continue;
+                        }
+
+                        ggml_cuda_mm_fusion_args_host fusion_data{};
+                        fusion_data.x_bias = bias_tensor;
+
+                        if (ggml_cuda_should_fuse_mul_mat_vec_f(mm_node)) {
+                            ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+                            fused_mul_mat_vec = true;
+                            fused_node_count = 2;
+                            break;
+                        }
+
+                        if (ggml_cuda_should_fuse_mul_mat_vec_q(mm_node)) {
+                            ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+                            fused_mul_mat_vec = true;
+                            fused_node_count = 2;
+                            break;
+                        }
+                    }
+
+                    if (fused_mul_mat_vec) {
+                        i += fused_node_count - 1;
+                        continue;
+                    }
 
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ADD}, {})) {
                         ggml_cuda_op_rms_norm_fused_add(*cuda_ctx, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
@@ -3120,7 +3490,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
     if (use_cuda_graph) {
         cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
 
-        use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph, use_cuda_graph);
+        use_cuda_graph = check_node_graph_compatibility(cgraph, use_cuda_graph);
 
         // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
         if (use_cuda_graph && cuda_graph_update_required) {
@@ -3147,10 +3517,6 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
-    if (!use_cuda_graph) {
-        cuda_ctx->cuda_graph->use_cpy_indirection = false;
-    }
-
 #else
     bool use_cuda_graph = false;
     bool cuda_graph_update_required = false;
@@ -3491,6 +3857,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                        op->src[0]->type == GGML_TYPE_F32 &&
                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
             } break;
+        case GGML_OP_SET:
+            {
+                const ggml_type t = op->type;
+                return (t == GGML_TYPE_F32 || t == GGML_TYPE_I32) &&
+                    t == op->src[0]->type &&
+                    t == op->src[1]->type;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -3645,12 +4018,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
-        case GGML_OP_SUM:
         case GGML_OP_ACC:
             return true;
+        case GGML_OP_SUM:
+            return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_ARGSORT:
-            // TODO: Support arbitrary column width
+#ifndef GGML_CUDA_USE_CUB
             return op->src[0]->ne[0] <= 1024;
+#else
+            return true;
+#endif
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:

ggml/src/ggml-cuda/mmf.cu

@@ -1,5 +1,7 @@
 #include "ggml.h"
 #include "mmf.cuh"
+#include "mmid.cuh"
+
 
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
@@ -37,6 +39,12 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
     const int64_t ids_s0 = ids ? ids->nb[0] / ggml_type_size(ids->type) : 0;
     const int64_t ids_s1 = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
 
+    mmf_ids_data ids_info{};
+    mmf_ids_data * ids_info_ptr = nullptr;
+    ggml_cuda_pool_alloc<int32_t> ids_src_compact_dev;
+    ggml_cuda_pool_alloc<int32_t> ids_dst_compact_dev;
+    ggml_cuda_pool_alloc<int32_t> expert_bounds_dev;
+
     // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
     const int64_t ncols_dst          = ids ? ne2  : ne1;
     const int64_t nchannels_dst      = ids ? ne1 : ne2;
@@ -54,6 +62,33 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
         nchannels_y      = ids->ne[0];
     }
 
+    if (ids && ncols_dst > 16) {
+        const int64_t n_expert_used = ids->ne[0];
+        const int64_t n_experts     = ne02;
+        const int64_t n_tokens      = ne12;
+        const int64_t ne_get_rows   = n_tokens * n_expert_used;
+
+        ids_src_compact_dev.alloc(ctx.pool(), ne_get_rows);
+        ids_dst_compact_dev.alloc(ctx.pool(), ne_get_rows);
+        expert_bounds_dev.alloc(ctx.pool(), n_experts + 1);
+
+        const int si1  = static_cast<int>(ids_s1);
+        const int sis1 = static_cast<int>(src1->nb[2] / src1->nb[1]);
+
+        GGML_ASSERT(sis1 > 0);
+
+        ggml_cuda_launch_mm_ids_helper(ids_d, ids_src_compact_dev.get(), ids_dst_compact_dev.get(), expert_bounds_dev.get(),
+            static_cast<int>(n_experts), static_cast<int>(n_tokens), static_cast<int>(n_expert_used), static_cast<int>(ne11), si1, sis1, ctx.stream());
+        CUDA_CHECK(cudaGetLastError());
+
+        ids_info.ids_src_compact   = ids_src_compact_dev.get();
+        ids_info.ids_dst_compact   = ids_dst_compact_dev.get();
+        ids_info.expert_bounds_dev = expert_bounds_dev.get();
+        ids_info.n_experts         = static_cast<int>(n_experts);
+        ids_info.sis1              = sis1;
+        ids_info_ptr = &ids_info;
+    }
+
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
@@ -61,7 +96,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         case GGML_TYPE_F16: {
             const half2 * src0_d = (const half2 *) src0->data;
@@ -69,7 +104,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat162 * src0_d = (const nv_bfloat162 *) src0->data;
@@ -77,7 +112,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -98,10 +133,9 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
     }
 
     if (mul_mat_id) {
-        if (type == GGML_TYPE_F32 && src1_ncols > 32) {
+        if (src0_ne[1] <= 1024 && src1_ncols > 512) {
             return false;
-        }
-        if ((type == GGML_TYPE_F16 || type == GGML_TYPE_BF16) && src1_ncols > 64) {
+        } else if(src0_ne[1] > 1024 && src1_ncols > 128) {
             return false;
         }
     } else {

ggml/src/ggml-cuda/mmf.cuh

@@ -7,6 +7,14 @@ using namespace ggml_cuda_mma;
 
 #define MMF_ROWS_PER_BLOCK 32
 
+struct mmf_ids_data {
+    const int32_t * ids_src_compact = nullptr;
+    const int32_t * ids_dst_compact = nullptr;
+    const int32_t * expert_bounds_dev = nullptr;
+    int n_experts = 0;
+    int sis1 = 0;
+};
+
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
 bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const int src1_ncols, bool mul_mat_id);
@@ -224,6 +232,250 @@ static __global__ void mul_mat_f(
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
+
+//This kernel is for larger batch sizes of mul_mat_id
+template <typename T, int rows_per_block, int cols_per_block, int nwarps>
+__launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
+static __global__ void mul_mat_f_ids(
+        const T * __restrict__ x, const float * __restrict__ y,
+        const int32_t * __restrict__ ids_src_compact, const int32_t * __restrict__ ids_dst_compact,
+        const int32_t * __restrict__ expert_bounds, float * __restrict__ dst,
+        const int ncols, const int ncols_dst_total, const int nchannels_dst, const int stride_row, const int stride_col_y, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const uint3 sis1_fd, const uint3 nch_fd) {
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    typedef tile<16, 8, T>     tile_A;
+    typedef tile< 8, 8, T>     tile_B;
+    typedef tile<16, 8, float> tile_C;
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    constexpr int tile_k_padded = warp_size + 4;
+    constexpr int ntA = rows_per_block / tile_A::I;
+    constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;
+
+    const int row0        = blockIdx.x * rows_per_block;
+
+    const int expert_idx = blockIdx.y;
+    const int expert_start = expert_bounds[expert_idx];
+    const int expert_end   = expert_bounds[expert_idx + 1];
+    const int ncols_expert = expert_end - expert_start;
+
+    const int tiles_for_expert = (ncols_expert + cols_per_block - 1) / cols_per_block;
+    const int tile_idx = blockIdx.z;
+    if (tile_idx >= tiles_for_expert) {
+        return;
+    }
+
+    const int col_base = tile_idx * cols_per_block;
+
+    GGML_UNUSED(channel_ratio);
+
+    const int channel_x   = expert_idx;
+    const int sample_dst  = 0;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
+
+    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x  + row0*stride_row;
+    y   += int64_t(sample_y)  *stride_sample_y;
+    dst += int64_t(sample_dst)*stride_sample_dst;
+
+    const int32_t * ids_src_expert = ids_src_compact + expert_start;
+    const int32_t * ids_dst_expert = ids_dst_compact + expert_start;
+
+    extern __shared__ char data_mmv[];
+    char * compute_base = data_mmv;
+
+    //const float2 * y2 = (const float2 *) y;
+
+    tile_C C[ntA][ntB];
+
+    T * tile_xy = (T *) compute_base + threadIdx.y*(tile_A::I * tile_k_padded);
+
+    for (int col = threadIdx.y*warp_size + threadIdx.x; col < ncols; col += nwarps*warp_size) {
+        tile_A A[ntA][warp_size / tile_A::J];
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int i = 0; i < tile_A::I; ++i) {
+                tile_xy[i*tile_k_padded + threadIdx.x] = x[(itA*tile_A::I + i)*stride_row  + col];
+            }
+#pragma unroll
+            for (int k0 = 0; k0 < warp_size; k0 += tile_A::J) {
+                load_ldmatrix(A[itA][k0/tile_A::J], tile_xy + k0, tile_k_padded);
+            }
+        }
+
+        if constexpr (std::is_same_v<T, float>) {
+            float vals_buf[2][tile_B::I];
+            auto gather_tile = [&](int tile_idx_local, float *vals) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + tile_idx_local*tile_B::I;
+                    const int global_j = col_base + j;
+                    float val = 0.0f;
+                    if (j < cols_per_block && global_j < ncols_expert) {
+                        const int src_entry = ids_src_expert[global_j];
+                        const uint2 qrm = fast_div_modulo((uint32_t) src_entry, sis1_fd);
+                        const int token   = (int) qrm.x;
+                        const int channel = (int) qrm.y;
+                        if (token < ncols_dst_total) {
+                            val = y[channel*stride_channel_y + token*stride_col_y + col];
+                        }
+                    }
+                    vals[j0] = val;
+                }
+            };
+
+            gather_tile(0, vals_buf[0]);
+
+            int curr_buf = 0;
+            int next_buf = 1;
+#pragma unroll
+            for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = vals_buf[curr_buf][j0];
+                }
+
+                if (itB + 1 < ntB) {
+                    gather_tile(itB + 1, vals_buf[next_buf]);
+                }
+
+#pragma unroll
+                for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
+                    tile_B B;
+                    load_ldmatrix(B, tile_xy + k0, tile_k_padded);
+#pragma unroll
+                    for (int itA = 0; itA < ntA; ++itA) {
+                        mma(C[itA][itB], A[itA][k0/tile_B::J], B);
+                    }
+                }
+
+                if (itB + 1 < ntB) {
+                    curr_buf ^= 1;
+                    next_buf ^= 1;
+                }
+            }
+        } else if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
+            float2 vals_buf[2][tile_B::I];
+            auto gather_tile = [&](int tile_idx_local, float2 *vals) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + tile_idx_local*tile_B::I;
+                    const int global_j = col_base + j;
+                    float2 tmp = make_float2(0.0f, 0.0f);
+                    if (j < cols_per_block && global_j < ncols_expert) {
+                        const int src_entry = ids_src_expert[global_j];
+                        const uint2 qrm = fast_div_modulo((uint32_t) src_entry, sis1_fd);
+                        const int token   = (int) qrm.x;
+                        const int channel = (int) qrm.y;
+                        if (token < ncols_dst_total) {
+                            tmp = *(const float2*) &y[channel*stride_channel_y + 2*(token*stride_col_y + col)];
+                        }
+                    }
+                    vals[j0] = tmp;
+                }
+            };
+
+            if (ntB > 0) {
+                gather_tile(0, vals_buf[0]);
+            }
+
+            int curr_buf = 0;
+            int next_buf = 1;
+#pragma unroll
+            for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const float2 tmp = vals_buf[curr_buf][j0];
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                }
+
+                if (itB + 1 < ntB) {
+                    gather_tile(itB + 1, vals_buf[next_buf]);
+                }
+
+#pragma unroll
+                for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
+                    tile_B B;
+                    load_ldmatrix(B, tile_xy + k0, tile_k_padded);
+#pragma unroll
+                    for (int itA = 0; itA < ntA; ++itA) {
+                        mma(C[itA][itB], A[itA][k0/tile_B::J], B);
+                    }
+                }
+
+                if (itB + 1 < ntB) {
+                    curr_buf ^= 1;
+                    next_buf ^= 1;
+                }
+            }
+        } else {
+            static_assert(std::is_same_v<T, void>, "unsupported type");
+        }
+    }
+
+    float * buf_iw = (float *) compute_base;
+    constexpr int kiw = nwarps*rows_per_block + 4;
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+#pragma unroll
+    for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int l = 0; l < tile_C::ne; ++l) {
+                const int i = threadIdx.y*rows_per_block + itA*tile_C::I + tile_C::get_i(l);
+                const int j = itB*tile_C::J + tile_C::get_j(l);
+                buf_iw[j*kiw + i] = C[itA][itB].x[l];
+            }
+        }
+    }
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < cols_per_block; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j0 + nwarps > cols_per_block && j >= cols_per_block) {
+            return;
+        }
+
+        float sum = 0.0f;
+        static_assert(rows_per_block == warp_size, "need loop/check");
+#pragma unroll
+        for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) {
+            const int i = i0 + threadIdx.x;
+
+            sum += buf_iw[j*kiw + i];
+        }
+
+        const int global_j = col_base + j;
+        if (j < cols_per_block && global_j < ncols_expert && nchannels_dst > 0) {
+            const int dst_entry = ids_dst_expert[global_j];
+            const uint2 qrm = fast_div_modulo((uint32_t) dst_entry, nch_fd);
+            const int token = (int) qrm.x;
+            if (token < ncols_dst_total) {
+                const int slot = (int) qrm.y;
+                dst[slot*stride_channel_dst + token*stride_col_dst + row0 + threadIdx.x] = sum;
+            }
+        }
+    }
+#else
+    GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, expert_bounds, dst,
+        ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, sis1_fd, nch_fd);
+    NO_DEVICE_CODE;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
 template<typename T, int cols_per_block, int nwarps>
 static inline void mul_mat_f_switch_ids(
         const T * x, const float * y, const int32_t * ids, float * dst,
@@ -232,13 +484,35 @@ static inline void mul_mat_f_switch_ids(
         const int64_t stride_col_id, const int64_t stride_row_id,
         const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
         const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream) {
-    if (ids) {
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream,
+        const mmf_ids_data * ids_data) {
+    const bool has_ids_data = ids_data && ids_data->ids_src_compact;
+
+    // Use the compact-ids kernel only for larger tiles; for small ncols_dst (< 16)
+    // we prefer the normal mul_mat_f path with has_ids=true.
+    if (has_ids_data && ncols_dst > 16) {
+        const int max_tiles = (int) ((ncols_dst + cols_per_block - 1) / cols_per_block);
+        if (max_tiles == 0) {
+            return;
+        }
+        dim3 block_nums_ids(block_nums.x, ids_data->n_experts, max_tiles);
+
+        const uint3 sis1_fd = ids_data->sis1 > 0 ? init_fastdiv_values((uint32_t) ids_data->sis1) : make_uint3(0, 0, 1);
+        const uint3 nch_fd  = init_fastdiv_values((uint32_t) nchannels_dst);
+
+        mul_mat_f_ids<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
+            (x, y, ids_data->ids_src_compact, ids_data->ids_dst_compact, ids_data->expert_bounds_dev, dst,
+            ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+            channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+            sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst,
+            sis1_fd, nch_fd);
+    } else if (ids) {
         const int64_t col_tiles = (ncols_dst + cols_per_block - 1) / cols_per_block;
         dim3 block_nums_ids = block_nums;
         block_nums_ids.y *= col_tiles;
+
         mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, true><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
-             (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+            (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
              stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
              sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     } else {
@@ -258,7 +532,7 @@ void mul_mat_f_cuda(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        cudaStream_t stream, const mmf_ids_data * ids_data) {
     typedef tile<16, 8, T>     tile_A;
     typedef tile< 8, 8, T>     tile_B;
 
@@ -290,7 +564,7 @@ void mul_mat_f_cuda(
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;
     const int nbytes_shared_total = nbytes_shared + nbytes_slotmap;
-    const int64_t grid_y = ids ? nchannels_x : nchannels_dst; // per expert when ids present
+    const int64_t grid_y = ids ? nchannels_x : nchannels_dst;
 
     const dim3 block_nums(nrows_x/rows_per_block, grid_y, nsamples_dst);
     const dim3 block_dims(warp_size, nwarps_best, 1);
@@ -300,49 +574,57 @@ void mul_mat_f_cuda(
             mul_mat_f_switch_ids<T, cols_per_block, 1>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 2: {
             mul_mat_f_switch_ids<T, cols_per_block, 2>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 3: {
             mul_mat_f_switch_ids<T, cols_per_block, 3>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 4: {
             mul_mat_f_switch_ids<T, cols_per_block, 4>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 5: {
             mul_mat_f_switch_ids<T, cols_per_block, 5>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 6: {
             mul_mat_f_switch_ids<T, cols_per_block, 6>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 7: {
             mul_mat_f_switch_ids<T, cols_per_block, 7>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 8: {
             mul_mat_f_switch_ids<T, cols_per_block, 8>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -361,7 +643,7 @@ static void mul_mat_f_switch_cols_per_block(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        cudaStream_t stream, const mmf_ids_data * ids_data) {
 
     const int ncols_case = (ids && ncols_dst > 16) ? 16 : ncols_dst;
 
@@ -371,82 +653,82 @@ static void mul_mat_f_switch_cols_per_block(
         case  1: {
             mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  2: {
             mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  3: {
             mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  4: {
             mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  5: {
             mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y,  stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y,  stride_sample_dst, stream, ids_data);
         } break;
         case  6: {
             mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  7: {
             mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  8: {
             mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  9: {
             mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 10: {
             mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 11: {
             mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 12: {
             mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 13: {
             mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 14: {
             mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 15: {
             mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 16: {
             mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -462,7 +744,7 @@ static void mul_mat_f_switch_cols_per_block(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, \
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,\
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, \
-        cudaStream_t stream);
+        cudaStream_t stream, const mmf_ids_data * ids_data);
 
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 #define DECL_MMF_CASE_EXTERN(ncols_dst) \

ggml/src/ggml-cuda/mmid.cu

@@ -0,0 +1,164 @@
+#include "common.cuh"
+#include "mmid.cuh"
+
+// To reduce shared memory use, store "it" and "iex_used" with 22/10 bits each.
+struct mm_ids_helper_store {
+    uint32_t data;
+
+    __device__ mm_ids_helper_store(const uint32_t it, const uint32_t iex_used) {
+        data = (it & 0x003FFFFF) | (iex_used << 22);
+    }
+
+    __device__ uint32_t it() const {
+        return data & 0x003FFFFF;
+    }
+
+    __device__ uint32_t iex_used() const {
+        return data >> 22;
+    }
+};
+static_assert(sizeof(mm_ids_helper_store) == 4, "unexpected size for mm_ids_helper_store");
+
+// Helper function for mul_mat_id, converts ids to a more convenient format.
+// ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
+// ids_dst describes the same mapping but for the dst tensor.
+// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
+template <int n_expert_used_template>
+__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
+    const int expert = blockIdx.x;
+
+    extern __shared__ char data_mm_ids_helper[];
+    mm_ids_helper_store * store = (mm_ids_helper_store *) data_mm_ids_helper;
+
+    int nex_prev   = 0; // Number of columns for experts with a lower index.
+    int it_compact = 0; // Running index for the compact slice of this expert.
+
+    if constexpr (n_expert_used_template == 0) {
+        // Generic implementation:
+        for (int it = 0; it < n_tokens; ++it) {
+            int iex_used = -1; // The index at which the expert is used, if any.
+            for (int iex = threadIdx.x; iex < n_expert_used; iex += warp_size) {
+                const int expert_used = ids[it*si1 + iex];
+                nex_prev += expert_used < expert;
+                if (expert_used == expert) {
+                    iex_used = iex;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact] = mm_ids_helper_store(it, iex_used);
+            }
+
+            if (warp_reduce_any<warp_size>(iex_used != -1)) {
+                it_compact++;
+            }
+        }
+    } else {
+        // Implementation optimized for specific numbers of experts used:
+        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
+        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
+        for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
+            const int it = it0 + threadIdx.x / neu_padded;
+
+            const int iex = threadIdx.x % neu_padded; // The index at which the expert is used, if any.
+            const int expert_used = (neu_padded == n_expert_used || iex < n_expert_used) && it < n_tokens ?
+                ids[it*si1 + iex] : INT_MAX;
+            const int iex_used = expert_used == expert ? iex : -1;
+            nex_prev += expert_used < expert;
+
+            // Whether the threads at this token position have used the expert:
+            const int it_compact_add_self = warp_reduce_any<neu_padded>(iex_used != -1);
+
+            // Do a scan over threads at lower token positions in warp to get the correct index for writing data:
+            int it_compact_add_lower = 0;
+#pragma unroll
+            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
+                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
+                if (threadIdx.x >= static_cast<unsigned int>(offset)) {
+                    it_compact_add_lower += tmp;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact + it_compact_add_lower] = mm_ids_helper_store(it, iex_used);
+            }
+
+            // The thread with the highest index in the warp always has the sum over the whole warp, use it to increment all threads:
+            it_compact += __shfl_sync(0xFFFFFFFF, it_compact_add_lower + it_compact_add_self, warp_size - 1, warp_size);
+        }
+    }
+    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
+
+    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
+        const mm_ids_helper_store store_it = store[itc];
+        const int it       = store_it.it();
+        const int iex_used = store_it.iex_used();
+        ids_src1[nex_prev + itc] = it*sis1          + iex_used % nchannels_y;
+        ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
+    }
+
+    if (threadIdx.x != 0) {
+        return;
+    }
+
+    expert_bounds[expert] = nex_prev;
+
+    if (expert < static_cast<int>(gridDim.x) - 1) {
+        return;
+    }
+
+    expert_bounds[gridDim.x] = nex_prev + it_compact;
+}
+
+template <int n_expert_used_template>
+static void launch_mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
+    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mm_ids_helper_store");
+    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mm_ids_helper_store");
+
+    const int id = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[id].warp_size;
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+    CUDA_SET_SHARED_MEMORY_LIMIT(mm_ids_helper<n_expert_used_template>, smpbo);
+
+    const dim3 num_blocks(n_experts, 1, 1);
+    const dim3 block_size(warp_size, 1, 1);
+    const size_t nbytes_shared = n_tokens*sizeof(mm_ids_helper_store);
+    GGML_ASSERT(nbytes_shared <= smpbo);
+    mm_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
+        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
+}
+
+void ggml_cuda_launch_mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_experts, const int n_tokens, const int n_expert_used, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
+    switch (n_expert_used) {
+        case  2:
+            launch_mm_ids_helper< 2>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  4:
+            launch_mm_ids_helper< 4>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  6:
+            launch_mm_ids_helper< 6>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  8:
+            launch_mm_ids_helper< 8>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case 16:
+            launch_mm_ids_helper<16>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case 32:
+            launch_mm_ids_helper<32>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        default:
+            launch_mm_ids_helper< 0>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+    }
+}

ggml/src/ggml-cuda/mmid.cuh

@@ -0,0 +1,5 @@
+#pragma once
+
+void ggml_cuda_launch_mm_ids_helper(
+        const int32_t * ids, int32_t * ids_src1, int32_t * ids_dst, int32_t * expert_bounds,
+        int n_experts, int n_tokens, int n_expert_used, int nchannels_y, int si1, int sis1, cudaStream_t stream);

ggml/src/ggml-cuda/mmq.cu

@@ -1,141 +1,6 @@
 #include "mmq.cuh"
 #include "quantize.cuh"
-
-#include <vector>
-
-// To reduce shared memory use, store "it" and "iex_used" with 22/10 bits each.
-struct mmq_ids_helper_store {
-    uint32_t data;
-
-    __device__ mmq_ids_helper_store(const uint32_t it, const uint32_t iex_used) {
-        data = (it & 0x003FFFFF) | (iex_used << 22);
-    }
-
-    __device__ uint32_t it() const {
-        return data & 0x003FFFFF;
-    }
-
-    __device__ uint32_t iex_used() const {
-        return data >> 22;
-    }
-};
-static_assert(sizeof(mmq_ids_helper_store) == 4, "unexpected size for mmq_ids_helper_store");
-
-// Helper function for mul_mat_id, converts ids to a more convenient format.
-// ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
-// ids_dst describes the same mapping but for the dst tensor.
-// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
-template <int n_expert_used_template>
-__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
-static __global__ void mmq_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
-    const int expert = blockIdx.x;
-
-    extern __shared__ char data_mmq_ids_helper[];
-    mmq_ids_helper_store * store = (mmq_ids_helper_store *) data_mmq_ids_helper;
-
-    int nex_prev   = 0; // Number of columns for experts with a lower index.
-    int it_compact = 0; // Running index for the compact slice of this expert.
-
-    if constexpr (n_expert_used_template == 0) {
-        // Generic implementation:
-        for (int it = 0; it < n_tokens; ++it) {
-            int iex_used = -1; // The index at which the expert is used, if any.
-            for (int iex = threadIdx.x; iex < n_expert_used; iex += warp_size) {
-                const int expert_used = ids[it*si1 + iex];
-                nex_prev += expert_used < expert;
-                if (expert_used == expert) {
-                    iex_used = iex;
-                }
-            }
-
-            if (iex_used != -1) {
-                store[it_compact] = mmq_ids_helper_store(it, iex_used);
-            }
-
-            if (warp_reduce_any<warp_size>(iex_used != -1)) {
-                it_compact++;
-            }
-        }
-    } else {
-        // Implementation optimized for specific numbers of experts used:
-        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
-        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
-        for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
-            const int it = it0 + threadIdx.x / neu_padded;
-
-            const int iex = threadIdx.x % neu_padded; // The index at which the expert is used, if any.
-            const int expert_used = (neu_padded == n_expert_used || iex < n_expert_used) && it < n_tokens ?
-                ids[it*si1 + iex] : INT_MAX;
-            const int iex_used = expert_used == expert ? iex : -1;
-            nex_prev += expert_used < expert;
-
-            // Whether the threads at this token position have used the expert:
-            const int it_compact_add_self = warp_reduce_any<neu_padded>(iex_used != -1);
-
-            // Do a scan over threads at lower token positions in warp to get the correct index for writing data:
-            int it_compact_add_lower = 0;
-#pragma unroll
-            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
-                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
-                if (threadIdx.x >= static_cast<unsigned int>(offset)) {
-                    it_compact_add_lower += tmp;
-                }
-            }
-
-            if (iex_used != -1) {
-                store[it_compact + it_compact_add_lower] = mmq_ids_helper_store(it, iex_used);
-            }
-
-            // The thread with the highest index in the warp always has the sum over the whole warp, use it to increment all threads:
-            it_compact += __shfl_sync(0xFFFFFFFF, it_compact_add_lower + it_compact_add_self, warp_size - 1, warp_size);
-        }
-    }
-    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
-
-    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
-        const mmq_ids_helper_store store_it = store[itc];
-        const int it       = store_it.it();
-        const int iex_used = store_it.iex_used();
-        ids_src1[nex_prev + itc] = it*sis1          + iex_used % nchannels_y;
-        ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
-    }
-
-    if (threadIdx.x != 0) {
-        return;
-    }
-
-    expert_bounds[expert] = nex_prev;
-
-    if (expert < static_cast<int>(gridDim.x) - 1) {
-        return;
-    }
-
-    expert_bounds[gridDim.x] = nex_prev + it_compact;
-}
-
-template <int n_expert_used_template>
-static void launch_mmq_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
-    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mmq_ids_helper_store");
-    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mmq_ids_helper_store");
-
-    const int id = ggml_cuda_get_device();
-    const int warp_size = ggml_cuda_info().devices[id].warp_size;
-    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
-    CUDA_SET_SHARED_MEMORY_LIMIT(mmq_ids_helper<n_expert_used_template>, smpbo);
-
-    const dim3 num_blocks(n_experts, 1, 1);
-    const dim3 block_size(warp_size, 1, 1);
-    const size_t nbytes_shared = n_tokens*sizeof(mmq_ids_helper_store);
-    GGML_ASSERT(nbytes_shared <= smpbo);
-    mmq_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
-        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
-}
+#include "mmid.cuh"
 
 static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
     switch (args.type_x) {
@@ -293,36 +158,8 @@ void ggml_cuda_mul_mat_q(
         const int si1  = ids->nb[1] / ggml_element_size(ids);
         const int sis1 = nb12 / nb11;
 
-        switch (n_expert_used) {
-            case  2:
-                launch_mmq_ids_helper< 2> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  4:
-                launch_mmq_ids_helper< 4> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  6:
-                launch_mmq_ids_helper< 6> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  8:
-                launch_mmq_ids_helper< 8> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case 16:
-                launch_mmq_ids_helper<16> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case 32:
-                launch_mmq_ids_helper<32> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            default:
-                launch_mmq_ids_helper< 0> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-        }
+        ggml_cuda_launch_mm_ids_helper((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+            ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
         CUDA_CHECK(cudaGetLastError());
     }
 

ggml/src/ggml-cuda/mmvf.cu

@@ -1,20 +1,21 @@
 #include "ggml.h"
 #include "common.cuh"
-#include "convert.cuh"
+#include "unary.cuh"
 #include "mmvf.cuh"
+#include "convert.cuh"
 
-template <typename T, typename type_acc, int ncols_dst, int block_size>
+template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false>
 static __global__ void mul_mat_vec_f(
-        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
         const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
-        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+        const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
     const int row         = blockIdx.x;
     const int channel_dst = blockIdx.y;
-    const int channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int channel_x   = ids ? ids[channel_dst]          : fastdiv((uint32_t) channel_dst, channel_ratio);
     const int channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
     const int sample_dst  = blockIdx.z;
-    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_x    = fastdiv((uint32_t) sample_dst, sample_ratio);
     const int sample_y    = sample_dst;
     const int tid         = threadIdx.x;
 
@@ -24,58 +25,164 @@ static __global__ void mul_mat_vec_f(
     y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
     dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
 
+    bool use_gate = false;
+    bool use_bias = false;
+    bool use_gate_bias = false;
+    ggml_glu_op glu_op = ggml_glu_op::GGML_GLU_OP_SWIGLU;
+    const T * gate_x = nullptr;
+    const float * x_bias = nullptr;
+    const float * gate_bias = nullptr;
+
+    if constexpr (has_fusion) {
+        use_gate = fusion.gate != nullptr;
+        use_bias = fusion.x_bias != nullptr;
+        use_gate_bias = fusion.gate_bias != nullptr;
+        glu_op = fusion.glu_op;
+
+        if (use_gate) {
+            gate_x = static_cast<const T *>(fusion.gate);
+        }
+        if (use_bias) {
+            x_bias = static_cast<const float *>(fusion.x_bias);
+        }
+        if (use_gate_bias) {
+            gate_bias = static_cast<const float *>(fusion.gate_bias);
+            use_gate_bias = use_gate;
+        } else {
+            use_gate_bias = false;
+        }
+    }
+
+    if (use_gate) {
+        gate_x += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    }
+    if constexpr (has_fusion) {
+        const int channel_bias = ids ? channel_x : channel_dst;
+        if (use_bias) {
+            x_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
+        }
+        if (use_gate_bias) {
+            gate_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
+        }
+    }
+
     const float2 * y2 = (const float2 *) y;
 
     extern __shared__ char data_mmv[];
     float * buf_iw = (float *) data_mmv;
+    float * buf_iw_gate = nullptr;
+    if constexpr (has_fusion) {
+        buf_iw_gate = (float *) (data_mmv + warp_size*sizeof(float));
+    }
 
     if (block_size > warp_size) {
         if (tid < warp_size) {
             buf_iw[tid] = 0.0f;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    buf_iw_gate[tid] = 0.0f;
+                }
+            }
         }
         __syncthreads();
     }
 
     float sumf[ncols_dst] = {0.0f};
+    float sumf_gate[ncols_dst];
+    if constexpr (has_fusion) {
+#pragma unroll
+        for (int j = 0; j < ncols_dst; ++j) {
+            sumf_gate[j] = 0.0f;
+        }
+    }
 
     if constexpr (std::is_same_v<T, float>) {
         const float2 * x2 = (const float2 *) x;
+        const float2 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const float2 *) gate_x;
+            }
+        }
 
         for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const float2 tmpx = x2[col2];
+            float2 tmpx_gate = make_float2(0.0f, 0.0f);
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
 
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
-                sumf[j] += tmpx.x*tmpy.x;
-                sumf[j] += tmpx.y*tmpy.y;
+                ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
+                ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                    }
+                }
             }
         }
     } else if constexpr (std::is_same_v<T, half>) {
         const half2 * x2 = (const half2 *) x;
+        const half2 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const half2 *) gate_x;
+            }
+        }
 
         if (std::is_same_v<type_acc, float>) {
             for (int col2 = tid; col2 < ncols2; col2 += block_size) {
                 const float2 tmpx = __half22float2(x2[col2]);
-
+                float2 tmpx_gate = make_float2(0.0f, 0.0f);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmpx_gate = __half22float2(gate_x2[col2]);
+                    }
+                }
 #pragma unroll
                 for (int j = 0; j < ncols_dst; ++j) {
                     const float2 tmpy = y2[j*stride_col_y2 + col2];
-                    sumf[j] += tmpx.x * tmpy.x;
-                    sumf[j] += tmpx.y * tmpy.y;
+                    ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
+                    ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                    if constexpr (has_fusion) {
+                        if (use_gate) {
+                            ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                            ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                        }
+                    }
                 }
             }
         } else {
 #ifdef FP16_AVAILABLE
             half2 sumh2[ncols_dst] = {{0.0f, 0.0f}};
+            half2 sumh2_gate[ncols_dst] = {{0.0f, 0.0f}};
 
             for (int col2 = tid; col2 < ncols2; col2 += block_size) {
                 const half2 tmpx = x2[col2];
-
+                half2 tmpx_gate = make_half2(0.0f, 0.0f);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmpx_gate = gate_x2[col2];
+                    }
+                }
 #pragma unroll
                 for (int j = 0; j < ncols_dst; ++j) {
                     const float2 tmpy = y2[j*stride_col_y2 + col2];
                     sumh2[j] += tmpx * make_half2(tmpy.x, tmpy.y);
+
+                    if constexpr (has_fusion) {
+                        if (use_gate) {
+                            sumh2_gate[j] += tmpx_gate * make_half2(tmpy.x, tmpy.y);
+                        }
+                    }
                 }
             }
 
@@ -83,21 +190,86 @@ static __global__ void mul_mat_vec_f(
             for (int j = 0; j < ncols_dst; ++j) {
                 sumf[j] = __low2float(sumh2[j]) + __high2float(sumh2[j]);
             }
+
+            if constexpr (has_fusion) {
+                if (use_gate) {
+#pragma unroll
+                    for (int j = 0; j < ncols_dst; ++j) {
+                        sumf_gate[j] = __low2float(sumh2_gate[j]) + __high2float(sumh2_gate[j]);
+                    }
+                }
+            }
 #else
             NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
         }
     } else if constexpr (std::is_same_v<T, nv_bfloat16>) {
+//TODO: add support for ggml_cuda_mad for hip_bfloat162
+#if defined(GGML_USE_HIP)
         const int * x2 = (const int *) x;
+        const int * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const int *) gate_x;
+            }
+        }
         for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const int tmpx = x2[col2];
+            int tmpx_gate = 0;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
-                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
-                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+                const float tmpx0 = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]);
+                const float tmpx1 = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]);
+                ggml_cuda_mad(sumf[j], tmpx0, tmpy.x);
+                ggml_cuda_mad(sumf[j], tmpx1, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        const float tmpx0_gate = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx_gate)[0]);
+                        const float tmpx1_gate = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx_gate)[1]);
+                        ggml_cuda_mad(sumf_gate[j], tmpx0_gate, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx1_gate, tmpy.y);
+                    }
+                }
             }
         }
+#else
+        const nv_bfloat162 * x2 = (const nv_bfloat162 *) x;
+        const nv_bfloat162 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const nv_bfloat162 *) gate_x;
+            }
+        }
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+            const nv_bfloat162 tmpx = x2[col2];
+            nv_bfloat162 tmpx_gate;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                const float2 tmpy = y2[j*stride_col_y2 + col2];
+                ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
+                ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                    }
+                }
+            }
+        }
+#endif
     } else {
         static_assert(std::is_same_v<T, void>, "unsupported type");
     }
@@ -106,13 +278,31 @@ static __global__ void mul_mat_vec_f(
     for (int j = 0; j < ncols_dst; ++j) {
         sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
 
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                sumf_gate[j] = warp_reduce_sum<warp_size>(sumf_gate[j]);
+            }
+        }
+
         if (block_size > warp_size) {
             buf_iw[tid/warp_size] = sumf[j];
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    buf_iw_gate[tid/warp_size] = sumf_gate[j];
+                }
+            }
             __syncthreads();
             if (tid < warp_size) {
                 sumf[j] = buf_iw[tid];
                 sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        sumf_gate[j] = buf_iw_gate[tid];
+                        sumf_gate[j] = warp_reduce_sum<warp_size>(sumf_gate[j]);
+                    }
+                }
             }
+
             if (j < ncols_dst) {
                 __syncthreads();
             }
@@ -123,12 +313,74 @@ static __global__ void mul_mat_vec_f(
         return;
     }
 
-    dst[tid*stride_col_dst + row] = sumf[tid];
+    float value = sumf[tid];
+
+    if constexpr (has_fusion) {
+        if (use_bias) {
+            value += x_bias[tid*stride_col_dst + row];
+        }
+
+        if (use_gate) {
+            float gate_value = sumf_gate[tid];
+            if (use_gate_bias) {
+                gate_value += gate_bias[tid*stride_col_dst + row];
+            }
+            switch (glu_op) {
+                case GGML_GLU_OP_SWIGLU:
+                    value *= ggml_cuda_op_silu_single(gate_value);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                    value *= ggml_cuda_op_gelu_single(gate_value);
+                    break;
+                case GGML_GLU_OP_SWIGLU_OAI: {
+                    value = ggml_cuda_op_swiglu_oai_single(gate_value, value);
+                    break;
+                }
+                default:
+                    break;
+            }
+        }
+    }
+
+    dst[tid*stride_col_dst + row] = value;
+
+    if constexpr (!has_fusion) {
+        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, glu_op, gate_x, x_bias, gate_bias, sumf_gate);
+    }
+}
+
+template<typename T, typename type_acc, int ncols_dst, int block_size>
+static void mul_mat_vec_f_switch_fusion(
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
+        const int64_t ncols, const int64_t nrows,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const cudaStream_t stream) {
+
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    if constexpr (ncols_dst == 1) {
+        if (has_fusion) {
+            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            return;
+       }
+    }
+
+    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
+
+    mul_mat_vec_f<T, type_acc, ncols_dst, block_size><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+
 }
 
 template <typename T, typename type_acc, int ncols_dst>
-static void launch_mul_mat_vec_f_cuda(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+void launch_mul_mat_vec_f_cuda(
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows,
         const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
@@ -140,8 +392,8 @@ static void launch_mul_mat_vec_f_cuda(
     GGML_ASSERT(stride_col_y % 2 == 0);
     GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
     GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
-    const int64_t channel_ratio = nchannels_dst / nchannels_x;
-    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
+    const uint3 channel_ratio_fd = ids ? make_uint3(0, 0, 0) : init_fastdiv_values(nchannels_dst / nchannels_x);
+    const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);
 
     const int device = ggml_cuda_get_device();
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
@@ -160,57 +412,59 @@ static void launch_mul_mat_vec_f_cuda(
         }
     }
 
-    const int nbytes_shared = warp_size*sizeof(float);
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+
+    const int nbytes_shared = warp_size*sizeof(float) + (has_fusion ? warp_size*sizeof(float) : 0);
     const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  32><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case   64: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  64><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case   96: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  96><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  128: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 128><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  160: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 160><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  192: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 192><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  224: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 224><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  256: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 256><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -220,7 +474,7 @@ static void launch_mul_mat_vec_f_cuda(
 
 template <typename T, typename type_acc>
 static void mul_mat_vec_f_cuda_switch_ncols_dst(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
         const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
@@ -230,49 +484,49 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
     switch (ncols_dst) {
         case 1:
             launch_mul_mat_vec_f_cuda<T, type_acc, 1>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 2:
             launch_mul_mat_vec_f_cuda<T, type_acc, 2>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 3:
             launch_mul_mat_vec_f_cuda<T, type_acc, 3>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 4:
             launch_mul_mat_vec_f_cuda<T, type_acc, 4>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 5:
             launch_mul_mat_vec_f_cuda<T, type_acc, 5>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 6:
             launch_mul_mat_vec_f_cuda<T, type_acc, 6>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 7:
             launch_mul_mat_vec_f_cuda<T, type_acc, 7>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 8:
             launch_mul_mat_vec_f_cuda<T, type_acc, 8>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
@@ -284,29 +538,31 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
 
 template<typename T>
 static void mul_mat_vec_f_cuda(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
         const int64_t stride_row, const int64_t stride_col_y, const int stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         enum ggml_prec prec, cudaStream_t stream) {
+
     if constexpr(std::is_same_v<T, half>) {
         if (prec == GGML_PREC_DEFAULT) {
             mul_mat_vec_f_cuda_switch_ncols_dst<T, half>
-                (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
-                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             return;
         }
     }
     mul_mat_vec_f_cuda_switch_ncols_dst<T, float>
-        (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
-         nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-         stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+        (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+        nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+        stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
 }
 
-void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+    const ggml_cuda_mm_fusion_args_host * fusion) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
     GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
     GGML_ASSERT(         dst->type == GGML_TYPE_F32);
@@ -332,6 +588,30 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
     float         *  dst_d =       (float         *)  dst->data;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
+
+    if (fusion) {
+        GGML_ASSERT( !ids || dst->ne[2] == 1);
+        GGML_ASSERT(  ids || dst->ne[1] == 1);
+        if (fusion->x_bias) {
+            GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->x_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->x_bias->ne[1] == src0->ne[2]);
+            fusion_local.x_bias = fusion->x_bias->data;
+        }
+        if (fusion->gate) {
+            GGML_ASSERT(fusion->gate->type == src0->type && ggml_are_same_stride(fusion->gate, src0));
+            fusion_local.gate = fusion->gate->data;
+        }
+        if (fusion->gate_bias) {
+            GGML_ASSERT(fusion->gate_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->gate_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
+            fusion_local.gate_bias = fusion->gate_bias->data;
+        }
+        fusion_local.glu_op = fusion->glu_op;
+    }
+
     const int64_t s01 = src0->nb[1] / ts_src0;
     const int64_t s11 = src1->nb[1] / ts_src1;
     const int64_t s1  =  dst->nb[1] / ts_dst;
@@ -354,19 +634,19 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
@@ -393,7 +673,6 @@ void ggml_cuda_op_mul_mat_vec_f(
     const int cc = ggml_cuda_info().devices[id].cc;
     const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
 
-
     // ggml_cuda_op provides single, contiguous matrices
     const int64_t stride_row         = ne00;
     const int64_t stride_col_y       = ne10;
@@ -410,22 +689,23 @@ void ggml_cuda_op_mul_mat_vec_f(
     const int64_t stride_sample_y    = 0;
     const int64_t stride_sample_dst  = 0;
 
+    ggml_cuda_mm_fusion_args_device empty{};
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;

ggml/src/ggml-cuda/mmvf.cuh

@@ -1,6 +1,7 @@
 #include "common.cuh"
 
-void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+    const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
 
 void ggml_cuda_op_mul_mat_vec_f(
     ggml_backend_cuda_context & ctx,

ggml/src/ggml-cuda/mmvq.cu

@@ -1,5 +1,6 @@
 #include "mmvq.cuh"
 #include "quantize.cuh"
+#include "unary.cuh"
 #include "vecdotq.cuh"
 
 #include <cstdint>
@@ -82,7 +83,7 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     return MMVQ_PARAMETERS_GENERIC;
 }
 
-static constexpr __host__ __device__ int calc_nwarps(int ncols_dst,  mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
         switch (ncols_dst) {
             case 1:
@@ -136,11 +137,11 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-template <ggml_type type, int ncols_dst>
 // tell the compiler to use as many registers as it wants, see nwarps definition below
+template <ggml_type type, int ncols_dst, bool has_fusion>
 __launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
-        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
@@ -169,8 +170,38 @@ static __global__ void mul_mat_vec_q(
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
 
+    bool use_gate = false;
+    bool use_bias = false;
+    bool use_gate_bias = false;
+    const void * vgate = nullptr;
+    const float * x_bias = nullptr;
+    const float * gate_bias = nullptr;
+    ggml_glu_op active_glu;
+
+    if constexpr (has_fusion) {
+        use_gate      = fusion.gate      != nullptr;
+        use_bias      = fusion.x_bias    != nullptr;
+        use_gate_bias = fusion.gate_bias != nullptr && use_gate;
+        vgate         = fusion.gate;
+        x_bias        = (const float *) fusion.x_bias;
+        gate_bias     = (const float *) fusion.gate_bias;
+        active_glu    = fusion.glu_op;
+    }
+
+    const uint32_t channel_bias = ids ? channel_x : channel_dst;
+
+    if constexpr (has_fusion) {
+        if (use_bias) {
+            x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+        }
+        if (use_gate_bias) {
+            gate_bias = gate_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+        }
+    }
+
     // partial sum for each thread
     float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
+    float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
     const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
@@ -187,17 +218,35 @@ static __global__ void mul_mat_vec_q(
             for (int i = 0; i < rows_per_cuda_block; ++i) {
                 tmp[j][i] += vec_dot_q_cuda(
                     vx, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_gate[j][i] += vec_dot_q_cuda(
+                            vgate, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
+                    }
+                }
             }
         }
     }
 
     __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
+    __shared__ float tmp_shared_gate[(has_fusion && (nwarps-1 > 0)) ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
+    if constexpr (!has_fusion) {
+        (void) tmp_shared_gate;
+    } else if (!use_gate) {
+        (void) tmp_shared_gate;
+    }
+
     if (threadIdx.y > 0) {
 #pragma unroll
         for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
             for (int i = 0; i < rows_per_cuda_block; ++i) {
                 tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_shared_gate[threadIdx.y-1][j][i][threadIdx.x] = tmp_gate[j][i];
+                    }
+                }
             }
         }
     }
@@ -216,14 +265,55 @@ static __global__ void mul_mat_vec_q(
 #pragma unroll
             for (int l = 0; l < nwarps-1; ++l) {
                 tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_gate[j][i] += tmp_shared_gate[l][j][i][threadIdx.x];
+                    }
+                }
             }
             tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmp_gate[j][i] = warp_reduce_sum<warp_size>(tmp_gate[j][i]);
+                }
+            }
         }
 
         if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
-            dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
+            float result = tmp[j][threadIdx.x];
+            if constexpr (has_fusion) {
+                if (use_bias) {
+                    result += x_bias[j*stride_col_dst + threadIdx.x];
+                }
+                if (use_gate) {
+                    float gate_value = tmp_gate[j][threadIdx.x];
+                    if (use_gate_bias) {
+                        gate_value += gate_bias[j*stride_col_dst + threadIdx.x];
+                    }
+                    switch (active_glu) {
+                        case GGML_GLU_OP_SWIGLU:
+                            result *= ggml_cuda_op_silu_single(gate_value);
+                            break;
+                        case GGML_GLU_OP_GEGLU:
+                            result *= ggml_cuda_op_gelu_single(gate_value);
+                            break;
+                        case GGML_GLU_OP_SWIGLU_OAI: {
+                            result = ggml_cuda_op_swiglu_oai_single(gate_value, result);
+                            break;
+                        }
+                        default:
+                            result = result * gate_value;
+                            break;
+                    }
+                }
+            }
+            dst[j*stride_col_dst + threadIdx.x] = result;
         }
     }
+
+    if constexpr (!has_fusion) {
+        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, active_glu, gate_bias, x_bias, tmp_gate);
+    }
 }
 
 static std::pair<dim3, dim3> calc_launch_params(
@@ -235,9 +325,37 @@ static std::pair<dim3, dim3> calc_launch_params(
     return {block_nums, block_dims};
 }
 
+template<ggml_type type, int c_ncols_dst>
+static void mul_mat_vec_q_switch_fusion(
+        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
+        const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
+        const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
+        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared, cudaStream_t stream) {
+
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    if constexpr (c_ncols_dst == 1) {
+        if (has_fusion) {
+            mul_mat_vec_q<type, c_ncols_dst, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            return;
+        }
+    }
+
+    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
+
+    mul_mat_vec_q<type, c_ncols_dst, false><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+}
+
 template <ggml_type type>
 static void mul_mat_vec_q_switch_ncols_dst(
-        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int ncols_x, const int nrows_x, const int ncols_dst,
         const int stride_row_x, const int stride_col_y, const int stride_col_dst,
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
@@ -256,80 +374,83 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
     const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
 
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+
     GGML_ASSERT(!ids || ncols_dst == 1);
     switch (ncols_dst) {
         case 1: {
             constexpr int c_ncols_dst = 1;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 2: {
             constexpr int c_ncols_dst = 2;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 3: {
             constexpr int c_ncols_dst = 3;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 4: {
             constexpr int c_ncols_dst = 4;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 5: {
             constexpr int c_ncols_dst = 5;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 6: {
             constexpr int c_ncols_dst = 6;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 7: {
             constexpr int c_ncols_dst = 7;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 8: {
             constexpr int c_ncols_dst = 8;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         default:
             GGML_ABORT("fatal error");
             break;
     }
-}
 
+    GGML_UNUSED(has_fusion);
+}
 static void mul_mat_vec_q_switch_type(
-        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, float * dst,
+        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int ncols_x, const int nrows_x, const int ncols_dst,
         const int stride_row_x, const int stride_col_y, const int stride_col_dst,
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
@@ -339,143 +460,123 @@ static void mul_mat_vec_q_switch_type(
     switch (type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q4_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q8_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_MXFP4:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q2_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q3_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q4_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q6_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ3_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ1_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ1_M:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ4_NL:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ4_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ3_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -484,7 +585,8 @@ static void mul_mat_vec_q_switch_type(
 }
 
 void ggml_cuda_mul_mat_vec_q(
-        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+        const ggml_cuda_mm_fusion_args_host * fusion) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
     GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
     GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
@@ -508,6 +610,31 @@ void ggml_cuda_mul_mat_vec_q(
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
     float         *  dst_d =       (float         *)  dst->data;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
+
+    if (fusion) {
+        GGML_ASSERT( !ids || dst->ne[2] == 1);
+        GGML_ASSERT(  ids || dst->ne[1] == 1);
+
+        if (fusion->x_bias) {
+            GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->x_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->x_bias->ne[1] == src0->ne[2]);
+            fusion_local.x_bias = fusion->x_bias->data;
+        }
+        if (fusion->gate) {
+            GGML_ASSERT(fusion->gate->type == src0->type && ggml_are_same_stride(fusion->gate, src0));
+            fusion_local.gate = fusion->gate->data;
+        }
+        if (fusion->gate_bias) {
+            GGML_ASSERT(fusion->gate_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->gate_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
+            fusion_local.gate_bias = fusion->gate_bias->data;
+        }
+        fusion_local.glu_op = fusion->glu_op;
+    }
+
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
         const size_t size_data  = ggml_nbytes(src0);
@@ -549,10 +676,10 @@ void ggml_cuda_mul_mat_vec_q(
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
     mul_mat_vec_q_switch_type(
-        src0->data, src0->type, src1_q8_1.get(), ids_d, dst_d, ne00,
+        src0->data, src0->type, src1_q8_1.get(), ids_d, fusion_local, dst_d, ne00,
         ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
         ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-        ne03,              ne3,           s03, s13,              s3,                 stream);
+        ne03,              ne3,           s03, s13,              s3,               stream);
 }
 
 void ggml_cuda_op_mul_mat_vec_q(
@@ -578,8 +705,9 @@ void ggml_cuda_op_mul_mat_vec_q(
     const int stride_row_x = ne00 / ggml_blck_size(src0->type);
     const int stride_col_y = src1_padded_row_size / QK8_1;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
     mul_mat_vec_q_switch_type(
-        src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
+        src0_dd_i, src0->type, src1_ddq_i, nullptr, fusion_local, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
         1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
 
     GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);

ggml/src/ggml-cuda/mmvq.cuh

@@ -3,7 +3,7 @@
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
 
 void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
 
 void ggml_cuda_op_mul_mat_vec_q(
     ggml_backend_cuda_context & ctx,

ggml/src/ggml-cuda/set.cu

@@ -0,0 +1,39 @@
+#include "set.cuh"
+#include "cpy.cuh"
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32));
+    GGML_ASSERT(src1->type == src0->type);
+    GGML_ASSERT(dst ->type == src0->type);
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const size_t nb1    = ((int32_t *) dst->op_params)[0];
+    const size_t nb2    = ((int32_t *) dst->op_params)[1];
+    const size_t nb3    = ((int32_t *) dst->op_params)[2];
+    const size_t offset = ((int32_t *) dst->op_params)[3];
+    const bool   inplace= (bool)     ((int32_t *) dst->op_params)[4];
+
+    if (!inplace) {
+        ggml_cuda_cpy(ctx, src0, dst);
+    }
+
+    ggml_tensor dst_view = *dst;
+    dst_view.data  = (void *)((char *)dst->data + offset);
+    dst_view.ne[0] = src1->ne[0];
+    dst_view.ne[1] = src1->ne[1];
+    dst_view.ne[2] = src1->ne[2];
+    dst_view.ne[3] = src1->ne[3];
+
+    dst_view.nb[0] = ggml_element_size(dst);
+    dst_view.nb[1] = nb1;
+    dst_view.nb[2] = nb2;
+    dst_view.nb[3] = nb3;
+
+    ggml_cuda_cpy(ctx, src1, &dst_view);
+}

ggml/src/ggml-cuda/set.cuh

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.cuh"
+
+#define CUDA_SET_BLOCK_SIZE 256
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/topk-moe.cu

@@ -2,23 +2,70 @@
 #include "ggml.h"
 #include "topk-moe.cuh"
 
+#include <cmath>
 #include <initializer_list>
 
+// Warp-local softmax used for both the pre-top-k logits and the post-top-k delayed path.
+template <int experts_per_thread, bool use_limit>
+__device__ void softmax_warp_inplace(float (&vals)[experts_per_thread], const int limit, const int lane) {
+    float max_val = -INFINITY;
+
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        const int  idx    = lane + i * WARP_SIZE;
+        const bool active = !use_limit || (idx < limit);
+        if (active) {
+            max_val = max(max_val, vals[i]);
+        }
+    }
+
+    max_val = warp_reduce_max(max_val);
+
+    float sum = 0.f;
+
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        const int  idx    = lane + i * WARP_SIZE;
+        const bool active = !use_limit || (idx < limit);
+        if (active) {
+            const float val = expf(vals[i] - max_val);
+            vals[i]         = val;
+            sum += val;
+        } else {
+            vals[i] = 0.f;
+        }
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    const float inv_sum = 1.0f / sum;
+
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        const int  idx    = lane + i * WARP_SIZE;
+        const bool active = !use_limit || (idx < limit);
+        if (active) {
+            vals[i] *= inv_sum;
+        }
+    }
+}
+
 /*
     This kernel does the following:
-    1. softmax over the logits per token [n_experts, n_tokens]
+    1. optionally softmax over the logits per token [n_experts, n_tokens]
     2. argmax reduce over the top-k (n_experts_used) logits
     3. write weights + ids to global memory
-    4. optionally normalize the weights
+    4. optionally normalize the weights or apply softmax over the selected logits
 
     It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models
 */
-template <int n_experts, bool with_norm>
+template <int n_experts, bool with_norm, bool delayed_softmax = false>
 __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * logits,
                                                                   float *       weights,
                                                                   int32_t *     ids,
                                                                   const int     n_rows,
-                                                                  const int     n_expert_used) {
+                                                                  const int     n_expert_used,
+                                                                  const float   clamp_val) {
     const int row = blockIdx.x * blockDim.y + threadIdx.y;
     if (row >= n_rows) {
         return;
@@ -30,51 +77,30 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
 
     constexpr int experts_per_thread = (n_experts > WARP_SIZE) ? n_experts / WARP_SIZE : 1;
 
-    float logits_r[experts_per_thread];
+    float wt[experts_per_thread];
 
 #pragma unroll
     for (int i = 0; i < n_experts; i += WARP_SIZE) {
-        const int expert        = i + threadIdx.x;
-        logits_r[i / WARP_SIZE] = n_experts % WARP_SIZE == 0 || expert < n_experts ? logits[expert] : -INFINITY;
+        const int expert  = i + threadIdx.x;
+        wt[i / WARP_SIZE] = (n_experts % WARP_SIZE == 0 || expert < n_experts) ? logits[expert] : -INFINITY;
     }
 
-    float max_val = logits_r[0];
-
-#pragma unroll
-    for (int i = 1; i < experts_per_thread; i++) {
-        const float val = logits_r[i];
-        max_val         = max(val, max_val);
+    if constexpr (!delayed_softmax) {
+        softmax_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
     }
 
-    max_val = warp_reduce_max(max_val);
-
-    float wt[experts_per_thread];
-    float tmp = 0.f;
-
-#pragma unroll
-    for (int i = 0; i < experts_per_thread; i++) {
-        const float val = logits_r[i];
-        wt[i]           = expf(val - max_val);
-        tmp += wt[i];
-    }
-
-    tmp = warp_reduce_sum(tmp);
-
-    const float inv_sum = 1.0f / tmp;
-
-#pragma unroll
-    for (int i = 0; i < experts_per_thread; i++) {
-        wt[i] = wt[i] * inv_sum;
-    }
-
-    //at this point, each thread holds a portion of softmax,
-    //we do the argmax reduce over n_expert_used, each time marking
+    //at this point, each thread holds either a portion of the softmax distribution
+    //or the raw logits. We do the argmax reduce over n_expert_used, each time marking
     //the expert weight as -inf to exclude from the next iteration
 
     float wt_sum = 0.f;
 
-    extern __shared__ float data_topk_shared[];
-    float *                 wt_shared_ptr = data_topk_shared + threadIdx.y * n_expert_used;
+    float output_weights[experts_per_thread];
+
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        output_weights[i] = 0.f;
+    }
 
     for (int k = 0; k < n_expert_used; k++) {
         float max_val    = wt[0];
@@ -99,11 +125,14 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
             }
         }
 
+        if ((k & (WARP_SIZE - 1)) == threadIdx.x) {
+            output_weights[k / WARP_SIZE] = max_val;
+        }
+
         if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) {
             wt[max_expert / WARP_SIZE] = -INFINITY;
 
-            wt_shared_ptr[k] = max_val;
-            ids[k]           = max_expert;
+            ids[k] = max_expert;
             if constexpr (with_norm) {
                 wt_sum += max_val;
             }
@@ -112,73 +141,86 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
 
     if constexpr (with_norm) {
         wt_sum              = warp_reduce_sum(wt_sum);
+        wt_sum              = max(wt_sum, clamp_val);
         const float inv_sum = 1.0f / wt_sum;
 
-        for (int i = threadIdx.x; i < n_expert_used; i += WARP_SIZE) {
-            wt_shared_ptr[i] = wt_shared_ptr[i] * inv_sum;
+        for (int i = 0; i < experts_per_thread; i++) {
+            output_weights[i] *= inv_sum;
         }
     }
 
-    for (int i = threadIdx.x; i < n_expert_used; i += WARP_SIZE) {
-        weights[i] = wt_shared_ptr[i];
+    if constexpr (delayed_softmax) {
+        softmax_warp_inplace<experts_per_thread, true>(output_weights, n_expert_used, threadIdx.x);
+    }
+
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        const int idx = i * WARP_SIZE + threadIdx.x;
+        if (idx < n_expert_used) {
+            weights[idx] = output_weights[i];
+        }
+    }
+
+    if (!with_norm) {
+        GGML_UNUSED(clamp_val);
     }
 }
 
-template <bool with_norm>
+template <bool with_norm, bool delayed_softmax = false>
 static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
                                  const float *               logits,
                                  float *                     weights,
                                  int32_t *                   ids,
                                  const int                   n_rows,
                                  const int                   n_expert,
-                                 const int                   n_expert_used) {
+                                 const int                   n_expert_used,
+                                 const float                 clamp_val) {
+    static_assert(!(with_norm && delayed_softmax), "delayed softmax is not supported with weight normalization");
     const int    rows_per_block = 4;
     dim3         grid_dims((n_rows + rows_per_block - 1) / rows_per_block, 1, 1);
     dim3         block_dims(WARP_SIZE, rows_per_block, 1);
     cudaStream_t stream = ctx.stream();
 
-    const int nbytes_shared = n_expert_used * rows_per_block * sizeof(float);
-
     switch (n_expert) {
         case 1:
-            topk_moe_cuda<1, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<1, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 2:
-            topk_moe_cuda<2, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<2, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 4:
-            topk_moe_cuda<4, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<4, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 8:
-            topk_moe_cuda<8, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<8, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 16:
-            topk_moe_cuda<16, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<16, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 32:
-            topk_moe_cuda<32, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<32, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 64:
-            topk_moe_cuda<64, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<64, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 128:
-            topk_moe_cuda<128, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<128, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 256:
-            topk_moe_cuda<256, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<256, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 512:
-            topk_moe_cuda<512, with_norm>
-                <<<grid_dims, block_dims, nbytes_shared, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+            topk_moe_cuda<512, with_norm, delayed_softmax>
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         default:
             GGML_ASSERT(false && "fatal error");
@@ -190,7 +232,9 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
                            const ggml_tensor *         logits,
                            ggml_tensor *               weights,
                            ggml_tensor *               ids,
-                           const bool                  with_norm) {
+                           const bool                  with_norm,
+                           const bool                  delayed_softmax,
+                           ggml_tensor *               clamp) {
     GGML_ASSERT(logits->type == GGML_TYPE_F32);
     GGML_ASSERT(weights->type == GGML_TYPE_F32);
     GGML_ASSERT(ids->type == GGML_TYPE_I32);
@@ -198,7 +242,7 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
     const int n_experts = logits->ne[0];
     const int n_rows    = logits->ne[1];
 
-    const float * logits_d  = (const float *) logits->src[0]->data;
+    const float * logits_d  = (const float *) logits->data;
     float *       weights_d = (float *) weights->data;
     int32_t *     ids_d     = (int32_t *) ids->data;
 
@@ -206,14 +250,25 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
 
     const int n_expert_used = weights->ne[1];
 
+    float clamp_val = -INFINITY;
     if (with_norm) {
-        launch_topk_moe_cuda<true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used);
+        if (clamp) {
+            clamp_val = ggml_get_op_params_f32(clamp, 0);
+        }
+        launch_topk_moe_cuda<true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used, clamp_val);
     } else {
-        launch_topk_moe_cuda<false>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used);
+        GGML_ASSERT(clamp == nullptr);
+        if (delayed_softmax) {
+            launch_topk_moe_cuda<false, true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used,
+                                              clamp_val);
+        } else {
+            launch_topk_moe_cuda<false, false>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used,
+                                               clamp_val);
+        }
     }
 }
 
-bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights) {
+bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights, const ggml_tensor * clamp) {
     float scale    = 1.0f;
     float max_bias = 0.0f;
 
@@ -239,19 +294,43 @@ bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tenso
         return false;
     }
 
+    if (clamp) {
+        if (clamp->op != GGML_OP_CLAMP) {
+            return false;
+        }
+        float max_val = ggml_get_op_params_f32(clamp, 1);
+
+        if (max_val != INFINITY) {
+            return false;
+        }
+    }
+
+
     return true;
 }
 
-std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool norm) {
+std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool norm, bool delayed_softmax) {
     static std::initializer_list<enum ggml_op> norm_ops = { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
                                                             GGML_OP_VIEW,     GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
-                                                            GGML_OP_SUM_ROWS, GGML_OP_DIV,      GGML_OP_RESHAPE };
+                                                            GGML_OP_SUM_ROWS, GGML_OP_CLAMP,    GGML_OP_DIV,
+                                                            GGML_OP_RESHAPE };
 
     static std::initializer_list<enum ggml_op> no_norm_ops = { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE, GGML_OP_ARGSORT,
                                                                GGML_OP_VIEW, GGML_OP_GET_ROWS };
 
+    static std::initializer_list<enum ggml_op> delayed_softmax_ops = { GGML_OP_ARGSORT,  GGML_OP_VIEW,
+                                                                       GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
+                                                                       GGML_OP_SOFT_MAX, GGML_OP_RESHAPE };
+
+    GGML_ASSERT(!norm || !delayed_softmax);
+
+    if (delayed_softmax) {
+        return delayed_softmax_ops;
+    }
+
     if (norm) {
         return norm_ops;
     }
+
     return no_norm_ops;
 }

ggml/src/ggml-cuda/topk-moe.cuh

@@ -6,9 +6,11 @@
 void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
                            const ggml_tensor *         logits,
                            ggml_tensor *               weights,
-                           ggml_tensor *               top_k,
-                           const bool                  with_norm);
+                           ggml_tensor *               ids,
+                           const bool                  with_norm,
+                           const bool                  delayed_softmax = false,
+                           ggml_tensor *               weight_clamp    = nullptr);
 
-bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights);
+bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights, const ggml_tensor * clamp = nullptr);
 
-std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool with_norm);
+std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool with_norm, bool delayed_softmax = false);

ggml/src/ggml-cuda/unary.cu

@@ -18,10 +18,7 @@ static __device__ __forceinline__ float op_step(float x) {
 }
 
 static __device__ __forceinline__ float op_gelu(float x) {
-    const float GELU_COEF_A    = 0.044715f;
-    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
-
-    return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+    return ggml_cuda_op_gelu_single(x);
 }
 
 static __device__ __forceinline__ float op_gelu_erf(float x) {
@@ -37,7 +34,7 @@ static __device__ __forceinline__ float op_gelu_quick(float x) {
 }
 
 static __device__ __forceinline__ float op_silu(float x) {
-    return x / (1.0f + expf(-x));
+    return ggml_cuda_op_silu_single(x);
 }
 
 static __device__ __forceinline__ float op_tanh(float x) {
@@ -317,13 +314,8 @@ static __global__ void swiglu_oai_kernel(const T * x, const T * g, T * dst, cons
 
     float xi = x[j0];
     float gi = g[j1];
-    xi = fminf(xi, limit);
-    gi = fmaxf(fminf(gi, limit), -limit);
 
-    float out_glu = xi / (1.0f + expf(-xi * alpha));
-    out_glu = out_glu * (1.0f + gi);
-
-    dst[i] = out_glu;
+    dst[i] = ggml_cuda_op_swiglu_oai_single(xi, gi, alpha, limit);
 }
 
 template <typename T>

ggml/src/ggml-cuda/unary.cuh

@@ -1,3 +1,4 @@
+#pragma once
 #include "common.cuh"
 
 #define CUDA_NEG_BLOCK_SIZE 256
@@ -75,3 +76,23 @@ void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_xielu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+__device__ __forceinline__ float ggml_cuda_op_silu_single(float x) {
+    return x / (1.0f + expf(-x));
+}
+
+__device__ __forceinline__ float ggml_cuda_op_gelu_single(float x) {
+    const float GELU_COEF_A    = 0.044715f;
+    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+
+    return 0.5f * x * (1.0f + tanhf(SQRT_2_OVER_PI * x * (1.0f + GELU_COEF_A * x * x)));
+}
+
+__device__ __forceinline__ float ggml_cuda_op_swiglu_oai_single(float x, float g, float alpha = 1.702f, float limit = 7.0f) {
+    x = fminf(x, limit);
+    g = fmaxf(fminf(g, limit), -limit);
+
+    float out_glu = x / (1.0f + expf(-x * alpha));
+    out_glu = out_glu * (1.0f + g);
+    return out_glu;
+}

ggml/src/ggml-cuda/upscale.cu

@@ -126,8 +126,8 @@ void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
         float pixel_offset = 0.5f;
         if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
-            sf0          = (float)(dst->ne[0] - 1) / (src0->ne[0] - 1);
-            sf1          = (float)(dst->ne[1] - 1) / (src0->ne[1] - 1);
+            sf0          = dst->ne[0] > 1 && src0->ne[0] > 1 ? (float)(dst->ne[0] - 1) / (src0->ne[0] - 1) : sf0;
+            sf1          = dst->ne[1] > 1 && src0->ne[1] > 1 ? (float)(dst->ne[1] - 1) / (src0->ne[1] - 1) : sf1;
             pixel_offset = 0.0f;
         }
         upscale_f32_bilinear_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],

ggml/src/ggml-hexagon/CMakeLists.txt

@@ -0,0 +1,68 @@
+include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
+include(ExternalProject)
+
+option(GGML_HEXAGON_HTP_DEBUG "ggml-hexagon: enable HTP debug output" OFF)
+
+add_library(htp_iface OBJECT
+    ${CMAKE_CURRENT_BINARY_DIR}/htp_iface_stub.c)
+
+set_target_properties(htp_iface PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_include_directories(htp_iface PUBLIC
+    ${HEXAGON_SDK_ROOT}/incs
+    ${HEXAGON_SDK_ROOT}/incs/stddef
+    ${HEXAGON_SDK_ROOT}/utils/examples
+    ${CMAKE_CURRENT_SOURCE_DIR}/htp
+    ${CMAKE_CURRENT_BINARY_DIR})
+
+build_idl(htp/htp_iface.idl htp_iface)
+
+if (CMAKE_SYSTEM_NAME MATCHES Android)
+    target_link_options(htp_iface PUBLIC -llog -ldl)
+elseif (CMAKE_SYSTEM_NAME MATCHES Windows)
+    target_precompile_headers(htp_iface PUBLIC <sal.h>)
+else()
+    target_link_options(htp_iface PUBLIC -ldl)
+endif()
+
+link_custom_library(htp_iface cdsprpc)
+link_custom_library(htp_iface rpcmem)
+
+set(TARGET_NAME ggml-hexagon)
+ggml_add_backend_library(${TARGET_NAME}
+    ggml-hexagon.cpp htp-utils.c htp-utils.h ../../include/ggml-hexagon.h)
+
+target_link_libraries(${TARGET_NAME} PRIVATE htp_iface)
+target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/htp ${CMAKE_CURRENT_BINARY_DIR})
+
+# Build HTP bits
+set(HTP_CMAKE_ARGS
+    -DCMAKE_TOOLCHAIN_FILE=${CMAKE_CURRENT_SOURCE_DIR}/htp/cmake-toolchain.cmake
+    -DCMAKE_BUILD_TYPE=Release
+    -DCMAKE_INSTALL_LIBDIR=${CMAKE_CURRENT_BINARY_DIR}
+    -DHEXAGON_SDK_ROOT=$ENV{HEXAGON_SDK_ROOT}
+    -DHEXAGON_TOOLS_ROOT=$ENV{HEXAGON_TOOLS_ROOT}
+    -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG})
+
+ExternalProject_Add(htp-v73
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v73 -DPREBUILT_LIB_DIR="toolv19_v73")
+
+ExternalProject_Add(htp-v75
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v75 -DPREBUILT_LIB_DIR="toolv19_v75")
+
+ExternalProject_Add(htp-v79
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v79 -DPREBUILT_LIB_DIR="toolv19_v79")
+
+ExternalProject_Add(htp-v81
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v81 -DPREBUILT_LIB_DIR="toolv19_v81")
+
+# Install Hexagon skels required at runtime
+install(FILES
+    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v73.so
+    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v75.so
+    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v79.so
+    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v81.so
+    TYPE LIB)

ggml/src/ggml-hexagon/htp-utils.c

@@ -0,0 +1,448 @@
+
+#pragma clang diagnostic ignored "-Wgnu-anonymous-struct"
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+#pragma clang diagnostic ignored "-Wsign-compare"
+
+#define GGML_COMMON_IMPL_C
+#include "ggml-backend-impl.h"
+#include "ggml-common.h"
+#include "ggml-hexagon.h"
+#include "ggml-impl.h"
+
+#include "htp-utils.h"
+
+#include <domain.h>
+#include <remote.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+domain * get_domain(int domain_id) {
+    int i    = 0;
+    int size = sizeof(supported_domains) / sizeof(domain);
+
+    for (i = 0; i < size; i++) {
+        if (supported_domains[i].id == domain_id) {
+            return &supported_domains[i];
+        }
+    }
+
+    return NULL;
+}
+
+bool is_valid_domain_id(int domain_id, int compute_only) {
+    int i    = 0;
+    int size = sizeof(supported_domains) / sizeof(domain);
+
+    if (compute_only) {
+        return is_CDSP(domain_id);
+    }
+
+    for (i = 0; i < size; i++) {
+        if (supported_domains[i].id == domain_id) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+int get_domains_info(char * domain_type, int * num_domains, fastrpc_domain ** domains_info) {
+    int nErr    = AEE_SUCCESS;
+    int ss_info = 0;
+    if (domain_type != NULL) {
+        if (strcmp(domain_type, "LPASS") == 0) {
+            ss_info = FASTRPC_LPASS;
+        } else if (strcmp(domain_type, "HPASS") == 0) {
+            ss_info = FASTRPC_HPASS;
+        } else {
+            ss_info = FASTRPC_NSP;
+        }
+    }
+    system_req_payload req  = { 0 };
+    req.id                  = FASTRPC_GET_DOMAINS;
+    req.sys.domains         = NULL;
+    fastrpc_domain * domain = NULL;
+    if (ss_info != 0) {
+        req.sys.flags = DOMAINS_LIST_FLAGS_SET_TYPE(req.sys.flags, ss_info);
+    } else {
+        req.sys.flags = 0;
+    }
+#ifdef _WIN32
+    nErr = AEE_EUNSUPPORTED;
+    goto bail;
+#endif
+    if (remote_system_request) {
+        nErr = remote_system_request(&req);
+        if (nErr != AEE_SUCCESS) {
+            GGML_LOG_ERROR("Failure in remote_system_request call: %d.\n", nErr);
+            goto bail;
+        }
+        // Allocate memory for domain-info array
+        req.sys.max_domains = req.sys.num_domains;
+        if ((req.sys.domains = calloc(req.sys.num_domains, sizeof(fastrpc_domain))) == NULL) {
+            nErr = AEE_ENOMEMORY;
+            GGML_LOG_ERROR("Unable to allocate memory for req.sys.domains");
+            goto bail;
+        }
+
+        nErr = remote_system_request(&req);
+        if (nErr != AEE_SUCCESS) {
+            GGML_LOG_ERROR("Failure in remote_system_request call: %d.\n", nErr);
+            goto bail;
+        }
+
+        for (int i = 0; i < req.sys.num_domains; i++) {
+            // Verify that only requested type domains were returned
+            domain = &req.sys.domains[i];
+            if (domain->type != ss_info && domain_type != NULL) {
+                nErr = -1;
+                GGML_LOG_ERROR("Incorrect data received from remote_system_request.\n");
+                goto bail;
+            }
+        }
+        *domains_info = req.sys.domains;
+        *num_domains  = req.sys.num_domains;
+    } else {
+        nErr = AEE_EUNSUPPORTED;
+        goto bail;
+    }
+bail:
+    if (nErr && !req.sys.domains) {
+        free(req.sys.domains);
+    }
+    return nErr;
+}
+
+int get_effective_domain_id(char * domain_name, int session_id, int * effec_domain_id) {
+    int                              err  = 0;
+    remote_rpc_effective_domain_id_t sess = { 0 };
+
+    sess.domain_name     = domain_name;
+    sess.domain_name_len = strlen(domain_name);
+    sess.session_id      = session_id;
+
+    err = remote_session_control(FASTRPC_GET_EFFECTIVE_DOMAIN_ID, &sess, sizeof(sess));
+    if (err) {
+        GGML_LOG_ERROR("Error 0x%x: failed to get effective domain id for %s, session id %d\n", err, sess.domain_name,
+               session_id);
+        return err;
+    }
+
+    *effec_domain_id = sess.effective_domain_id;
+    return err;
+}
+
+int get_dsp_support(int * domain) {
+    int nErr = AEE_SUCCESS;
+    *domain  = CDSP_DOMAIN_ID;  // DSP domain default value is CDSP_DOMAIN_ID
+
+    if (remote_handle_control) {
+        struct remote_dsp_capability dsp_capability_domain = { CDSP_DOMAIN_ID, DOMAIN_SUPPORT, 0 };
+        nErr = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_domain, sizeof(struct remote_dsp_capability));
+        if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
+            GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
+            goto bail;
+        }
+
+        if (dsp_capability_domain.capability == 0) {
+            dsp_capability_domain.domain       = ADSP_DOMAIN_ID;  // Check for ADSP support.
+            dsp_capability_domain.attribute_ID = DOMAIN_SUPPORT;
+            dsp_capability_domain.capability   = 0;
+            nErr                               = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_domain,
+                                                                       sizeof(struct remote_dsp_capability));
+            if (dsp_capability_domain.capability) {
+                *domain = ADSP_DOMAIN_ID;  // For targets like Agatti (not having cDSP), domain is ADSP_DOMAIN_ID
+            }
+        }
+
+        if (nErr != AEE_SUCCESS) {
+            GGML_LOG_ERROR("\nget_dsp_support failed with Error 0x%x\n", nErr);
+            goto bail;
+        }
+    } else {
+        nErr = AEE_EUNSUPPORTEDAPI;
+        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
+    }
+
+bail:
+    return nErr;
+}
+
+int get_vtcm_info(int domain, uint32_t * capability, uint32_t attr) {
+    int nErr    = AEE_SUCCESS;
+    *capability = 0;
+
+    if (attr == VTCM_PAGE || attr == VTCM_COUNT) {
+    } else {
+        nErr = AEE_EBADPARM;
+        GGML_LOG_ERROR("Unsupported attr. Only VTCM_PAGE and VTCM_COUNT supported\n");
+        goto bail;
+    }
+    if (remote_handle_control) {
+        if (domain == ADSP_DOMAIN_ID || domain == CDSP_DOMAIN_ID) {
+            /*
+            * Query the DSP for VTCM information
+            * Since the ADSP does not have a dedicated VTCM, we expect the output to be 0
+            */
+            struct remote_dsp_capability dsp_capability_vtcm_dsp;
+            dsp_capability_vtcm_dsp.domain       = (uint32_t) domain;
+            dsp_capability_vtcm_dsp.attribute_ID = attr;
+            dsp_capability_vtcm_dsp.capability   = (uint32_t) 0;
+            nErr                                 = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_vtcm_dsp,
+                                                                         sizeof(struct remote_dsp_capability));
+            if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
+                GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
+                GGML_LOG_ERROR("Running the usecase without checking the capability\n");
+                nErr = AEE_SUCCESS;
+                goto bail;
+            } else if (nErr == AEE_SUCCESS) {
+                *capability = dsp_capability_vtcm_dsp.capability;
+            } else {
+                GGML_LOG_ERROR("\nget_vtcm_info failed with Error 0x%x\n", nErr);
+                goto bail;
+            }
+        } else {
+            nErr = AEE_EUNSUPPORTED;
+            GGML_LOG_ERROR("Unsupported domain %d\n", domain);
+            goto bail;
+        }
+    } else {
+        nErr = AEE_EUNSUPPORTEDAPI;
+        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
+    }
+
+bail:
+    return nErr;
+}
+
+bool is_unsignedpd_supported(int domain_id) {
+    int nErr = AEE_SUCCESS;
+    if (remote_handle_control) {
+        struct remote_dsp_capability dsp_capability_domain = { domain_id, UNSIGNED_PD_SUPPORT, 0 };
+        nErr = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_domain, sizeof(struct remote_dsp_capability));
+        if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
+            GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device. Falling back to signed pd.\n");
+            return false;
+        }
+        if (nErr) {
+            GGML_LOG_ERROR("\nERROR 0x%x: FastRPC Capability API failed. Falling back to signed pd.", nErr);
+            return false;
+        }
+        if (dsp_capability_domain.capability == 1) {
+            return true;
+        }
+    } else {
+        nErr = AEE_EUNSUPPORTEDAPI;
+        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device. Falling back to signed pd.\n");
+        return false;
+    }
+    return false;
+}
+
+bool get_unsignedpd_support(void) {
+    return is_unsignedpd_supported(CDSP_DOMAIN_ID);
+}
+
+bool is_async_fastrpc_supported(int domain) {
+    int nErr = AEE_SUCCESS;
+    if (remote_handle_control) {
+        if (domain == CDSP_DOMAIN_ID) {
+            /*
+            * Query the DSP for ASYNC_FASTRPC_SUPPORT information
+            * Async fastrpc is supported only on CDSP
+            */
+            struct remote_dsp_capability dsp_capability_async_support;
+            dsp_capability_async_support.domain       = (uint32_t) domain;
+            dsp_capability_async_support.attribute_ID = ASYNC_FASTRPC_SUPPORT;
+            dsp_capability_async_support.capability   = (uint32_t) 0;
+            nErr = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_async_support,
+                                         sizeof(struct remote_dsp_capability));
+            if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
+                GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
+                GGML_LOG_ERROR("Running the usecase without checking the capability\n");
+                nErr = AEE_SUCCESS;
+                goto bail;
+            } else if (dsp_capability_async_support.capability == 1) {
+                return true;
+            }
+            if (nErr != AEE_SUCCESS) {
+                GGML_LOG_ERROR("\nis_async_fastrpc_supported failed with Error 0x%x\n", nErr);
+                goto bail;
+            }
+        } else {
+            nErr = AEE_EUNSUPPORTED;
+            GGML_LOG_ERROR("Async fastrpc is not supported on domain %d\n", domain);
+            goto bail;
+        }
+    } else {
+        nErr = AEE_EUNSUPPORTEDAPI;
+        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
+    }
+
+bail:
+    return false;
+}
+
+bool is_status_notification_supported(int domain) {
+    int nErr = AEE_SUCCESS;
+
+    if (remote_handle_control) {
+        /*
+        * Query the DSP for STATUS_NOTIFICATION_SUPPORT information
+        * DSP User PD status notification Support
+        */
+        struct remote_dsp_capability dsp_capability_status_notification_support;
+        dsp_capability_status_notification_support.domain       = (uint32_t) domain;
+        dsp_capability_status_notification_support.attribute_ID = STATUS_NOTIFICATION_SUPPORT;
+        dsp_capability_status_notification_support.capability   = (uint32_t) 0;
+        nErr = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_status_notification_support,
+                                     sizeof(struct remote_dsp_capability));
+        if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
+            GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
+            GGML_LOG_ERROR("Running the usecase without checking the capability\n");
+            nErr = AEE_SUCCESS;
+            goto bail;
+        } else if (dsp_capability_status_notification_support.capability == 1) {
+            return true;
+        }
+        if (nErr != AEE_SUCCESS) {
+            GGML_LOG_ERROR("\nis_status_notification_supported failed with Error 0x%x\n", nErr);
+            goto bail;
+        }
+    } else {
+        nErr = AEE_EUNSUPPORTEDAPI;
+        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
+    }
+
+bail:
+    return false;
+}
+
+int get_hmx_support_info(int domain, uint32_t * capability, uint32_t attr) {
+    int nErr    = AEE_SUCCESS;
+    *capability = 0;
+
+    if (attr != HMX_SUPPORT_SPATIAL && attr != HMX_SUPPORT_DEPTH) {
+        nErr = AEE_EBADPARM;
+        GGML_LOG_ERROR("Unsupported attr. Only HMX_SUPPORT_SPATIAL and HMX_SUPPORT_DEPTH supported\n");
+        goto bail;
+    }
+    if (remote_handle_control) {
+        if (domain == CDSP_DOMAIN_ID) {
+            /*
+            * Query the DSP for HMX SUPPORT information
+            * HMX is supported on CDSP only
+            */
+            struct remote_dsp_capability dsp_capability_hmx_dsp;
+            dsp_capability_hmx_dsp.domain       = (uint32_t) domain;
+            dsp_capability_hmx_dsp.attribute_ID = attr;
+            dsp_capability_hmx_dsp.capability   = (uint32_t) 0;
+            nErr                                = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_hmx_dsp,
+                                                                        sizeof(struct remote_dsp_capability));
+            if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
+                GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
+                GGML_LOG_ERROR("Running the usecase without checking the capability\n");
+                nErr = AEE_SUCCESS;
+                goto bail;
+            } else if (nErr == AEE_SUCCESS) {
+                *capability = dsp_capability_hmx_dsp.capability;
+            } else {
+                GGML_LOG_ERROR("\nget_hmx_support_info failed with Error 0x%x\n", nErr);
+                goto bail;
+            }
+        } else {
+            nErr = AEE_EUNSUPPORTED;
+            GGML_LOG_ERROR("HMX support is not there for domain %d\n", domain);
+            goto bail;
+        }
+    } else {
+        nErr = AEE_EUNSUPPORTEDAPI;
+        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
+    }
+
+bail:
+    return nErr;
+}
+
+int get_hex_arch_ver(int domain, int * arch) {
+    if (!remote_handle_control) {
+        GGML_LOG_ERROR("ggml-hex: remote_handle_control is not supported on this device\n");
+        return AEE_EUNSUPPORTEDAPI;
+    }
+
+    struct remote_dsp_capability arch_ver;
+    arch_ver.domain       = (uint32_t) domain;
+    arch_ver.attribute_ID = ARCH_VER;
+    arch_ver.capability   = (uint32_t) 0;
+
+    int err = remote_handle_control(DSPRPC_GET_DSP_INFO, &arch_ver, sizeof(arch_ver));
+    if ((err & 0xff) == (AEE_EUNSUPPORTEDAPI & 0xff)) {
+        GGML_LOG_ERROR("ggml-hex: FastRPC capability API is not supported on this device\n");
+        return AEE_EUNSUPPORTEDAPI;
+    }
+
+    if (err != AEE_SUCCESS) {
+        GGML_LOG_ERROR("ggml-hex: FastRPC capability query failed (err %d)\n", err);
+        return err;
+    }
+
+    switch (arch_ver.capability & 0xff) {
+        case 0x73:
+            *arch = 73;
+            return 0;
+        case 0x75:
+            *arch = 75;
+            return 0;
+        case 0x79:
+            *arch = 79;
+            return 0;
+        case 0x81:
+            *arch = 81;
+            return 0;
+    }
+    return -1;
+}
+
+int get_hvx_support_info(int domain, uint32_t * capability, uint32_t attr) {
+    int nErr    = AEE_SUCCESS;
+    *capability = 0;
+
+    if (remote_handle_control) {
+        if (domain == CDSP_DOMAIN_ID) {
+            /*
+            * Query the DSP for HVX SUPPORT information
+            * HVX is supported on CDSP only
+            */
+            struct remote_dsp_capability dsp_capability_hvx_dsp;
+            dsp_capability_hvx_dsp.domain       = (uint32_t) domain;
+            dsp_capability_hvx_dsp.attribute_ID = attr;
+            dsp_capability_hvx_dsp.capability   = (uint32_t) 0;
+            nErr                                = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_hvx_dsp,
+                                                                        sizeof(struct remote_dsp_capability));
+            if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
+                GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
+                GGML_LOG_ERROR("Running the usecase without checking the capability\n");
+                nErr = AEE_SUCCESS;
+                goto bail;
+            } else if (nErr == AEE_SUCCESS) {
+                *capability = dsp_capability_hvx_dsp.capability;
+            } else {
+                GGML_LOG_ERROR("\nget_hvx_support_info failed with Error 0x%x\n", nErr);
+                goto bail;
+            }
+        } else {
+            nErr = AEE_EUNSUPPORTED;
+            GGML_LOG_ERROR("HVX support is not available on domain %d\n", domain);
+            goto bail;
+        }
+    } else {
+        nErr = AEE_EUNSUPPORTEDAPI;
+        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
+    }
+
+bail:
+    return nErr;
+}

ggml/src/ggml-hexagon/htp-utils.h

@@ -0,0 +1,219 @@
+#ifndef HTP_UTILS_H
+#define HTP_UTILS_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <AEEStdErr.h>
+#include <inttypes.h>
+#include <remote.h>
+#include <stdbool.h>
+
+/* Offset to differentiate HLOS and Hexagon error codes.
+   Stores the value of AEE_EOFFSET for Hexagon. */
+#ifndef DSP_OFFSET
+#    define DSP_OFFSET 0x80000400
+#endif
+
+/* Errno for connection reset by peer. */
+#ifndef ECONNRESET
+#    ifdef __hexagon__
+#        define ECONNRESET 104
+#    endif
+#endif
+
+/* Abstraction of different OS specific sleep APIs.
+   SLEEP accepts input in seconds. */
+#ifndef SLEEP
+#    ifdef __hexagon__
+#        define SLEEP(x)                      \
+            { /* Do nothing for simulator. */ \
+            }
+#    else
+#        ifdef _WINDOWS
+#            define SLEEP(x) Sleep(1000 * x) /* Sleep accepts input in milliseconds. */
+#        else
+#            define SLEEP(x) sleep(x)        /* sleep accepts input in seconds. */
+#        endif
+#    endif
+#endif
+
+/* Include windows specific header files. */
+#ifdef _WINDOWS
+#    include <sysinfoapi.h>
+#    include <windows.h>
+#    define _CRT_SECURE_NO_WARNINGS         1
+#    define _WINSOCK_DEPRECATED_NO_WARNINGS 1
+/* Including this file for custom implementation of getopt function. */
+#    include "getopt_custom.h"
+#endif
+
+/* Includes and defines for all HLOS except windows */
+#if !defined(__hexagon__) && !defined(_WINDOWS)
+#    include "unistd.h"
+
+#    include <sys/time.h>
+#endif
+
+/* Includes and defines for Hexagon and all HLOS except Windows. */
+#if !defined(_WINDOWS)
+/* Weak reference to remote symbol for compilation. */
+#    pragma weak remote_session_control
+#    pragma weak remote_handle_control
+#    pragma weak remote_handle64_control
+#    pragma weak fastrpc_mmap
+#    pragma weak fastrpc_munmap
+#endif
+
+#if !defined(_WINDOWS)
+#    pragma weak remote_system_request
+#endif
+/**
+ * Wrapper for FastRPC Capability API: query DSP support.
+ *
+ * @param[out]  domain pointer to supported domain.
+ * @return      0          if query is successful.
+ *              non-zero   if error, return value points to the error.
+ */
+int get_dsp_support(int * domain);
+
+/**
+ * Wrapper for FastRPC Capability API: query VTCM information.
+ *
+ * @param[in]   domain value of domain in the queried.
+ * @param[out]  capability capability value of the attribute queried.
+ * @param[in]   attr value of the attribute to the queried.
+ * @return      0          if query is successful.
+ *              non-zero   if error, return value points to the error.
+ */
+int get_vtcm_info(int domain, uint32_t * capability, uint32_t attr);
+
+/**
+ * Wrapper for FastRPC Capability API: query unsigned pd support on CDSP domain.
+ *
+ * @return      true          if unsigned pd is supported.
+ *              false         if unsigned pd is not supported, capability query failed.
+ */
+
+bool get_unsignedpd_support(void);
+
+/**
+ * Wrapper for FastRPC Capability API: query unsigned pd support.
+ *
+ * @param[in]   domain value of domain in the queried.
+ * @return      true          if unsigned pd is supported.
+ *              false         if unsigned pd is not supported, capability query failed.
+ */
+
+bool is_unsignedpd_supported(int domain_id);
+
+/**
+ * is_valid_domain_id API: query a domain id is valid.
+ *
+ * @param[in]   domain value of domain in the queried.
+ * @param[in]   compute_only value of domain is only compared with CDSP domains supported by the target when enabled.
+ * @return      true          if value of domain is valid.
+ *              false         if value of domain is not valid.
+ */
+
+bool is_valid_domain_id(int domain_id, int compute_only);
+
+/**
+ * get_domain API: get domain struct from domain value.
+ *
+ * @param[in]  domain value of a domain
+ * @return     Returns domain struct of the domain if it is supported or else
+ *             returns NULL.
+ *
+ */
+
+domain * get_domain(int domain_id);
+
+/**
+ * get_domains_info API: get information for all the domains available on the device
+ *
+ * @param[in]  domain_type pointer to domain type
+ * @param[in]  num_domains pointer to number of domains
+ * @param[in]  domains_info pointer to save discovered domains information.
+ * @return     0 if query is successful.
+ *              non-zero if error, return value points to the error.
+ *
+ * It is user's responsibility to free the memory used to store the domains info whose address is present in domains_info before closing the application.
+ *
+ */
+
+int get_domains_info(char * domain_type, int * num_domains, fastrpc_domain ** domains_info);
+
+/**
+ * get_effective_domain_id API: get effective domain id for given session id
+ *
+ * @param[in]  domain_name pointer to domain name
+ * @param[in]  session_id
+ * @param[in]  effec_domain_id pointer to save obtained effective domain id.
+ * @return     0 if query is successful.
+ *              non-zero if error, return value points to the error.
+ *
+ */
+
+int get_effective_domain_id(char * domain_name, int session_id, int * effec_domain_id);
+
+/**
+ * is_async_fastrpc_supported API: query a domain id has async fastrpc supported or not
+ *
+ * @param[in]  domain_id value of a domain
+ * @return     Returns true or false stating support of Async FastRPC
+ *
+ */
+
+bool is_async_fastrpc_supported(int domain_id);
+
+/**
+ * is_status_notification_supported API: query the DSP for STATUS_NOTIFICATION_SUPPORT information
+ *
+ * @param[in]  domain_id value of a domain
+ * @return     Returns true or false stating status notification support information
+ *
+ */
+bool is_status_notification_supported(int domain_id);
+
+/**
+ * get_hmx_support_info API: query the DSP for HMX SUPPORT information
+ *
+ * @param[in]   domain_id value of a domain
+ * @param[out]  capability capability value of the attribute queried.
+ * @param[in]   attr value of the attribute to the queried.
+ * @return      0 if query is successful.
+ *              non-zero if error, return value points to the error.
+ *
+ */
+int get_hmx_support_info(int domain, uint32_t * capability, uint32_t attr);
+
+/**
+ * get_hex_arch_ver API: query the Hexagon processor architecture version information
+ *
+ * @param[in]   domain_id value of a domain
+ * @param[out]  Arch version (73, 75, ...)
+ * @return      0 if query is successful.
+ *              non-zero if error, return value points to the error.
+ *
+ */
+int get_hex_arch_ver(int domain, int * arch);
+
+/**
+ * get_hvx_support_info API: query the DSP for HVX SUPPORT information
+ *
+ * @param[in]   domain_id value of a domain
+ * @param[out]  capability capability value of the attribute queried.
+ * @param[in]   attr value of the attribute to the queried.
+ * @return      0 if query is successful.
+ *              non-zero if error, return value points to the error.
+ *
+ */
+int get_hvx_support_info(int domain, uint32_t * capability, uint32_t attr);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  //DSP_CAPABILITIES_UTILS_H

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -0,0 +1,40 @@
+cmake_minimum_required(VERSION 3.22.2)
+project(ggml-htp C CXX ASM)
+
+include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
+
+include_directories(
+    ${HEXAGON_SDK_ROOT}/incs
+    ${HEXAGON_SDK_ROOT}/incs/stddef
+    ${CMAKE_CURRENT_SOURCE_DIR}/../..
+    ${CMAKE_CURRENT_SOURCE_DIR}/..
+    ${CMAKE_CURRENT_SOURCE_DIR}
+    ${CMAKE_CURRENT_BINARY_DIR})
+
+set(HTP_LIB ggml-htp-${DSP_VERSION})
+
+add_library(${HTP_LIB} SHARED
+    main.c
+    htp_iface_skel.c
+    worker-pool.c
+    htp-dma.c
+    hvx-sigmoid.c
+    hvx-inverse.c
+    hvx-exp.c
+    hvx-utils.c
+    matmul-ops.c
+    binary-ops.c
+    unary-ops.c
+    softmax-ops.c
+    act-ops.c
+    rope-ops.c
+)
+
+target_compile_definitions(${HTP_LIB} PRIVATE
+    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,HTP_DEBUG=1,NDEBUG=1>)
+
+build_idl(htp_iface.idl ${HTP_LIB})
+
+set_target_properties(${HTP_LIB} PROPERTIES EXPORT_COMPILE_COMMANDS ON)
+
+install(TARGETS ${HTP_LIB})

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

@@ -0,0 +1,448 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#ifdef HTP_DEBUG
+#    define FARF_HIGH 1
+#endif
+#include <HAP_farf.h>
+#include <HAP_mem.h>
+#include <HAP_perf.h>
+#include <HAP_ps.h>
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <qurt_thread.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+#include "ops-utils.h"
+
+#define htp_act_preamble3              \
+    const uint32_t ne00 = src0->ne[0]; \
+    const uint32_t ne01 = src0->ne[1]; \
+    const uint32_t ne02 = src0->ne[2]; \
+    const uint32_t ne03 = src0->ne[3]; \
+                                       \
+    const uint32_t ne10 = src1->ne[0]; \
+    const uint32_t ne11 = src1->ne[1]; \
+    const uint32_t ne12 = src1->ne[2]; \
+    const uint32_t ne13 = src1->ne[3]; \
+                                       \
+    const uint32_t ne0 = dst->ne[0];   \
+    const uint32_t ne1 = dst->ne[1];   \
+    const uint32_t ne2 = dst->ne[2];   \
+    const uint32_t ne3 = dst->ne[3];   \
+                                       \
+    const uint32_t nb00 = src0->nb[0]; \
+    const uint32_t nb01 = src0->nb[1]; \
+    const uint32_t nb02 = src0->nb[2]; \
+    const uint32_t nb03 = src0->nb[3]; \
+                                       \
+    const uint32_t nb10 = src1->nb[0]; \
+    const uint32_t nb11 = src1->nb[1]; \
+    const uint32_t nb12 = src1->nb[2]; \
+    const uint32_t nb13 = src1->nb[3]; \
+                                       \
+    const uint32_t nb0 = dst->nb[0];   \
+    const uint32_t nb1 = dst->nb[1];   \
+    const uint32_t nb2 = dst->nb[2];   \
+    const uint32_t nb3 = dst->nb[3];
+
+#define htp_act_preamble2              \
+    const uint32_t ne00 = src0->ne[0]; \
+    const uint32_t ne01 = src0->ne[1]; \
+    const uint32_t ne02 = src0->ne[2]; \
+    const uint32_t ne03 = src0->ne[3]; \
+                                       \
+    const uint32_t ne0 = dst->ne[0];   \
+    const uint32_t ne1 = dst->ne[1];   \
+    const uint32_t ne2 = dst->ne[2];   \
+    const uint32_t ne3 = dst->ne[3];   \
+                                       \
+    const uint32_t nb00 = src0->nb[0]; \
+    const uint32_t nb01 = src0->nb[1]; \
+    const uint32_t nb02 = src0->nb[2]; \
+    const uint32_t nb03 = src0->nb[3]; \
+                                       \
+    const uint32_t nb0 = dst->nb[0];   \
+    const uint32_t nb1 = dst->nb[1];   \
+    const uint32_t nb2 = dst->nb[2];   \
+    const uint32_t nb3 = dst->nb[3];
+
+static void glu_swiglu_fp32_per_thread(const struct htp_tensor * src0,
+                                       const struct htp_tensor * src1,
+                                       struct htp_tensor *       dst,
+                                       const int32_t *           op_params,
+                                       struct htp_spad *         src0_spad,
+                                       struct htp_spad *         src1_spad,
+                                       struct htp_spad *         dst_spad,
+                                       uint32_t                  nth,
+                                       uint32_t                  ith,
+                                       uint32_t                  src0_nrows_per_thread) {
+    htp_act_preamble3;
+
+    size_t src0_row_size = nb01;
+    size_t src1_row_size = nb11;
+    size_t dst_row_size  = nb1;
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    int is_aligned = 1;
+    int opt_path   = 0;
+    if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) {
+        is_aligned = 0;
+        FARF(HIGH, "swiglu-f32: unaligned addresses in elementwise op, possibly slower execution\n");
+    }
+    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
+    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
+    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+
+    bool src1_valid = src1->ne[0];
+    if (!src1_valid) {
+        data_src1     = data_src0;
+        src1_row_size = src0_row_size;
+    }
+
+    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_row_size);
+    uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_row_size);
+    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_row_size);
+
+    const int32_t swapped = op_params[1];
+
+    const int nc = (src1_valid) ? ne0 : ne0 / 2;
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
+        const float * restrict src0 = (float *) (data_src0 + (ir * src0_row_size));
+        const float * restrict src1 = (float *) (data_src1 + (ir * src1_row_size));
+        float * restrict dst        = (float *) (data_dst + (ir * dst_row_size));
+
+        if (ir + 1 < src0_end_row) {
+            htp_l2fetch(src0 + src0_row_size, 1, src0_row_size, src0_row_size);
+        }
+
+        if (!src1_valid) {
+            src0 += swapped ? nc : 0;
+            src1 += swapped ? 0 : nc;
+        }
+
+        if (1 == opt_path) {
+            hvx_fast_sigmoid_f32((const uint8_t *) src0, (uint8_t *) src0_spad_data, nc);
+            hvx_mul_mul_f32_opt((const uint8_t *) src0, (const uint8_t *) src0_spad_data, (const uint8_t *) src1,
+                                (uint8_t *) dst, nc);
+        } else {
+            hvx_exp_f32((const uint8_t *) src0, src0_spad_data, nc, true);
+            hvx_add_scalar_f32(src0_spad_data, 1.0, src1_spad_data, nc);
+            hvx_inverse_f32(src1_spad_data, src0_spad_data, nc);
+
+            hvx_mul_f32((const uint8_t *) src0, src0_spad_data, dst_spad_data, nc);
+            hvx_mul_f32(dst_spad_data, (const uint8_t *) src1, (uint8_t *) dst, nc);
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "swiglu-f32 %d/%d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth, opt_path,
+         ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3,
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static void glu_swiglu_oai_fp32_per_thread(const struct htp_tensor * src0,
+                                           const struct htp_tensor * src1,
+                                           struct htp_tensor *       dst,
+                                           const int32_t *           op_params,
+                                           struct htp_spad *         src0_spad,
+                                           struct htp_spad *         src1_spad,
+                                           struct htp_spad *         dst_spad,
+                                           uint32_t                  nth,
+                                           uint32_t                  ith,
+                                           uint32_t                  src0_nrows_per_thread) {
+    htp_act_preamble3;
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const size_t src0_row_size = nb01;
+    const size_t src1_row_size = nb11;
+    const size_t dst_row_size  = nb1;
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) {
+        FARF(HIGH, "act-f32: unaligned addresses in activations op, possibly slower execution\n");
+    }
+
+    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
+    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
+    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+
+    bool src1_valid = src1->ne[0];
+    if (!src1_valid) {
+        data_src1 = data_src0;
+    }
+
+    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_row_size);
+    uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_row_size);
+    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_row_size);
+
+    const int32_t swapped = op_params[1];
+    const float   alpha   = ((const float *) (op_params))[2];
+    const float   limit   = ((const float *) (op_params))[3];
+
+    const int nc = (src1_valid) ? ne0 : ne0 / 2;
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
+        const float * restrict src0 = (float *) (data_src0 + (ir * src0_row_size));
+        const float * restrict src1 = (float *) (data_src1 + (ir * src1_row_size));
+        float * restrict dst        = (float *) (data_dst + (ir * dst_row_size));
+
+        if (ir + 1 < src0_end_row) {
+            htp_l2fetch(src0 + src0_row_size, 1, src0_row_size, src0_row_size);
+        }
+
+        if (!src1) {
+            src0 += swapped ? nc : 0;
+            src1 += swapped ? 0 : nc;
+        }
+
+        // x (src0_spad_data) = std::min(src0_p[k], limit);
+        hvx_min_scalar_f32((const uint8_t *) src0, limit, src0_spad_data, nc);
+        // y1 (src1_spad_data) = std::clamp(src1_p[k], -limit, limit);
+        hvx_clamp_scalar_f32((const uint8_t *) src1, limit, limit, src1_spad_data, nc);
+        // y (src1_spad_data)  = y1 + 1.f
+        hvx_add_scalar_f32(src1_spad_data, 1.0, src1_spad_data, nc);
+        // x1 (dst_spad_data) = alpha * (x)
+        hvx_mul_scalar_f32(src0_spad_data, alpha, dst_spad_data, nc);
+        // x2 (dst_spad_data) = expf(-x1)
+        hvx_exp_f32(dst_spad_data, dst_spad_data, nc, true);
+        // x3 (dst_spad_data) = x2 + 1.f
+        hvx_add_scalar_f32(dst_spad_data, 1.0, dst_spad_data, nc);
+        // x4 (dst_spad_data) = 1 / x3
+        hvx_inverse_f32(dst_spad_data, dst_spad_data, nc);
+        // out_glu(dst_spad_data) = x * x4
+        hvx_mul_f32(src0_spad_data, dst_spad_data, dst_spad_data, nc);
+        // out = out_glu * (y + 1.f);
+        hvx_mul_f32(dst_spad_data, src1_spad_data, (uint8_t *) dst, nc);
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "swiglu-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth, src0->ne[0],
+         src0->ne[1], src0->ne[2], src0->ne[3], src0_start_row, src0_end_row, src1->ne[0], src1->ne[1], src1->ne[2],
+         src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static void unary_silu_fp32_per_thread(const struct htp_tensor * src0,
+                                       struct htp_tensor *       dst,
+                                       const int32_t *           op_params,
+                                       struct htp_spad *         src0_spad,
+                                       struct htp_spad *         dst_spad,
+                                       uint32_t                  nth,
+                                       uint32_t                  ith,
+                                       uint32_t                  src0_nrows_per_thread) {
+    htp_act_preamble2;
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const size_t src0_row_size = nb01;
+    const size_t dst_row_size  = nb1;
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    int is_aligned = 1;
+    int opt_path   = 0;
+    if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) {
+        is_aligned = 0;
+        FARF(HIGH, "silu-f32: unaligned addresses in elementwise op, possibly slower execution\n");
+    }
+    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
+    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+
+    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_row_size);
+    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_row_size);
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
+        const float * restrict src0 = (float *) (data_src0 + (ir * src0_row_size));
+        float * restrict dst        = (float *) (data_dst + (ir * dst_row_size));
+
+        if (ir + 1 < src0_end_row) {
+            htp_l2fetch(src0 + src0_row_size, 1, src0_row_size, src0_row_size);
+        }
+
+        if (1 == opt_path) {
+            hvx_fast_sigmoid_f32((const uint8_t *) src0, (uint8_t *) src0_spad_data, ne0);
+            hvx_mul_f32_opt((const uint8_t *) src0, src0_spad_data, (uint8_t *) dst, ne0);
+        } else {
+            hvx_exp_f32((const uint8_t *) src0, src0_spad_data, ne0, true);
+            hvx_add_scalar_f32(src0_spad_data, 1.0, dst_spad_data, ne0);
+            hvx_inverse_f32(dst_spad_data, src0_spad_data, ne0);
+
+            hvx_mul_f32((const uint8_t *) src0, src0_spad_data, (uint8_t *) dst, ne0);
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "silu-f32 %d/%d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, opt_path, ne00, ne01, ne02,
+         ne03, src0_start_row, src0_end_row, ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static void unary_silu_fp32(unsigned int n, unsigned int i, void * data) {
+    struct htp_ops_context * octx = (struct htp_ops_context *) data;
+    unary_silu_fp32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i,
+                               octx->src0_nrows_per_thread);
+}
+
+static void glu_swiglu_fp32(unsigned int n, unsigned int i, void * data) {
+    struct htp_ops_context * octx = (struct htp_ops_context *) data;
+    glu_swiglu_fp32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
+                               &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread);
+}
+
+static void glu_swiglu_oai_fp32(unsigned int n, unsigned int i, void * data) {
+    struct htp_ops_context * octx = (struct htp_ops_context *) data;
+    glu_swiglu_oai_fp32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
+                                   &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread);
+}
+
+static int execute_op_activations_fp32(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    if (((src0->ne[0] * SIZEOF_FP32) != src0->nb[1]) || ((dst->ne[0] * SIZEOF_FP32) != dst->nb[1])) {
+        FARF(ERROR, "Non-contiguous tensors are not supported at this time \n");
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    worker_callback_t act_op_func;
+    const char *      op_type = NULL;
+
+    switch (octx->op) {
+        case HTP_OP_UNARY_SILU:
+            act_op_func = unary_silu_fp32;
+            op_type     = "silu-f32";
+            break;
+
+        case HTP_OP_GLU_SWIGLU:
+            act_op_func = glu_swiglu_fp32;
+            op_type     = "swiglu-f32";
+            break;
+
+        case HTP_OP_GLU_SWIGLU_OAI:
+            act_op_func = glu_swiglu_oai_fp32;
+            op_type     = "swiglu-oai-f32";
+            break;
+
+        default:
+            FARF(ERROR, "Unsupported activations Op %u\n", octx->op);
+            return HTP_STATUS_NO_SUPPORT;
+    }
+
+    const uint32_t n_threads  = octx->n_threads;
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+
+    const size_t src0_row_size = src0->nb[1];
+    const size_t src1_row_size = src1->ne[0] ? src1->nb[1] : src0->nb[1];
+    const size_t dst_row_size  = dst->nb[1];
+
+    // VTCM scratchpads for all tensors
+    // N rows per thread, padded to HVX vector size
+    octx->dst_spad.size  = htp_round_up(dst_row_size, 128) * octx->n_threads;
+    octx->src0_spad.size = htp_round_up(src0_row_size, 128) * octx->n_threads;
+    octx->src1_spad.size = htp_round_up(src1_row_size, 128) * octx->n_threads;
+
+    size_t spad_size = octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;
+
+    if (src1->ne[0]) {
+        FARF(HIGH,
+             "%s: %ux%ux%ux%u x %ux%ux%ux%u -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n",
+             op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
+             src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], octx->src0_spad.size, octx->src1_spad.size,
+             octx->dst_spad.size);
+    } else {
+        FARF(HIGH, "%s: %ux%ux%ux%u -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n", op_type,
+             src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+             octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
+    }
+
+    // Make sure the reserved vtcm size is sufficient
+    if (octx->ctx->vtcm_size < spad_size) {
+        FARF(ERROR, "act-%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
+             spad_size);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
+
+    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
+        uint32_t n_jobs = MIN(n_threads, src0_nrows);
+
+        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+        worker_pool_run_func(octx->ctx->worker_pool, act_op_func, octx, n_jobs);
+    }
+
+    return err;
+}
+
+int op_activations(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    switch (octx->src0.type) {
+        case HTP_TYPE_F32:
+            err = execute_op_activations_fp32(octx);
+            break;
+
+        default:
+            err = HTP_STATUS_NO_SUPPORT;
+            break;
+    }
+
+    return err;
+}

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

@@ -0,0 +1,344 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#ifdef HTP_DEBUG
+#    define FARF_HIGH 1
+#endif
+
+#include <HAP_farf.h>
+#include <HAP_mem.h>
+#include <HAP_perf.h>
+#include <HAP_ps.h>
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <qurt_thread.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+#include "ops-utils.h"
+
+typedef void (*hvx_elemwise_f32_func)(const uint8_t * src0,
+                                      const uint8_t * src1,
+                                      uint8_t *       data_dst,
+                                      const int       num_elems);
+
+static hvx_elemwise_f32_func func_table_HVX[]     = { hvx_mul_f32, hvx_add_f32, hvx_sub_f32 };
+static hvx_elemwise_f32_func func_table_HVX_opt[] = { hvx_mul_f32_opt, hvx_add_f32_opt, hvx_sub_f32_opt };
+
+#define htp_binary_preamble            \
+    const uint32_t ne00 = src0->ne[0]; \
+    const uint32_t ne01 = src0->ne[1]; \
+    const uint32_t ne02 = src0->ne[2]; \
+    const uint32_t ne03 = src0->ne[3]; \
+                                       \
+    const uint32_t ne10 = src1->ne[0]; \
+    const uint32_t ne11 = src1->ne[1]; \
+    const uint32_t ne12 = src1->ne[2]; \
+    const uint32_t ne13 = src1->ne[3]; \
+                                       \
+    const uint32_t ne0 = dst->ne[0];   \
+    const uint32_t ne1 = dst->ne[1];   \
+    const uint32_t ne2 = dst->ne[2];   \
+    const uint32_t ne3 = dst->ne[3];   \
+                                       \
+    const uint32_t nb00 = src0->nb[0]; \
+    const uint32_t nb01 = src0->nb[1]; \
+    const uint32_t nb02 = src0->nb[2]; \
+    const uint32_t nb03 = src0->nb[3]; \
+                                       \
+    const uint32_t nb10 = src1->nb[0]; \
+    const uint32_t nb11 = src1->nb[1]; \
+    const uint32_t nb12 = src1->nb[2]; \
+    const uint32_t nb13 = src1->nb[3]; \
+                                       \
+    const uint32_t nb0 = dst->nb[0];   \
+    const uint32_t nb1 = dst->nb[1];   \
+    const uint32_t nb2 = dst->nb[2];   \
+    const uint32_t nb3 = dst->nb[3];
+
+static void binary_job_f32_per_thread(const struct htp_tensor * src0,
+                                      const struct htp_tensor * src1,
+                                      struct htp_tensor *       dst,
+                                      uint8_t *                 spad_data,
+                                      uint32_t                  nth,
+                                      uint32_t                  ith,
+                                      uint32_t                  src0_nrows_per_thread,
+                                      enum htp_op               op) {
+    htp_binary_preamble;
+
+    const size_t src0_row_size = nb01;
+    const size_t src1_row_size = nb11;
+    const size_t dst_row_size  = nb1;
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+    const uint32_t src1_nrows = ne11 * ne12 * ne13;  // src1 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    int is_aligned = 1;
+    int opt_path   = 0;
+    if ((0 == htp_is_aligned((void *) src0->data, VLEN)) || (0 == htp_is_aligned((void *) src1->data, VLEN)) ||
+        (0 == htp_is_aligned((void *) dst->data, VLEN))) {
+        FARF(HIGH, "binary-f32: unaligned addresses in elementwise op, possibly slower execution\n");
+        is_aligned = 0;
+    }
+    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    hvx_elemwise_f32_func func_HVX = (1 == opt_path) ? func_table_HVX_opt[op] : func_table_HVX[op];
+
+    uint8_t * restrict spad_data_th = spad_data + (ith * src0_row_size);
+
+    const uint32_t nr0 = ne00 / ne10;
+
+    const uint8_t * restrict src0_ptr = (const uint8_t *) src0->data + (src0_start_row * src0_row_size);
+    uint8_t * restrict dst_ptr        = (uint8_t *) dst->data + (src0_start_row * dst_row_size);
+
+    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
+    const uint8_t * restrict src1_ptr  = NULL;
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
+        src1_ptr = data_src1 + (ir % src1_nrows) * src1_row_size;
+
+        if (ir + 1 < src0_end_row) {
+            htp_l2fetch(src0_ptr + ne00, 1, src0_row_size, src0_row_size);
+            if (src1_row_size == src0_row_size) {
+                htp_l2fetch(src1_ptr, 1, src1_row_size, src1_row_size);
+            }
+        }
+
+        if (nr0 > 1) {
+            if ((1 == is_aligned) && (nr0 == ne00)) {
+                hvx_bcast_fp32_a(spad_data_th, *(float *) src1_ptr, nr0);
+            } else {
+                for (uint32_t r = 0; r < nr0; r++) {
+                    memcpy(spad_data_th + r * nb11, (const uint8_t *) src1_ptr, nb11);
+                }
+            }
+            func_HVX((const uint8_t *) src0_ptr, (const uint8_t *) spad_data_th, (uint8_t *) dst_ptr, ne00);
+        } else {
+            func_HVX((const uint8_t *) src0_ptr, (const uint8_t *) src1_ptr, (uint8_t *) dst_ptr, ne00);
+        }
+
+        src0_ptr += src0_row_size;
+        dst_ptr += dst_row_size;
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "binary-f32 %d/%d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth, opt_path,
+         ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3,
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static void binary_add_id_job_f32_per_thread(const struct htp_tensor * src0,
+                                             const struct htp_tensor * src1,
+                                             const struct htp_tensor * src2,
+                                             struct htp_tensor *       dst,
+                                             uint8_t *                 spad_data,
+                                             uint32_t                  nth,
+                                             uint32_t                  ith,
+                                             uint32_t                  src0_nrows_per_thread,
+                                             hvx_elemwise_f32_func     func_HVX) {
+    htp_binary_preamble;
+
+    const size_t src0_row_size = nb01;
+    const size_t src1_row_size = nb11;
+    const size_t dst_row_size  = nb1;
+
+    const uint32_t ne02_ne01  = ne02 * ne01;
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    if ((0 == htp_is_aligned((void *) src0->data, VLEN)) || (0 == htp_is_aligned((void *) src1->data, VLEN)) ||
+        (0 == htp_is_aligned((void *) dst->data, VLEN))) {
+        FARF(HIGH, "add-id-f32: unaligned addresses, possibly slower execution\n");
+    }
+
+    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
+    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
+    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
+        // src0 indices
+        const uint32_t i03 = ir / ne02_ne01;
+        const uint32_t i02 = (ir - i03 * ne02_ne01) / ne01;
+        const uint32_t i01 = (ir - i03 * ne02_ne01 - i02 * ne01);
+
+        // src1 indices
+        const int i11 = *(int32_t *) ((char *) src2->data + i01 * src2->nb[0] + i02 * src2->nb[1]);
+        assert(i11 >= 0 && i11 < ne11);
+
+        float * restrict dst_ptr        = (float *) (data_dst + i03 * nb3 + i02 * nb2 + i01 * nb1);
+        const float * restrict src0_ptr = (const float *) (data_src0 + i03 * nb03 + i02 * nb02 + i01 * nb01);
+        const float * restrict src1_ptr = (const float *) (data_src1 + 0 + 0 + i11 * nb11);
+
+        if (ir + 1 < src0_end_row) {
+            htp_l2fetch(src0_ptr + ne00, 1, src0_row_size, src0_row_size);
+            if (src1_row_size == src0_row_size) {
+                htp_l2fetch(src1_ptr + ne10, 1, src1_row_size, src1_row_size);
+            }
+        }
+
+        const uint32_t nr0 = ne00 / ne10;
+        if (nr0 > 1) {
+            for (uint32_t r = 0; r < nr0; r++) {
+                memcpy(spad_data + r * nb10, (const uint8_t *) src1_ptr, nb10);
+            }
+            func_HVX((const uint8_t *) src0_ptr, (const uint8_t *) spad_data, (uint8_t *) dst_ptr, ne00);
+        } else {
+            func_HVX((const uint8_t *) src0_ptr, (const uint8_t *) src1_ptr, (uint8_t *) dst_ptr, ne00);
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "add-id-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u (%ux%ux%ux%u) -> %ux%ux%ux%u usec %u\n", ith, nth,
+         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src0_start_row, src0_end_row, src1->ne[0], src1->ne[1],
+         src1->ne[2], src1->ne[3], src2->ne[0], src2->ne[1], src2->ne[2], src2->ne[3], dst->ne[0], dst->ne[1],
+         dst->ne[2], dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static void binary_job_dispatcher_f32(unsigned int n, unsigned int i, void * data) {
+    struct htp_ops_context * octx = (struct htp_ops_context *) data;
+
+    switch (octx->op) {
+        case HTP_OP_MUL:
+        case HTP_OP_ADD:
+        case HTP_OP_SUB:
+            binary_job_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->src1_spad.data, n, i,
+                                      octx->src0_nrows_per_thread, octx->op);
+            break;
+
+        case HTP_OP_ADD_ID:
+            binary_add_id_job_f32_per_thread(&octx->src0, &octx->src1, &octx->src2, &octx->dst, octx->src0_spad.data, n,
+                                             i, octx->src0_nrows_per_thread, hvx_add_f32);
+            break;
+
+        default:
+            FARF(ERROR, "Unknown Binary Op %u", octx->op);
+            break;
+    }
+}
+
+static int execute_op_binary_f32(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    worker_callback_t binary_op_func;
+    const char *      op_type = NULL;
+
+    switch (octx->op) {
+        case HTP_OP_MUL:
+            binary_op_func = binary_job_dispatcher_f32;
+            op_type        = "mul-f32";
+            break;
+
+        case HTP_OP_ADD:
+            binary_op_func = binary_job_dispatcher_f32;
+            op_type        = "add-f32";
+            break;
+
+        case HTP_OP_SUB:
+            binary_op_func = binary_job_dispatcher_f32;
+            op_type        = "sub-f32";
+            break;
+
+        case HTP_OP_ADD_ID:
+            binary_op_func = binary_job_dispatcher_f32;
+            op_type        = "add-id-f32";
+            break;
+
+        default:
+            FARF(ERROR, "Unsupported binary-Op %u\n", octx->op);
+            return HTP_STATUS_NO_SUPPORT;
+    }
+
+    const int      n_threads  = octx->n_threads;
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+
+    const size_t src0_row_size = src0->nb[1];
+    const size_t src1_row_size = src1->nb[1];
+    const size_t dst_row_size  = dst->nb[1];
+
+    // VTCM scratchpads for all tensors
+    octx->dst_spad.size  = htp_round_up(dst_row_size, 128) * n_threads;
+    octx->src0_spad.size = htp_round_up(src0_row_size, 128) * n_threads;
+    octx->src1_spad.size = htp_round_up(src1_row_size, 128) * n_threads;
+
+    size_t spad_size = octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;
+
+    FARF(HIGH,
+         "%s: (%ux%ux%ux%u) * (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n",
+         op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
+         src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], octx->src0_spad.size, octx->src1_spad.size,
+         octx->dst_spad.size);
+
+    // Make sure the reserved vtcm size is sufficient
+    if (octx->ctx->vtcm_size < spad_size) {
+        FARF(ERROR, "binary-%s : current VTCM reservation %zu is too small, needed %zu\n", op_type,
+             octx->ctx->vtcm_size, spad_size);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
+
+    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
+        uint32_t n_jobs = MIN(n_threads, src0_nrows);
+
+        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+
+        worker_pool_run_func(octx->ctx->worker_pool, binary_op_func, octx, n_jobs);
+    }
+
+    return err;
+}
+
+int op_binary(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    switch (octx->src0.type) {
+        case HTP_TYPE_F32:
+            err = execute_op_binary_f32(octx);
+            break;
+
+        default:
+            err = HTP_STATUS_NO_SUPPORT;
+            break;
+    }
+
+    return err;
+}

ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake

@@ -0,0 +1,157 @@
+if (HEXAGON_TOOLCHAIN_INCLUDED)
+  return()
+endif()
+set(HEXAGON_TOOLCHAIN_INCLUDED true)
+
+#Cross Compiling for Hexagon
+set(HEXAGON TRUE)
+set(CMAKE_SYSTEM_NAME QURT)
+set(CMAKE_SYSTEM_PROCESSOR Hexagon)
+set(CMAKE_SYSTEM_VERSION "1") #${HEXAGON_PLATFORM_LEVEL})
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
+set(CUSTOM_RUNELF_PATH "")
+
+#To fix backward compatibility with EAI addon.
+if (NOT HEXAGON_SDK_ROOT)
+    set(HEXAGON_SDK_ROOT $ENV{HEXAGON_SDK_ROOT})
+endif()
+
+if (NOT HEXAGON_TOOLS_ROOT)
+    if (DEFINED ENV{HEXAGON_TOOLS_ROOT})
+        set(HEXAGON_TOOLS_ROOT $ENV{HEXAGON_TOOLS_ROOT})
+    endif()
+    if(NOT HEXAGON_TOOLS_ROOT)
+        set(HEXAGON_TOOLS_ROOT $ENV{DEFAULT_HEXAGON_TOOLS_ROOT})
+    endif()
+endif()
+
+file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
+file(TO_CMAKE_PATH "${HEXAGON_SDK_ROOT}"   HEXAGON_SDK_ROOT)
+
+#Get the Binary extension of the Hexagon Toolchain
+if(CMAKE_HOST_SYSTEM_NAME STREQUAL Windows)
+    set(HEXAGON_TOOLCHAIN_SUFFIX .exe)
+endif()
+message(DEBUG "CMAKE_HOST_SYSTEM_NAME:${CMAKE_HOST_SYSTEM_NAME}")
+
+include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_arch.cmake)
+
+set(HEXAGON_TOOLCHAIN ${HEXAGON_TOOLS_ROOT})
+set(HEXAGON_LIB_DIR "${HEXAGON_TOOLCHAIN}/Tools/target/hexagon/lib")
+set(HEXAGON_ISS_DIR ${HEXAGON_TOOLCHAIN}/Tools/lib/iss)
+
+set(CMAKE_TRY_COMPILE_PLATFORM_VARIABLES
+    HEXAGON_SDK_ROOT
+    HEXAGON_TOOLS_ROOT
+)
+
+#QURT Related includes and linker flags
+set(V_ARCH ${HEXAGON_ARCH})
+set(_QURT_INSTALL_DIR "${HEXAGON_SDK_ROOT}/rtos/qurt/ADSP${V_ARCH}MP${V_ARCH_EXTN}")
+set(_QURT_INSTALL_DIR "${HEXAGON_SDK_ROOT}/rtos/qurt/compute${V_ARCH}${V_ARCH_EXTN}")
+
+if( ${TREE} MATCHES PAKMAN )
+    set(_QURT_INSTALL_DIR "${QURT_IMAGE_DIR}/compute${V_ARCH}${V_ARCH_EXTN}")
+endif()
+message(DEBUG "_QURT_INSTALL_DIR:${_QURT_INSTALL_DIR}")
+set(RTOS_DIR ${_QURT_INSTALL_DIR})
+set(QCC_DIR "${HEXAGON_QCC_DIR}/${V_ARCH}/G0")
+set(TARGET_DIR "${HEXAGON_LIB_DIR}/${V_ARCH}/G0")
+
+include_directories(
+    ${_QURT_INSTALL_DIR}/include
+    ${_QURT_INSTALL_DIR}/include/qurt
+    ${_QURT_INSTALL_DIR}/include/posix
+    )
+
+set(QURT_START_LINK_LIBS)
+set(QURT_START_LINK_LIBS
+    "${TARGET_DIR}/init.o"
+    "${RTOS_DIR}/lib/crt1.o"
+    "${RTOS_DIR}/lib/debugmon.o"
+    "${RTOS_DIR}/lib/libqurt.a"
+    "${TARGET_DIR}/libc.a"
+    "${TARGET_DIR}/libqcc.a"
+    "${TARGET_DIR}/libhexagon.a"
+    "${RTOS_DIR}/lib/libqurtcfs.a"
+    "${RTOS_DIR}/lib/libtimer_island.a"
+    "${RTOS_DIR}/lib/libtimer_main.a"
+    "${RTOS_DIR}/lib/libposix.a"
+    )
+STRING(REPLACE ";" " " QURT_START_LINK_LIBS "${QURT_START_LINK_LIBS}")
+
+set(QURT_END_LINK_LIBS
+    ${TARGET_DIR}/fini.o
+    )
+
+#Non QURT related includes and linker flags
+
+set(TARGET_DIR_NOOS "${HEXAGON_TOOLCHAIN}/Tools/target/hexagon/lib/${HEXAGON_ARCH}")
+
+if (NOT NO_WRAP_MEM_API)
+    set(WRAP_MALLOC   -Wl,--wrap=malloc)
+    set(WRAP_CALLOC   -Wl,--wrap=calloc)
+    set(WRAP_FREE     -Wl,--wrap=free)
+    set(WRAP_REALLOC  -Wl,--wrap=realloc)
+    set(WRAP_MEMALIGN -Wl,--wrap=memalign)
+endif()
+
+set(PIC_SHARED_LD_FLAGS
+    -mcpu=${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH}
+    -G0
+    -fpic
+    -Wl,-Bsymbolic
+    -Wl,-L${TARGET_DIR_NOOS}/G0/pic
+    -Wl,-L${HEXAGON_TOOLCHAIN}/Tools/target/hexagon/lib/
+    -Wl,--no-threads ${WRAP_MALLOC} ${WRAP_CALLOC} ${WRAP_FREE} ${WRAP_REALLOC} ${WRAP_MEMALIGN}
+    -shared
+    "-o <TARGET> <SONAME_FLAG><TARGET_SONAME>"
+    "<LINK_FLAGS>"
+    -Wl,--start-group
+    "<OBJECTS>"
+    "<LINK_LIBRARIES>"
+    -Wl,--end-group
+    -lc
+    )
+STRING(REPLACE ";" " " PIC_SHARED_LD_FLAGS "${PIC_SHARED_LD_FLAGS}")
+
+set(HEXAGON_PIC_SHARED_LINK_OPTIONS "${PIC_SHARED_LD_FLAGS}")
+
+#System include paths
+include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/incs)
+include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/incs/stddef)
+include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/ipc/fastrpc/incs)
+
+#LLVM toolchain setup
+#Compiler paths, options and architecture
+set(CMAKE_C_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang${HEXAGON_TOOLCHAIN_SUFFIX})
+set(CMAKE_CXX_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang++${HEXAGON_TOOLCHAIN_SUFFIX})
+set(CMAKE_AR ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-ar${HEXAGON_TOOLCHAIN_SUFFIX})
+set(CMAKE_ASM_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang++${HEXAGON_TOOLCHAIN_SUFFIX})
+set(HEXAGON_LINKER ${CMAKE_C_COMPILER})
+set(CMAKE_PREFIX_PATH ${HEXAGON_TOOLCHAIN}/Tools/target/hexagon)
+
+set(CMAKE_SHARED_LIBRARY_SONAME_C_FLAG   "-Wl,-soname,")
+set(CMAKE_SHARED_LIBRARY_SONAME_CXX_FLAG "-Wl,-soname,")
+
+#Compiler Options
+set(COMMON_FLAGS "-mcpu=hexagon${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH} -fvectorize -Wall -Werror -fno-zero-initialized-in-bss -G0 -fdata-sections -fpic ${XQF_ARGS}")
+
+set(CMAKE_CXX_FLAGS_DEBUG          "${COMMON_FLAGS} -O0 -D_DEBUG -g")
+set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS} -O3 -g")
+set(CMAKE_CXX_FLAGS_RELEASE        "${COMMON_FLAGS} -O3")
+
+set(CMAKE_C_FLAGS_DEBUG            "${COMMON_FLAGS} -O0 -D_DEBUG -g")
+set(CMAKE_C_FLAGS_RELWITHDEBINFO   "${COMMON_FLAGS} -O3 -g")
+set(CMAKE_C_FLAGS_RELEASE          "${COMMON_FLAGS} -O3")
+
+set(CMAKE_ASM_FLAGS_DEBUG          "${COMMON_FLAGS} ${CMAKE_CXX_FLAGS_DEBUG}")
+set(CMAKE_ASM_FLAGS_RELEASE        "${COMMON_FLAGS} ${CMAKE_CXX_FLAGS_RELEASE}")
+set(CMAKE_ASM_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS} ${CMAKE_CXX_FLAGS_RELWITHDEBINFO}" )
+
+#Linker Options
+set(CMAKE_C_CREATE_SHARED_LIBRARY   "${HEXAGON_LINKER} ${HEXAGON_PIC_SHARED_LINK_OPTIONS}")
+set(CMAKE_CXX_CREATE_SHARED_LIBRARY "${HEXAGON_LINKER} ${HEXAGON_PIC_SHARED_LINK_OPTIONS}")

ggml/src/ggml-hexagon/htp/htp-ctx.h

@@ -0,0 +1,40 @@
+#ifndef HTP_CTX_H
+#define HTP_CTX_H
+
+#include "htp-dma.h"
+#include "worker-pool.h"
+
+#include <assert.h>
+#include <dspqueue.h>
+#include <stdatomic.h>
+#include <stdint.h>
+
+#define HTP_MAX_NTHREADS 10
+
+// FIXME: move these into matmul-ops
+#define HTP_SPAD_SRC0_NROWS 16
+#define HTP_SPAD_SRC1_NROWS 16
+#define HTP_SPAD_DST_NROWS  2
+
+// Main context for htp DSP backend
+struct htp_context {
+    dspqueue_t            queue;
+    dma_queue *           dma[HTP_MAX_NTHREADS];
+    worker_pool_context_t worker_pool;
+    uint32_t              n_threads;
+
+    int thread_id;
+    int thread_prio;
+
+    uint8_t * vtcm_base;
+    size_t    vtcm_size;
+    uint32_t  vtcm_rctx;
+
+    atomic_bool vtcm_valid;
+    atomic_bool vtcm_inuse;
+    atomic_bool vtcm_needs_release;
+
+    uint32_t opmask;
+};
+
+#endif /* HTP_CTX_H */

ggml/src/ggml-hexagon/htp/htp-dma.c

@@ -0,0 +1,69 @@
+#include "htp-dma.h"
+
+#include <stdbool.h>
+#include <stdlib.h>
+#include <string.h>
+
+#pragma clang diagnostic ignored "-Wunused-function"
+
+static inline uint32_t pow2_ceil(uint32_t x) {
+    if (x <= 1) {
+        return 1;
+    }
+    int p = 2;
+    x--;
+    while (x >>= 1) {
+        p <<= 1;
+    }
+    return p;
+}
+
+dma_queue * dma_queue_create(size_t capacity) {
+    dma_queue * q = (dma_queue *) memalign(32, sizeof(dma_queue));
+    if (q == NULL) {
+        FARF(ERROR, "%s: failed to allocate DMA queue\n", __FUNCTION__);
+        return NULL;
+    }
+
+    capacity = pow2_ceil(capacity);
+
+    memset(q, 0, sizeof(dma_queue));
+    q->capacity = capacity;
+    q->idx_mask = capacity - 1;
+
+    q->desc = (hexagon_udma_descriptor_type1_t *) memalign(64, capacity * sizeof(hexagon_udma_descriptor_type1_t));
+    memset(q->desc, 0, capacity * sizeof(hexagon_udma_descriptor_type1_t));
+
+    q->dst = (void **) memalign(4, capacity * sizeof(void *));
+    memset(q->dst, 0, capacity * sizeof(void *));
+
+    q->tail = &q->desc[capacity - 1];
+
+    if (!q->desc && !q->dst) {
+        FARF(ERROR, "%s: failed to allocate DMA queue items\n", __FUNCTION__);
+        return NULL;
+    }
+
+    FARF(HIGH, "dma-queue: capacity %u\n", capacity);
+
+    return q;
+}
+
+void dma_queue_delete(dma_queue * q) {
+    if (!q) {
+        return;
+    }
+    free(q->desc);
+    free(q->dst);
+    free(q);
+}
+
+void dma_queue_flush(dma_queue * q) {
+    while (1) {
+        uint32_t s = dmwait() & 0x3;
+        if (s == HEXAGON_UDMA_DM0_STATUS_IDLE) {
+            break;
+        }
+    }
+    q->tail = NULL;
+}

ggml/src/ggml-hexagon/htp/htp-dma.h

@@ -0,0 +1,119 @@
+#ifndef HTP_DMA_H
+#define HTP_DMA_H
+
+#include <HAP_farf.h>
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <stdbool.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+    hexagon_udma_descriptor_type1_t * desc;  // descriptor pointers
+    hexagon_udma_descriptor_type1_t * tail;  // tail pointer
+    void **                           dst;   // dst pointers
+    uint32_t                          push_idx;
+    uint32_t                          pop_idx;
+    uint32_t                          capacity;
+    uint32_t                          idx_mask;
+} dma_queue;
+
+dma_queue * dma_queue_create(size_t capacity);
+void        dma_queue_delete(dma_queue * q);
+void        dma_queue_flush(dma_queue * q);
+
+// TODO: technically we don't need these and could use Q6_dmstart/wait/etc instead
+// but those do not seem to always compiler properly.
+static inline void dmstart(void * next) {
+    asm volatile(" release(%0):at" : : "r"(next));
+    asm volatile(" dmstart(%0)" : : "r"(next));
+}
+
+static inline void dmlink(void * cur, void * next) {
+    asm volatile(" release(%0):at" : : "r"(next));
+    asm volatile(" dmlink(%0, %1)" : : "r"(cur), "r"(next));
+}
+
+static inline unsigned int dmpoll(void) {
+    unsigned int ret = 0;
+    asm volatile(" %0 = dmpoll" : "=r"(ret) : : "memory");
+    return ret;
+}
+
+static inline unsigned int dmwait(void) {
+    unsigned int ret = 0;
+    asm volatile(" %0 = dmwait" : "=r"(ret) : : "memory");
+    return ret;
+}
+
+static inline bool dma_queue_push(dma_queue *  q,
+                                  void *       dst,
+                                  const void * src,
+                                  size_t       dst_row_size,
+                                  size_t       src_row_size,
+                                  size_t       nrows) {
+    if (((q->push_idx + 1) & q->idx_mask) == q->pop_idx) {
+        return false;
+    }
+
+    hexagon_udma_descriptor_type1_t * desc = &q->desc[q->push_idx];
+
+    desc->next           = NULL;
+    desc->length         = 0;
+    desc->desctype       = HEXAGON_UDMA_DESC_DESCTYPE_TYPE1;
+    desc->dstbypass      = 1;
+    desc->srcbypass      = 1;
+    desc->order          = 0;
+    desc->dstate         = HEXAGON_UDMA_DESC_DSTATE_INCOMPLETE;
+    desc->src            = (void *) src;
+    desc->dst            = (void *) dst;
+    desc->allocation     = 0;
+    desc->padding        = 0;
+    desc->roiwidth       = src_row_size;
+    desc->roiheight      = nrows;
+    desc->srcstride      = src_row_size;
+    desc->dststride      = dst_row_size;
+    desc->srcwidthoffset = 0;
+    desc->dstwidthoffset = 0;
+
+    q->dst[q->push_idx] = dst;
+
+    dmlink(q->tail, desc);
+    q->tail = desc;
+
+    // FARF(ERROR, "dma-push: i %u len %u dst %p src %p\n", q->push_idx, len, dst, src);
+    q->push_idx = (q->push_idx + 1) & q->idx_mask;
+    return true;
+}
+
+static inline uint8_t * dma_queue_pop(dma_queue * q) {
+    if (q->push_idx == q->pop_idx) {
+        return NULL;
+    }
+
+    hexagon_udma_descriptor_type1_t * desc = &q->desc[q->pop_idx];
+
+    // Wait for desc to complete
+    while (1) {
+        dmpoll();
+        if (desc->dstate == HEXAGON_UDMA_DESC_DSTATE_COMPLETE) {
+            break;
+        }
+        // FARF(ERROR, "dma-pop: waiting for DMA : %u\n", q->pop_idx);
+    }
+
+    uint8_t * dst = (uint8_t *) q->dst[q->pop_idx];
+
+    // FARF(ERROR, "dma-pop: i %u dst %p\n", q->pop_idx, dst);
+    q->pop_idx = (q->pop_idx + 1) & q->idx_mask;
+    return dst;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* HTP_DMA_H */

ggml/src/ggml-hexagon/htp/htp-msg.h

@@ -0,0 +1,156 @@
+#ifndef HTP_MSG_H
+#define HTP_MSG_H
+
+#include <assert.h>
+
+// ggml-common.h must be included prio to this header
+
+// Mask to enable various stages of the Ops.
+// Used for debugging and profiling.
+enum {
+    HTP_OPMASK_QUEUE    = (1 << 0),  // Enable Queueing (ie calls into the DSP)
+    HTP_OPMASK_QUANTIZE = (1 << 1),  // Enable Quantize
+    HTP_OPMASK_COMPUTE  = (1 << 2),  // Enable Compute
+};
+
+// Op flags
+enum {
+    HTP_OPFLAGS_SKIP_QUANTIZE = (1 << 0),  // Skip dynamic quantization (reuse quantized tensors)
+    HTP_OPFLAGS_SKIP_COMPUTE  = (1 << 1),  // Skip actual computation (used for profiling)
+    HTP_OPFLAGS_EARLY_WAKEUP  = (1 << 2)   // Send early wakeup notification
+};
+
+enum htp_status {
+    HTP_STATUS_OK             = 1,
+    HTP_STATUS_INTERNAL_ERR   = 2,
+    HTP_STATUS_NO_SUPPORT     = 3,
+    HTP_STATUS_INVAL_PARAMS   = 4,
+    HTP_STATUS_VTCM_TOO_SMALL = 5,
+};
+
+// The values must match the ggml_type.
+// Duplicated here because we can't include full ggml.h in the htp build.
+// We have some static_asserts in the cpp code to ensure things are in sync.
+enum htp_data_type {
+    HTP_TYPE_F32   = 0,
+    HTP_TYPE_F16   = 1,
+    HTP_TYPE_Q4_0  = 2,
+    HTP_TYPE_Q8_0  = 8,
+    HTP_TYPE_MXFP4 = 39,
+    HTP_TYPE_COUNT
+};
+
+// These values are manually translated over to HTP
+// !!!! DO NOT ALTER THE ORDER OF THE FIRST FOUR ENUMS !!!!
+enum htp_op {
+    HTP_OP_MUL            = 0,
+    HTP_OP_ADD            = 1,
+    HTP_OP_SUB            = 2,
+    HTP_OP_DIV            = 3,
+    HTP_OP_MUL_MAT        = 4,
+    HTP_OP_MUL_MAT_ID     = 5,
+    HTP_OP_RMS_NORM       = 6,
+    HTP_OP_UNARY_SILU     = 7,
+    HTP_OP_GLU_SWIGLU     = 8,
+    HTP_OP_GLU_SWIGLU_OAI = 9,
+    HTP_OP_SOFTMAX        = 10,
+    HTP_OP_ADD_ID         = 11,
+    HTP_OP_ROPE           = 12,
+    INVALID
+};
+
+static inline size_t htp_type_block_size(uint32_t t) {
+    switch (t) {
+        case HTP_TYPE_F32:
+            return 1;
+        case HTP_TYPE_F16:
+            return 1;
+        case HTP_TYPE_Q4_0:
+            return QK4_0;
+        case HTP_TYPE_Q8_0:
+            return QK8_0;
+        case HTP_TYPE_MXFP4:
+            return QK_MXFP4;
+        default:
+            assert(0 && "unsupported HTP data type");
+    }
+    return 0;
+}
+
+static inline size_t htp_type_nbytes(uint32_t t) {
+    switch (t) {
+        case HTP_TYPE_F32:
+            return 4;
+        case HTP_TYPE_F16:
+            return 2;
+        case HTP_TYPE_Q4_0:
+            return sizeof(block_q4_0);
+        case HTP_TYPE_Q8_0:
+            return sizeof(block_q8_0);
+        case HTP_TYPE_MXFP4:
+            return sizeof(block_mxfp4);
+        default:
+            assert(0 && "unsupported HTP data type");
+    }
+    return 0;
+}
+
+static const char * htp_type_name(uint32_t t) {
+    switch (t) {
+        case HTP_TYPE_F32:
+            return "fp32";
+        case HTP_TYPE_F16:
+            return "fp16";
+        case HTP_TYPE_Q4_0:
+            return "q4_0";
+        case HTP_TYPE_Q8_0:
+            return "q8_0";
+        case HTP_TYPE_MXFP4:
+            return "mxfp4";
+    }
+    return 0;
+}
+
+// Internal types
+#define QK_Q4_0x4x2  256  // 4x Q4_0 blocks packed with next 4x Q4_0 blocks (size in bytes 128)
+#define QK_Q8_0x4x2  256  // 4x Q8_0 blocks concat with next 4x Q8_0 blocks
+#define QK_MXFP4x4x2 256  // 4x MXFP4 blocks concat with next 4x MXFP4 blocks
+
+#define HTP_MAX_DIMS 4
+
+struct htp_tensor {
+    uint32_t data;              // Buffer offset in the messages, and data pointer on the NSP
+    uint32_t type;              // Data type
+    uint32_t ne[HTP_MAX_DIMS];  // Number of elements
+    uint32_t nb[HTP_MAX_DIMS];  // Stride in bytes (see ggml.h ggml_tensor)
+};
+
+#define HTP_MAX_OP_PARAMS 64
+
+struct htp_general_req {
+    uint32_t op;  // GGML/HTP Op
+    int32_t  op_params[HTP_MAX_OP_PARAMS / sizeof(int32_t)];
+    // Params for the op, e.g. epsilon of RMS norm
+    uint32_t flags;          // Request flags
+
+    struct htp_tensor src0;  // Input0 tensor
+    struct htp_tensor src1;  // Input1 tensor
+    struct htp_tensor src2;  // Input2 tensor
+    struct htp_tensor dst;   // Output tensor
+
+    // should be multiple of 64 bytes (cacheline)
+};
+
+struct htp_general_rsp {
+    uint32_t op;           // GGML/HTP Op
+    uint32_t status;       // HTP_STATUS_...
+    uint32_t prof_usecs;   // Number of usec per request
+    uint32_t prof_cycles;  // Number of cycles per request
+    uint32_t prof_pkts;    // Number of instruction packets per request
+    uint8_t  unused[44];   // Pad to 64 bytes
+};
+
+#define HTP_MAX_MESSAGE_SIZE   sizeof(struct htp_general_req)
+#define HTP_MAX_PACKET_BUFFERS 4
+
+#endif /* HTP_MSG_H */

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -0,0 +1,53 @@
+#ifndef HTP_OPS_H
+#define HTP_OPS_H
+
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "worker-pool.h"
+
+#include <assert.h>
+#include <stdint.h>
+
+// ggml-common.h must be included prior to this header
+
+struct htp_spad {
+    uint8_t * data;
+    size_t    size;
+    size_t    size_per_thread;
+};
+
+struct htp_ops_context {
+    struct htp_context * ctx;
+
+    enum htp_op op;
+    int32_t     op_params[HTP_MAX_OP_PARAMS / sizeof(int32_t)];
+
+    struct htp_tensor src0;
+    struct htp_tensor src1;
+    struct htp_tensor src2;
+    struct htp_tensor dst;
+
+    struct htp_spad src0_spad;
+    struct htp_spad src1_spad;
+    struct htp_spad src2_spad;
+    struct htp_spad dst_spad;
+
+    worker_pool_context_t * wpool;      // worker pool
+    uint32_t                n_threads;  // num threads
+
+    uint32_t src0_nrows_per_thread;
+    uint32_t src1_nrows_per_thread;
+
+    uint32_t flags;
+};
+
+int op_matmul(struct htp_ops_context * octx);
+int op_matmul_id(struct htp_ops_context * octx);
+int op_binary(struct htp_ops_context * octx);
+int op_unary(struct htp_ops_context * octx);
+int op_activations(struct htp_ops_context * octx);
+int op_softmax(struct htp_ops_context * octx);
+int op_add_id(struct htp_ops_context * octx);
+int op_rope(struct htp_ops_context * octx);
+
+#endif /* HTP_OPS_H */

ggml/src/ggml-hexagon/htp/htp_iface.idl

@@ -0,0 +1,16 @@
+// FastRPC IDL interface for GGML HTP
+
+#ifndef HTP_IDL
+#define HTP_IDL
+
+#include "AEEStdDef.idl"
+#include "remote.idl"
+
+interface htp_iface : remote_handle64 {
+    AEEResult start(in uint32 sess_id, in uint64 dsp_queue_id, in uint32 n_hvx);
+    AEEResult stop();
+    AEEResult enable_etm();
+    AEEResult disable_etm();
+};
+
+#endif /* HTP_IDL */

ggml/src/ggml-hexagon/htp/hvx-exp.c

@@ -0,0 +1,80 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+#include "ops-utils.h"
+
+void hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems, bool negate) {
+    int left_over       = num_elems & (VLEN_FP32 - 1);
+    int num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_exp_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+    // assert((0 == unaligned_addr) || (0 == num_elems_whole));
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_exp_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    HVX_Vector vec_out = Q6_V_vzero();
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * p_vec_in1 = (HVX_Vector *) src;
+        HVX_Vector * p_vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            if (true == negate) {
+                HVX_Vector neg_vec_in = hvx_vec_neg_fp32(*p_vec_in1++);
+                *p_vec_out++          = hvx_vec_exp_fp32(neg_vec_in);
+            } else {
+                *p_vec_out++ = hvx_vec_exp_fp32(*p_vec_in1++);
+            }
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
+
+            if (true == negate) {
+                HVX_Vector neg_vec_in                    = hvx_vec_neg_fp32(in);
+                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32(neg_vec_in);
+            } else {
+                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32(in);
+            }
+        }
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (float *) src + num_elems_whole;
+        float *       dstf = (float *) dst + num_elems_whole;
+
+        HVX_Vector in = *(HVX_UVector *) srcf;
+
+        if (true == negate) {
+            HVX_Vector neg_vec_in = hvx_vec_neg_fp32(in);
+
+            vec_out = hvx_vec_exp_fp32(neg_vec_in);
+        } else {
+            vec_out = hvx_vec_exp_fp32(in);
+        }
+
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, vec_out);
+    }
+}

ggml/src/ggml-hexagon/htp/hvx-inverse.c

@@ -0,0 +1,60 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+#include "ops-utils.h"
+
+void hvx_inverse_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems) {
+    int left_over       = num_elems & (VLEN_FP32 - 1);
+    int num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_inverse_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+    // assert((0 == unaligned_addr) || (0 == num_elems_whole));
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_inverse_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * p_vec_in  = (HVX_Vector *) src;
+        HVX_Vector * p_vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            *p_vec_out++ = hvx_vec_inverse_fp32(*p_vec_in++);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in                            = *(HVX_UVector *) (src + i * SIZEOF_FP32);
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_inverse_fp32(in);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (float *) src + num_elems_whole;
+        float *       dstf = (float *) dst + num_elems_whole;
+
+        HVX_Vector in  = *(HVX_UVector *) srcf;
+        HVX_Vector out = hvx_vec_inverse_fp32(in);
+
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, out);
+    }
+}

ggml/src/ggml-hexagon/htp/hvx-sigmoid.c

@@ -0,0 +1,49 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+#include "ops-utils.h"
+
+#if 0
+// Reference algo used in hvx-utils
+static void fast_sigmoid_f32(const float*  restrict src, float* restrict dst, const int num_elems)
+{
+    const float c1 = 0.03138777;
+    const float c2 = 0.276281267;
+    const float c_log2f = 1.442695022;
+
+    int32_t store_ints[32];
+    float store_floats[3][32];
+
+    for (int i = 0; i < num_elems; i++)
+    {
+        float v = src0[i];
+
+        v *= c_log2f*0.5;
+        int intPart = (int)v;
+        float x = (v - intPart);
+        float xx = x * x;
+        float v1 = c_log2f + c2 * xx;
+        float v2 = x + xx * c1 * x;
+        float v3 = (v2 + v1);
+        *((int*)&v3) += intPart << 24;
+        float v4 = v2 - v1;
+        float v5 = v3 - v4;
+        float res = v3 / v5;
+
+        dst[i] = res;
+    }
+}
+#endif

ggml/src/ggml-hexagon/htp/hvx-utils.c

@@ -0,0 +1,947 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#ifdef HTP_DEBUG
+#    define FARF_HIGH 1
+#endif
+
+#include <HAP_farf.h>
+#include <HAP_mem.h>
+#include <HAP_perf.h>
+#include <HAP_ps.h>
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "hvx-utils.h"
+
+#define htp_binary_ops_preamble                                                                                \
+    int step_of_4 = num_elems >> 7;                                                                            \
+    int step_of_2 = (num_elems - step_of_4 * VLEN_FP32 * 4) >> 6;                                              \
+    int step_of_1 = (num_elems - step_of_4 * VLEN_FP32 * 4 - step_of_2 * VLEN_FP32 * 2) >> 5;                  \
+    int remaining = num_elems - step_of_4 * VLEN_FP32 * 4 - step_of_2 * VLEN_FP32 * 2 - step_of_1 * VLEN_FP32; \
+                                                                                                               \
+    const uint8_t * restrict src0_curr = src0;                                                                 \
+    const uint8_t * restrict src1_curr = src1;                                                                 \
+    uint8_t * restrict dst_curr        = dst;
+
+void hvx_mul_f32(const uint8_t * restrict src0,
+                 const uint8_t * restrict src1,
+                 uint8_t * restrict dst,
+                 const int num_elems) {
+    int left_over       = num_elems & (VLEN_FP32 - 1);
+    int num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src0, VLEN)) || (0 == htp_is_aligned((void *) src1, VLEN)) ||
+        (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_mul_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_mul_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src0;
+        HVX_Vector * restrict vec_in2 = (HVX_Vector *) src1;
+        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*vec_in1++, *vec_in2++);
+            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in1 = *(HVX_UVector *) (src0 + i * SIZEOF_FP32);
+            HVX_Vector in2 = *(HVX_UVector *) (src1 + i * SIZEOF_FP32);
+
+            HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(in1, in2);
+
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * src0f = (const float *) src0 + num_elems_whole;
+        const float * src1f = (const float *) src1 + num_elems_whole;
+        float *       dstf  = (float *) dst + num_elems_whole;
+
+        HVX_Vector in1 = *(HVX_UVector *) src0f;
+        HVX_Vector in2 = *(HVX_UVector *) src1f;
+
+        HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(in1, in2);
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
+    }
+}
+
+void hvx_mul_f32_opt(const uint8_t * restrict src0,
+                     const uint8_t * restrict src1,
+                     uint8_t * restrict dst,
+                     const int num_elems) {
+    htp_binary_ops_preamble;
+
+    for (int i = 0; i < step_of_4; i++) {
+        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
+
+        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
+
+        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
+
+        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(v1a, v1b);
+
+        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
+
+        HVX_Vector v3a = *(HVX_Vector *) (src0_curr + 2 * VLEN);
+
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v2a, v2b);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
+
+        HVX_Vector v3b = *(HVX_Vector *) (src1_curr + 2 * VLEN);
+
+        HVX_Vector v4a = *(HVX_Vector *) (src0_curr + 3 * VLEN);
+
+        src0_curr += 4 * VLEN;
+
+        HVX_Vector v3 = Q6_Vqf32_vmpy_VsfVsf(v3a, v3b);
+
+        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
+
+        HVX_Vector v4b = *(HVX_Vector *) (src1_curr + 3 * VLEN);
+
+        *(HVX_Vector *) (dst_curr + 2 * VLEN) = Q6_Vsf_equals_Vqf32(v3);
+
+        HVX_Vector v4 = Q6_Vqf32_vmpy_VsfVsf(v4a, v4b);
+
+        src1_curr += 4 * VLEN;
+
+        *(HVX_Vector *) (dst_curr + 3 * VLEN) = Q6_Vsf_equals_Vqf32(v4);
+
+        dst_curr += 4 * VLEN;
+    }
+
+    for (int i = 0; i < step_of_2; i++) {
+        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
+
+        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
+
+        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
+
+        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(v1a, v1b);
+
+        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
+
+        src0_curr += 2 * VLEN;
+
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v2a, v2b);
+
+        src1_curr += 2 * VLEN;
+
+        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
+
+        dst_curr += 2 * VLEN;
+    }
+
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector va = *(HVX_Vector *) src0_curr;
+
+        src0_curr += VLEN;
+
+        HVX_Vector vb = *(HVX_Vector *) src1_curr;
+
+        src1_curr += VLEN;
+
+        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(va, vb);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v);
+
+        dst_curr += VLEN;
+    }
+
+    if (remaining > 0) {
+        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*(HVX_Vector *) src0_curr, *(HVX_Vector *) src1_curr);
+        hvx_vec_store_u((void *) dst_curr, remaining * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(v));
+    }
+}
+
+void hvx_mul_mul_f32_opt(const uint8_t * restrict src0,
+                         const uint8_t * restrict src1,
+                         const uint8_t * restrict src2,
+                         uint8_t * restrict dst,
+                         const int num_elems) {
+    const uint8_t * restrict src0_curr = src0;
+    const uint8_t * restrict src1_curr = src1;
+    const uint8_t * restrict src2_curr = src2;
+    uint8_t * restrict dst_curr        = dst;
+
+    int step_of_2 = num_elems >> 6;
+    int step_of_1 = (num_elems - step_of_2 * VLEN_FP32 * 2) >> 5;
+    int remaining = num_elems - step_of_2 * VLEN_FP32 * 2 - step_of_1 * VLEN_FP32;
+
+    for (int i = 0; i < step_of_2; i++) {
+        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
+        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
+        HVX_Vector v1c = *(HVX_Vector *) src2_curr;
+
+        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
+
+        HVX_Vector v1_ = Q6_Vqf32_vmpy_VsfVsf(v1a, v1b);
+        HVX_Vector v1  = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v1_), v1c);
+
+        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
+
+        HVX_Vector v2c = *(HVX_Vector *) (src2_curr + VLEN);
+
+        src0_curr += 2 * VLEN;
+
+        HVX_Vector v2_ = Q6_Vqf32_vmpy_VsfVsf(v2a, v2b);
+        HVX_Vector v2  = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v2_), v2c);
+
+        src1_curr += 2 * VLEN;
+        src2_curr += 2 * VLEN;
+
+        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
+
+        dst_curr += 2 * VLEN;
+    }
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector va = *(HVX_Vector *) src0_curr;
+        src0_curr += VLEN;
+
+        HVX_Vector vb = *(HVX_Vector *) src1_curr;
+        src1_curr += VLEN;
+
+        HVX_Vector vc = *(HVX_Vector *) src2_curr;
+        src2_curr += VLEN;
+
+        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(va, vb);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v1), vc);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v2);
+        dst_curr += VLEN;
+    }
+    if (remaining > 0) {
+        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(*(HVX_Vector *) src0_curr, *(HVX_Vector *) src1_curr);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v1), *(HVX_Vector *) src2_curr);
+        hvx_vec_store_u((void *) dst_curr, remaining * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(v2));
+    }
+}
+
+void hvx_add_f32(const uint8_t * restrict src0,
+                 const uint8_t * restrict src1,
+                 uint8_t * restrict dst,
+                 const int num_elems) {
+    int left_over       = num_elems & (VLEN_FP32 - 1);
+    int num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src0, VLEN)) || (0 == htp_is_aligned((void *) src1, VLEN)) ||
+        (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_add_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_add_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src0;
+        HVX_Vector * restrict vec_in2 = (HVX_Vector *) src1;
+        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(*vec_in1++, *vec_in2++);
+            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in1 = *(HVX_UVector *) (src0 + i * SIZEOF_FP32);
+            HVX_Vector in2 = *(HVX_UVector *) (src1 + i * SIZEOF_FP32);
+
+            HVX_Vector out = Q6_Vqf32_vadd_VsfVsf(in1, in2);
+
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * src0f = (const float *) src0 + num_elems_whole;
+        const float * src1f = (const float *) src1 + num_elems_whole;
+        float *       dstf  = (float *) dst + num_elems_whole;
+
+        HVX_Vector in1 = *(HVX_UVector *) src0f;
+        HVX_Vector in2 = *(HVX_UVector *) src1f;
+
+        HVX_Vector out = Q6_Vqf32_vadd_VsfVsf(in1, in2);
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
+    }
+}
+
+void hvx_add_f32_opt(const uint8_t * restrict src0,
+                     const uint8_t * restrict src1,
+                     uint8_t * restrict dst,
+                     const int num_elems) {
+    htp_binary_ops_preamble;
+
+    for (int i = 0; i < step_of_4; i++) {
+        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
+
+        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
+
+        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
+
+        HVX_Vector v1 = Q6_Vqf32_vadd_VsfVsf(v1a, v1b);
+
+        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
+
+        HVX_Vector v3a = *(HVX_Vector *) (src0_curr + 2 * VLEN);
+
+        HVX_Vector v2 = Q6_Vqf32_vadd_VsfVsf(v2a, v2b);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
+
+        HVX_Vector v3b = *(HVX_Vector *) (src1_curr + 2 * VLEN);
+
+        HVX_Vector v4a = *(HVX_Vector *) (src0_curr + 3 * VLEN);
+
+        src0_curr += 4 * VLEN;
+
+        HVX_Vector v3 = Q6_Vqf32_vadd_VsfVsf(v3a, v3b);
+
+        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
+
+        HVX_Vector v4b = *(HVX_Vector *) (src1_curr + 3 * VLEN);
+
+        *(HVX_Vector *) (dst_curr + 2 * VLEN) = Q6_Vsf_equals_Vqf32(v3);
+
+        HVX_Vector v4 = Q6_Vqf32_vadd_VsfVsf(v4a, v4b);
+
+        src1_curr += 4 * VLEN;
+
+        *(HVX_Vector *) (dst_curr + 3 * VLEN) = Q6_Vsf_equals_Vqf32(v4);
+
+        dst_curr += 4 * VLEN;
+    }
+    for (int i = 0; i < step_of_2; i++) {
+        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
+
+        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
+
+        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
+
+        HVX_Vector v1 = Q6_Vqf32_vadd_VsfVsf(v1a, v1b);
+
+        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
+
+        src0_curr += 2 * VLEN;
+
+        HVX_Vector v2 = Q6_Vqf32_vadd_VsfVsf(v2a, v2b);
+
+        src1_curr += 2 * VLEN;
+
+        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
+
+        dst_curr += 2 * VLEN;
+    }
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector va = *(HVX_Vector *) src0_curr;
+
+        src0_curr += VLEN;
+
+        HVX_Vector vb = *(HVX_Vector *) src1_curr;
+
+        src1_curr += VLEN;
+
+        HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(va, vb);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v);
+
+        dst_curr += VLEN;
+    }
+    if (remaining > 0) {
+        HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(*(HVX_Vector *) src0_curr, *(HVX_Vector *) src1_curr);
+        hvx_vec_store_u((void *) dst_curr, remaining * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(v));
+    }
+}
+
+void hvx_add_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems) {
+    size_t left_over       = num_elems & (VLEN_FP32 - 1);
+    size_t num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_add_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_add_scalar_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    HVX_Vector val_vec = hvx_vec_splat_fp32(val);
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
+        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(*vec_in1++, val_vec);
+            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
+
+            HVX_Vector out = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
+
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+        float *       dstf = (float *) dst + num_elems_whole;
+
+        HVX_Vector in = *(HVX_UVector *) srcf;
+
+        HVX_Vector out = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
+    }
+}
+
+void hvx_mul_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems) {
+    size_t left_over       = num_elems & (VLEN_FP32 - 1);
+    size_t num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_mul_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_mul_scalar_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    HVX_Vector val_vec = hvx_vec_splat_fp32(val);
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
+        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*vec_in1++, val_vec);
+            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
+
+            HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(in, val_vec);
+
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+        float *       dstf = (float *) dst + num_elems_whole;
+
+        HVX_Vector in = *(HVX_UVector *) srcf;
+
+        HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(in, val_vec);
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
+    }
+}
+
+void hvx_sub_f32(const uint8_t * restrict src0,
+                 const uint8_t * restrict src1,
+                 uint8_t * restrict dst,
+                 const int num_elems) {
+    size_t left_over       = num_elems & (VLEN_FP32 - 1);
+    size_t num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src0, VLEN)) || (0 == htp_is_aligned((void *) src1, VLEN)) ||
+        (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_sub_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_sub_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src0;
+        HVX_Vector * restrict vec_in2 = (HVX_Vector *) src1;
+        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector v = Q6_Vqf32_vsub_VsfVsf(*vec_in1++, *vec_in2++);
+            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in1 = *(HVX_UVector *) (src0 + i * SIZEOF_FP32);
+            HVX_Vector in2 = *(HVX_UVector *) (src1 + i * SIZEOF_FP32);
+
+            HVX_Vector out = Q6_Vqf32_vsub_VsfVsf(in1, in2);
+
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * src0f = (const float *) src0 + num_elems_whole;
+        const float * src1f = (const float *) src1 + num_elems_whole;
+        float *       dstf  = (float *) dst + num_elems_whole;
+
+        HVX_Vector in1 = *(HVX_UVector *) src0f;
+        HVX_Vector in2 = *(HVX_UVector *) src1f;
+
+        HVX_Vector out = Q6_Vqf32_vsub_VsfVsf(in1, in2);
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
+    }
+}
+
+void hvx_sub_f32_opt(const uint8_t * restrict src0,
+                     const uint8_t * restrict src1,
+                     uint8_t * restrict dst,
+                     const int num_elems) {
+    htp_binary_ops_preamble;
+
+    for (int i = 0; i < step_of_4; i++) {
+        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
+
+        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
+
+        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
+
+        HVX_Vector v1 = Q6_Vqf32_vsub_VsfVsf(v1a, v1b);
+
+        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
+
+        HVX_Vector v3a = *(HVX_Vector *) (src0_curr + 2 * VLEN);
+
+        HVX_Vector v2 = Q6_Vqf32_vsub_VsfVsf(v2a, v2b);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
+
+        HVX_Vector v3b = *(HVX_Vector *) (src1_curr + 2 * VLEN);
+
+        HVX_Vector v4a = *(HVX_Vector *) (src0_curr + 3 * VLEN);
+
+        src0_curr += 4 * VLEN;
+
+        HVX_Vector v3 = Q6_Vqf32_vsub_VsfVsf(v3a, v3b);
+
+        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
+
+        HVX_Vector v4b = *(HVX_Vector *) (src1_curr + 3 * VLEN);
+
+        *(HVX_Vector *) (dst_curr + 2 * VLEN) = Q6_Vsf_equals_Vqf32(v3);
+
+        HVX_Vector v4 = Q6_Vqf32_vsub_VsfVsf(v4a, v4b);
+
+        src1_curr += 4 * VLEN;
+
+        *(HVX_Vector *) (dst_curr + 3 * VLEN) = Q6_Vsf_equals_Vqf32(v4);
+
+        dst_curr += 4 * VLEN;
+    }
+    for (int i = 0; i < step_of_2; i++) {
+        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
+
+        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
+
+        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
+
+        HVX_Vector v1 = Q6_Vqf32_vsub_VsfVsf(v1a, v1b);
+
+        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
+
+        src0_curr += 2 * VLEN;
+
+        HVX_Vector v2 = Q6_Vqf32_vsub_VsfVsf(v2a, v2b);
+
+        src1_curr += 2 * VLEN;
+
+        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
+
+        dst_curr += 2 * VLEN;
+    }
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector va = *(HVX_Vector *) src0_curr;
+
+        src0_curr += VLEN;
+
+        HVX_Vector vb = *(HVX_Vector *) src1_curr;
+
+        src1_curr += VLEN;
+
+        HVX_Vector v = Q6_Vqf32_vsub_VsfVsf(va, vb);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v);
+
+        dst_curr += VLEN;
+    }
+    if (remaining > 0) {
+        HVX_Vector v = Q6_Vqf32_vsub_VsfVsf(*(HVX_Vector *) src0_curr, *(HVX_Vector *) src1_curr);
+        hvx_vec_store_u((void *) dst_curr, remaining * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(v));
+    }
+}
+
+void hvx_sub_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems) {
+    size_t left_over       = num_elems & (VLEN_FP32 - 1);
+    size_t num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_sub_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_sub_scalar_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    HVX_Vector val_vec = hvx_vec_splat_fp32(val);
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
+        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector v = Q6_Vqf32_vsub_VsfVsf(*vec_in1++, val_vec);
+            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
+
+            HVX_Vector out = Q6_Vqf32_vsub_VsfVsf(in, val_vec);
+
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+        float *       dstf = (float *) dst + num_elems_whole;
+
+        HVX_Vector in = *(HVX_UVector *) srcf;
+
+        HVX_Vector out = Q6_Vqf32_vsub_VsfVsf(in, val_vec);
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
+    }
+}
+
+float hvx_sum_of_squares_f32(const uint8_t * restrict src, const int num_elems) {
+    int left_over       = num_elems & (VLEN_FP32 - 1);
+    int num_elems_whole = num_elems - left_over;
+
+    if (0 == htp_is_aligned((void *) src, VLEN)) {
+        FARF(HIGH, "hvx_sum_of_squares_f32: unaligned address in hvx op, possibly slower execution\n");
+    }
+
+    assert((1 == htp_is_aligned((void *) src, VLEN)) || (0 == num_elems_whole));
+
+    HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
+
+    HVX_Vector sum_vec_acc = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector zero_vec    = Q6_V_vsplat_R(0x00000000);
+
+    #pragma unroll(4)
+    for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*vec_in1, *vec_in1);
+        sum_vec_acc  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_vec_acc, v);
+        vec_in1++;
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+
+        HVX_Vector vec_left = *(HVX_UVector *) srcf;
+
+        HVX_Vector vec_left_sq = Q6_Vqf32_vmpy_VsfVsf(vec_left, vec_left);
+        HVX_Vector vec_tmp     = Q6_V_valign_VVR(vec_left_sq, zero_vec, left_over * SIZEOF_FP32);
+
+        sum_vec_acc = Q6_Vqf32_vadd_Vqf32Vqf32(sum_vec_acc, vec_tmp);
+    }
+
+    HVX_Vector v = hvx_vec_qf32_reduce_sum(sum_vec_acc);
+    return hvx_vec_get_fp32(Q6_Vsf_equals_Vqf32(v));
+}
+
+float hvx_self_sum_f32(const uint8_t * restrict src, const int num_elems) {
+    int left_over       = num_elems & (VLEN_FP32 - 1);
+    int num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if (0 == htp_is_aligned((void *) src, VLEN)) {
+        FARF(HIGH, "hvx_self_sum_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_self_sum_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    HVX_Vector sum_vec  = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector zero_vec = Q6_V_vsplat_R(0x00000000);
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * vec_in = (HVX_Vector *) src;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            // sum_vec = Q6_Vqf32_vadd_Vqf32Vsf(sum_vec, *vec_in++);
+            sum_vec = Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(sum_vec), *vec_in++);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
+
+            sum_vec = Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(sum_vec), in);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+
+        HVX_Vector vec_left = *(HVX_UVector *) srcf;
+        HVX_Vector vec_tmp  = Q6_V_valign_VVR(vec_left, zero_vec, left_over * SIZEOF_FP32);
+        // sum_vec = Q6_Vqf32_vadd_Vqf32Vsf(sum_vec, vec_tmp);
+        sum_vec             = Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(sum_vec), vec_tmp);
+    }
+
+    HVX_Vector v = hvx_vec_qf32_reduce_sum(sum_vec);
+    return hvx_vec_get_fp32(Q6_Vsf_equals_Vqf32(v));
+}
+
+void hvx_scale_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems, const float scale) {
+    int left_over       = num_elems & (VLEN_FP32 - 1);
+    int num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_scale_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_scale_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    HVX_Vector scale_vec = hvx_vec_splat_fp32(scale);
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * vec_in1 = (HVX_Vector *) src;
+        HVX_Vector * vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*vec_in1++, scale_vec);
+            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
+
+            HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(in, scale_vec);
+
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+        float *       dstf = (float *) dst + num_elems_whole;
+
+        HVX_Vector in = *(HVX_UVector *) srcf;
+
+        HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(in, scale_vec);
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
+    }
+}
+
+float hvx_self_max_f32(const uint8_t * restrict src, const int num_elems) {
+    int left_over       = num_elems & (VLEN_FP32 - 1);
+    int num_elems_whole = num_elems - left_over;
+
+    int unaligned_addr = 0;
+    int unaligned_loop = 0;
+    if (0 == htp_is_aligned((void *) src, VLEN)) {
+        FARF(HIGH, "hvx_self_max_f32: unaligned address in hvx op, possibly slower execution\n");
+        unaligned_addr = 1;
+    }
+
+    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
+        unaligned_loop = 1;
+        FARF(HIGH, "hvx_self_max_f32: unaligned loop in hvx op, possibly slower execution\n");
+    }
+
+    HVX_Vector vec_max   = hvx_vec_splat_fp32(((const float *) src)[0]);
+    HVX_Vector vec_first = hvx_vec_splat_fp32(((const float *) src)[0]);
+
+    if (0 == unaligned_loop) {
+        HVX_Vector * restrict vec_in = (HVX_Vector *) src;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            vec_max = Q6_Vsf_vmax_VsfVsf(vec_max, *vec_in++);
+        }
+    } else {
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
+
+            vec_max = Q6_Vsf_vmax_VsfVsf(vec_max, in);
+        }
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+
+        HVX_Vector in = *(HVX_UVector *) srcf;
+
+        HVX_Vector temp = Q6_V_valign_VVR(in, vec_first, left_over * SIZEOF_FP32);
+        vec_max         = Q6_Vsf_vmax_VsfVsf(vec_max, temp);
+    }
+
+    HVX_Vector v = hvx_vec_reduce_max_fp32(vec_max);
+    return hvx_vec_get_fp32(v);
+}
+
+void hvx_min_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems) {
+    size_t left_over       = num_elems & (VLEN_FP32 - 1);
+    size_t num_elems_whole = num_elems - left_over;
+
+    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_min_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
+    }
+
+    assert((1 == htp_is_aligned((void *) src, VLEN)) || (0 == num_elems_whole));
+
+    const float * src_f = (const float *) src;
+
+    HVX_Vector vec_min = Q6_V_vsplat_R(val);
+
+    HVX_Vector * restrict vec_in  = (HVX_Vector *) src;
+    HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+    #pragma unroll(4)
+    for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+        vec_min    = Q6_Vsf_vmin_VsfVsf(vec_min, *vec_in++);
+        *vec_out++ = Q6_Vsf_equals_Vqf32(vec_min);
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+        float *       dstf = (float *) dst + num_elems_whole;
+
+        HVX_Vector in = *(HVX_UVector *) srcf;
+
+        vec_min = Q6_Vsf_vmin_VsfVsf(vec_min, in);
+
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(vec_min));
+    }
+}
+
+void hvx_clamp_scalar_f32(const uint8_t * restrict src,
+                          const float limit_left,
+                          const float limit_right,
+                          uint8_t * restrict dst,
+                          const int num_elems) {
+    size_t left_over       = num_elems & (VLEN_FP32 - 1);
+    size_t num_elems_whole = num_elems - left_over;
+
+    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
+        FARF(HIGH, "hvx_clamp_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
+    }
+
+    assert((1 == htp_is_aligned((void *) src, VLEN)) || (0 == num_elems_whole));
+
+    HVX_Vector * restrict vec_in  = (HVX_Vector *) src;
+    HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+    HVX_Vector range_left  = hvx_vec_splat_fp32(limit_left);
+    HVX_Vector range_right = hvx_vec_splat_fp32(limit_right);
+
+    #pragma unroll(4)
+    for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+        HVX_Vector in_vec = *vec_in++;
+        HVX_Vector temp_v = in_vec;
+
+        HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VsfVsf(in_vec, range_right);
+        HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VsfVsf(range_left, in_vec);
+
+        in_vec = Q6_V_vmux_QVV(pred_cap_right, range_right, temp_v);
+        in_vec = Q6_V_vmux_QVV(pred_cap_left, range_left, temp_v);
+
+        *vec_out++ = Q6_Vsf_equals_Vqf32(in_vec);
+    }
+
+    if (left_over > 0) {
+        const float * srcf = (const float *) src + num_elems_whole;
+        float *       dstf = (float *) dst + num_elems_whole;
+
+        HVX_Vector in = *(HVX_UVector *) srcf;
+
+        HVX_Vector temp_v = in;
+
+        HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VsfVsf(in, range_right);
+        HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VsfVsf(range_left, in);
+
+        in = Q6_V_vmux_QVV(pred_cap_right, range_right, temp_v);
+        in = Q6_V_vmux_QVV(pred_cap_left, range_left, temp_v);
+
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(in));
+    }
+}

ggml/src/ggml-hexagon/htp/hvx-utils.h

@@ -0,0 +1,998 @@
+#ifndef HVX_UTILS_H
+#define HVX_UTILS_H
+
+#include "ops-utils.h"
+
+#include <stdbool.h>
+#include <stdint.h>
+
+#define SIZEOF_FP32 (4)
+#define SIZEOF_FP16 (2)
+#define VLEN        (128)
+#define VLEN_FP32   (VLEN / SIZEOF_FP32)
+#define VLEN_FP16   (VLEN / SIZEOF_FP16)
+
+static inline HVX_Vector hvx_vec_splat_fp32(float i) {
+    union {
+        float   f;
+        int32_t i;
+    } fp32 = { .f = i };
+
+    return Q6_V_vsplat_R(fp32.i);
+}
+
+static inline void hvx_vec_store_u(void * addr, uint32_t n, HVX_Vector v) {
+    // Rotate as needed.
+    v = Q6_V_vlalign_VVR(v, v, (size_t) addr);
+
+    uint32_t left_off  = (size_t) addr & 127;
+    uint32_t right_off = left_off + n;
+
+    HVX_VectorPred ql_not = Q6_Q_vsetq_R((size_t) addr);
+    HVX_VectorPred qr     = Q6_Q_vsetq2_R(right_off);
+
+    if (right_off > 128) {
+        Q6_vmem_QRIV(qr, (HVX_Vector *) addr + 1, v);
+        // all 1's
+        qr = Q6_Q_vcmp_eq_VbVb(v, v);
+    }
+
+    ql_not = Q6_Q_or_QQn(ql_not, qr);
+    Q6_vmem_QnRIV(ql_not, (HVX_Vector *) addr, v);
+}
+
+static inline void hvx_vec_store_a(void * ptr, size_t n, HVX_Vector v) {
+    assert((unsigned long) ptr % 128 == 0);
+
+    HVX_VectorPred ql_not = Q6_Q_vsetq_R((size_t) ptr);
+    HVX_VectorPred qr     = Q6_Q_vsetq2_R(n);
+    ql_not                = Q6_Q_or_QQn(ql_not, qr);
+    Q6_vmem_QnRIV(ql_not, (HVX_Vector *) ptr, v);
+}
+
+static inline HVX_Vector hvx_vec_repl4(HVX_Vector v) {
+    // vdelta control to replicate first 4 bytes across all elements
+    static const uint8_t __attribute__((aligned(128))) repl[128] = {
+        0x00, 0x00, 0x00, 0x00, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
+        0x10, 0x10, 0x10, 0x10, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
+        0x20, 0x20, 0x20, 0x20, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
+        0x10, 0x10, 0x10, 0x10, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
+        0x40, 0x40, 0x40, 0x40, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
+        0x10, 0x10, 0x10, 0x10, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
+        0x20, 0x20, 0x20, 0x20, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
+        0x10, 0x10, 0x10, 0x10, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
+    };
+
+    HVX_Vector ctrl = *(HVX_Vector *) repl;
+    return Q6_V_vdelta_VV(v, ctrl);
+}
+
+// copy n fp16 elements : source and destination are aligned to HVX Vector (128)
+static inline void hvx_copy_fp16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    HVX_Vector * restrict vdst = (HVX_Vector *) dst;
+    HVX_Vector * restrict vsrc = (HVX_Vector *) src;
+
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+
+    uint32_t nvec = n / 64;
+    uint32_t nloe = n % 64;
+
+    uint32_t i = 0;
+
+    #pragma unroll(4)
+    for (; i < nvec; i++) {
+        HVX_Vector v = vsrc[i];
+        vdst[i]      = v;
+    }
+
+    if (nloe) {
+        HVX_Vector v = vsrc[i];
+        hvx_vec_store_u((void *) &vdst[i], nloe * sizeof(__fp16), v);
+    }
+}
+
+// copy n fp16 elements : source is aligned, destination is potentially unaligned
+static inline void hvx_copy_fp16_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    HVX_UVector * restrict vdst = (HVX_UVector *) dst;
+    HVX_Vector * restrict vsrc  = (HVX_Vector *) src;
+
+    assert((unsigned long) src % 128 == 0);
+
+    uint32_t nvec = n / 64;
+    uint32_t nloe = n % 64;
+
+    uint32_t i = 0;
+
+    #pragma unroll(4)
+    for (; i < nvec; i++) {
+        HVX_Vector v = vsrc[i];
+        vdst[i]      = v;
+    }
+
+    if (nloe) {
+        HVX_Vector v = vsrc[i];
+        hvx_vec_store_u((void *) &vdst[i], nloe * sizeof(__fp16), v);
+    }
+}
+
+// copy n fp16 elements : source is aligned, destination is potentially unaligned
+static inline void hvx_copy_fp16_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    HVX_Vector * restrict vdst  = (HVX_Vector *) dst;
+    HVX_UVector * restrict vsrc = (HVX_UVector *) src;
+
+    assert((unsigned long) dst % 128 == 0);
+
+    uint32_t nvec = n / 64;
+    uint32_t nloe = n % 64;
+
+    uint32_t i = 0;
+
+    #pragma unroll(4)
+    for (; i < nvec; i++) {
+        HVX_Vector v = vsrc[i];
+        vdst[i]      = v;
+    }
+
+    if (nloe) {
+        HVX_Vector v = vsrc[i];
+        hvx_vec_store_u((void *) &vdst[i], nloe * sizeof(__fp16), v);
+    }
+}
+
+// copy n fp32 elements : source and destination are aligned to HVX Vector (128)
+static inline void hvx_copy_fp32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    HVX_Vector * restrict vdst = (HVX_Vector *) dst;
+    HVX_Vector * restrict vsrc = (HVX_Vector *) src;
+
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+
+    uint32_t nvec = n / 32;
+    uint32_t nloe = n % 32;
+
+    uint32_t i = 0;
+
+    #pragma unroll(4)
+    for (; i < nvec; i++) {
+        HVX_Vector v = vsrc[i];
+        vdst[i]      = v;
+    }
+
+    if (nloe) {
+        HVX_Vector v = vsrc[i];
+        hvx_vec_store_u((void *) &vdst[i], nloe * sizeof(float), v);
+    }
+}
+
+// copy n fp32 elements : source is aligned, destination is unaligned
+static inline void hvx_copy_fp32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    HVX_UVector * restrict vdst = (HVX_UVector *) dst;
+    HVX_Vector * restrict vsrc  = (HVX_Vector *) src;
+
+    assert((unsigned long) src % 128 == 0);
+
+    uint32_t nvec = n / 32;
+    uint32_t nloe = n % 32;
+
+    uint32_t i = 0;
+
+    #pragma unroll(4)
+    for (; i < nvec; i++) {
+        HVX_Vector v = vsrc[i];
+        vdst[i]      = v;
+    }
+
+    if (nloe) {
+        HVX_Vector v = vsrc[i];
+        hvx_vec_store_u((void *) &vdst[i], nloe * sizeof(float), v);
+    }
+}
+
+// copy n fp32 elements : source is unaligned, destination is aligned
+static inline void hvx_copy_fp32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    HVX_Vector * restrict vdst  = (HVX_Vector *) dst;
+    HVX_UVector * restrict vsrc = (HVX_UVector *) src;
+
+    assert((unsigned long) dst % 128 == 0);
+
+    uint32_t nvec = n / 32;
+    uint32_t nloe = n % 32;
+
+    uint32_t i = 0;
+
+    #pragma unroll(4)
+    for (; i < nvec; i++) {
+        HVX_Vector v = vsrc[i];
+        vdst[i]      = v;
+    }
+
+    if (nloe) {
+        HVX_Vector v = vsrc[i];
+        hvx_vec_store_u((void *) &vdst[i], nloe * sizeof(float), v);
+    }
+}
+
+// bcast 1 fp32 element from source to n fp32 elements in destination : destination is aligned
+static inline void hvx_bcast_fp32_a(uint8_t * restrict dst, float elem, uint32_t n) {
+    HVX_Vector * restrict vdst = (HVX_Vector *) dst;
+
+    HVX_Vector velem = hvx_vec_splat_fp32(elem);
+
+    assert((unsigned long) dst % 128 == 0);
+
+    uint32_t nvec = n / 32;
+    uint32_t nloe = n % 32;
+
+    uint32_t i = 0;
+
+    #pragma unroll(4)
+    for (; i < nvec; i++) {
+        vdst[i] = velem;
+    }
+
+    if (nloe) {
+        hvx_vec_store_u((void *) &vdst[i], nloe * sizeof(float), velem);
+    }
+}
+
+static __attribute__((always_inline)) int32_t is_in_one_chunk(void * addr, uint32_t n, uint32_t chunk_size) {
+    uint32_t left_off  = (size_t) addr & (chunk_size - 1);
+    uint32_t right_off = left_off + n;
+    return right_off <= chunk_size;
+}
+
+static void hvx_vec_dump_fp16_n(char * pref, HVX_Vector v, uint32_t n) {
+    union {
+        HVX_Vector v;
+        __fp16 d[64];
+    } u = { .v = v };
+
+    const uint32_t n0 = n / 16;
+    const uint32_t n1 = n % 16;
+    int            i  = 0;
+    for (; i < n0; i++) {
+        htp_dump_fp16_line(pref, u.d + (16 * i), 16);
+    }
+    if (n1) {
+        htp_dump_fp16_line(pref, u.d + (16 * i), n1);
+    }
+}
+
+static void hvx_vec_dump_fp16(char * pref, HVX_Vector v) {
+    hvx_vec_dump_fp16_n(pref, v, 64);
+}
+
+static void hvx_vec_dump_fp32_n(char * pref, HVX_Vector v, uint32_t n) {
+    union {
+        HVX_Vector v;
+        float      d[32];
+    } u = { .v = v };
+
+    const uint32_t n0 = n / 16;
+    const uint32_t n1 = n % 16;
+    int            i  = 0;
+    for (; i < n0; i++) {
+        htp_dump_fp32_line(pref, u.d + (16 * i), 16);
+    }
+    if (n1) {
+        htp_dump_fp32_line(pref, u.d + (16 * i), n1);
+    }
+}
+
+static void hvx_vec_dump_fp32_hmt(char * pref, HVX_Vector v) {
+    union {
+        HVX_Vector v;
+        float      d[32];
+    } u = { .v = v };
+
+    FARF(HIGH, "%s: %.6f %.6f %.6f %.6f ...  %.6f %.6f %.6f %.6f ... %.6f %.6f %.6f %.6f\n", pref, u.d[0], u.d[1],
+         u.d[2], u.d[3], u.d[12], u.d[13], u.d[14], u.d[15], u.d[28], u.d[29], u.d[30], u.d[31]);
+}
+
+static void hvx_vec_dump_fp32(char * pref, HVX_Vector v) {
+    hvx_vec_dump_fp32_n(pref, v, 32);
+}
+
+static void hvx_vec_dump_int32(char * pref, HVX_Vector v) {
+    union {
+        HVX_Vector v;
+        int32_t    d[32];
+    } u = { .v = v };
+
+    for (int i = 0; i < 32 / 16; i++) {
+        htp_dump_int32_line(pref, u.d + (16 * i), 16);
+    }
+}
+
+static void hvx_vec_dump_int32_hmt(char * pref, HVX_Vector v) {
+    union {
+        HVX_Vector v;
+        int32_t    d[32];
+    } u = { .v = v };
+
+    FARF(HIGH, "%s: %d %d %d %d ... %d %d %d %d ... %d %d %d %d\n", pref, u.d[0], u.d[1], u.d[2], u.d[3], u.d[12],
+         u.d[13], u.d[14], u.d[15], u.d[28], u.d[29], u.d[30], u.d[31]);
+}
+
+static void hvx_vec_dump_int8_hmt(char * pref, HVX_Vector v) {
+    union {
+        HVX_Vector v;
+        int8_t     d[128];
+    } u = { .v = v };
+
+    FARF(HIGH, "%s: %d %d %d %d ... %d %d %d %d ... %d %d %d %d\n", pref, u.d[0], u.d[1], u.d[2], u.d[3], u.d[60],
+         u.d[61], u.d[62], u.d[63], u.d[124], u.d[125], u.d[126], u.d[127]);
+}
+
+static void hvx_vec_dump_int8(char * pref, HVX_Vector v) {
+    union {
+        HVX_Vector v;
+        int8_t     d[128];
+    } u = { .v = v };
+
+    for (int i = 0; i < 128 / 16; i++) {
+        htp_dump_int8_line(pref, u.d + (16 * i), 16);
+    }
+}
+
+static void hvx_vec_dump_uint8(char * pref, HVX_Vector v) {
+    union {
+        HVX_Vector v;
+        uint8_t    d[128];
+    } u = { .v = v };
+
+    for (int i = 0; i < 128 / 16; i++) {
+        htp_dump_uint8_line(pref, u.d + (16 * i), 16);
+    }
+}
+
+static bool hvx_vec_eq(HVX_Vector v0, HVX_Vector v1, size_t n) {
+    typedef union {
+        HVX_Vector v;
+        int8_t     d[128];
+    } U;
+
+    U u0 = { .v = v0 };
+    U u1 = { .v = v1 };
+
+    for (int i = 0; i < n; i++) {
+        if (u0.d[i] != u1.d[i]) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static inline float hvx_vec_get_fp32(HVX_Vector v) {
+    float __attribute__((aligned(128))) x;
+    hvx_vec_store_a(&x, 4, v);
+    return x;
+}
+
+static inline HVX_Vector hvx_vec_int32_reduce_sum_n(HVX_Vector in, unsigned int n) {
+    unsigned int total = n * 4;  // total vec nbytes
+    unsigned int width = 4;      // int32
+
+    HVX_Vector sum = in, sum_t;
+    while (width < total) {
+        sum_t = Q6_V_vror_VR(sum, width);     // rotate right
+        sum   = Q6_Vw_vadd_VwVw(sum_t, sum);  // elementwise sum
+        width = width << 1;
+    }
+    return sum;
+}
+
+static inline HVX_Vector hvx_vec_int32_reduce_sum(HVX_Vector in) {
+    return hvx_vec_int32_reduce_sum_n(in, 32);
+}
+
+static inline HVX_Vector hvx_vec_qf32_reduce_sum_n(HVX_Vector in, unsigned int n) {
+    unsigned int total = n * 4;  // total vec nbytes
+    unsigned int width = 4;      // fp32 nbytes
+
+    HVX_Vector sum = in, sum_t;
+    while (width < total) {
+        sum_t = Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum), width);  // rotate right
+        sum   = Q6_Vqf32_vadd_Vqf32Vsf(sum, sum_t);             // elementwise sum
+        width = width << 1;
+    }
+    return sum;
+}
+
+static inline HVX_Vector hvx_vec_qf32_reduce_sum(HVX_Vector in) {
+    return hvx_vec_qf32_reduce_sum_n(in, 32);
+}
+
+static inline HVX_Vector hvx_vec_fp32_reduce_sum_n(HVX_Vector in, unsigned int n) {
+    unsigned int total = n * 4;  // total vec nbytes
+    unsigned int width = 4;      // fp32 nbytes
+
+    HVX_Vector sum = in, sum_t;
+    while (width < total) {
+        sum_t = Q6_V_vror_VR(sum, width);       // rotate right
+        sum   = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(sum, sum_t)); // elementwise sum
+        width = width << 1;
+    }
+    return sum;
+}
+
+static inline HVX_Vector hvx_vec_fp32_reduce_sum(HVX_Vector in) {
+    return hvx_vec_fp32_reduce_sum_n(in, 32);
+}
+
+static inline HVX_Vector hvx_vec_reduce_max_fp16(HVX_Vector in) {
+    unsigned total = 128;  // total vec nbytes
+    unsigned width = 2;    // fp16 nbytes
+
+    HVX_Vector _max = in, _max_t;
+    while (width < total) {
+        _max_t = Q6_V_vror_VR(_max, width);         // rotate right
+        _max   = Q6_Vhf_vmax_VhfVhf(_max_t, _max);  // elementwise max
+        width  = width << 1;
+    }
+
+    return _max;
+}
+
+static inline HVX_Vector hvx_vec_reduce_max2_fp16(HVX_Vector in, HVX_Vector _max) {
+    unsigned total = 128;  // total vec nbytes
+    unsigned width = 2;    // fp32 nbytes
+
+    HVX_Vector _max_t;
+
+    _max = Q6_Vhf_vmax_VhfVhf(in, _max);
+    while (width < total) {
+        _max_t = Q6_V_vror_VR(_max, width);         // rotate right
+        _max   = Q6_Vhf_vmax_VhfVhf(_max_t, _max);  // elementwise max
+        width  = width << 1;
+    }
+
+    return _max;
+}
+
+static inline HVX_Vector hvx_vec_reduce_max_fp32(HVX_Vector in) {
+    unsigned total = 128;  // total vec nbytes
+    unsigned width = 4;    // fp32 nbytes
+
+    HVX_Vector _max = in, _max_t;
+    while (width < total) {
+        _max_t = Q6_V_vror_VR(_max, width);         // rotate right
+        _max   = Q6_Vsf_vmax_VsfVsf(_max_t, _max);  // elementwise max
+        width  = width << 1;
+    }
+
+    return _max;
+}
+
+static inline HVX_Vector hvx_vec_reduce_max2_fp32(HVX_Vector in, HVX_Vector _max) {
+    unsigned total = 128;  // total vec nbytes
+    unsigned width = 4;    // fp32 nbytes
+
+    HVX_Vector _max_t;
+
+    _max = Q6_Vsf_vmax_VsfVsf(in, _max);
+    while (width < total) {
+        _max_t = Q6_V_vror_VR(_max, width);         // rotate right
+        _max   = Q6_Vsf_vmax_VsfVsf(_max_t, _max);  // elementwise max
+        width  = width << 1;
+    }
+
+    return _max;
+}
+
+static inline HVX_Vector hvx_vec_abs_fp16(HVX_Vector v) {
+    // abs by clearing the fp16 sign bit
+    HVX_Vector mask = Q6_Vh_vsplat_R(0x7fff);
+    return Q6_V_vand_VV(v, mask);
+}
+
+static inline HVX_Vector hvx_vec_neg_fp16(HVX_Vector v) {
+    // neg by setting the fp16 sign bit
+    HVX_Vector mask = Q6_Vh_vsplat_R(0x8000);
+    return Q6_V_vor_VV(v, mask);
+}
+
+static inline HVX_Vector hvx_vec_abs_fp32(HVX_Vector v) {
+    // abs by clearing the fp32 sign bit
+    HVX_Vector mask = Q6_V_vsplat_R(0x7fffffff);
+    return Q6_V_vand_VV(v, mask);
+}
+
+static inline HVX_Vector hvx_vec_neg_fp32(HVX_Vector v) {
+#if __HTP_ARCH__ > 75
+    return Q6_Vsf_vfneg_Vsf(v);
+#else
+    // neg by setting the fp32 sign bit
+    HVX_Vector mask = Q6_V_vsplat_R(0x80000000);
+    return Q6_V_vor_VV(v, mask);
+#endif  // __HTP_ARCH__ > 75
+}
+
+// ====================================================
+// FUNCTION: 1/(x+1)     y(0) = 1,  y(0.5) = 0.6667, y(1) = 0.5
+// Order:3; continuity: True; Ends forced: True
+// Mode: unsigned;   Result fractional bits: 14
+// Peak Error: 1.1295e-04  Rms Error: 2.8410e-05   Mean Error: 1.1370e-05
+//      32769  -32706   31252  -10589
+//      32590  -30635   22793   -4493
+//      32066  -27505   16481   -2348
+//      31205  -24054   11849   -1306
+
+static inline HVX_Vector hvx_vec_recip_xp1_O3_unsigned(HVX_Vector vx) {
+    // input is 0..0xffff representing 0.0  .. 1.0
+    HVX_Vector p;
+    p = Q6_Vh_vlut4_VuhPh(vx, 0xFAE6F6D4EE73D6A3ull);
+    p = Q6_Vh_vmpa_VhVhVuhPuh_sat(p, vx, 0x2E49406159097A14ull);
+    p = Q6_Vh_vmps_VhVhVuhPuh_sat(p, vx, 0x5DF66B7177AB7FC2ull);
+    p = Q6_Vh_vmpa_VhVhVuhPuh_sat(p, vx, 0x79E57D427F4E8001ull);
+    return p;  // signed result, 14 fractional bits
+}
+
+// Find reciprocal of fp16.
+// (1) first, convert to fp32, multiplying by 1.0; this is done to
+//    handle denormals. Ignoring sign and zero, result should be at
+//    least 5.9604645e-08 (32-bit code 0x33800000) and at most 131008 (0x47ffe000)
+//    (exponent in range [103,143])
+// (2) extract the mantissa into 16-bit unsigned; find reciprocal using a fitted poly
+// (3) put this, along with '253-exp' (exp from (1)) together to make an qf32
+// (4) convert that to fp16
+// (5) put sign back in. Also, if the original value (w/o sign) was <0x81, replace
+//     the result with the max value.
+static inline HVX_Vector hvx_vec_inverse_fp16(HVX_Vector vals) {
+    HVX_Vector     em_mask  = Q6_Vh_vsplat_R(0x7FFF);
+    HVX_Vector     avals    = Q6_V_vand_VV(vals, em_mask);
+    HVX_VectorPred is_neg   = Q6_Q_vcmp_gt_VhVh(avals, vals);
+    // is too small to 1/x ? for 'standard' fp16, this would be 0x101
+    HVX_VectorPred is_small = Q6_Q_vcmp_gt_VhVh(Q6_Vh_vsplat_R(0x101), avals);
+
+    HVX_VectorPair to_qf32  = Q6_Wqf32_vmpy_VhfVhf(avals, Q6_Vh_vsplat_R(0x3C00));  // *1.0
+    HVX_Vector     to_f32_0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(to_qf32));
+    HVX_Vector     to_f32_1 = Q6_Vsf_equals_Vqf32(Q6_V_hi_W(to_qf32));
+
+    // bits 22..13 contain the mantissa now (w/o hidden bit); move to bit 14..5 of a 16-bit vector
+    HVX_Vector mant_u16 = Q6_Vh_vshuffo_VhVh(Q6_Vw_vasl_VwR(to_f32_1, 9), Q6_Vw_vasl_VwR(to_f32_0, 9));
+    // likewise extract the upper 16 from each, containing the exponents in range 103..142
+    HVX_Vector exp_u16  = Q6_Vh_vshuffo_VhVh(to_f32_1, to_f32_0);
+    //Get exponent in IEEE 32-bit representation
+    exp_u16             = Q6_Vuh_vlsr_VuhR(exp_u16, 7);
+
+    // so, mant_u16 contains an unbiased mantissa in upper 10 bits of each u16 lane
+    // We can consider it to be x-1.0, with 16 fractional bits, where 'x' is in range [1.0,2.0)
+    // Use poly to transform to 1/x, with 14 fractional bits
+    //
+    HVX_Vector rm = hvx_vec_recip_xp1_O3_unsigned(mant_u16);
+
+    HVX_Vector vcl0 = Q6_Vuh_vcl0_Vuh(rm);  //count leading zeros
+
+    // Get mantissa for 16-bit represenation
+    HVX_Vector mant_recip = Q6_V_vand_VV(Q6_Vh_vasr_VhR(Q6_Vh_vasl_VhVh(rm, vcl0), 5), Q6_Vh_vsplat_R(0x03FF));
+
+    //Compute Reciprocal Exponent
+    HVX_Vector exp_recip =
+        Q6_Vh_vsub_VhVh(Q6_Vh_vsub_VhVh(Q6_Vh_vsplat_R(254), exp_u16), Q6_Vh_vsub_VhVh(vcl0, Q6_Vh_vsplat_R(1)));
+    //Convert it for 16-bit representation
+    exp_recip = Q6_Vh_vadd_VhVh_sat(Q6_Vh_vsub_VhVh(exp_recip, Q6_Vh_vsplat_R(127)), Q6_Vh_vsplat_R(15));
+    exp_recip = Q6_Vh_vasl_VhR(exp_recip, 10);
+
+    //Merge exponent and mantissa for reciprocal
+    HVX_Vector recip = Q6_V_vor_VV(exp_recip, mant_recip);
+    // map 'small' inputs to standard largest value 0x7bff
+    recip            = Q6_V_vmux_QVV(is_small, Q6_Vh_vsplat_R(0x7bff), recip);
+    // add sign back
+    recip            = Q6_V_vandor_VQR(recip, is_neg, 0x80008000);
+    return recip;
+}
+
+#define IEEE_VSF_EXPLEN   (8)
+#define IEEE_VSF_EXPBIAS  (127)
+#define IEEE_VSF_EXPMASK  (0xFF)
+#define IEEE_VSF_MANTLEN  (23)
+#define IEEE_VSF_MANTMASK (0x7FFFFF)
+#define IEEE_VSF_MIMPMASK (0x800000)
+
+static inline HVX_Vector hvx_vec_truncate_fp32(HVX_Vector in_vec) {
+    HVX_Vector mask_mant_v  = Q6_V_vsplat_R(IEEE_VSF_MANTMASK);
+    HVX_Vector mask_impl_v  = Q6_V_vsplat_R(IEEE_VSF_MIMPMASK);
+    HVX_Vector const_zero_v = Q6_V_vzero();
+
+    HVX_VectorPred q_negative = Q6_Q_vcmp_gt_VwVw(const_zero_v, in_vec);
+
+    HVX_Vector expval_v = in_vec >> IEEE_VSF_MANTLEN;
+    expval_v &= IEEE_VSF_EXPMASK;
+    expval_v -= IEEE_VSF_EXPBIAS;
+
+    // negative exp == fractional value
+    HVX_VectorPred q_negexp = Q6_Q_vcmp_gt_VwVw(const_zero_v, expval_v);
+
+    HVX_Vector rshift_v = IEEE_VSF_MANTLEN - expval_v;         // fractional bits - exp shift
+
+    HVX_Vector mant_v = in_vec & mask_mant_v;                  // obtain mantissa
+    HVX_Vector vout   = Q6_Vw_vadd_VwVw(mant_v, mask_impl_v);  // add implicit 1.0
+
+    vout = Q6_Vw_vasr_VwVw(vout, rshift_v);                    // shift to obtain truncated integer
+    vout = Q6_V_vmux_QVV(q_negexp, const_zero_v, vout);        // expval<0 -> 0
+
+    HVX_Vector neg_vout = -vout;
+
+    vout = Q6_V_vmux_QVV(q_negative, neg_vout, vout);  // handle negatives
+
+    return (vout);
+}
+
+static inline HVX_Vector hvx_vec_floor_fp32(HVX_Vector in_vec) {
+    HVX_Vector mask_mant_v    = Q6_V_vsplat_R(IEEE_VSF_MANTMASK);
+    HVX_Vector mask_impl_v    = Q6_V_vsplat_R(IEEE_VSF_MIMPMASK);
+    HVX_Vector const_mnlen_v  = Q6_V_vsplat_R(IEEE_VSF_MANTLEN);
+    HVX_Vector const_zero_v   = Q6_V_vzero();
+    HVX_Vector const_negone_v = Q6_V_vsplat_R(0xbf800000);  // -1 IEEE vsf
+
+    HVX_VectorPred q_negative = Q6_Q_vcmp_gt_VwVw(const_zero_v, in_vec);
+
+    HVX_Vector expval_v = in_vec >> IEEE_VSF_MANTLEN;
+    expval_v &= IEEE_VSF_EXPMASK;
+    expval_v -= IEEE_VSF_EXPBIAS;
+
+    HVX_VectorPred q_negexp     = Q6_Q_vcmp_gt_VwVw(const_zero_v, expval_v);
+    HVX_VectorPred q_expltmn    = Q6_Q_vcmp_gt_VwVw(const_mnlen_v, expval_v);
+    HVX_VectorPred q_negexp_pos = Q6_Q_vcmp_gtand_QVwVw(q_negexp, in_vec, const_zero_v);
+    HVX_VectorPred q_negexp_neg = Q6_Q_vcmp_gtand_QVwVw(q_negexp, const_zero_v, in_vec);
+
+    // if expval < 0 (q_negexp)         // <0, floor is 0
+    //    if vin > 0
+    //       floor = 0
+    //    if vin < 0
+    //       floor = -1
+    // if expval < mant_len (q_expltmn) // >0, but fraction may exist
+    //    get sign (q_negative)
+    //    mask >> expval                // fraction bits to mask off
+    //    vout = ~(mask)                // apply mask to remove fraction
+    //    if (qneg)                     // negative floor is one less (more, sign bit for neg)
+    //      vout += ((impl_mask) >> expval)
+    //    if (mask && vin)
+    //      vout = vin
+    // else                             // already an integer
+    //    ;                             // no change
+
+    // compute floor
+    mask_mant_v >>= expval_v;
+    HVX_Vector neg_addin_v    = mask_impl_v >> expval_v;
+    HVX_Vector vout_neg_addin = Q6_Vw_vadd_VwVw(in_vec, neg_addin_v);
+    HVX_Vector vout           = Q6_V_vmux_QVV(q_negative, vout_neg_addin, in_vec);
+
+    HVX_Vector     mask_chk_v = Q6_V_vand_VV(in_vec, mask_mant_v);  // chk if bits set
+    HVX_VectorPred q_integral = Q6_Q_vcmp_eq_VwVw(const_zero_v, mask_chk_v);
+
+    HVX_Vector not_mask_v = Q6_V_vnot_V(mask_mant_v);        // frac bits to clear
+    HVX_Vector vfrfloor_v = Q6_V_vand_VV(vout, not_mask_v);  // clear frac bits
+
+    vout = in_vec;
+    vout = Q6_V_vmux_QVV(q_expltmn, vfrfloor_v, vout);         // expval<mant
+    vout = Q6_V_vmux_QVV(q_integral, in_vec, vout);            // integral values
+    vout = Q6_V_vmux_QVV(q_negexp_pos, const_zero_v, vout);    // expval<0 x>0 -> 0
+    vout = Q6_V_vmux_QVV(q_negexp_neg, const_negone_v, vout);  // expval<0 x<0 -> -1
+
+    return vout;
+}
+
+static inline HVX_Vector hvx_vec_i16_from_hf_rnd_sat(HVX_Vector vin) {
+    // This looks complicated.
+    // Ideally should just be Q6_Vh_equals_Vhf(vin)
+    // but that instruction does not do proper rounding.
+
+    // convert to qf32, multiplying by 1.0 in the process.
+    HVX_VectorPair v32 = Q6_Wqf32_vmpy_VhfVhf(vin, Q6_Vh_vsplat_R(0x3C00));
+
+    // 'in-range' values are +/32752.
+    // add 192K to it, convert to sf
+    HVX_Vector v192K = Q6_V_vsplat_R(0x48400000);
+    HVX_Vector vsf_0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_V_lo_W(v32), v192K));
+    HVX_Vector vsf_1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_V_hi_W(v32), v192K));
+
+    // for in-range cases, result is {163858... 229360} so the exponent is always 144.
+    // if we extract bits 21..0 as a signed quantity, and round 6 bits off, that will be the answer.
+    // Start by <<10 to get the final 'sign' bit in bit 15...
+    vsf_0 = Q6_Vw_vasl_VwR(vsf_0, 10);
+    vsf_1 = Q6_Vw_vasl_VwR(vsf_1, 10);
+
+    // now round down to 16
+    return Q6_Vh_vround_VwVw_sat(vsf_1, vsf_0);
+}
+
+static inline HVX_Vector hvx_vec_inverse_fp32(HVX_Vector v_sf) {
+    HVX_Vector inv_aprox_sf = Q6_V_vsplat_R(0x7EEEEBB3);
+    HVX_Vector two_sf       = hvx_vec_splat_fp32(2.0);
+
+    // First approximation
+    HVX_Vector i_sf = Q6_Vw_vsub_VwVw(inv_aprox_sf, v_sf);
+
+    HVX_Vector r_qf;
+
+    // Refine
+    r_qf = Q6_Vqf32_vmpy_VsfVsf(
+        i_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(two_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(i_sf, v_sf)))));
+    r_qf = Q6_Vqf32_vmpy_Vqf32Vqf32(
+        r_qf, Q6_Vqf32_vsub_VsfVsf(two_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(r_qf), v_sf))));
+    r_qf = Q6_Vqf32_vmpy_Vqf32Vqf32(
+        r_qf, Q6_Vqf32_vsub_VsfVsf(two_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(r_qf), v_sf))));
+
+    return Q6_Vsf_equals_Vqf32(r_qf);
+}
+
+#define FAST_SIGMOID_LOG2F (0x3fb8aa3b)  // 1.442695022
+#define FAST_SIGMOID_C1    (0x3d009076)  // 0.03138777
+#define FAST_SIGMOID_C2    (0x3e8d74bd)  // 0.276281267
+#define FAST_SIGMOID_C3    (0x3f000000)  // 0.5
+
+static inline HVX_Vector hvx_vec_fast_sigmoid_fp32(HVX_Vector v) {
+    v = Q6_Vqf32_vmpy_VsfVsf(v, Q6_V_vsplat_R(FAST_SIGMOID_LOG2F));
+    v = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v), Q6_V_vsplat_R(FAST_SIGMOID_C3));
+
+    HVX_Vector in_int = hvx_vec_truncate_fp32(Q6_Vsf_equals_Vqf32(v));
+    HVX_Vector x      = Q6_Vqf32_vsub_Vqf32Vsf(v, Q6_Vsf_equals_Vw(in_int));
+    HVX_Vector xx     = Q6_Vqf32_vmpy_Vqf32Vqf32(x, x);
+
+    HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(xx), Q6_V_vsplat_R(FAST_SIGMOID_C2));
+    v1            = Q6_Vqf32_vadd_Vqf32Vsf(v1, Q6_V_vsplat_R(FAST_SIGMOID_LOG2F));
+
+    HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(x), Q6_V_vsplat_R(FAST_SIGMOID_C1));
+    v2            = Q6_Vqf32_vmpy_Vqf32Vqf32(v2, xx);
+    v2            = Q6_Vqf32_vadd_Vqf32Vqf32(v2, x);
+
+    HVX_Vector v3          = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vqf32(v2, v1));
+    HVX_Vector v3_exponent = Q6_Vw_vasl_VwR(v3, 1);
+    v3_exponent            = Q6_Vuw_vlsr_VuwR(v3_exponent, 24);
+    v3_exponent            = Q6_Vw_vadd_VwVw(in_int, v3_exponent);
+    v3                     = Q6_Vw_vaslacc_VwVwR(v3, in_int, 24);
+
+    HVX_Vector v4 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_Vqf32Vqf32(v2, v1));
+    HVX_Vector v5 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(v3, v4));
+
+    HVX_Vector res = hvx_vec_inverse_fp32(v5);
+    res            = Q6_Vqf32_vmpy_VsfVsf(v3, res);
+
+    return Q6_Vsf_equals_Vqf32(res);
+}
+
+#define EXP_COEFF_5 (0x39506967)  // 0.000198757 = 1/(7!)
+#define EXP_COEFF_4 (0x3AB743CE)  // 0.0013982   = 1/(6!)
+#define EXP_COEFF_3 (0x3C088908)  // 0.00833345  = 1/(5!)
+#define EXP_COEFF_2 (0x3D2AA9C1)  // 0.416658    = 1/(4!)
+#define EXP_COEFF_1 (0x3E2AAAAA)  // 0.16666667  = 1/(3!)
+#define EXP_COEFF_0 (0x3F000000)  // 0.5         = 1/(2!)
+#define EXP_LOGN2   (0x3F317218)  // ln(2)   = 0.6931471805
+#define EXP_LOG2E   (0x3FB8AA3B)  // log2(e) = 1/ln(2) = 1.4426950408
+#define EXP_ONE     (0x3f800000)  // 1.0
+#define EXP_RANGE_R (0x41a00000)  // 20.0
+#define EXP_RANGE_L (0xc1a00000)  // -20.0
+
+static inline HVX_Vector hvx_vec_exp_fp32(HVX_Vector in_vec) {
+    HVX_Vector z_qf32_v;
+    HVX_Vector x_v;
+    HVX_Vector x_qf32_v;
+    HVX_Vector y_v;
+    HVX_Vector k_v;
+    HVX_Vector f_v;
+    HVX_Vector epsilon_v;
+    HVX_Vector log2e = Q6_V_vsplat_R(EXP_LOG2E);
+    HVX_Vector logn2 = Q6_V_vsplat_R(EXP_LOGN2);
+    HVX_Vector E_const;
+    HVX_Vector zero_v = Q6_V_vzero();
+
+    // exp(x) is approximated as follows:
+    //   f = floor(x/ln(2)) = floor(x*log2(e))
+    //   epsilon = x - f*ln(2)
+    //   exp(x) = exp(epsilon+f*ln(2))
+    //          = exp(epsilon)*exp(f*ln(2))
+    //          = exp(epsilon)*2^f
+    //
+    //   Since epsilon is close to zero, it can be approximated with its Taylor series:
+    //            exp(x) ~= 1+x+x^2/2!+x^3/3!+...+x^n/n!+...
+    //   Preserving the first eight elements, we get:
+    //            exp(x) ~= 1+x+e0*x^2+e1*x^3+e2*x^4+e3*x^5+e4*x^6+e5*x^7
+    //                   =  1+x+(E0+(E1+(E2+(E3+(E4+E5*x)*x)*x)*x)*x)*x^2
+
+    HVX_Vector temp_v = in_vec;
+
+    // Clamp inputs to (-20.0, 20.0)
+    HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VsfVsf(in_vec, Q6_V_vsplat_R(EXP_RANGE_R));
+    HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VsfVsf(Q6_V_vsplat_R(EXP_RANGE_L), in_vec);
+
+    in_vec = Q6_V_vmux_QVV(pred_cap_right, Q6_V_vsplat_R(EXP_RANGE_R), temp_v);
+    in_vec = Q6_V_vmux_QVV(pred_cap_left, Q6_V_vsplat_R(EXP_RANGE_L), temp_v);
+
+    epsilon_v = Q6_Vqf32_vmpy_VsfVsf(log2e, in_vec);
+    epsilon_v = Q6_Vsf_equals_Vqf32(epsilon_v);
+
+    //    f_v is the floating point result and k_v is the integer result
+    f_v = hvx_vec_floor_fp32(epsilon_v);
+    k_v = hvx_vec_truncate_fp32(f_v);
+
+    x_qf32_v = Q6_Vqf32_vadd_VsfVsf(in_vec, zero_v);
+
+    //  x = x - f_v * logn2;
+    epsilon_v = Q6_Vqf32_vmpy_VsfVsf(f_v, logn2);
+    x_qf32_v  = Q6_Vqf32_vsub_Vqf32Vqf32(x_qf32_v, epsilon_v);
+    // normalize before every QFloat's vmpy
+    x_qf32_v  = Q6_Vqf32_vadd_Vqf32Vsf(x_qf32_v, zero_v);
+
+    // z = x * x;
+    z_qf32_v = Q6_Vqf32_vmpy_Vqf32Vqf32(x_qf32_v, x_qf32_v);
+    z_qf32_v = Q6_Vqf32_vadd_Vqf32Vsf(z_qf32_v, zero_v);
+
+    x_v = Q6_Vsf_equals_Vqf32(x_qf32_v);
+
+    // y = E4 + E5 * x;
+    E_const = Q6_V_vsplat_R(EXP_COEFF_5);
+    y_v     = Q6_Vqf32_vmpy_VsfVsf(E_const, x_v);
+    E_const = Q6_V_vsplat_R(EXP_COEFF_4);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, E_const);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, zero_v);
+
+    // y = E3 + y * x;
+    E_const = Q6_V_vsplat_R(EXP_COEFF_3);
+    y_v     = Q6_Vqf32_vmpy_Vqf32Vqf32(y_v, x_qf32_v);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, E_const);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, zero_v);
+
+    // y = E2 + y * x;
+    E_const = Q6_V_vsplat_R(EXP_COEFF_2);
+    y_v     = Q6_Vqf32_vmpy_Vqf32Vqf32(y_v, x_qf32_v);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, E_const);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, zero_v);
+
+    // y = E1 + y * x;
+    E_const = Q6_V_vsplat_R(EXP_COEFF_1);
+    y_v     = Q6_Vqf32_vmpy_Vqf32Vqf32(y_v, x_qf32_v);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, E_const);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, zero_v);
+
+    // y = E0 + y * x;
+    E_const = Q6_V_vsplat_R(EXP_COEFF_0);
+    y_v     = Q6_Vqf32_vmpy_Vqf32Vqf32(y_v, x_qf32_v);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, E_const);
+    y_v     = Q6_Vqf32_vadd_Vqf32Vsf(y_v, zero_v);
+
+    // y = x + y * z;
+    y_v = Q6_Vqf32_vmpy_Vqf32Vqf32(y_v, z_qf32_v);
+    y_v = Q6_Vqf32_vadd_Vqf32Vqf32(y_v, x_qf32_v);
+    y_v = Q6_Vqf32_vadd_Vqf32Vsf(y_v, zero_v);
+
+    // y = y + 1.0;
+    y_v = Q6_Vqf32_vadd_Vqf32Vsf(y_v, Q6_V_vsplat_R(EXP_ONE));
+
+    // insert exponents
+    //        y = ldexpf(y, k);
+    //    y_v += k_v; // qf32
+    // modify exponent
+
+    y_v = Q6_Vsf_equals_Vqf32(y_v);
+
+    // add k_v to the exponent of y_v
+    HVX_Vector y_v_exponent = Q6_Vw_vasl_VwR(y_v, 1);
+
+    y_v_exponent = Q6_Vuw_vlsr_VuwR(y_v_exponent, IEEE_VSF_MANTLEN + 1);
+    y_v_exponent = Q6_Vw_vadd_VwVw(k_v, y_v_exponent);
+
+    // exponent cannot be negative; if overflow is detected, result is set to zero
+    HVX_VectorPred qy_v_negative_exponent = Q6_Q_vcmp_gt_VwVw(zero_v, y_v_exponent);
+
+    y_v = Q6_Vw_vaslacc_VwVwR(y_v, k_v, IEEE_VSF_MANTLEN);
+
+    y_v = Q6_V_vmux_QVV(qy_v_negative_exponent, zero_v, y_v);
+
+    return y_v;
+}
+
+#define RSQRT_CONST        0x5f3759df  // Constant for fast inverse square root calculation
+#define RSQRT_ONE_HALF     0x3f000000  // 0.5
+#define RSQRT_THREE_HALVES 0x3fc00000  // 1.5
+
+static inline HVX_Vector hvx_vec_rsqrt_fp32(HVX_Vector in_vec) {
+    //Algorithm :
+    //  x2 = input*0.5
+    //  y  = * (long *) &input
+    //  y  = 0x5f3759df - (y>>2)
+    //  y  = y*(threehalfs - x2*y*y)
+
+    HVX_Vector rsqrtconst = Q6_V_vsplat_R(RSQRT_CONST);
+    HVX_Vector onehalf    = Q6_V_vsplat_R(RSQRT_ONE_HALF);
+    HVX_Vector threehalfs = Q6_V_vsplat_R(RSQRT_THREE_HALVES);
+
+    HVX_Vector x2, y, ypower2, temp;
+
+    x2 = Q6_Vqf32_vmpy_VsfVsf(in_vec, onehalf);
+    x2 = Q6_Vqf32_vadd_Vqf32Vsf(x2, Q6_V_vzero());
+
+    y = Q6_Vw_vasr_VwR(in_vec, 1);
+    y = Q6_Vw_vsub_VwVw(rsqrtconst, y);
+
+    // 1st iteration
+    ypower2 = Q6_Vqf32_vmpy_VsfVsf(y, y);
+    ypower2 = Q6_Vqf32_vadd_Vqf32Vsf(ypower2, Q6_V_vzero());
+    temp    = Q6_Vqf32_vmpy_Vqf32Vqf32(x2, ypower2);
+    temp    = Q6_Vqf32_vsub_VsfVsf(threehalfs, Q6_Vsf_equals_Vqf32(temp));
+    temp    = Q6_Vqf32_vmpy_VsfVsf(y, Q6_Vsf_equals_Vqf32(temp));
+
+    // 2nd iteration
+    y       = Q6_Vqf32_vadd_Vqf32Vsf(temp, Q6_V_vzero());
+    ypower2 = Q6_Vqf32_vmpy_Vqf32Vqf32(y, y);
+    ypower2 = Q6_Vqf32_vadd_Vqf32Vsf(ypower2, Q6_V_vzero());
+    temp    = Q6_Vqf32_vmpy_Vqf32Vqf32(x2, ypower2);
+    temp    = Q6_Vqf32_vsub_VsfVsf(threehalfs, Q6_Vsf_equals_Vqf32(temp));
+    temp    = Q6_Vqf32_vmpy_Vqf32Vqf32(y, temp);
+
+    // 3rd iteration
+    y       = Q6_Vqf32_vadd_Vqf32Vsf(temp, Q6_V_vzero());
+    ypower2 = Q6_Vqf32_vmpy_Vqf32Vqf32(y, y);
+    ypower2 = Q6_Vqf32_vadd_Vqf32Vsf(ypower2, Q6_V_vzero());
+    temp    = Q6_Vqf32_vmpy_Vqf32Vqf32(x2, ypower2);
+    temp    = Q6_Vqf32_vsub_VsfVsf(threehalfs, Q6_Vsf_equals_Vqf32(temp));
+    temp    = Q6_Vqf32_vmpy_Vqf32Vqf32(y, temp);
+
+    return Q6_Vsf_equals_Vqf32(temp);
+}
+
+static inline void hvx_fast_sigmoid_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems) {
+    int step_of_1 = num_elems >> 5;
+    int remaining = num_elems - step_of_1 * VLEN_FP32;
+
+    assert(remaining == 0);
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        v_dst[i] = hvx_vec_fast_sigmoid_fp32(v_src[i]);
+    }
+}
+
+float hvx_sum_of_squares_f32(const uint8_t * restrict src, const int num_elems);
+void  hvx_mul_f32(const uint8_t * restrict src0,
+                  const uint8_t * restrict src1,
+                  uint8_t * restrict dst,
+                  const int num_elems);
+void  hvx_mul_f32_opt(const uint8_t * restrict src0,
+                      const uint8_t * restrict src1,
+                      uint8_t * restrict dst,
+                      const int num_elems);
+void  hvx_mul_mul_f32_opt(const uint8_t * restrict src0,
+                          const uint8_t * restrict src1,
+                          const uint8_t * restrict src2,
+                          uint8_t * restrict dst,
+                          const int num_elems);
+void  hvx_mul_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems);
+void  hvx_add_f32(const uint8_t * restrict src0,
+                  const uint8_t * restrict src1,
+                  uint8_t * restrict dst,
+                  const int num_elems);
+void  hvx_add_f32_opt(const uint8_t * restrict src0,
+                      const uint8_t * restrict src1,
+                      uint8_t * restrict dst,
+                      const int num_elems);
+void  hvx_add_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems);
+void  hvx_sub_f32(const uint8_t * restrict src0,
+                  const uint8_t * restrict src1,
+                  uint8_t * restrict dst,
+                  const int num_elems);
+void  hvx_sub_f32_opt(const uint8_t * restrict src0,
+                      const uint8_t * restrict src1,
+                      uint8_t * restrict dst,
+                      const int num_elems);
+void  hvx_sub_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems);
+void  hvx_scale_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems, const float scale);
+void  hvx_inverse_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems);
+void  hvx_sigmoid_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems);
+void  hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems, bool negate);
+float hvx_self_max_f32(const uint8_t * restrict src, const int num_elems);
+float hvx_self_sum_f32(const uint8_t * restrict src, const int num_elems);
+void  hvx_min_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems);
+void  hvx_clamp_scalar_f32(const uint8_t * restrict src,
+                           const float limit_left,
+                           const float limit_right,
+                           uint8_t * restrict dst,
+                           const int num_elems);
+
+#endif /* HVX_UTILS_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -0,0 +1,945 @@
+#pragma clang diagnostic ignored "-Wgnu-zero-variadic-macro-arguments"
+#pragma clang diagnostic ignored "-Wunused-function"
+
+#define FARF_ERROR  1
+#define FARF_HIGH   1
+#define FARF_MEDIUM 0
+#define FARF_LOW    0
+#include <AEEStdErr.h>
+#include <dspqueue.h>
+#include <HAP_compute_res.h>
+#include <HAP_etm_config.h>
+#include <HAP_farf.h>
+#include <HAP_mem.h>
+#include <HAP_perf.h>
+#include <HAP_power.h>
+#include <HAP_ps.h>
+#include <qurt.h>
+#include <qurt_thread.h>
+#include <remote.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "ops-utils.h"
+#include "worker-pool.h"
+
+AEEResult htp_iface_open(const char * uri, remote_handle64 * handle) {
+    struct htp_context * ctx;
+    int                  err = 0;
+
+    ctx = calloc(1, sizeof(*ctx));
+    if (ctx == NULL) {
+        return AEE_ENOMEMORY;
+    }
+
+    // Use the context structure as a handle
+    *handle = (remote_handle64) ctx;
+
+    // Enable FARF logs
+    HAP_setFARFRuntimeLoggingParams(0xffff, NULL, 0);
+
+    // Set client class
+    {
+        HAP_power_request_t request;
+        memset(&request, 0, sizeof(HAP_power_request_t));
+        request.type    = HAP_power_set_apptype;
+        request.apptype = HAP_POWER_COMPUTE_CLIENT_CLASS;
+
+        if ((err = HAP_power_set((void *) ctx, &request)) != 0) {
+            return err;
+        }
+    }
+
+    {
+        HAP_power_request_t request;
+        memset(&request, 0, sizeof(request));
+
+        request.type                              = HAP_power_set_DCVS_v3;
+        request.dcvs_v3.set_dcvs_enable           = TRUE;
+        request.dcvs_v3.dcvs_enable               = TRUE;
+        request.dcvs_v3.dcvs_option               = HAP_DCVS_V2_PERFORMANCE_MODE;
+        request.dcvs_v3.set_bus_params            = TRUE;
+        request.dcvs_v3.bus_params.min_corner     = HAP_DCVS_VCORNER_MAX;
+        request.dcvs_v3.bus_params.max_corner     = HAP_DCVS_VCORNER_MAX;
+        request.dcvs_v3.bus_params.target_corner  = HAP_DCVS_VCORNER_MAX;
+        request.dcvs_v3.set_core_params           = TRUE;
+        request.dcvs_v3.core_params.min_corner    = HAP_DCVS_VCORNER_MAX;
+        request.dcvs_v3.core_params.max_corner    = HAP_DCVS_VCORNER_MAX;
+        request.dcvs_v3.core_params.target_corner = HAP_DCVS_VCORNER_MAX;
+        request.dcvs_v3.set_sleep_disable         = TRUE;
+        request.dcvs_v3.sleep_disable             = TRUE;
+        if ((err = HAP_power_set((void *) ctx, &request)) != 0) {
+            return err;
+        }
+
+        memset(&request, 0, sizeof(request));
+        request.type         = HAP_power_set_HVX;
+        request.hvx.power_up = TRUE;
+        if ((err = HAP_power_set((void *) ctx, &request)) != 0) {
+            return err;
+        }
+    }
+
+    {
+        // Power on HMX
+        HAP_power_request_t request;
+        memset(&request, 0, sizeof(HAP_power_request_t));
+        request.type         = HAP_power_set_HMX;
+        request.hmx.power_up = TRUE;
+        FARF(ALWAYS, "Powering HMX on\n");
+        err = HAP_power_set((void *) &ctx, &request);
+        if (err != AEE_SUCCESS) {
+            FARF(ERROR, "Error powering on HMX.");
+            return err;
+        }
+    }
+
+    return AEE_SUCCESS;
+}
+
+AEEResult htp_iface_close(remote_handle64 handle) {
+    struct htp_context * ctx = (struct htp_context *) handle;
+
+    if (!ctx) {
+        return AEE_EBADPARM;
+    }
+
+    if (ctx->queue) {
+        FARF(ERROR, "Closing handle with queue still open");
+        return AEE_EITEMBUSY;
+    }
+
+    free(ctx);
+    return AEE_SUCCESS;
+}
+
+AEEResult htp_iface_enable_etm(remote_handle64 handle) {
+    int err = HAP_user_etm_enable();
+    if (err) {
+        if (err == AEE_EVERSIONNOTSUPPORT) {
+            FARF(ERROR, "API HAP_user_etm_enable is not supported\n");
+        } else {
+            FARF(ERROR, "Error executing HAP_user_etm_enable with error code : 0x%x\n", err);
+        }
+    }
+    return err;
+}
+
+AEEResult htp_iface_disable_etm(remote_handle64 handle) {
+    int err = HAP_user_etm_disable();
+    if (err) {
+        if (err == AEE_EVERSIONNOTSUPPORT) {
+            FARF(ERROR, "API HAP_user_etm_disable is not supported\n");
+        } else {
+            FARF(ERROR, "Error executing HAP_user_etm_disable with error code : 0x%x\n", err);
+        }
+    }
+    return err;
+}
+
+static int vtcm_acquire(struct htp_context * ctx) {
+    if (!ctx->vtcm_valid) {
+        // Temporarily bump thread priority to make sure it's higher than other sessions.
+        // This way the resource manager will notify the other thread to release VTCM.
+        // Note that we need to reaquire VTCM at normal priority for this to work next time.
+        qurt_thread_set_priority(qurt_thread_get_id(), ctx->thread_prio - 10);
+        HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 1000000);
+        HAP_compute_res_release_cached(ctx->vtcm_rctx);
+        qurt_thread_set_priority(qurt_thread_get_id(), ctx->thread_prio);
+
+        HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 1000000);
+        ctx->vtcm_valid = true;
+    }
+
+    ctx->vtcm_inuse = true;
+    return 0;
+}
+
+static int vtcm_release(struct htp_context * ctx) {
+    ctx->vtcm_inuse = false;
+
+    if (ctx->vtcm_valid && ctx->vtcm_needs_release) {
+        ctx->vtcm_valid         = false;
+        ctx->vtcm_needs_release = false;
+        HAP_compute_res_release_cached(ctx->vtcm_rctx);
+    }
+
+    return 0;
+}
+
+static int vtcm_release_callback(unsigned int rctx, void * state) {
+    struct htp_context * ctx = (struct htp_context *) state;
+
+    if (!ctx || ctx->vtcm_rctx != rctx) {
+        return AEE_EBADPARM;
+    }
+
+    // If VTCM is not inuse (not processing Ops) release it right here
+    // otherwise we'll release it once we're done with the current Op.
+
+    if (ctx->vtcm_inuse) {
+        ctx->vtcm_needs_release = false;
+        return 0;
+    }
+
+    ctx->vtcm_valid = false;
+    HAP_compute_res_release_cached(ctx->vtcm_rctx);
+
+    return 0;
+}
+
+static int vtcm_alloc(struct htp_context * ctx) {
+    unsigned int vtcm_size = 8 * 1024 * 1024;  // 8MB default
+    HAP_compute_res_query_VTCM(0, &vtcm_size, NULL, NULL, NULL);
+
+    compute_res_attr_t attr;
+    HAP_compute_res_attr_init(&attr);
+    HAP_compute_res_attr_set_serialize(&attr, 0);
+    HAP_compute_res_attr_set_cache_mode(&attr, 1);
+    HAP_compute_res_attr_set_vtcm_param_v2(&attr, vtcm_size, vtcm_size, vtcm_size);
+    HAP_compute_res_attr_set_release_callback(&attr, vtcm_release_callback, (void *) ctx);
+    HAP_compute_res_attr_set_hmx_param(&attr, 1);
+
+    // Allocate VTCM for scratch pads
+    uint32_t rctx = HAP_compute_res_acquire(&attr, 1000000 /* timeout */);
+    if (!rctx) {
+        FARF(ERROR, "failed to allocate %zu bytes VTCM\n", ctx->vtcm_size);
+        return AEE_ENOMEMORY;
+    }
+
+    void * vtcm_ptr;
+    if (HAP_compute_res_attr_get_vtcm_ptr_v2(&attr, &vtcm_ptr, &vtcm_size) != 0) {
+        HAP_compute_res_release(rctx);
+        FARF(ERROR, "failed to allocate %zu bytes VTCM (new)\n", ctx->vtcm_size);
+        return AEE_ENOMEMORY;
+    }
+
+    ctx->vtcm_base          = (uint8_t *) vtcm_ptr;
+    ctx->vtcm_size          = vtcm_size;
+    ctx->vtcm_rctx          = rctx;
+    ctx->vtcm_valid         = false;
+    ctx->vtcm_inuse         = false;
+    ctx->vtcm_needs_release = false;
+
+    return 0;
+}
+
+static void vtcm_free(struct htp_context * ctx) {
+    if (ctx->vtcm_rctx) {
+        HAP_compute_res_release(ctx->vtcm_rctx);
+        ctx->vtcm_base = 0;
+        ctx->vtcm_rctx = 0;
+    }
+}
+
+static void htp_packet_callback(dspqueue_t queue, int error, void * context);
+static void htp_error_callback(dspqueue_t queue, int error, void * context);
+
+AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_queue_id, uint32 n_hvx) {
+    struct htp_context * ctx = (struct htp_context *) handle;
+
+    if (!ctx) {
+        return AEE_EBADPARM;
+    }
+
+    if (ctx->queue) {
+        FARF(ERROR, "Queue already open");
+        return AEE_EITEMBUSY;
+    }
+
+    // Import queue created on the CPU
+    int err = dspqueue_import(dsp_queue_id,         // Queue ID from dspqueue_export
+                              htp_packet_callback,  // Packet callback
+                              htp_error_callback,   // Error callback; no errors expected on the DSP
+                              (void *) ctx,         // Callback context
+                              &ctx->queue);
+
+    if (err) {
+        FARF(ERROR, "Queue import failed with 0x%08x", (unsigned) err);
+        return err;
+    }
+
+    ctx->thread_id   = qurt_thread_get_id();
+    ctx->thread_prio = qurt_thread_get_priority(ctx->thread_id);
+
+    // allocate VTCM
+    err = vtcm_alloc(ctx);
+    if (err != AEE_SUCCESS) {
+        FARF(ERROR, "Unable to allocate VTCM");
+        return AEE_ENOMEMORY;
+    }
+
+    qurt_sysenv_max_hthreads_t hw_threads;
+    qurt_sysenv_get_max_hw_threads(&hw_threads);
+    uint32_t hw_nhvx = (qurt_hvx_get_units() >> 8) & 0xFF;
+
+    if (n_hvx == 0) {
+        n_hvx = hw_nhvx;
+    }
+    if (n_hvx > hw_threads.max_hthreads) {
+        n_hvx = hw_threads.max_hthreads;
+    }
+    if (n_hvx > HTP_MAX_NTHREADS) {
+        n_hvx = HTP_MAX_NTHREADS;
+    }
+
+    ctx->n_threads = n_hvx;
+    for (int i = 0; i < ctx->n_threads; i++) {
+        ctx->dma[i] = dma_queue_create(HTP_SPAD_SRC0_NROWS * 2);
+    }
+
+    // init worker pool
+    err = worker_pool_init(&ctx->worker_pool, n_hvx);
+    if (err != AEE_SUCCESS) {
+        FARF(ERROR, "Unable to create worker pool");
+        return err;
+    }
+
+    FARF(HIGH, "session %u started: n-hvx %u vtcm-size %zu vtcm-rctx %u n-threads %u thread-id %d thread-prio %d \n",
+         sess_id, hw_nhvx, ctx->vtcm_size, ctx->vtcm_rctx, ctx->n_threads, ctx->thread_id, ctx->thread_prio);
+
+    return AEE_SUCCESS;
+}
+
+AEEResult htp_iface_stop(remote_handle64 handle) {
+    struct htp_context * ctx = (struct htp_context *) handle;
+    if (!ctx) {
+        return AEE_EBADPARM;
+    }
+
+    if (!ctx->queue) {
+        FARF(ERROR, "Queue not open");
+        return AEE_EBADSTATE;
+    }
+
+    // Close queue. dspqueue_close() will also wait for callbacks to finish.
+    int err    = dspqueue_close(ctx->queue);
+    ctx->queue = NULL;
+    if (err != 0) {
+        FARF(ERROR, "Queue close failed with 0x%08x", (unsigned) err);
+        return err;
+    }
+
+    if (ctx->worker_pool) {
+        // Release worker pool
+        worker_pool_release(&ctx->worker_pool);
+    }
+
+    for (int i = 0; i < ctx->n_threads; i++) {
+        dma_queue_delete(ctx->dma[i]);
+    }
+
+    vtcm_free(ctx);
+
+    return AEE_SUCCESS;
+}
+
+static void htp_error_callback(dspqueue_t queue, int error, void * context) {
+    // No errors expected on the DSP.
+    FARF(ERROR, "Error callback: 0x%08x", (unsigned) error);
+}
+
+struct profile_data {
+    uint64_t usecs;
+    uint64_t cycles;
+    uint64_t pkts;
+};
+
+static inline void profile_start(struct profile_data * d) {
+    d->usecs  = HAP_perf_get_qtimer_count();
+    d->cycles = htp_get_cycles();
+    d->pkts   = htp_get_pktcnt();
+}
+
+static inline void profile_stop(struct profile_data * d) {
+    d->usecs  = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - d->usecs);
+    d->cycles = htp_get_cycles() - d->cycles;
+    d->pkts   = htp_get_pktcnt() - d->pkts;
+}
+
+static int send_htp_rsp(struct htp_context *     c,
+                        uint32_t                 op,
+                        uint32_t                 status,
+                        struct dspqueue_buffer * bufs,
+                        size_t                   n_bufs,
+                        struct profile_data *    prof) {
+    // Prep response struct
+    struct htp_general_rsp rsp;
+    rsp.op          = op;
+    rsp.status      = status;
+    rsp.prof_usecs  = prof->usecs;
+    rsp.prof_cycles = prof->cycles;
+    rsp.prof_pkts   = prof->pkts;
+
+    int err = dspqueue_write(c->queue,
+                             0,                       // Flags
+                             n_bufs,
+                             bufs,                    // Buffer references
+                             sizeof(rsp),
+                             (const uint8_t *) &rsp,  // Message
+                             DSPQUEUE_TIMEOUT_NONE);
+
+    if (err != 0) {
+        FARF(ERROR, "dspqueue_write failed: 0x%08x", (unsigned) err);
+    }
+
+    return err;
+}
+
+static void proc_matmul_req(struct htp_context *     ctx,
+                            struct htp_general_req * req,
+                            struct dspqueue_buffer * bufs,
+                            size_t                   n_bufs) {
+    // Prep response buffer structs (needed for error responses, etc)
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+    memset(rsp_bufs, 0, sizeof(rsp_bufs));
+    rsp_bufs[0].fd     = bufs[0].fd;
+    rsp_bufs[0].ptr    = bufs[0].ptr;
+    rsp_bufs[0].size   = bufs[0].size;
+    rsp_bufs[0].offset = bufs[0].offset;
+    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    rsp_bufs[1].fd     = bufs[1].fd;
+    rsp_bufs[1].ptr    = bufs[1].ptr;
+    rsp_bufs[1].size   = bufs[1].size;
+    rsp_bufs[1].offset = bufs[1].offset;
+    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[2].fd     = bufs[2].fd;
+    rsp_bufs[2].ptr    = bufs[2].ptr;
+    rsp_bufs[2].size   = bufs[2].size;
+    rsp_bufs[2].offset = bufs[2].offset;
+    rsp_bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
+                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.src1                   = req->src1;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.src1.data = (uint32_t) bufs[1].ptr;
+    octx.dst.data  = (uint32_t) bufs[2].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_matmul(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 3, &prof);
+}
+
+static void proc_matmul_id_req(struct htp_context *     ctx,
+                               struct htp_general_req * req,
+                               struct dspqueue_buffer * bufs,
+                               size_t                   n_bufs) {
+    // Prep response buffer structs (needed for error responses, etc)
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+    memset(rsp_bufs, 0, sizeof(rsp_bufs));
+    rsp_bufs[0].fd     = bufs[0].fd;
+    rsp_bufs[0].ptr    = bufs[0].ptr;
+    rsp_bufs[0].size   = bufs[0].size;
+    rsp_bufs[0].offset = bufs[0].offset;
+    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    rsp_bufs[1].fd     = bufs[1].fd;
+    rsp_bufs[1].ptr    = bufs[1].ptr;
+    rsp_bufs[1].size   = bufs[1].size;
+    rsp_bufs[1].offset = bufs[1].offset;
+    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    rsp_bufs[2].fd     = bufs[2].fd;
+    rsp_bufs[2].ptr    = bufs[2].ptr;
+    rsp_bufs[2].size   = bufs[2].size;
+    rsp_bufs[2].offset = bufs[2].offset;
+    rsp_bufs[2].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[3].fd     = bufs[3].fd;
+    rsp_bufs[3].ptr    = bufs[3].ptr;
+    rsp_bufs[3].size   = bufs[3].size;
+    rsp_bufs[3].offset = bufs[3].offset;
+    rsp_bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
+                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.src1                   = req->src1;
+    octx.src2                   = req->src2;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.src1.data = (uint32_t) bufs[1].ptr;
+    octx.src2.data = (uint32_t) bufs[2].ptr;
+    octx.dst.data  = (uint32_t) bufs[3].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_matmul_id(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 4, &prof);
+}
+
+static void proc_binary_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+    memset(rsp_bufs, 0, sizeof(rsp_bufs));
+
+    rsp_bufs[0].fd     = bufs[0].fd;
+    rsp_bufs[0].ptr    = bufs[0].ptr;
+    rsp_bufs[0].offset = bufs[0].offset;
+    rsp_bufs[0].size   = bufs[0].size;
+    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    rsp_bufs[1].fd     = bufs[1].fd;
+    rsp_bufs[1].ptr    = bufs[1].ptr;
+    rsp_bufs[1].offset = bufs[1].offset;
+    rsp_bufs[1].size   = bufs[1].size;
+    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[2].fd     = bufs[2].fd;
+    rsp_bufs[2].ptr    = bufs[2].ptr;
+    rsp_bufs[2].offset = bufs[2].offset;
+    rsp_bufs[2].size   = bufs[2].size;
+    rsp_bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
+                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.src1                   = req->src1;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.src1.data = (uint32_t) bufs[1].ptr;
+    octx.dst.data  = (uint32_t) bufs[2].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_binary(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 3, &prof);
+}
+
+static void proc_add_id_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+    memset(rsp_bufs, 0, sizeof(rsp_bufs));
+
+    rsp_bufs[0].fd     = bufs[0].fd;
+    rsp_bufs[0].ptr    = bufs[0].ptr;
+    rsp_bufs[0].offset = bufs[0].offset;
+    rsp_bufs[0].size   = bufs[0].size;
+    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    rsp_bufs[1].fd     = bufs[1].fd;
+    rsp_bufs[1].ptr    = bufs[1].ptr;
+    rsp_bufs[1].offset = bufs[1].offset;
+    rsp_bufs[1].size   = bufs[1].size;
+    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    rsp_bufs[2].fd     = bufs[2].fd;
+    rsp_bufs[2].ptr    = bufs[2].ptr;
+    rsp_bufs[2].offset = bufs[2].offset;
+    rsp_bufs[2].size   = bufs[2].size;
+    rsp_bufs[2].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[3].fd     = bufs[3].fd;
+    rsp_bufs[3].ptr    = bufs[3].ptr;
+    rsp_bufs[3].offset = bufs[3].offset;
+    rsp_bufs[3].size   = bufs[3].size;
+    rsp_bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
+                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.src1                   = req->src1;
+    octx.src2                   = req->src2;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.src1.data = (uint32_t) bufs[1].ptr;
+    octx.src2.data = (uint32_t) bufs[2].ptr;
+    octx.dst.data  = (uint32_t) bufs[3].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_binary(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 4, &prof);
+}
+
+static void proc_unary_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+    memset(rsp_bufs, 0, sizeof(rsp_bufs));
+
+    rsp_bufs[0].fd     = bufs[0].fd;
+    rsp_bufs[0].ptr    = bufs[0].ptr;
+    rsp_bufs[0].offset = bufs[0].offset;
+    rsp_bufs[0].size   = bufs[0].size;
+    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[1].fd     = bufs[1].fd;
+    rsp_bufs[1].ptr    = bufs[1].ptr;
+    rsp_bufs[1].offset = bufs[1].offset;
+    rsp_bufs[1].size   = bufs[1].size;
+    rsp_bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
+                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.dst.data  = (uint32_t) bufs[1].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_unary(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 2, &prof);
+}
+
+static void proc_activations_req(struct htp_context *     ctx,
+                                 struct htp_general_req * req,
+                                 struct dspqueue_buffer * bufs,
+                                 uint32_t                 n_bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+    memset(rsp_bufs, 0, sizeof(rsp_bufs));
+
+    rsp_bufs[0].fd     = bufs[0].fd;
+    rsp_bufs[0].ptr    = bufs[0].ptr;
+    rsp_bufs[0].offset = bufs[0].offset;
+    rsp_bufs[0].size   = bufs[0].size;
+    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    int write_idx = 1;
+    if (3 == n_bufs) {
+        rsp_bufs[1].fd     = bufs[1].fd;
+        rsp_bufs[1].ptr    = bufs[1].ptr;
+        rsp_bufs[1].offset = bufs[1].offset;
+        rsp_bufs[1].size   = bufs[1].size;
+        rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+        write_idx = 2;
+    }
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[write_idx].fd     = bufs[write_idx].fd;
+    rsp_bufs[write_idx].ptr    = bufs[write_idx].ptr;
+    rsp_bufs[write_idx].offset = bufs[write_idx].offset;
+    rsp_bufs[write_idx].size   = bufs[write_idx].size;
+    rsp_bufs[write_idx].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
+                                 DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+                                 DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    if (3 == n_bufs) {
+        octx.src1 = req->src1;
+    }
+    octx.dst   = req->dst;
+    octx.flags = req->flags;
+    octx.op    = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    if (3 == n_bufs) {
+        octx.src1.data = (uint32_t) bufs[1].ptr;
+        octx.dst.data  = (uint32_t) bufs[2].ptr;
+    } else {
+        octx.dst.data = (uint32_t) bufs[1].ptr;
+    }
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        if (octx.op == HTP_OP_SOFTMAX) {
+            rsp_status = op_softmax(&octx);
+        } else {
+            rsp_status = op_activations(&octx);
+        }
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, n_bufs, &prof);
+}
+
+static void proc_rope_req(struct htp_context *     ctx,
+                          struct htp_general_req * req,
+                          struct dspqueue_buffer * bufs,
+                          uint32_t                 n_bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+    memset(rsp_bufs, 0, sizeof(rsp_bufs));
+
+    rsp_bufs[0].fd     = bufs[0].fd;
+    rsp_bufs[0].ptr    = bufs[0].ptr;
+    rsp_bufs[0].offset = bufs[0].offset;
+    rsp_bufs[0].size   = bufs[0].size;
+    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    rsp_bufs[1].fd     = bufs[1].fd;
+    rsp_bufs[1].ptr    = bufs[1].ptr;
+    rsp_bufs[1].offset = bufs[1].offset;
+    rsp_bufs[1].size   = bufs[1].size;
+    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+    int write_idx = 2;
+    if (4 == n_bufs) {
+        rsp_bufs[write_idx].fd     = bufs[write_idx].fd;
+        rsp_bufs[write_idx].ptr    = bufs[write_idx].ptr;
+        rsp_bufs[write_idx].offset = bufs[write_idx].offset;
+        rsp_bufs[write_idx].size   = bufs[write_idx].size;
+        rsp_bufs[write_idx].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+
+        write_idx++;
+    }
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[write_idx].fd     = bufs[write_idx].fd;
+    rsp_bufs[write_idx].ptr    = bufs[write_idx].ptr;
+    rsp_bufs[write_idx].offset = bufs[write_idx].offset;
+    rsp_bufs[write_idx].size   = bufs[write_idx].size;
+    rsp_bufs[write_idx].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
+                                 DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+                                 DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.src1                   = req->src1;
+    if (4 == n_bufs) {
+        octx.src2 = req->src2;
+    }
+    octx.dst   = req->dst;
+    octx.flags = req->flags;
+    octx.op    = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.src1.data = (uint32_t) bufs[1].ptr;
+    if (4 == n_bufs) {
+        octx.src2.data = (uint32_t) bufs[2].ptr;
+        octx.dst.data  = (uint32_t) bufs[3].ptr;
+    } else {
+        octx.dst.data = (uint32_t) bufs[2].ptr;
+    }
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_rope(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, n_bufs, &prof);
+}
+
+static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
+    struct htp_context * ctx = (struct htp_context *) context;
+
+    // Repeatedly read packets from the queue until it's empty. We don't
+    // necessarily get a separate callback for each packet, and new packets
+    // may arrive while we're processing the previous one. This ensures we
+    // keep the DSP busy as much as possible and avoid waiting for the CPU.
+
+    while (1) {
+        struct htp_general_req req;
+        uint32_t               req_size;
+
+        struct dspqueue_buffer bufs[HTP_MAX_PACKET_BUFFERS];
+        uint32_t               n_bufs;
+        uint32_t               flags;
+
+        // Read packet from queue
+        int err = dspqueue_read_noblock(queue, &flags,
+                                        HTP_MAX_PACKET_BUFFERS,  // Maximum number of buffer references
+                                        &n_bufs,                 // Number of buffer references
+                                        bufs,                    // Buffer references
+                                        sizeof(req),             // Max message length
+                                        &req_size,               // Message length
+                                        (uint8_t *) &req);       // Message
+
+        if (err == AEE_EWOULDBLOCK) {
+            // Consumed all packets available for now
+            return;
+        }
+
+        if (err != 0) {
+            FARF(ERROR, "dspqueue_read_noblock failed: 0x%08x", (unsigned) err);
+            return;
+        }
+
+        if (req_size != sizeof(req)) {
+            FARF(ERROR, "Invalid request size");
+            continue;
+        }
+
+        if (req.flags & HTP_OPFLAGS_EARLY_WAKEUP) {
+            // Host wants early notification
+            dspqueue_write_early_wakeup_noblock(ctx->queue, 10, 0);
+        }
+
+        // Process packet based on its message type
+        switch (req.op) {
+            case HTP_OP_MUL_MAT:
+                if (n_bufs != 3) {
+                    FARF(ERROR, "Bad matmul-req buffer list");
+                    continue;
+                }
+                proc_matmul_req(ctx, &req, bufs, n_bufs);
+                break;
+
+            case HTP_OP_MUL_MAT_ID:
+                if (n_bufs != 4) {
+                    FARF(ERROR, "Bad matmul-id-req buffer list");
+                    continue;
+                }
+                proc_matmul_id_req(ctx, &req, bufs, n_bufs);
+                break;
+
+            case HTP_OP_MUL:
+            case HTP_OP_ADD:
+            case HTP_OP_SUB:
+                if (n_bufs != 3) {
+                    FARF(ERROR, "Bad binary-req buffer list");
+                    continue;
+                }
+                proc_binary_req(ctx, &req, bufs);
+                break;
+
+            case HTP_OP_RMS_NORM:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad unary-req buffer list");
+                    continue;
+                }
+
+                proc_unary_req(ctx, &req, bufs);
+                break;
+
+            case HTP_OP_UNARY_SILU:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad act-req buffer list");
+                    continue;
+                }
+                proc_activations_req(ctx, &req, bufs, n_bufs);
+                break;
+
+            case HTP_OP_GLU_SWIGLU:
+            case HTP_OP_SOFTMAX:
+                if ((n_bufs != 2) && (n_bufs != 3)) {
+                    FARF(ERROR, "Bad act-req buffer list");
+                    continue;
+                }
+                proc_activations_req(ctx, &req, bufs, n_bufs);
+                break;
+
+            case HTP_OP_ADD_ID:
+                if (n_bufs != 4) {
+                    FARF(ERROR, "Bad add-id-req buffer list");
+                    continue;
+                }
+                proc_add_id_req(ctx, &req, bufs);
+                break;
+
+            case HTP_OP_ROPE:
+                if ((n_bufs != 3) && (n_bufs != 4)) {
+                    FARF(ERROR, "Bad rope-req buffer list");
+                    continue;
+                }
+                proc_rope_req(ctx, &req, bufs, n_bufs);
+                break;
+
+            default:
+                FARF(ERROR, "Unknown Op %u", req.op);
+                break;
+        }
+    }
+}

ggml/src/ggml-hexagon/htp/ops-utils.h

@@ -0,0 +1,116 @@
+#ifndef OPS_UTILS_H
+#define OPS_UTILS_H
+
+#include "htp-msg.h"
+
+#ifndef MAX
+#    define MAX(a, b) ((a) > (b) ? (a) : (b))
+#endif
+
+#ifndef MIN
+#    define MIN(a, b) ((a) < (b) ? (a) : (b))
+#endif
+
+static inline uint64_t htp_get_cycles() {
+    uint64_t cycles = 0;
+    asm volatile(" %0 = c15:14\n" : "=r"(cycles));
+    return cycles;
+}
+
+static inline uint64_t htp_get_pktcnt() {
+    uint64_t pktcnt;
+    asm volatile(" %0 = c19:18\n" : "=r"(pktcnt));
+    return pktcnt;
+}
+
+static inline int32_t htp_is_aligned(void * addr, uint32_t align) {
+    return ((size_t) addr & (align - 1)) == 0;
+}
+
+static inline uint32_t htp_round_up(uint32_t n, uint32_t m) {
+    return m * ((n + m - 1) / m);
+}
+
+static inline void htp_l2fetch(const void * p, uint32_t height, uint32_t width, uint32_t stride) {
+    const uint64_t control = Q6_P_combine_RR(stride, Q6_R_combine_RlRl(width, height));
+    asm volatile(" l2fetch(%0,%1) " : : "r"(p), "r"(control));
+}
+
+static inline int32_t htp_is_one_chunk(void * addr, uint32_t n, uint32_t chunk_size) {
+    uint32_t left_off  = (size_t) addr & (chunk_size - 1);
+    uint32_t right_off = left_off + n;
+    return right_off <= chunk_size;
+}
+
+static inline void htp_dump_int8_line(char * pref, const int8_t * x, int n) {
+    char str[1024], *p = str;
+    p += sprintf(p, "%s: ", pref);
+    for (int i = 0; i < 16; i++) {
+        p += sprintf(p, "%d, ", x[i]);
+    }
+    FARF(HIGH, "%s\n", str);
+}
+
+static inline void htp_dump_uint8_line(char * pref, const uint8_t * x, uint32_t n) {
+    char str[1024], *p = str;
+    p += sprintf(p, "%s: ", pref);
+    for (int i = 0; i < n; i++) {
+        p += sprintf(p, "%d, ", x[i]);
+    }
+    FARF(HIGH, "%s\n", str);
+}
+
+static inline void htp_dump_int32_line(char * pref, const int32_t * x, uint32_t n) {
+    char str[1024], *p = str;
+    p += sprintf(p, "%s: ", pref);
+    for (int i = 0; i < n; i++) {
+        p += sprintf(p, "%d, ", (int) x[i]);
+    }
+    FARF(HIGH, "%s\n", str);
+}
+
+static inline void htp_dump_fp16_line(char * pref, const __fp16 * x, uint32_t n) {
+    char str[1024], *p = str;
+    p += sprintf(p, "%s: ", pref);
+    for (int i = 0; i < n; i++) {
+        p += sprintf(p, "%.6f, ", (float) x[i]);
+    }
+    FARF(HIGH, "%s\n", str);
+}
+
+static inline void htp_dump_fp32_line(char * pref, const float * x, uint32_t n) {
+    char str[1024], *p = str;
+    p += sprintf(p, "%s: ", pref);
+    for (int i = 0; i < n; i++) {
+        p += sprintf(p, "%.6f, ", x[i]);
+    }
+    FARF(HIGH, "%s\n", str);
+}
+
+static inline void htp_dump_f32(char * pref, const float * x, uint32_t n) {
+    uint32_t n0 = n / 16;
+    uint32_t n1 = n % 16;
+
+    uint32_t i = 0;
+    for (; i < n0; i++) {
+        htp_dump_fp32_line(pref, x + (16 * i), 16);
+    }
+    if (n1) {
+        htp_dump_fp32_line(pref, x + (16 * i), n1);
+    }
+}
+
+static inline void htp_dump_f16(char * pref, const __fp16 * x, uint32_t n) {
+    uint32_t n0 = n / 16;
+    uint32_t n1 = n % 16;
+
+    uint32_t i = 0;
+    for (; i < n0; i++) {
+        htp_dump_fp16_line(pref, x + (16 * i), 16);
+    }
+    if (n1) {
+        htp_dump_fp16_line(pref, x + (16 * i), n1);
+    }
+}
+
+#endif /* OPS_UTILS_H */

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

@@ -0,0 +1,418 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#ifdef HTP_DEBUG
+#    define FARF_HIGH 1
+#endif
+#include <HAP_farf.h>
+#include <HAP_mem.h>
+#include <HAP_perf.h>
+#include <HAP_ps.h>
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <qurt_thread.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+#include "ops-utils.h"
+
+#define htp_rope_preamble              \
+    const uint32_t ne00 = src0->ne[0]; \
+    const uint32_t ne01 = src0->ne[1]; \
+    const uint32_t ne02 = src0->ne[2]; \
+    const uint32_t ne03 = src0->ne[3]; \
+                                       \
+    const uint32_t ne0 = dst->ne[0];   \
+    const uint32_t ne1 = dst->ne[1];   \
+    const uint32_t ne2 = dst->ne[2];   \
+    const uint32_t ne3 = dst->ne[3];   \
+                                       \
+    const uint32_t nb00 = src0->nb[0]; \
+    const uint32_t nb01 = src0->nb[1]; \
+    const uint32_t nb02 = src0->nb[2]; \
+    const uint32_t nb03 = src0->nb[3]; \
+                                       \
+    const uint32_t nb0 = dst->nb[0];   \
+    const uint32_t nb1 = dst->nb[1];   \
+    const uint32_t nb2 = dst->nb[2];   \
+    const uint32_t nb3 = dst->nb[3];
+
+struct rope_th_ctx {
+    int32_t n_dims;
+    int32_t mode;
+    int32_t n_ctx_orig;
+    int32_t sections[4];
+
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+    float theta_scale;
+    float corr_dims[2];
+
+    struct htp_ops_context * octx;
+};
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / MAX(0.001f, high - low);
+
+    return (1 - MIN(1, MAX(0, y)));
+}
+
+static void rope_cache_init(const float   theta_base,
+                            float         freq_scale,
+                            const float * freq_factors,
+                            float *       corr_dims,
+                            uint32_t      ne0,
+                            float         ext_factor,
+                            float         mscale,
+                            float *       cache,
+                            float         theta_scale) {
+    // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
+    float theta = theta_base;
+
+    for (uint32_t i0 = 0; i0 < ne0; i0 += 2) {
+        const float ff = freq_factors ? freq_factors[i0 / 2] : 1.0f;
+
+        float theta_extrap = theta / ff;
+
+        // Get n-d rotational scaling corrected for extrapolation
+        float theta_interp = freq_scale * theta_extrap;
+        float theta2       = theta_interp;
+
+        if (ext_factor != 0.0f) {
+            float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+            theta2         = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+            // Get n-d magnitude scaling corrected for interpolation
+            mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
+        }
+
+        cache[i0 + 0] = cosf(theta2) * mscale;
+        cache[i0 + 1] = sinf(theta2) * mscale;
+
+        theta *= theta_scale;
+    }
+}
+
+#define M_PI 3.1415926535897932384626433
+
+static void rope_corr_dims(int     n_dims,
+                           int     n_ctx_orig,
+                           float   freq_base,
+                           float   beta_fast,
+                           float   beta_slow,
+                           float * dims) {
+    float start = floorf(n_dims * logf(n_ctx_orig / (beta_fast * 2 * (float) M_PI)) / (2 * logf(freq_base)));
+    float end   = ceilf(n_dims * logf(n_ctx_orig / (beta_slow * 2 * (float) M_PI)) / (2 * logf(freq_base)));
+    dims[0]     = MAX(0, start);
+    dims[1]     = MIN(n_dims - 1, end);
+}
+
+static void init_rope_ctx(struct rope_th_ctx * rope_ctx, struct htp_ops_context * octx) {
+    memset(rope_ctx, 0, sizeof(struct rope_th_ctx));
+
+    const int32_t * op_params = &octx->op_params[0];
+
+    rope_ctx->n_dims     = ((const int32_t *) op_params)[1];
+    rope_ctx->mode       = ((const int32_t *) op_params)[2];
+    rope_ctx->n_ctx_orig = ((const int32_t *) op_params)[4];
+
+    memcpy(&rope_ctx->freq_base, (int32_t *) op_params + 5, sizeof(float));
+    memcpy(&rope_ctx->freq_scale, (int32_t *) op_params + 6, sizeof(float));
+    memcpy(&rope_ctx->ext_factor, (int32_t *) op_params + 7, sizeof(float));
+    memcpy(&rope_ctx->attn_factor, (int32_t *) op_params + 8, sizeof(float));
+    memcpy(&rope_ctx->beta_fast, (int32_t *) op_params + 9, sizeof(float));
+    memcpy(&rope_ctx->beta_slow, (int32_t *) op_params + 10, sizeof(float));
+    memcpy(&rope_ctx->sections, (int32_t *) op_params + 11, sizeof(int) * 4);
+
+    rope_ctx->theta_scale = powf(rope_ctx->freq_base, -2.0f / rope_ctx->n_dims);
+
+    rope_corr_dims(rope_ctx->n_dims, rope_ctx->n_ctx_orig, rope_ctx->freq_base, rope_ctx->beta_fast,
+                   rope_ctx->beta_slow, rope_ctx->corr_dims);
+
+    rope_ctx->octx = octx;
+    FARF(HIGH, "rope-f32 n_dims:%d, ext_factor:%.6f, theta_scale:%.6f, attn_factor:%.6f\n", rope_ctx->n_dims,
+         rope_ctx->ext_factor, rope_ctx->theta_scale, rope_ctx->attn_factor);
+}
+
+static void hvx_calc_rope_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[1];
+
+    //dst[0] = x0*cos_theta - x1*sin_theta;
+    //dst[1] = x0*sin_theta + x1*cos_theta;
+
+    //src += 2;
+    //dst += 2;
+    // }
+
+    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
+
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v0 = *(HVX_Vector *) src0_curr;
+        HVX_Vector v1 = *(HVX_Vector *) (src0_curr + VLEN);
+
+        HVX_Vector v2 = *(HVX_Vector *) theta_curr;
+        HVX_Vector v3 = *(HVX_Vector *) (theta_curr + VLEN);
+
+        HVX_VectorPair vx0_x1   = Q6_W_vdeal_VVR(v1, v0, -4);  // vx0_x1[0] = x0, vx0_x1[1] = x1
+        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(Q6_V_lo_W(vx0_x1), Q6_V_lo_W(vcos_sin));
+        HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_hi_W(vcos_sin));
+        HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_lo_W(vcos_sin));
+        HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), 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_VectorPair vstore = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(v5), Q6_Vsf_equals_Vqf32(v4), -4);
+
+        *(HVX_Vector *) dst_curr          = Q6_V_lo_W(vstore);
+        *(HVX_Vector *) (dst_curr + VLEN) = Q6_V_hi_W(vstore);
+
+        src0_curr += 2 * VLEN;
+        theta_curr += 2 * VLEN;
+        dst_curr += 2 * VLEN;
+    }
+}
+
+static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
+                         const uint32_t       ir0,
+                         const uint32_t       ir1,
+                         int                  nth,
+                         int                  ith,
+                         int                  opt_path) {
+    struct htp_ops_context * octx = rope_ctx->octx;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+    const struct htp_tensor * src2 = &octx->src2;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    htp_rope_preamble;
+
+    const int32_t * pos = (const int32_t *) src1->data;
+
+    float * wp0 = (float *) (octx->src0_spad.data + (ith * nb01));
+
+    const float * freq_factors = NULL;
+    if (src2 != NULL) {
+        freq_factors = (const float *) src2->data;
+    }
+
+    int ir = 0;
+
+    for (uint32_t i3 = 0; i3 < ne3; i3++) {      // batch
+        for (uint32_t i2 = 0; i2 < ne2; i2++) {  // seq-len
+            const int32_t p = pos[i2];
+
+            rope_cache_init(p, rope_ctx->freq_scale, freq_factors, rope_ctx->corr_dims, ne0, rope_ctx->ext_factor,
+                            rope_ctx->attn_factor, wp0, rope_ctx->theta_scale);
+
+            for (uint32_t i1 = 0; i1 < ne1; i1++) {  // attn-heads
+                if (ir++ < ir0) {
+                    continue;
+                }
+                if (ir > ir1) {
+                    break;
+                }
+
+                const float * src      = (float *) ((char *) src0->data + i3 * nb03 + i2 * nb02 + i1 * nb01);
+                float *       dst_data = (float *) ((char *) dst->data + i3 * nb3 + i2 * nb2 + i1 * nb1);
+
+                const float * src_loc      = src;
+                float *       dst_data_loc = dst_data;
+
+                if (1 == opt_path) {
+                    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];
+
+                        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;
+                    }
+                }
+
+                for (uint32_t i0 = rope_ctx->n_dims; i0 < ne0; i0 += 2) {
+                    dst_data_loc[0] = src_loc[0];
+                    dst_data_loc[1] = src_loc[1];
+
+                    src_loc += 2;
+                    dst_data_loc += 2;
+                }
+            }
+        }
+    }
+}
+
+static void rope_job_f32_per_thread(struct rope_th_ctx * rope_ctx, int nth, int ith) {
+    struct htp_ops_context * octx = rope_ctx->octx;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    htp_rope_preamble;
+
+    const uint32_t src0_nrows            = ne01 * ne02 * ne03;  // src0 rows
+    const uint32_t src0_nrows_per_thread = octx->src0_nrows_per_thread;
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    int is_aligned = 1;
+    int opt_path   = 0;
+    if ((0 == htp_is_aligned((void *) src0->data, VLEN)) || (0 == htp_is_aligned((void *) src1->data, VLEN)) ||
+        (0 == htp_is_aligned((void *) dst->data, VLEN))) {
+        FARF(HIGH, "rope-f32: unaligned addresses in rope op, possibly slower execution\n");
+        is_aligned = 0;
+    }
+    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    rope_hex_f32(rope_ctx, src0_start_row, src0_end_row, nth, ith, opt_path);
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "rope-f32: %d/%d/%d: (%u:%u) usec %u\n", ith, nth, opt_path, src0_start_row, src0_end_row,
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static void rope_job_dispatcher_f32(unsigned int n, unsigned int i, void * data) {
+    struct rope_th_ctx * rope_ctx = (struct rope_th_ctx *) data;
+
+    rope_job_f32_per_thread(rope_ctx, n, i);
+}
+
+static int execute_op_rope_f32(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+    const struct htp_tensor * src2 = &octx->src2;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    worker_callback_t op_func;
+    const char *      op_type = NULL;
+
+    struct rope_th_ctx rope_ctx;
+
+    switch (octx->op) {
+        case HTP_OP_ROPE:
+            op_func = rope_job_dispatcher_f32;
+            op_type = "rope-f32";
+
+            init_rope_ctx(&rope_ctx, octx);
+            break;
+
+        default:
+            FARF(ERROR, "Unsupported Op %u\n", octx->op);
+            return HTP_STATUS_NO_SUPPORT;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
+    const size_t src0_row_size = src0->nb[1];
+    const size_t src1_row_size = src0_row_size;
+    const size_t dst_row_size  = dst->nb[1];
+
+    // VTCM scratchpads for all tensors
+    // N rows per thread, padded to HVX vector size
+    octx->dst_spad.size  = htp_round_up(dst_row_size, 128) * n_threads;
+    octx->src0_spad.size = htp_round_up(src0_row_size, 128) * n_threads;
+    octx->src1_spad.size = htp_round_up(src1_row_size, 128) * n_threads;
+
+    size_t spad_size = octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;
+
+    if (src2->ne[0]) {
+        FARF(HIGH,
+             "%s: %ux%ux%ux%u (x %ux%ux%ux%u x %ux%ux%ux%u) -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u "
+             "dst-spad-size %u\n",
+             op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
+             src1->ne[3], src2->ne[0], src2->ne[1], src2->ne[2], src2->ne[3], dst->ne[0], dst->ne[1], dst->ne[2],
+             dst->ne[3], octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
+    } else {
+        FARF(HIGH,
+             "%s: %ux%ux%ux%u (%ux%ux%ux%u) -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n",
+             op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
+             src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], octx->src0_spad.size, octx->src1_spad.size,
+             octx->dst_spad.size);
+    }
+
+    // Make sure the reserved vtcm size is sufficient
+    if (octx->ctx->vtcm_size < spad_size) {
+        FARF(ERROR, "%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
+             spad_size);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
+
+    uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+
+    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
+        uint32_t n_jobs             = MIN(n_threads, src0_nrows);
+        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+        worker_pool_run_func(octx->ctx->worker_pool, op_func, &rope_ctx, n_jobs);
+    }
+
+    return err;
+}
+
+int op_rope(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    switch (octx->src0.type) {
+        case HTP_TYPE_F32:
+            err = execute_op_rope_f32(octx);
+            break;
+
+        default:
+            err = HTP_STATUS_NO_SUPPORT;
+            break;
+    }
+
+    return err;
+}

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

@@ -0,0 +1,402 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#ifdef HTP_DEBUG
+#    define FARF_HIGH 1
+#endif
+#include <HAP_farf.h>
+#include <HAP_mem.h>
+#include <HAP_perf.h>
+#include <HAP_ps.h>
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <qurt_thread.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+#include "ops-utils.h"
+
+#define htp_softmax_preamble3                              \
+    const uint32_t ne00 = src0->ne[0];                     \
+    const uint32_t ne01 = src0->ne[1];                     \
+    const uint32_t ne02 = src0->ne[2];                     \
+    const uint32_t ne03 = src0->ne[3];                     \
+                                                           \
+    const uint32_t nb00 = src0->nb[0];                     \
+    const uint32_t nb01 = src0->nb[1];                     \
+    const uint32_t nb02 = src0->nb[2];                     \
+    const uint32_t nb03 = src0->nb[3];                     \
+                                                           \
+    const uint32_t ne10 = (src1->ne[0]) ? src1->ne[0] : 1; \
+    const uint32_t ne11 = (src1->ne[0]) ? src1->ne[1] : 1; \
+    const uint32_t ne12 = (src1->ne[0]) ? src1->ne[2] : 1; \
+    const uint32_t ne13 = (src1->ne[0]) ? src1->ne[3] : 1; \
+                                                           \
+    const uint32_t nb10 = (src1->ne[0]) ? src1->nb[0] : 1; \
+    const uint32_t nb11 = (src1->ne[0]) ? src1->nb[1] : 1; \
+    const uint32_t nb12 = (src1->ne[0]) ? src1->nb[2] : 1; \
+    const uint32_t nb13 = (src1->ne[0]) ? src1->nb[3] : 1; \
+                                                           \
+    const uint32_t ne0 = dst->ne[0];                       \
+    const uint32_t ne1 = dst->ne[1];                       \
+    const uint32_t ne2 = dst->ne[2];                       \
+    const uint32_t ne3 = dst->ne[3];                       \
+                                                           \
+    const uint32_t nb0 = dst->nb[0];                       \
+    const uint32_t nb1 = dst->nb[1];                       \
+    const uint32_t nb2 = dst->nb[2];                       \
+    const uint32_t nb3 = dst->nb[3];
+
+struct softmax_th_ctx {
+    bool     use_f16;
+    bool     use_src1;
+    uint32_t n_head;
+    uint32_t n_head_log2;
+
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+
+    struct htp_ops_context * octx;
+};
+
+static void init_softmax_ctx(struct softmax_th_ctx * softmax_ctx, struct htp_ops_context * octx) {
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+
+    memset(softmax_ctx, 0, sizeof(struct softmax_th_ctx));
+
+    memcpy(&softmax_ctx->scale, (float *) octx->op_params, sizeof(float));
+    memcpy(&softmax_ctx->max_bias, (float *) octx->op_params + 1, sizeof(float));
+
+    softmax_ctx->n_head      = src0->ne[2];
+    softmax_ctx->n_head_log2 = 1u << (uint32_t) floor(log2(softmax_ctx->n_head));
+
+    softmax_ctx->m0 = powf(2.0f, -(softmax_ctx->max_bias) / softmax_ctx->n_head_log2);
+    softmax_ctx->m1 = powf(2.0f, -(softmax_ctx->max_bias / 2.0f) / softmax_ctx->n_head_log2);
+
+    softmax_ctx->use_src1 = (src1->ne[0] != 0);
+    softmax_ctx->use_f16  = (src1->ne[0] != 0) && (src1->type == HTP_TYPE_F16);
+
+    softmax_ctx->octx = octx;
+}
+
+static void hvx_fast_softmax_prep_f32(const uint8_t * restrict src,
+                                      uint8_t * restrict dst,
+                                      const int num_elems,
+                                      float     scale,
+                                      const uint8_t * restrict mask,
+                                      float slope) {
+    const uint8_t * restrict src_curr  = src;
+    uint8_t * restrict dst_curr        = dst;
+    const uint8_t * restrict mask_curr = mask;
+
+    HVX_Vector scale_vec = hvx_vec_splat_fp32(scale);
+    HVX_Vector slope_vec = hvx_vec_splat_fp32(slope);
+
+    int step_of_1 = num_elems >> 5;
+
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v1 = *(HVX_Vector *) src_curr;
+
+        HVX_Vector v3 = *(HVX_Vector *) mask_curr;
+
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_vec);
+
+        HVX_Vector v4 = Q6_Vqf32_vmpy_VsfVsf(v3, slope_vec);
+
+        HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(v2, v4);
+
+        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v5);
+
+        src_curr += VLEN;
+        dst_curr += VLEN;
+        mask_curr += VLEN;
+    }
+}
+
+static void hvx_fast_softmax_f32(const uint8_t * restrict src,
+                                 uint8_t * restrict dst,
+                                 uint8_t * restrict pad,
+                                 const int num_elems) {
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_pad       = (HVX_Vector *) pad;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    HVX_Vector sum_vec = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector max_vec = hvx_vec_splat_fp32(((const float *) src)[0]);
+    HVX_Vector zero_v  = Q6_V_vzero();
+    HVX_Vector one_v   = hvx_vec_splat_fp32(1.0);
+
+    int step_of_1 = num_elems >> 5;
+
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v1 = v_src[i];
+        max_vec       = Q6_Vsf_vmax_VsfVsf(max_vec, v1);
+    }
+
+    HVX_Vector v = hvx_vec_reduce_max_fp32(max_vec);
+    max_vec      = hvx_vec_repl4(v);
+
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vsub_VsfVsf(v1, max_vec);
+
+        HVX_Vector v3 = hvx_vec_exp_fp32(Q6_Vsf_equals_Vqf32(v2));
+
+        sum_vec = Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(sum_vec), v3);
+
+        v_pad[i] = v3;
+    }
+
+    v       = hvx_vec_qf32_reduce_sum(sum_vec);
+    sum_vec = hvx_vec_repl4(Q6_Vsf_equals_Vqf32(v));
+
+    HVX_VectorPred pos_sum   = Q6_Q_vcmp_gt_VwVw(sum_vec, zero_v);
+    HVX_Vector     v4        = hvx_vec_inverse_fp32(sum_vec);
+    HVX_Vector     scale_vec = Q6_V_vmux_QVV(pos_sum, v4, one_v);
+
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v1 = v_pad[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_vec);
+        v_dst[i]      = Q6_Vsf_equals_Vqf32(v2);
+    }
+}
+
+static float hvx_softmax_f32(const uint8_t * restrict src,
+                             uint8_t * restrict dst,
+                             uint8_t * restrict spad,
+                             const int   num_elems,
+                             const float max) {
+    hvx_sub_scalar_f32(src, max, spad, num_elems);
+
+    hvx_exp_f32(spad, dst, num_elems, false);
+
+    float sum = hvx_self_sum_f32(dst, num_elems);
+
+    return sum;
+}
+
+static void softmax_htp_f32(int nth, int ith, struct softmax_th_ctx * softmax_ctx, int opt_path) {
+    struct htp_ops_context * octx = softmax_ctx->octx;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+    const struct htp_tensor * dst  = &octx->dst;
+
+    htp_softmax_preamble3;
+
+    uint8_t * src0_spad_data = octx->src0_spad.data + (ith * nb01);
+    uint8_t * src1_spad_data = octx->src1_spad.data + (ith * nb01);
+    uint8_t * dst_spad_data  = octx->dst_spad.data + (ith * nb1);
+
+    float * wp0 = (float *) src0_spad_data;
+    float * wp1 = (float *) src1_spad_data;
+    float * wp2 = (float *) dst_spad_data;
+
+    for (uint32_t i03 = 0; i03 < ne03; i03++) {
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {
+            for (uint32_t i01 = ith; i01 < ne01; i01 += nth) {
+                const uint32_t i11 = i01;
+                const uint32_t i12 = i02 % ne12;
+                const uint32_t i13 = i03 % ne13;
+
+                // ALiBi
+                const uint32_t h = i02;  // head
+
+                const float slope = (softmax_ctx->max_bias > 0.0f) ?
+                                        h < softmax_ctx->n_head_log2 ?
+                                        powf(softmax_ctx->m0, h + 1) :
+                                        powf(softmax_ctx->m1, 2 * (h - softmax_ctx->n_head_log2) + 1) :
+                                        1.0f;
+
+                float * sp = (float *) ((char *) octx->src0.data + i01 * nb01 + i02 * nb02 + i03 * nb03);
+                float * dp = (float *) ((char *) octx->dst.data + i01 * nb1 + i02 * nb2 + i03 * nb3);
+
+                // broadcast the mask across rows
+                __fp16 * mp_f16 = (softmax_ctx->use_src1) ?
+                                      (__fp16 *) ((char *) octx->src1.data + i11 * nb11 + i12 * nb12 + i13 * nb13) :
+                                      NULL;
+                float *  mp_f32 = (softmax_ctx->use_src1) ?
+                                      (float *) ((char *) octx->src1.data + i11 * nb11 + i12 * nb12 + i13 * nb13) :
+                                      NULL;
+
+                if ((1 == opt_path) && (mp_f32) && !(softmax_ctx->use_f16)) {
+                    hvx_fast_softmax_prep_f32((const uint8_t *) sp, (uint8_t *) wp0, ne00, softmax_ctx->scale,
+                                              (const uint8_t *) mp_f32, slope);
+                } else {
+                    hvx_scale_f32((const uint8_t *) sp, (uint8_t *) wp0, ne00, softmax_ctx->scale);
+                    if (mp_f32) {
+                        if (softmax_ctx->use_f16) {
+                            for (int i = 0; i < ne00; ++i) {
+                                wp0[i] += slope * (float) mp_f16[i];
+                            }
+                        } else {
+                            for (int i = 0; i < ne00; ++i) {
+                                wp0[i] += slope * mp_f32[i];
+                            }
+                        }
+                    }
+                }
+
+                if (1 == opt_path) {
+                    hvx_fast_softmax_f32((const uint8_t *) wp0, (uint8_t *) dp, (uint8_t *) wp1, ne00);
+                } else {
+                    float max = hvx_self_max_f32((const uint8_t *) wp0, ne00);
+                    float sum = hvx_softmax_f32((const uint8_t *) wp0, (uint8_t *) wp2, (uint8_t *) wp1, ne00, max);
+                    sum       = sum > 0.0 ? (1.0 / sum) : 1;
+                    hvx_scale_f32((const uint8_t *) wp2, (uint8_t *) dp, ne00, sum);
+                }
+            }
+        }
+    }
+}
+
+static void softmax_job_f32_per_thread(struct softmax_th_ctx * softmax_ctx, int nth, int ith) {
+    struct htp_ops_context * octx = softmax_ctx->octx;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    htp_softmax_preamble3;
+
+    const uint32_t src0_nrows            = ne01 * ne02 * ne03;  // src0 rows
+    const uint32_t src0_nrows_per_thread = octx->src0_nrows_per_thread;
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    int is_aligned = 1;
+    int opt_path   = 0;
+    if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) {
+        is_aligned = 0;
+        FARF(HIGH, "softmax-f32: unaligned addresses in elementwise op, possibly slower execution\n");
+    }
+    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    softmax_htp_f32(nth, ith, softmax_ctx, opt_path);
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "softmax-f32 %d/%d/%d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         softmax_ctx->use_f16, opt_path, ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13,
+         ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static void softmax_job_dispatcher_f32(unsigned int n, unsigned int i, void * p_data) {
+    struct softmax_th_ctx * p_softmax_ctx = (struct softmax_th_ctx *) p_data;
+    softmax_job_f32_per_thread(p_softmax_ctx, n, i);
+}
+
+static int execute_op_softmax_f32(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * src1 = &octx->src1;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    worker_callback_t op_func;
+    const char *      op_type = NULL;
+
+    struct softmax_th_ctx softmax_ctx;
+
+    switch (octx->op) {
+        case HTP_OP_SOFTMAX:
+            op_func = softmax_job_dispatcher_f32;
+            op_type = "softmax-f32";
+
+            init_softmax_ctx(&softmax_ctx, octx);
+            break;
+
+        default:
+            FARF(ERROR, "Unsupported Op %u\n", octx->op);
+            return HTP_STATUS_NO_SUPPORT;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
+    const size_t src0_row_size = src0->nb[1];
+    const size_t src1_row_size = src0_row_size;
+    const size_t dst_row_size  = dst->nb[1];
+
+    // VTCM scratchpads for all tensors
+    // N rows per thread, padded to HVX vector size
+    octx->dst_spad.size  = htp_round_up(dst_row_size, 128) * n_threads;
+    octx->src0_spad.size = htp_round_up(src0_row_size, 128) * n_threads;
+    octx->src1_spad.size = htp_round_up(src1_row_size, 128) * n_threads;
+
+    size_t spad_size = octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;
+
+    if (src1->ne[0]) {
+        FARF(HIGH,
+             "%s: %ux%ux%ux%u x %ux%ux%ux%u -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n",
+             op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
+             src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], octx->src0_spad.size, octx->src1_spad.size,
+             octx->dst_spad.size);
+    } else {
+        FARF(HIGH, "%s: %ux%ux%ux%u -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n", op_type,
+             src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+             octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
+    }
+
+    // Make sure the reserved vtcm size is sufficient
+    if (octx->ctx->vtcm_size < spad_size) {
+        FARF(ERROR, "%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
+             spad_size);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
+
+    uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+
+    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
+        uint32_t n_jobs             = MIN(n_threads, src0_nrows);
+        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+        worker_pool_run_func(octx->ctx->worker_pool, op_func, &softmax_ctx, n_jobs);
+    }
+
+    return err;
+}
+
+int op_softmax(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    switch (octx->src0.type) {
+        case HTP_TYPE_F32:
+            err = execute_op_softmax_f32(octx);
+            break;
+
+        default:
+            err = HTP_STATUS_NO_SUPPORT;
+            break;
+    }
+
+    return err;
+}

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

@@ -0,0 +1,255 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#ifdef HTP_DEBUG
+#    define FARF_HIGH 1
+#endif
+
+#include <HAP_farf.h>
+#include <HAP_mem.h>
+#include <HAP_perf.h>
+#include <HAP_ps.h>
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <qurt_thread.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+#include "ops-utils.h"
+
+#define htp_unary_preamble            \
+    const uint32_t ne00 = src->ne[0]; \
+    const uint32_t ne01 = src->ne[1]; \
+    const uint32_t ne02 = src->ne[2]; \
+    const uint32_t ne03 = src->ne[3]; \
+                                      \
+    const uint32_t ne0 = dst->ne[0];  \
+    const uint32_t ne1 = dst->ne[1];  \
+    const uint32_t ne2 = dst->ne[2];  \
+    const uint32_t ne3 = dst->ne[3];  \
+                                      \
+    const uint32_t nb00 = src->nb[0]; \
+    const uint32_t nb01 = src->nb[1]; \
+    const uint32_t nb02 = src->nb[2]; \
+    const uint32_t nb03 = src->nb[3]; \
+                                      \
+    const uint32_t nb0 = dst->nb[0];  \
+    const uint32_t nb1 = dst->nb[1];  \
+    const uint32_t nb2 = dst->nb[2];  \
+    const uint32_t nb3 = dst->nb[3];
+
+static void hvx_fast_rms_norm_f32(const uint8_t * restrict src,
+                                  uint8_t * restrict dst,
+                                  uint8_t * restrict pad,
+                                  const int num_elems,
+                                  float     epsilon) {
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    HVX_Vector sum_v     = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_fp32(epsilon);
+
+    int step_of_1 = num_elems >> 5;
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v         = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
+    }
+
+    HVX_Vector reduced_sum = hvx_vec_qf32_reduce_sum(sum_v);
+    sum_v                  = hvx_vec_repl4(Q6_Vsf_equals_Vqf32(reduced_sum));
+
+    HVX_Vector t_v            = hvx_vec_splat_fp32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_fp32(t_v);
+    HVX_Vector mean_v         = Q6_Vqf32_vmpy_VsfVsf(sum_v, denom_v);
+    HVX_Vector mean_epsilon_v = Q6_Vqf32_vadd_Vqf32Vsf(mean_v, epsilon_v);
+
+    HVX_Vector scale_v = hvx_vec_rsqrt_fp32(Q6_Vsf_equals_Vqf32(mean_epsilon_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_v);
+        v_dst[i]      = Q6_Vsf_equals_Vqf32(v2);
+    }
+}
+
+static void rms_norm_htp_f32(const float * restrict src,
+                             float * restrict dst,
+                             uint8_t * restrict spad,
+                             const uint32_t num_rows,
+                             const uint32_t row_elems,
+                             const size_t   row_size,
+                             int32_t *      op_params,
+                             int            opt_path) {
+    float epsilon = 0.f;
+    memcpy(&epsilon, op_params, sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const float * restrict src_local = src + (ir * row_elems);
+        float * restrict dst_local       = dst + (ir * row_elems);
+
+        if (ir + 1 < num_rows) {
+            htp_l2fetch(src_local + row_elems, 1, row_size, row_size);
+        }
+
+        if (1 == opt_path) {
+            hvx_fast_rms_norm_f32((const uint8_t *) src_local, (uint8_t *) dst_local, spad, row_elems, epsilon);
+        } else {
+            float sum = hvx_sum_of_squares_f32((const uint8_t *) src_local, row_elems);
+
+            const float mean  = sum / row_elems;
+            const float scale = 1.0f / sqrtf(mean + epsilon);
+
+            hvx_scale_f32((const uint8_t *) src_local, (uint8_t *) dst_local, row_elems, scale);
+        }
+    }
+}
+
+static void unary_job_f32_per_thread(const struct htp_tensor * src,
+                                     struct htp_tensor *       dst,
+                                     uint8_t *                 spad,
+                                     int                       htp_op,
+                                     int32_t *                 op_params,
+                                     uint32_t                  nth,
+                                     uint32_t                  ith,
+                                     uint32_t                  src0_nrows_per_thread) {
+    htp_unary_preamble;
+
+    const size_t src0_row_size = nb01;
+    const size_t dst_row_size  = nb1;
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    int is_aligned = 1;
+    int opt_path   = 0;
+    if ((0 == htp_is_aligned((void *) src->data, VLEN)) || (0 == htp_is_aligned((void *) dst->data, VLEN))) {
+        is_aligned = 0;
+        FARF(HIGH, "unary-f32: unaligned addresses in unary op, possibly slower execution\n");
+    }
+    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    const uint8_t * restrict data_src = (const uint8_t *) src->data;
+    uint8_t * restrict data_dst       = (uint8_t *) dst->data;
+
+    const float * restrict src_th = (float *) (data_src + (src0_start_row * src0_row_size));
+    float * restrict dst_th       = (float *) (data_dst + (src0_start_row * dst_row_size));
+    uint8_t * restrict spad_th    = (uint8_t *) spad + (ith * nb01);
+
+    switch (htp_op) {
+        case HTP_OP_RMS_NORM:
+            rms_norm_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
+            break;
+
+        default:
+            break;
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "unary-f32 %d/%d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, opt_path, src->ne[0],
+         src->ne[1], src->ne[2], src->ne[3], src0_start_row, src0_end_row, dst->ne[0], dst->ne[1], dst->ne[2],
+         dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static void unary_job_dispatcher_f32(unsigned int n, unsigned int i, void * data) {
+    struct htp_ops_context * octx = (struct htp_ops_context *) data;
+
+    unary_job_f32_per_thread(&octx->src0, &octx->dst, octx->src0_spad.data, octx->op, octx->op_params, n, i,
+                             octx->src0_nrows_per_thread);
+}
+
+static int execute_op_unary_f32(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    worker_callback_t unary_op_func;
+    const char *      op_type = NULL;
+
+    switch (octx->op) {
+        case HTP_OP_RMS_NORM:
+            unary_op_func = unary_job_dispatcher_f32;
+            op_type       = "rmsnorm-f32";
+            break;
+
+        default:
+            FARF(ERROR, "Unsupported unary Op %u\n", octx->op);
+            return HTP_STATUS_NO_SUPPORT;
+    }
+
+    const int      n_threads  = octx->n_threads;
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+
+    const size_t src0_row_size = src0->nb[1];
+    const size_t dst_row_size  = dst->nb[1];
+
+    // VTCM scratchpads for all tensors
+    octx->dst_spad.size  = htp_round_up(dst_row_size, 128) * n_threads;
+    octx->src0_spad.size = htp_round_up(src0_row_size, 128) * n_threads;
+
+    size_t spad_size = octx->src0_spad.size + octx->dst_spad.size;
+
+    FARF(HIGH, "%s: (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n", op_type,
+         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+         octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
+
+    // Make sure the reserved vtcm size is sufficient
+    if (octx->ctx->vtcm_size < spad_size) {
+        FARF(ERROR, "unary-%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
+             spad_size);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->dst_spad.data  = octx->src0_spad.data + octx->src0_spad.size;
+
+    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
+        uint32_t n_jobs = MIN(n_threads, src0_nrows);
+
+        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+
+        worker_pool_run_func(octx->ctx->worker_pool, unary_op_func, octx, n_jobs);
+    }
+
+    return err;
+}
+
+int op_unary(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    switch (octx->src0.type) {
+        case HTP_TYPE_F32:
+            err = execute_op_unary_f32(octx);
+            break;
+
+        default:
+            err = HTP_STATUS_NO_SUPPORT;
+            break;
+    }
+
+    return err;
+}

ggml/src/ggml-hexagon/htp/worker-pool.c

@@ -0,0 +1,297 @@
+#include "worker-pool.h"
+
+#include <qurt.h>
+#include <stdatomic.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#ifdef HTP_DEBUG
+#    define FARF_HIGH 1
+#endif
+
+#include "HAP_farf.h"
+
+#define WORKER_THREAD_STACK_SZ  (2 * 16384)
+#define LOWEST_USABLE_QURT_PRIO (254)
+
+struct worker_pool_s;
+
+// internal structure kept in thread-local storage per instance of worker pool
+typedef struct {
+    struct worker_pool_s * pool;
+    unsigned int           id;
+} worker_context_t;
+
+// internal structure kept in thread-local storage per instance of worker pool
+typedef struct worker_pool_s {
+    worker_pool_job_t job[MAX_NUM_WORKERS];      // list of job descriptors
+    qurt_thread_t     thread[MAX_NUM_WORKERS];   // thread ID's of the workers
+    worker_context_t  context[MAX_NUM_WORKERS];  // worker contexts
+    void *            stack[MAX_NUM_WORKERS];    // thread stack pointers
+    unsigned int      n_threads;                 // number of workers in this pool
+
+    atomic_uint seqn;                            // seqno used to detect new jobs
+    atomic_uint next_job;                        // next job index
+    atomic_uint n_pending;                       // number of pending jobs
+    atomic_uint n_jobs;                          // number of current jobs
+    atomic_bool killed;                          // threads need to exit
+} worker_pool_t;
+
+static void worker_pool_main(void * context) {
+    worker_context_t * me   = (worker_context_t *) context;
+    worker_pool_t *    pool = me->pool;
+
+    FARF(HIGH, "worker-pool: thread %u started", me->id);
+
+    unsigned int prev_seqn = 0;
+    while (!atomic_load(&pool->killed)) {
+        unsigned int seqn = atomic_load(&pool->seqn);
+        if (seqn == prev_seqn) {
+            // Nothing to do
+            qurt_futex_wait(&pool->seqn, prev_seqn);
+            continue;
+        }
+
+        // New job
+        prev_seqn = seqn;
+
+        unsigned int n = atomic_load(&pool->n_jobs);
+        unsigned int i = atomic_fetch_add(&pool->next_job, 1);
+        if (i >= n) {
+            // Spurios wakeup
+            continue;
+        }
+
+        pool->job[i].func(n, i, pool->job[i].data);
+
+        atomic_fetch_sub(&pool->n_pending, 1);
+    }
+
+    FARF(HIGH, "worker-pool: thread %u stopped", me->id);
+}
+
+AEEResult worker_pool_init_with_stack_size(worker_pool_context_t * context, uint32_t n_threads, uint32_t stack_size) {
+    int err = 0;
+
+    if (NULL == context) {
+        FARF(ERROR, "NULL context passed to worker_pool_init().");
+        return AEE_EBADPARM;
+    }
+
+    // Allocations
+    int size = (stack_size * n_threads) + (sizeof(worker_pool_t));
+
+    unsigned char * mem_blob = (unsigned char *) malloc(size);
+    if (!mem_blob) {
+        FARF(ERROR, "Could not allocate memory for worker pool!!");
+        return AEE_ENOMEMORY;
+    }
+
+    worker_pool_t * me = (worker_pool_t *) (mem_blob + stack_size * n_threads);
+
+    // name for the first worker, useful in debugging threads
+    char name[19];
+    snprintf(name, 12, "0x%8x:", (int) me);
+    strcat(name, "worker0");
+    me->n_threads = n_threads;
+
+    // initializations
+    for (unsigned int i = 0; i < me->n_threads; i++) {
+        me->stack[i]  = NULL;
+        me->thread[i] = 0;
+
+        me->context[i].id   = i;
+        me->context[i].pool = me;
+    }
+
+    // initialize job queue
+    me->n_pending = 0;
+    me->n_jobs    = 0;
+    me->next_job  = 0;
+    me->seqn      = 0;
+    me->killed    = 0;
+
+    // launch the workers
+    qurt_thread_attr_t attr;
+    qurt_thread_attr_init(&attr);
+
+    for (unsigned int i = 0; i < me->n_threads; i++) {
+        // set up stack
+        me->stack[i] = mem_blob;
+        mem_blob += stack_size;
+        qurt_thread_attr_set_stack_addr(&attr, me->stack[i]);
+        qurt_thread_attr_set_stack_size(&attr, stack_size);
+
+        // set up name
+        qurt_thread_attr_set_name(&attr, name);
+        name[17] = (name[17] + 1);
+        // name threads context:worker0, context:worker1, .. (recycle at 9, but num threads should be less than that anyway)
+        if (name[17] > '9') {
+            name[17] = '0';
+        }
+
+        // set up priority - by default, match the creating thread's prio
+        int prio = qurt_thread_get_priority(qurt_thread_get_id());
+
+        if (prio < 1) {
+            prio = 1;
+        }
+        if (prio > LOWEST_USABLE_QURT_PRIO) {
+            prio = LOWEST_USABLE_QURT_PRIO;
+        }
+
+        qurt_thread_attr_set_priority(&attr, prio);
+
+        // launch
+        err = qurt_thread_create(&me->thread[i], &attr, worker_pool_main, (void *) &me->context[i]);
+        if (err) {
+            FARF(ERROR, "Could not launch worker threads!");
+            worker_pool_release((worker_pool_context_t *) &me);
+            return AEE_EQURTTHREADCREATE;
+        }
+    }
+    *context = (worker_pool_context_t *) me;
+    return AEE_SUCCESS;
+}
+
+AEEResult worker_pool_init(worker_pool_context_t * context, uint32_t n_threads) {
+    return worker_pool_init_with_stack_size(context, n_threads, WORKER_THREAD_STACK_SZ);
+}
+
+// clean up worker pool
+void worker_pool_release(worker_pool_context_t * context) {
+    worker_pool_t * me = (worker_pool_t *) *context;
+
+    // if no worker pool exists, return error.
+    if (NULL == me) {
+        return;
+    }
+
+    atomic_store(&me->killed, 1);
+    atomic_fetch_add(&me->seqn, 1);
+    qurt_futex_wake(&me->seqn, me->n_threads);
+
+    // de-initializations
+    for (unsigned int i = 0; i < me->n_threads; i++) {
+        if (me->thread[i]) {
+            int status;
+            (void) qurt_thread_join(me->thread[i], &status);
+        }
+    }
+
+    // free allocated memory (were allocated as a single buffer starting at stack[0])
+    if (me->stack[0]) {
+        free(me->stack[0]);
+    }
+
+    *context = NULL;
+}
+
+// run jobs
+AEEResult worker_pool_run_jobs(worker_pool_context_t context, worker_pool_job_t * job, unsigned int n) {
+    worker_pool_t * me = (worker_pool_t *) context;
+    if (NULL == me) {
+        FARF(ERROR, "worker-pool: invalid context");
+        return AEE_EBADPARM;
+    }
+
+    if (n > me->n_threads) {
+        FARF(ERROR, "worker-pool: invalid number of jobs %u for n-threads %u", n, me->n_threads);
+        return AEE_EBADPARM;
+    }
+
+    memcpy(me->job, job, sizeof(worker_pool_job_t) * n);
+
+    if (n > 1) {
+        atomic_store(&me->next_job, 1);
+        atomic_store(&me->n_jobs, n);
+        atomic_store(&me->n_pending, n - 1);
+
+        // wake up workers
+        atomic_fetch_add(&me->seqn, 1);
+        qurt_futex_wake(&me->seqn, n - 1);
+    }
+
+    // main thread runs job #0
+    me->job[0].func(n, 0, me->job[0].data);
+
+    if (n > 1) {
+        while (atomic_load(&me->n_pending))
+            ;
+    }
+
+    return 0;
+}
+
+// run func
+AEEResult worker_pool_run_func(worker_pool_context_t context, worker_callback_t func, void * data, unsigned int n) {
+    worker_pool_job_t job[n];
+
+    for (unsigned int i = 0; i < n; i++) {
+        job[i].func = func;
+        job[i].data = data;
+    }
+
+    return worker_pool_run_jobs(context, job, n);
+}
+
+AEEResult worker_pool_set_thread_priority(worker_pool_context_t context, unsigned int prio) {
+    worker_pool_t * me = (worker_pool_t *) context;
+
+    // if no worker pool exists, return error.
+    if (!me) {
+        return AEE_ENOMORE;
+    }
+
+    int result = AEE_SUCCESS;
+    if (prio < 1) {
+        prio = 1;
+    }
+    if (prio > LOWEST_USABLE_QURT_PRIO) {
+        prio = LOWEST_USABLE_QURT_PRIO;
+    }
+
+    for (unsigned int i = 0; i < me->n_threads; i++) {
+        int res = qurt_thread_set_priority(me->thread[i], (unsigned short) prio);
+        if (0 != res) {
+            result = AEE_EBADPARM;
+            FARF(ERROR, "QURT failed to set priority of thread %d, ERROR = %d", me->thread[i], res);
+        }
+    }
+
+    return result;
+}
+
+AEEResult worker_pool_retrieve_thread_id(worker_pool_context_t context, unsigned int * tids) {
+    worker_pool_t * me = (worker_pool_t *) context;
+    if (!me) {
+        FARF(ERROR, "worker-pool: invalid context");
+        return AEE_EBADPARM;
+        ;
+    }
+
+    for (int i = 0; i < me->n_threads; i++) {
+        tids[i] = me->thread[i];
+    }
+
+    return AEE_SUCCESS;
+}
+
+AEEResult worker_pool_get_thread_priority(worker_pool_context_t context, unsigned int * prio) {
+    worker_pool_t * me = (worker_pool_t *) context;
+    if (!me) {
+        FARF(ERROR, "worker-pool: invalid context");
+        return AEE_EBADPARM;
+    }
+
+    int priority = qurt_thread_get_priority(me->thread[0]);
+    if (priority > 0) {
+        *prio = priority;
+        return 0;
+    } else {
+        *prio = 0;
+        return AEE_EBADSTATE;
+    }
+}

ggml/src/ggml-hexagon/htp/worker-pool.h

@@ -0,0 +1,57 @@
+#ifndef HTP_WORKER_POOL_H
+#define HTP_WORKER_POOL_H
+
+// MACRO enables function to be visible in shared-library case.
+#define WORKERPOOL_API __attribute__((visibility("default")))
+
+#include <AEEStdDef.h>
+#include <AEEStdErr.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// signature of callbacks to be invoked by worker threads
+typedef void (*worker_callback_t)(unsigned int n, unsigned int i, void *);
+
+/// Typedef of worker_pool context
+typedef void * worker_pool_context_t;
+
+/// descriptor for requested callback
+typedef struct {
+    worker_callback_t func;
+    void *            data;
+} worker_pool_job_t;
+
+/// Maximum supported number of worker threads.
+#define MAX_NUM_WORKERS 10
+
+// Initialize worker pool.
+WORKERPOOL_API AEEResult worker_pool_init(worker_pool_context_t * context, uint32_t n_threads);
+
+// Initialize worker pool with custom stack size
+WORKERPOOL_API AEEResult worker_pool_init_with_stack_size(worker_pool_context_t * context,
+                                                          uint32_t                n_threads,
+                                                          uint32_t                stack_size);
+
+// Kill worker threads and release worker pool resources
+WORKERPOOL_API void worker_pool_release(worker_pool_context_t * context);
+
+// Run jobs with the worker pool.
+WORKERPOOL_API AEEResult worker_pool_run_jobs(worker_pool_context_t context, worker_pool_job_t * job, unsigned int n);
+
+WORKERPOOL_API AEEResult worker_pool_run_func(worker_pool_context_t context,
+                                              worker_callback_t     func,
+                                              void *                data,
+                                              unsigned int          n);
+
+WORKERPOOL_API AEEResult worker_pool_set_thread_priority(worker_pool_context_t context, unsigned int prio);
+WORKERPOOL_API AEEResult worker_pool_get_thread_priority(worker_pool_context_t context, unsigned int * prio);
+WORKERPOOL_API AEEResult worker_pool_retrieve_thread_id(worker_pool_context_t context, unsigned int * tids);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // #ifndef HTP_WORKER_POOL_H

ggml/src/ggml-hip/CMakeLists.txt

@@ -28,9 +28,12 @@ if (CXX_IS_HIPCC)
                 " Prefer setting the HIP compiler directly. See README for details.")
     endif()
 else()
-    # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
-    if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
-        set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+    # Forward (AMD)GPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+    if(AMDGPU_TARGETS AND NOT GPU_TARGETS)
+        set(GPU_TARGETS ${AMDGPU_TARGETS})
+    endif()
+    if(GPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+        set(CMAKE_HIP_ARCHITECTURES ${GPU_TARGETS})
     endif()
     cmake_minimum_required(VERSION 3.21)
     enable_language(HIP)

ggml/src/ggml-impl.h

@@ -565,14 +565,23 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
 #define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
 #define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
 
+static inline int32_t ggml_node_get_use_count(const struct ggml_cgraph * cgraph, int node_idx) {
+    const struct ggml_tensor * node = cgraph->nodes[node_idx];
+
+    size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
+    if (!ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos)) {
+        return 0;
+    }
+    return cgraph->use_counts[hash_pos];
+}
+
 // return true if the node's results are only used by N other nodes
 // and can be fused into their calculations.
 static inline bool ggml_node_has_n_uses(const struct ggml_cgraph * cgraph, int node_idx, int32_t n_uses) {
     const struct ggml_tensor * node = cgraph->nodes[node_idx];
 
     // check the use count against how many we're replacing
-    size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
-    if (!ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos) || cgraph->use_counts[hash_pos] != n_uses) {
+    if (ggml_node_get_use_count(cgraph, node_idx) != n_uses) {
         return false;
     }
 
@@ -638,6 +647,36 @@ static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx
     return ggml_can_fuse_ext(cgraph, idxs, ops, num_ops);
 }
 
+GGML_API bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
+                                         const int *                node_idxs,
+                                         int                        count,
+                                         const enum ggml_op *       ops,
+                                         const int *                outputs,
+                                         int                        num_outputs);
+
+// Returns true if the subgraph formed by {node_idxs} can be fused
+// checks whethers all nodes which are not part of outputs can be elided
+// by checking if their num_uses are confined to the subgraph
+static inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph * cgraph,
+                                          int                        node_idx,
+                                          int                        count,
+                                          const enum ggml_op *       ops,
+                                          const int *                outputs,
+                                          int                        num_outputs) {
+    GGML_ASSERT(count < 32);
+    if (node_idx + count > cgraph->n_nodes) {
+        return false;
+    }
+
+    int idxs[32];
+
+    for (int i = 0; i < count; ++i) {
+        idxs[i] = node_idx + i;
+    }
+
+    return ggml_can_fuse_subgraph_ext(cgraph, idxs, count, ops, outputs, num_outputs);
+}
+
 #ifdef __cplusplus
 }
 #endif
@@ -651,6 +690,13 @@ inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::
     return ggml_can_fuse(cgraph, node_idx, ops.begin(), (int)ops.size());
 }
 
+inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph *          cgraph,
+                                   int                                 start_idx,
+                                   std::initializer_list<enum ggml_op> ops,
+                                   std::initializer_list<int>          outputs = {}) {
+    return ggml_can_fuse_subgraph(cgraph, start_idx, ops.size(), ops.begin(), outputs.begin(), outputs.size());
+}
+
 // expose GGUF internals for test code
 GGML_API size_t gguf_type_size(enum gguf_type type);
 GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);